{ "updated": "2026-10-14", "milestones": [ { "id": "dendral-structure-elucidation", "date": "1965", "date_basis": "announcement", "title": "DENDRAL program infers molecular structures from mass spectra", "domain": "chemistry-materials", "subfield": "structure elucidation / expert systems", "summary": "Stanford's DENDRAL, begun in 1965 by Joshua Lederberg, Edward Feigenbaum and Carl Djerassi, generated candidate molecular structures and pruned those that did not fit mass-spectrometry data. Once fully working it did this faster than an expert spectrometrist with comparable accuracy. Its offshoot Meta-DENDRAL inferred fragmentation rules from data (J. Am. Chem. Soc., 1976).", "before_after": null, "why_it_matters": "It was the first sustained attempt to make a computer form scientific hypotheses, and became the template for expert systems and later AI tools in chemistry.", "ai_systems": [ "DENDRAL", "Meta-DENDRAL" ], "organizations": [ "Stanford University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Artificial Intelligence journal, 1993 (DENDRAL case study by its creators)", "url": "https://doi.org/10.1016/0004-3702(93)90068-M" }, { "label": "J. Am. Chem. Soc., 1976 (Meta-DENDRAL rule discovery)", "url": "https://doi.org/10.1021/ja00436a017" }, { "label": "US National Library of Medicine, Lederberg Papers profile", "url": "https://profiles.nlm.nih.gov/spotlight/bb/feature/ai" } ], "foundation0": false }, { "id": "corey-wipke-ocss-computer-synthesis-planning", "date": "1969-10-10", "date_basis": "publication", "title": "Corey and Wipke's OCSS program plans organic syntheses by computer", "domain": "chemistry-materials", "subfield": "synthesis planning / retrosynthesis", "summary": "Harvard's E. J. Corey and W. Todd Wipke described OCSS, an interactive program applying retrosynthetic logic: from a target molecule it proposes bond disconnections and simpler precursors, working back toward known starting materials. It led to LHASA; Corey's 1990 Nobel citation noted his retrosynthetic analysis 'was amenable to computer programming'.", "before_after": null, "why_it_matters": "The origin of computer-aided synthesis planning, the line that runs through Chematica to today's deep-learning retrosynthesis tools.", "ai_systems": [ "OCSS", "LHASA" ], "organizations": [ "Harvard University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science 166:178-192, 1969-10-10", "url": "https://doi.org/10.1126/science.166.3902.178" }, { "label": "Nobel Prize in Chemistry 1990 press release", "url": "https://www.nobelprize.org/prizes/chemistry/1990/press-release/" } ], "foundation0": false }, { "id": "four-colour-theorem-computer-proof", "date": "1976", "date_basis": "publication", "title": "Four Colour Theorem becomes the first major computer-assisted proof", "domain": "mathematics", "subfield": "graph theory / computer-assisted proof", "summary": "Kenneth Appel and Wolfgang Haken (University of Illinois) proved that four colours suffice for any planar map, open since 1852, by having computers check well over a thousand reducible configurations, taking over 1,000 hours of computer time. In 2005 Georges Gonthier and Benjamin Werner machine-checked a proof in the Coq proof assistant.", "before_after": { "metric": "Status of the Four Colour Conjecture", "before": "open since 1852", "after": "proved (1976), machine-checked in Coq (2005)" }, "why_it_matters": "The first famous theorem whose proof depended on a computer. It began the debate over trusting machine-made proofs that formal proof assistants such as Coq and Lean now help settle.", "ai_systems": [ "Computer case-checking programs", "Coq proof assistant" ], "organizations": [ "University of Illinois", "Microsoft Research", "INRIA" ], "ai_role": "instrument", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Bulletin of the AMS 82:711-712, 1976", "url": "https://doi.org/10.1090/S0002-9904-1976-14122-5" }, { "label": "Gonthier, A computer-checked proof of the Four Colour Theorem (2005)", "url": "https://www.cl.cam.ac.uk/~lp15/Pages/4colproof.pdf" } ], "foundation0": false }, { "id": "bacon-rediscovers-empirical-laws", "date": "1979", "date_basis": "publication", "title": "BACON program rediscovers Kepler's third law and other laws from data", "domain": "science-of-science", "subfield": "computational scientific discovery", "summary": "Pat Langley's BACON, begun at Carnegie Mellon in 1977, is a production system that finds empirical laws by spotting constancies and trends in data and defining new terms. Its BACON.3 version (IJCAI 1979; Cognitive Science, 1981) rediscovered versions of the ideal gas law, Kepler's third law, Coulomb's law, Ohm's law and Galileo's laws.", "before_after": null, "why_it_matters": "Founded computational scientific discovery as a research field, the line that leads to today's AI-scientist systems.", "ai_systems": [ "BACON", "BACON.3" ], "organizations": [ "Carnegie Mellon University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Langley, Rediscovering Physics with BACON.3, IJCAI 1979", "url": "https://mlanthology.org/ijcai/1979/langley1979ijcai-rediscovering/" }, { "label": "Cognitive Science 5(1), 1981 (BACON.3)", "url": "https://doi.org/10.1111/j.1551-6708.1981.tb00869.x" }, { "label": "Langley, BACON, IJCAI-77 proceedings", "url": "https://www.ijcai.org/Proceedings/77-1/Papers/057.pdf" } ], "foundation0": false }, { "id": "internist-1-diagnostic-expert-system", "date": "1982-08-19", "date_basis": "publication", "title": "INTERNIST-1 expert system tackles hard diagnostic cases in internal medicine", "domain": "medicine", "subfield": "clinical decision support / expert systems", "summary": "University of Pittsburgh's INTERNIST-1 made multiple diagnoses across general internal medicine from a large knowledge base. On 19 published Massachusetts General Hospital clinicopathological cases, its performance was qualitatively similar to the hospital clinicians' but below the expert case discussants'. The authors judged it not yet reliable enough for clinical use.", "before_after": null, "why_it_matters": "One of the first broad, rigorously evaluated attempts to automate diagnosis, it exposed gaps such as no anatomical or temporal reasoning and no explanations that later systems had to close.", "ai_systems": [ "INTERNIST-1 (later CADUCEUS)" ], "organizations": [ "University of Pittsburgh" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "New England Journal of Medicine 307:468-476, 1982-08-19", "url": "https://doi.org/10.1056/NEJM198208193070803" }, { "label": "PubMed record 7048091", "url": "https://pubmed.ncbi.nlm.nih.gov/7048091/" } ], "foundation0": false }, { "id": "prospector-mount-tolman-molybdenum", "date": "1982-09-03", "date_basis": "publication", "title": "PROSPECTOR expert system locates hidden ore-grade molybdenum at Mount Tolman", "domain": "earth-climate", "subfield": "mineral exploration / expert systems", "summary": "SRI International's PROSPECTOR, using rules from a porphyry-molybdenum exploration specialist, was given maps of pre-drilling exploration data for Mount Tolman, Washington. It identified the location of previously unknown ore-grade mineralization; its authors wrote that this appeared to be the first such determination by a computer-based approach (Science, 1982).", "before_after": null, "why_it_matters": "An early AI program making a testable prediction about the physical world, showing that encoded expert reasoning could point to where resources lie underground.", "ai_systems": [ "PROSPECTOR" ], "organizations": [ "SRI International" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science 217:927-929, 1982-09-03", "url": "https://doi.org/10.1126/science.217.4563.927" }, { "label": "SRI International, PROSPECTOR history page", "url": "https://www.sri.com/hoi/prospector-computer-based-expert-system/" }, { "label": "Christian Science Monitor, 1982-09-30", "url": "https://www.csmonitor.com/1982/0930/093037.html" } ], "foundation0": false }, { "id": "graffiti-conjecture-average-distance", "date": "1988-06", "date_basis": "publication", "title": "Graffiti program's conjecture proved: average distance ≤ independence number", "domain": "mathematics", "subfield": "graph theory / automated conjecturing", "summary": "Siemion Fajtlowicz's Graffiti, a University of Houston program from the mid-1980s, generated conjectures by testing relations among graph invariants on a database of graphs. Its conjecture number 2, that a connected graph's average distance never exceeds its independence number, was proved by Fan Chung in 1988. Graffiti's conjectures prompted many published papers.", "before_after": null, "why_it_matters": "One of the first cases of a machine posing new mathematical conjectures that mathematicians then proved, a forerunner of today's AI-assisted conjecture generation.", "ai_systems": [ "Graffiti" ], "organizations": [ "University of Houston" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Chung, Journal of Graph Theory 12(2), 1988", "url": "https://doi.org/10.1002/jgt.3190120213" }, { "label": "Fajtlowicz, On conjectures of Graffiti, Discrete Mathematics 72, 1988", "url": "https://doi.org/10.1016/0012-365X(88)90199-9" }, { "label": "DeLaViña and Waller, Electronic J. Combinatorics 15, 2008 (Graffiti history)", "url": "https://www.combinatorics.org/ojs/index.php/eljc/article/view/v15i1r33" } ], "foundation0": false }, { "id": "neural-networks-protein-secondary-structure-psipred", "date": "1988-08-20", "date_basis": "publication", "title": "Neural networks learn protein secondary structure, reaching 78% by 1999", "domain": "biology", "subfield": "protein structure prediction", "summary": "Qian and Sejnowski trained a neural network on known structures to label each residue helix, sheet or coil, scoring 64.3% on proteins unrelated to its training set. In 1999 PSIPRED, a two-stage network fed evolutionary sequence profiles, scored 76.5% to 78.3% and, per its author, beat other methods in blind CASP3 testing.", "before_after": { "metric": "Three-state (helix, sheet, coil) accuracy; different test sets", "before": "64.3% (1988 network, proteins unrelated to its training set)", "after": "76.5% to 78.3% (PSIPRED, 1999, fold-level cross-validation)" }, "why_it_matters": "Secondary structure is a first step toward a protein's full shape, the problem AlphaFold later cracked. These networks learned it from solved structures, and adding evolutionary profiles lifted accuracy to about 78%.", "ai_systems": [ "Feed-forward neural network (Qian and Sejnowski, 1988)", "PSIPRED (two-stage neural network, 1999)" ], "organizations": [ "Johns Hopkins University", "University of Warwick" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Qian and Sejnowski, Journal of Molecular Biology, 1988-08-20", "url": "https://doi.org/10.1016/0022-2836(88)90564-5" }, { "label": "Jones (PSIPRED), Journal of Molecular Biology, 1999-09-17", "url": "https://doi.org/10.1006/jmbi.1999.3091" }, { "label": "Orengo et al., Proteins, 1999 (CASP3 assessment: best secondary-structure methods used neural networks)", "url": "https://pubmed.ncbi.nlm.nih.gov/10526364/" } ], "foundation0": false }, { "id": "robbins-conjecture-eqp", "date": "1996-10", "date_basis": "announcement", "title": "EQP automated prover proves the Robbins conjecture", "domain": "mathematics", "subfield": "Boolean algebra / automated reasoning", "summary": "William McCune's EQP theorem prover at Argonne National Laboratory proved, on 10 October 1996, the last missing step (Winker's condition) showing every Robbins algebra is Boolean, settling a question posed in the 1930s. The search took about eight days on an RS/6000 processor; the proof appeared in the Journal of Automated Reasoning in 1997.", "before_after": { "metric": "Status of the Robbins problem", "before": "Open for about 60 years", "after": "Proved by EQP (October 1996)" }, "why_it_matters": "The first famous open mathematical question settled by an automated reasoning program, showing a machine could find a proof that experienced mathematicians had missed.", "ai_systems": [ "EQP" ], "organizations": [ "Argonne National Laboratory" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "McCune, Robbins Algebras Are Boolean (posted October 1996)", "url": "https://www.cs.unm.edu/~mccune/papers/robbins/" }, { "label": "J. Automated Reasoning 19(3):263-276, 1997", "url": "https://doi.org/10.1023/A:1005843212881" } ], "foundation0": false }, { "id": "remote-agent-deep-space-1", "date": "1999-05-17", "date_basis": "deployment", "title": "Remote Agent AI takes command of NASA's Deep Space 1 spacecraft", "domain": "astronomy-space", "subfield": "spacecraft autonomy", "summary": "From 17 May 1999, Remote Agent, built by NASA Ames and JPL, commanded Deep Space 1 nearly 120 million km from Earth. Its onboard planner, executive and model-based fault diagnosis turned high-level goals into plans and handled simulated failures. A software deadlock halted the first run; a shorter rerun that week met all validation objectives.", "before_after": null, "why_it_matters": "The first time AI software had primary command of a spacecraft, opening the way to autonomous operations far from Earth, where signal delays rule out step-by-step control.", "ai_systems": [ "Remote Agent (Planner/Scheduler, Executive, Livingstone)" ], "organizations": [ "NASA Ames Research Center", "NASA Jet Propulsion Laboratory" ], "ai_role": "originator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "NASA press release, 1999-05-17", "url": "https://nssdc.gsfc.nasa.gov/planetary/text/ds1_pr_19990517.txt" }, { "label": "Nayak et al., Validating the DS1 Remote Agent Experiment (1999)", "url": "https://ai.jpl.nasa.gov/public/documents/papers/rax-results-isairas99.pdf" }, { "label": "Artificial Intelligence 103:5-47, 1998 (Remote Agent design)", "url": "https://doi.org/10.1016/S0004-3702(98)00068-X" } ], "foundation0": false }, { "id": "golem-evolved-robots-3d-printed", "date": "2000-08-31", "date_basis": "publication", "title": "Golem project evolves robot bodies and brains in simulation, then prints them", "domain": "engineering-robotics", "subfield": "evolutionary robotics", "summary": "Brandeis researchers evolved simple machines made of bars, actuators and artificial neurons inside a physics simulation, keeping the designs that moved best. The fittest designs were then fabricated with rapid-prototyping (3D printing) machines and crawled in the real world. The authors called them the first robots both designed and built by machines.", "before_after": null, "why_it_matters": "Robots had always been designed piece by piece by engineers. Letting a search process invent both body and controller started the line of work that leads to AI-designed robots and organisms.", "ai_systems": [ "Evolutionary algorithm with neural-network controllers (Golem project)" ], "organizations": [ "Brandeis University" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature, 2000-08-31", "url": "https://doi.org/10.1038/35023115" } ], "foundation0": false }, { "id": "top7-first-designed-novel-protein-fold", "date": "2003-11-21", "date_basis": "publication", "title": "Rosetta designs Top7, a protein with a fold not seen in nature", "domain": "biology", "subfield": "computational protein design (pre-deep-learning)", "summary": "Baker's group used Rosetta, a program that alternates between designing a sequence and predicting its structure, to design Top7, a 93-residue protein with a new topology. Top7 folded and was extremely stable. Its X-ray structure matched the computer design to 1.2 angstroms RMSD.", "before_after": null, "why_it_matters": "A computer invented a working protein from first principles instead of editing a natural one. The 2024 Chemistry Nobel background describes Top7 as the first protein entirely different from all known ones.", "ai_systems": [ "Rosetta" ], "organizations": [ "University of Washington", "Howard Hughes Medical Institute" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Kuhlman et al., Science, 2003-11-21", "url": "https://doi.org/10.1126/science.1089427" }, { "label": "Nobel Prize in Chemistry 2024, popular information (Top7)", "url": "https://www.nobelprize.org/prizes/chemistry/2024/popular-information/" } ], "foundation0": false }, { "id": "stanley-darpa-grand-challenge-2005", "date": "2005-10-08", "date_basis": "competition", "title": "Stanford's Stanley wins the DARPA Grand Challenge desert race without a driver", "domain": "engineering-robotics", "subfield": "autonomous vehicles", "summary": "Stanford's robot car Stanley drove the 132-mile desert course from Primm, Nevada, autonomously in 6 h 53 min 58 s, first of five finishers. In the 2004 race no robot covered more than 5% of the course. Its software relied on machine learning, including terrain and speed models learned from human driving.", "before_after": { "metric": "Best robot progress on DARPA desert course", "before": "under 5% of the 2004 course", "after": "132-mile course finished in 6 h 54 min" }, "why_it_matters": "A founding moment for self-driving vehicles: learned perception combined with probabilistic reasoning made long-distance off-road autonomy work for the first time.", "ai_systems": [ "Stanley (learned terrain classification, probabilistic robotics)" ], "organizations": [ "Stanford University", "Volkswagen Electronics Research Laboratory", "Intel Research" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Journal of Field Robotics 23:661-692, 2006", "url": "https://doi.org/10.1002/rob.20147" }, { "label": "Thrun et al., author copy (Stanford)", "url": "https://robots.stanford.edu/papers/thrun.stanley05.pdf" }, { "label": "NBC News (AP) report, October 2005", "url": "https://www.nbcnews.com/id/wbna9621761" } ], "foundation0": false }, { "id": "nasa-st5-evolved-antenna", "date": "2006-03-22", "date_basis": "deployment", "title": "Computer-evolved antenna flies on NASA's ST5, first evolved hardware in space", "domain": "engineering-robotics", "subfield": "antenna design", "summary": "NASA Ames used evolutionary algorithms and an antenna simulator to design an irregular bent-wire antenna for the three ST5 microsatellites, redesigning it within a month when the mission's orbit changed. On a spacecraft mock-up, two evolved antennas reached 93% efficiency versus 38% for two conventional helix antennas. It launched on 22 March 2006.", "before_after": { "metric": "Efficiency of the antenna pair on an ST5 spacecraft mock-up", "before": "38% (two conventional quadrifilar helix antennas)", "after": "93% (two evolved antennas)" }, "why_it_matters": "Automated search found a shape no engineer would draw, beat the hand-made design, and flew. It showed computer-invented designs could become flight-qualified hardware, not just paper ideas.", "ai_systems": [ "Evolutionary algorithms (NASA Ames evolved-antenna software)" ], "organizations": [ "NASA Ames Research Center" ], "ai_role": "originator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "AIAA Space 2006 conference paper (Hornby et al.), 2006-09-19", "url": "https://doi.org/10.2514/6.2006-7242" }, { "label": "Full text, NASA Technical Reports Server", "url": "https://ntrs.nasa.gov/citations/20060024675" } ], "foundation0": false }, { "id": "behler-parrinello-neural-network-potentials", "date": "2007-04-02", "date_basis": "publication", "title": "Neural-network potentials give DFT-level forces far faster than DFT", "domain": "chemistry-materials", "subfield": "machine-learned interatomic potentials", "summary": "Jörg Behler and Michele Parrinello (ETH Zurich) introduced a neural-network representation of density-functional potential-energy surfaces that gives energies and forces for systems of any size, built from each atom's local environment. Tested on bulk silicon, it showed high accuracy against DFT while running several orders of magnitude faster.", "before_after": null, "why_it_matters": "Made quantum-accurate molecular dynamics of large systems affordable. Machine-learned potentials built on this idea are now widely used across chemistry and materials simulation.", "ai_systems": [ "Behler-Parrinello neural network potential" ], "organizations": [ "ETH Zurich" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Physical Review Letters 98:146401, 2007-04-02", "url": "https://doi.org/10.1103/PhysRevLett.98.146401" }, { "label": "PubMed record 17501293", "url": "https://pubmed.ncbi.nlm.nih.gov/17501293/" } ], "foundation0": false }, { "id": "checkers-solved-chinook", "date": "2007-07-19", "date_basis": "publication", "title": "Checkers solved: perfect play by both sides is a draw", "domain": "computing", "subfield": "game solving / search", "summary": "After dozens of computers had worked almost continuously since 1989, Jonathan Schaeffer's University of Alberta team proved that checkers, with about 5×10^20 positions, is a draw with perfect play. The proof combined search with endgame databases of 39 trillion positions, replacing the heuristics of game programs with certainty.", "before_after": { "metric": "Game-theoretic value of checkers", "before": "unknown (experts conjectured a draw)", "after": "proved a draw" }, "why_it_matters": "The most complex popular game solved to date, about a million times harder than Connect Four, and a computational result that is now settled.", "ai_systems": [ "Chinook" ], "organizations": [ "University of Alberta" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science 317:1518-1522 (online 2007-07-19)", "url": "https://doi.org/10.1126/science.1144079" }, { "label": "IEEE Spectrum, July 2007", "url": "https://spectrum.ieee.org/checkers-solved" } ], "foundation0": false }, { "id": "single-top-quark-observation-multivariate", "date": "2009-03-04", "date_basis": "preprint", "title": "Neural networks and decision trees help Tevatron observe single top quarks", "domain": "physics", "subfield": "particle physics (top quark)", "summary": "Fermilab's D0 and CDF experiments each observed single top-quark production at 5.0 standard deviations. Signal was only 3-9% of D0's selected events and smaller than the background uncertainty, so D0 combined boosted decision trees, Bayesian neural networks and matrix-element methods, and CDF combined analyses through a neuro-evolved neural network that added 13% sensitivity.", "before_after": { "metric": "Single top-quark significance", "before": "evidence: 3.7σ (CDF), 3.6σ (D0)", "after": "observation: 5.0σ in each experiment" }, "why_it_matters": "An early particle-physics discovery that depended on machine-learned classifiers, establishing methods later used for the Higgs boson.", "ai_systems": [ "Boosted decision trees", "Bayesian neural networks", "Neuro-evolved neural network (CDF)" ], "organizations": [ "Fermilab", "D0 Collaboration", "CDF Collaboration" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "D0 preprint, arXiv 0903.0850, 2009-03-04", "url": "https://arxiv.org/abs/0903.0850" }, { "label": "D0, Physical Review Letters 103:092001 (2009)", "url": "https://doi.org/10.1103/PhysRevLett.103.092001" }, { "label": "CDF, Physical Review Letters 103:092002 (2009)", "url": "https://doi.org/10.1103/PhysRevLett.103.092002" } ], "foundation0": false }, { "id": "adam-robot-scientist-yeast-genes", "date": "2009-04-03", "date_basis": "publication", "title": "Robot scientist Adam autonomously discovers gene functions in yeast", "domain": "science-of-science", "subfield": "automated hypothesis and experiment", "summary": "Adam, a laboratory robot, generated hypotheses about which yeast genes encode 'orphan' enzymes, designed and ran the experiments itself and interpreted the results; humans then confirmed its conclusions by hand. The study logged 6.6 million biomass measurements. Its makers called it the first machine to discover new scientific knowledge independently.", "before_after": null, "why_it_matters": "It was the first closed loop of hypothesis, experiment and conclusion run by a machine, the template for today's self-driving labs and AI scientists.", "ai_systems": [ "Adam (Robot Scientist)" ], "organizations": [ "Aberystwyth University", "University of Cambridge" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Science, 2009-04-03", "url": "https://doi.org/10.1126/science.1165620" }, { "label": "PubMed record", "url": "https://pubmed.ncbi.nlm.nih.gov/19342587/" } ], "foundation0": false }, { "id": "schmidt-lipson-distilling-natural-laws", "date": "2009-04-03", "date_basis": "publication", "title": "Symbolic regression distils physical laws from raw motion data", "domain": "physics", "subfield": "automated law discovery / symbolic regression", "summary": "Michael Schmidt and Hod Lipson (Cornell) built an algorithm that searches for equations describing motion-tracking data and judges which relations are non-trivial. They reported that, without prior physics knowledge, it recovered Hamiltonians, Lagrangians and conservation laws for systems from simple oscillators to chaotic double pendulums.", "before_after": null, "why_it_matters": "A landmark in automating the step from data to equations, and a direct ancestor of later tools such as AI Feynman.", "ai_systems": [ "Symbolic regression (genetic programming)" ], "organizations": [ "Cornell University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "Hillar and Sommer (arXiv 1210.7273, 2012) argued the method implicitly builds in Hamilton's equations and Newton's second law, which undercuts the claim that it found these laws without prior physics knowledge.", "significance": 3, "sources": [ { "label": "Science 324:81-85, 2009-04-03", "url": "https://doi.org/10.1126/science.1165893" }, { "label": "Hillar and Sommer, comment, arXiv 1210.7273 (2012)", "url": "https://arxiv.org/abs/1210.7273" } ], "foundation0": false }, { "id": "higgs-discovery-multivariate-ml", "date": "2012-07-04", "date_basis": "announcement", "title": "Machine-learned classifiers help ATLAS and CMS discover the Higgs boson", "domain": "physics", "subfield": "particle physics (Higgs boson)", "summary": "ATLAS and CMS announced a new boson near 125 GeV on 4 July 2012; their papers reported 5.9 and 5.0 standard deviations. CMS's two-photon search used boosted decision trees to select and classify events, gaining 15% sensitivity over a simpler cross-check analysis. ATLAS combined photon-identification variables in a neural network for its 2011 data.", "before_after": { "metric": "CMS H→γγ search sensitivity", "before": "simpler cut-based cross-check analysis", "after": "15% higher with boosted-decision-tree analysis" }, "why_it_matters": "The Higgs discovery completed the Standard Model and underpinned the 2013 Nobel Prize in Physics. Machine-learned classifiers were part of the toolkit that pulled the faint signal out of background.", "ai_systems": [ "Boosted decision trees (TMVA)", "Neural-network photon identification (ATLAS)" ], "organizations": [ "CERN", "CMS Collaboration", "ATLAS Collaboration" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "CMS, Physics Letters B 716:30 (2012)", "url": "https://doi.org/10.1016/j.physletb.2012.08.021" }, { "label": "ATLAS, Physics Letters B 716:1 (2012)", "url": "https://doi.org/10.1016/j.physletb.2012.08.020" }, { "label": "Nobel Prize in Physics 2013 summary", "url": "https://www.nobelprize.org/prizes/physics/2013/summary/" } ], "foundation0": false }, { "id": "flyspeck-kepler-conjecture-formal-proof", "date": "2014-08-10", "date_basis": "announcement", "title": "Flyspeck completes a formal proof of the Kepler sphere-packing conjecture", "domain": "mathematics", "subfield": "discrete geometry / formal verification", "summary": "Thomas Hales's 1998 proof that no packing of equal spheres beats the familiar cannonball stacking relied on huge computer calculations that a panel of 12 referees could only call '99% certain'. The Flyspeck project, begun in 2003, completed a formal proof in the HOL Light and Isabelle proof assistants on 10 August 2014.", "before_after": { "metric": "Certainty of the Kepler conjecture proof", "before": "referees '99% certain' (2003)", "after": "machine-checked formal proof (2014)" }, "why_it_matters": "Showed proof assistants can certify a major computer-dependent proof that human refereeing could not, a direct forerunner of today's Lean-checked results.", "ai_systems": [ "HOL Light", "Isabelle" ], "organizations": [ "University of Pittsburgh", "Flyspeck project collaborators" ], "ai_role": "instrument", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Forum of Mathematics, Pi 5:e2, 2017", "url": "https://doi.org/10.1017/fmp.2017.1" }, { "label": "arXiv 1501.02155, 2015-01-09", "url": "https://arxiv.org/abs/1501.02155" }, { "label": "Experimental Mathematics blog, 2014-08", "url": "https://experimentalmath.info/blog/2014/08/formal-proof-completed-for-keplers-conjecture-on-sphere-packing/" } ], "foundation0": false }, { "id": "eve-robot-scientist-antimalarial-leads", "date": "2015-03-06", "date_basis": "publication", "title": "Robot scientist Eve automates drug screening and finds antimalarial leads", "domain": "medicine", "subfield": "drug repositioning, lab automation", "summary": "Eve, a lab robot that learns structure-activity models and picks which compounds to test next, was shown by econometric modelling to beat standard screening on cost. It found the anti-cancer compound TNP-470 potently inhibits malaria-parasite DHFR; a 2018 Eve screen found triclosan also hits DHFR, including pyrimethamine-resistant forms.", "before_after": null, "why_it_matters": "Closing the loop between hypothesis and experiment inside a machine lowers the cost of searching for drugs against tropical diseases that markets neglect.", "ai_systems": [ "Eve (Robot Scientist)" ], "organizations": [ "Aberystwyth University", "University of Cambridge", "University of Manchester" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "J. R. Soc. Interface, 2015-03-06", "url": "https://doi.org/10.1098/rsif.2014.1289" }, { "label": "Scientific Reports (triclosan), 2018-01-18", "url": "https://doi.org/10.1038/s41598-018-19549-x" }, { "label": "University of Cambridge news, 2018-01-18", "url": "https://cam.ac.uk/research/news/ai-scientist-finds-that-toothpaste-ingredient-may-help-fight-drug-resistant-malaria" } ], "foundation0": false }, { "id": "melvin-automated-quantum-experiment-design", "date": "2016-03-04", "date_basis": "publication", "title": "Algorithm Melvin designs quantum optics experiments that humans had not found", "domain": "physics", "subfield": "quantum optics", "summary": "The search program Melvin combined optical elements into new set-ups for complex entangled photon states, often using asymmetric tricks its authors found hard to understand intuitively. One of its designs was built in Vienna, giving the first entangled state where both the number of photons and the dimensions exceed two (three photons, 3×3×2).", "before_after": { "metric": "multi-photon entanglement with more than two particles and more than two dimensions", "before": "never demonstrated", "after": "three-photon 3×3×2 state built from a computer-found design" }, "why_it_matters": "Experiments are usually invented by human intuition, which misses unusual layouts. A program that searches designs can reach set-ups people overlooked, which humans then build and test.", "ai_systems": [ "Melvin (automated experiment-search algorithm)" ], "organizations": [ "University of Vienna", "Austrian Academy of Sciences (IQOQI Vienna)" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Physical Review Letters, 2016-03-04", "url": "https://doi.org/10.1103/PhysRevLett.116.090405" }, { "label": "Nature Photonics (experiment), 2016-02-29", "url": "https://doi.org/10.1038/nphoton.2016.12" }, { "label": "arXiv preprint 1509.02749, 2015-09-09", "url": "https://arxiv.org/abs/1509.02749" } ], "foundation0": false }, { "id": "alphago-move-37", "date": "2016-03-10", "date_basis": "competition", "title": "AlphaGo's move 37 surprises Go experts and helps beat Lee Sedol", "domain": "computing", "subfield": "game-playing AI / machine creativity", "summary": "In game two of its match with Lee Sedol, winner of 18 world titles, DeepMind's AlphaGo played move 37, a move its own model rated as having about a 1 in 10,000 chance of being played by a human. Commentators were startled; Lee later called AlphaGo creative. AlphaGo won the game and the match 4-1.", "before_after": null, "why_it_matters": "Widely seen as the first time a machine produced an idea that experts recognised as both original and good, challenging long-held human strategy in a game studied for millennia.", "ai_systems": [ "AlphaGo" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind, AlphaGo page", "url": "https://deepmind.google/research/breakthroughs/alphago/" }, { "label": "Nature (AlphaGo method), 2016-01-27", "url": "https://doi.org/10.1038/nature16961" } ], "foundation0": false }, { "id": "boolean-pythagorean-triples-sat", "date": "2016-05-03", "date_basis": "preprint", "title": "SAT solver settles the Boolean Pythagorean triples problem", "domain": "mathematics", "subfield": "Ramsey theory / SAT solving", "summary": "Heule, Kullmann and Marek showed the integers 1 to 7,824 can be coloured red and blue with no single-colour Pythagorean triple (a²+b²=c²), but 1 to 7,825 cannot. The cube-and-conquer SAT search ran about two days on 800 cores and produced a roughly 200 TB proof, checkable from a 68 GB certificate.", "before_after": { "metric": "Largest n where 1..n has a 2-colouring with no monochromatic Pythagorean triple", "before": "Unknown (Ronald Graham's question, 1980s)", "after": "7,824" }, "why_it_matters": "A decades-old combinatorics question was settled by automated reasoning, with a proof that independent checking software can verify even though no human can read it.", "ai_systems": [ "Cube-and-conquer SAT solving", "DRAT proof checking" ], "organizations": [ "University of Texas at Austin", "Swansea University", "University of Kentucky" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv:1605.00723 (SAT 2016, LNCS 9710)", "url": "https://arxiv.org/abs/1605.00723" }, { "label": "Nature news, 2016-05-26", "url": "https://doi.org/10.1038/nature.2016.19990" } ], "foundation0": false }, { "id": "deepmind-datacenter-cooling-40pct", "date": "2016-07-20", "date_basis": "announcement", "title": "Machine learning cuts energy used to cool a Google data centre by up to 40%", "domain": "energy", "subfield": "data-centre efficiency", "summary": "DeepMind trained deep neural networks on historical data from thousands of data-centre sensors to recommend cooling settings, tested in a live Google data centre. Google reported up to 40% less cooling energy, a 15% cut in PUE overhead. By 2018 an autonomous version ran cooling in several data centres, saving about 30% on average.", "before_after": { "metric": "energy used for cooling", "before": "existing control (baseline)", "after": "up to 40% lower (15% lower PUE overhead)" }, "why_it_matters": "Data centres use a growing share of the world's electricity, and cooling is a large part of it. Learned control can save energy in complex plants without new hardware.", "ai_systems": [ "DeepMind deep neural network ensemble" ], "organizations": [ "DeepMind", "Google" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "DeepMind blog, 2016-07-20", "url": "https://deepmind.google/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40/" }, { "label": "DeepMind blog (autonomous control), 2018-08-17", "url": "https://deepmind.google/blog/safety-first-ai-for-autonomous-data-centre-cooling-and-industrial-control/" }, { "label": "Google blog, 2016-07-20", "url": "https://blog.google/company-news/outreach-and-initiatives/environment/deepmind-ai-reduces-energy-used-for/" } ], "foundation0": false }, { "id": "oled-virtual-screening-1-6m", "date": "2016-08-08", "date_basis": "publication", "title": "ML-aided virtual screen of 1.6 million molecules finds efficient OLED emitters", "domain": "chemistry-materials", "subfield": "organic electronics / virtual screening", "summary": "Harvard, MIT and Samsung researchers combined quantum chemistry, cheminformatics and a neural network to explore 1.6 million candidate OLED molecules, screening over 400,000 with time-dependent DFT. The team jointly picked the best candidates; those synthesized reached external quantum efficiencies as high as 22%.", "before_after": null, "why_it_matters": "Display and lighting materials are usually found by slow trial and error. Computing hundreds of thousands of molecules first, then making a few, became a model for molecular discovery.", "ai_systems": [ "Neural-network property predictor within a TD-DFT screening funnel" ], "organizations": [ "Harvard University", "MIT", "Samsung Advanced Institute of Technology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Materials, 2016-08-08", "url": "https://www.nature.com/articles/nmat4717" } ], "foundation0": false }, { "id": "gulshan-deep-learning-diabetic-retinopathy", "date": "2016-11-29", "date_basis": "publication", "title": "Deep learning detects diabetic eye disease in photos at eye-doctor level", "domain": "medicine", "subfield": "ophthalmology / medical imaging", "summary": "Google researchers trained a deep convolutional network on 128,175 retinal photographs graded by a panel of 54 US ophthalmologists and senior residents. On two validation sets it detected referable diabetic retinopathy with area under the curve of 0.991 and 0.990; at a high-specificity setting, sensitivity was 90.3% and specificity 98.1% on EyePACS-1.", "before_after": null, "why_it_matters": "With the skin-cancer study weeks later, it opened the medical-imaging AI era and led to screening deployments in clinics short of eye specialists.", "ai_systems": [ "Deep convolutional neural network" ], "organizations": [ "Google", "EyePACS", "Aravind Eye Care System", "Sankara Nethralaya" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "JAMA 316:2402, published online 2016-11-29", "url": "https://doi.org/10.1001/jama.2016.17216" }, { "label": "Google Research blog, 2016-11-29", "url": "https://research.google/blog/deep-learning-for-detection-of-diabetic-eye-disease/" } ], "foundation0": false }, { "id": "esteva-dermatologist-level-skin-cancer", "date": "2017-01-25", "date_basis": "publication", "title": "Neural network classifies skin cancer on par with 21 dermatologists", "domain": "medicine", "subfield": "dermatology / medical imaging", "summary": "Stanford researchers trained a single convolutional neural network end to end on 129,450 clinical images covering 2,032 diseases. Tested against 21 board-certified dermatologists on biopsy-proven images, it performed on par with all of them at separating carcinomas from benign keratoses and melanomas from benign moles.", "before_after": null, "why_it_matters": "Showed one network trained only on pixels and labels could match specialists at a common visual diagnosis, suggesting phones could extend skin-cancer screening beyond the clinic.", "ai_systems": [ "Deep convolutional neural network" ], "organizations": [ "Stanford University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature 542:115-118, 2017-01-25", "url": "https://doi.org/10.1038/nature21056" }, { "label": "PubMed record 28117445", "url": "https://pubmed.ncbi.nlm.nih.gov/28117445/" } ], "foundation0": false }, { "id": "neural-network-quantum-states", "date": "2017-02-10", "date_basis": "publication", "title": "Neural networks used to represent quantum many-body wavefunctions", "domain": "physics", "subfield": "quantum many-body physics", "summary": "Carleo and Troyer showed that a neural network (a restricted Boltzmann machine) can serve as a variational wavefunction for interacting quantum spins. Trained with a reinforcement-learning scheme, it found ground states and simulated time evolution of prototypical 1D and 2D spin models with high accuracy, launching the 'neural quantum states' method.", "before_after": null, "why_it_matters": "The information needed to describe many interacting quantum particles grows exponentially. Neural networks give compact descriptions, a new general tool for magnets, materials and molecules.", "ai_systems": [ "Restricted Boltzmann machine (neural quantum state)" ], "organizations": [ "ETH Zurich", "Microsoft Research" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Science, 2017-02-10", "url": "https://doi.org/10.1126/science.aag2302" }, { "label": "arXiv preprint 1606.02318, 2016-06-07", "url": "https://arxiv.org/abs/1606.02318" } ], "foundation0": false }, { "id": "aegis-curiosity-autonomous-targeting", "date": "2017-06", "date_basis": "publication", "title": "AEGIS lets the Curiosity rover pick its own laser targets on Mars", "domain": "astronomy-space", "subfield": "planetary rover autonomy", "summary": "JPL's AEGIS software, in routine use on Curiosity since May 2016, finds rocks in navigation-camera images and fires the ChemCam laser spectrometer at those best matching scientists' criteria, with no Earth in the loop. Over 2.5 km of new terrain it chose the desired material over 93% of the time, against about 24% without it.", "before_after": { "metric": "Desired target material selected", "before": "about 24% (blind targeting)", "after": "over 93% with AEGIS" }, "why_it_matters": "Lets a rover do targeted science in hours that would otherwise be lost waiting for instructions from Earth, and changed how the mission's scientists explore.", "ai_systems": [ "AEGIS" ], "organizations": [ "NASA Jet Propulsion Laboratory", "Los Alamos National Laboratory" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science Robotics 2:eaan4582, June 2017", "url": "https://doi.org/10.1126/scirobotics.aan4582" }, { "label": "JPL news, 2017-06-21", "url": "https://www.jpl.nasa.gov/news/laser-targeting-ai-yields-more-mars-science/" } ], "foundation0": false }, { "id": "alphago-zero-tabula-rasa", "date": "2017-10-19", "date_basis": "publication", "title": "AlphaGo Zero masters Go from self-play alone and beats AlphaGo 100-0", "domain": "computing", "subfield": "reinforcement learning / game playing", "summary": "DeepMind's AlphaGo Zero learned Go with no human game records, starting from random play and given only the rules. After three days of self-play reinforcement learning it beat the version that defeated Lee Sedol by 100 games to 0, and after 40 days it surpassed AlphaGo Master.", "before_after": { "metric": "Result against AlphaGo Lee", "before": "AlphaGo needed human expert games to learn", "after": "100-0 win after 3 days of self-play" }, "why_it_matters": "Showed superhuman strategy can emerge without any human data, removing a key bottleneck and seeding the self-play methods later used in AlphaTensor, AlphaDev and AlphaProof.", "ai_systems": [ "AlphaGo Zero" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature 550:354, 2017-10-19", "url": "https://doi.org/10.1038/nature24270" }, { "label": "Google DeepMind blog, AlphaGo Zero", "url": "https://deepmind.google/blog/alphago-zero-starting-from-scratch/" } ], "foundation0": false }, { "id": "learned-index-structures", "date": "2017-12-04", "date_basis": "preprint", "title": "Learned models stand in for B-tree indexes, with faster lookups in tests", "domain": "computing", "subfield": "databases / data structures", "summary": "Google and MIT researchers showed that the indexes databases use to find records can be replaced by models that predict where a record sits. On several real datasets, their learned indexes ran up to 70% faster than cache-optimized B-trees while using about ten times less memory, on read-only data.", "before_after": { "metric": "Lookup speed and memory vs cache-optimized B-tree", "before": "B-tree baseline", "after": "Up to 70% faster, ~10x less memory" }, "why_it_matters": "Indexes sit under nearly every database and search engine. Treating them as prediction problems started a research field on letting the data itself shape core system components.", "ai_systems": [ "Learned index (neural network and recursive model index)" ], "organizations": [ "Google", "MIT" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "Thomas Neumann (TU Munich, Database Architects blog, 2017-12-23) argued the B-tree baselines were weak and that an interpolating B-tree matches the reported speed. A later independent benchmark (Marcus et al., VLDB 2020) found learned indexes do beat tuned traditional indexes, but on read-only, in-memory data.", "significance": 3, "sources": [ { "label": "arXiv preprint, 2017-12-04", "url": "https://arxiv.org/abs/1712.01208" }, { "label": "SIGMOD 2018 (peer-reviewed version), 2018-05-27", "url": "https://doi.org/10.1145/3183713.3196909" } ], "foundation0": false }, { "id": "alphazero-chess-shogi-go", "date": "2017-12-05", "date_basis": "preprint", "title": "AlphaZero learns chess, shogi and Go from scratch and beats top engines", "domain": "computing", "subfield": "reinforcement learning / game playing", "summary": "A single algorithm, given only the rules, reached superhuman play in chess, shogi and Go within 24 hours of self-play (2017 preprint). In the peer-reviewed 1,000-game match (Science, December 2018) it beat Stockfish, the 2016 computer chess champion, winning 155 games and losing 6. Grandmasters studied its dynamic, unorthodox style for new ideas.", "before_after": { "metric": "Chess match vs Stockfish (1,000 games)", "before": "engines built on decades of hand-tuned evaluation", "after": "155 wins, 6 losses, rest drawn" }, "why_it_matters": "One general learning method replaced decades of game-specific engineering, and its play gave human experts fresh ideas in a game studied for centuries.", "ai_systems": [ "AlphaZero" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv 1712.01815, 2017-12-05", "url": "https://arxiv.org/abs/1712.01815" }, { "label": "Science 362:1140, 2018-12-07", "url": "https://doi.org/10.1126/science.aar6404" }, { "label": "Google DeepMind blog, AlphaZero", "url": "https://deepmind.google/blog/alphazero-shedding-new-light-on-chess-shogi-and-go/" } ], "foundation0": false }, { "id": "kepler-90i-kepler-80g-neural-network", "date": "2017-12-14", "date_basis": "announcement", "title": "Neural network finds new planets around Kepler-90 and Kepler-80", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "A convolutional neural network trained on labelled Kepler signals searched weak, previously overlooked transit signals and found Kepler-90i and Kepler-80g, both statistically validated as planets. Kepler-90 became the first known system with eight planets, tying the Solar System. On test data the model ranked real planets above false positives 98.8% of the time.", "before_after": { "metric": "Known planets orbiting Kepler-90", "before": "7", "after": "8 (ties the Solar System)" }, "why_it_matters": "Telescopes record far more faint signals than people can inspect. A trained classifier can reopen old data and pull out real planets that human-tuned searches missed.", "ai_systems": [ "AstroNet (convolutional neural network)" ], "organizations": [ "Google", "University of Texas at Austin", "NASA" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "The Astronomical Journal 155:94 (Shallue & Vanderburg), 2018", "url": "https://doi.org/10.3847/1538-3881/aa9e09" }, { "label": "arXiv:1712.05044, 2017-12-13", "url": "https://arxiv.org/abs/1712.05044" }, { "label": "NASA/JPL news, 2017-12-14", "url": "https://www.jpl.nasa.gov/news/artificial-intelligence-nasa-data-used-to-discover-eighth-planet-circling-distant-star/" } ], "foundation0": false }, { "id": "chematica-eight-routes-executed", "date": "2018-03", "date_basis": "publication", "title": "Computer-planned syntheses of eight medicinal targets all succeed in the lab", "domain": "chemistry-materials", "subfield": "retrosynthesis planning", "summary": "The Chematica program (later sold as Synthia) autonomously designed routes to eight diverse targets: seven commercially valuable bioactive substances and one natural product. Chemists executed all eight routes in the lab; they improved yields or costs, offered alternatives to patented routes, or made targets never synthesized before.", "before_after": null, "why_it_matters": "It showed computer synthesis planning could do more than suggest single steps: complete routes worked at the bench and could beat human-designed ones on cost and yield.", "ai_systems": [ "Chematica (Synthia)" ], "organizations": [ "Polish Academy of Sciences", "UNIST / IBS Center for Soft and Living Matter", "Northwestern University", "MilliporeSigma" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Chem, March 2018", "url": "https://doi.org/10.1016/j.chempr.2018.02.002" } ], "foundation0": false }, { "id": "segler-3n-mcts-retrosynthesis", "date": "2018-03-29", "date_basis": "publication", "title": "Deep networks plus tree search plan syntheses chemists rate equal to published", "domain": "chemistry-materials", "subfield": "retrosynthesis planning", "summary": "Segler, Preuss and Waller trained deep neural networks on essentially all published organic reactions and combined them with Monte Carlo tree search. The system solved almost twice as many molecules, thirty times faster, than rule-based computer planning, and in a double-blind test chemists judged its routes equivalent to published ones.", "before_after": { "metric": "Target molecules solved vs. rule-based computer search", "before": "baseline (hand-coded rules and heuristics)", "after": "almost 2x as many, 30x faster" }, "why_it_matters": "Working out how to make a molecule is a core bottleneck in chemistry. Learning the rules from data instead of hand-coding them made automated synthesis planning practical.", "ai_systems": [ "3N-MCTS (expansion policy and filter networks with Monte Carlo tree search)" ], "organizations": [ "University of Münster", "Shanghai University", "BenevolentAI" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature, 2018-03-29", "url": "https://www.nature.com/articles/nature25978" } ], "foundation0": false }, { "id": "idx-dr-first-autonomous-ai-diagnostic", "date": "2018-04-11", "date_basis": "announcement", "title": "FDA authorizes first autonomous AI diagnostic system (diabetic eye disease)", "domain": "medicine", "subfield": "ophthalmology screening", "summary": "IDx-DR became the first FDA-authorized system that gives a screening decision without a clinician reading the images. In a 900-person pivotal trial in primary-care clinics it detected more-than-mild diabetic retinopathy with 87.2% sensitivity and 90.7% specificity, beating prespecified targets.", "before_after": { "metric": "Sensitivity / specificity vs pre-set trial targets", "before": "targets >85% / >82.5%", "after": "87.2% / 90.7%" }, "why_it_matters": "Diabetic eye disease causes preventable blindness. Letting primary-care clinics screen without an eye specialist brings screening to people who rarely see one.", "ai_systems": [ "IDx-DR" ], "organizations": [ "IDx LLC", "University of Iowa", "US FDA" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "FDA De Novo DEN180001 record (decision 2018-04-11)", "url": "https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN180001" }, { "label": "npj Digital Medicine, 2018-08-28", "url": "https://doi.org/10.1038/s41746-018-0040-6" } ], "foundation0": false }, { "id": "metallic-glasses-ml-high-throughput", "date": "2018-04-13", "date_basis": "publication", "title": "Machine learning plus fast experiments find new metallic-glass systems", "domain": "chemistry-materials", "subfield": "alloys / metallic glasses", "summary": "SLAC, Northwestern and NIST researchers trained a model on published data to predict where metallic glasses form, tested it with high-throughput experiments, and retrained on the results. Iterating led to three new glass-forming alloy systems and raised the hit rate from about 1 in 300-400 samples to about 1 in 2-3.", "before_after": { "metric": "Share of tested samples that were metallic glasses", "before": "about 1 in 300-400", "after": "about 1 in 2-3 (third round)" }, "why_it_matters": "Metallic glasses are strong, corrosion-resistant alloys that theory predicts poorly. Closing the loop between prediction and experiment lets learning compensate where physics models fall short.", "ai_systems": [ "Machine-learning glass-formation model, iteratively retrained" ], "organizations": [ "SLAC National Accelerator Laboratory", "Northwestern University", "NIST" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Science Advances, 2018-04-13", "url": "https://doi.org/10.1126/sciadv.aaq1566" }, { "label": "Northwestern University news, 2018-04-13", "url": "https://news.northwestern.edu/stories/2018/april/artificial-intelligence-accelerates-discovery-of-metallic-glass" } ], "foundation0": false }, { "id": "snapshot-serengeti-deep-learning-camera-traps", "date": "2018-06-05", "date_basis": "publication", "title": "Deep learning identifies wildlife in 3.2 million camera-trap images", "domain": "biology", "subfield": "ecology, camera-trap species identification", "summary": "Deep neural networks learned to identify, count and describe 48 species in the 3.2-million-image Snapshot Serengeti set with over 93.8% accuracy. Keeping only confident calls automated 99.3% of images at 96.6% accuracy, matching volunteer crowds and saving over 8.4 years of human labelling.", "before_after": { "metric": "Images labelled automatically at volunteer-level accuracy (96.6%)", "before": "0% (all labelled by volunteers)", "after": "99.3% of 3.2 million images" }, "why_it_matters": "Camera traps gather far more images than experts can label. Automatic species counts turn them into data on animal numbers and behaviour, cheap enough to monitor whole ecosystems.", "ai_systems": [ "Deep convolutional neural networks" ], "organizations": [ "University of Wyoming", "Auburn University", "University of Minnesota", "Uber AI Labs" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Norouzzadeh et al., PNAS, 2018-06-05", "url": "https://doi.org/10.1073/pnas.1719367115" } ], "foundation0": false }, { "id": "higgs-bottom-quark-decay-ml", "date": "2018-08-28", "date_basis": "announcement", "title": "Machine learning helps ATLAS and CMS observe Higgs decay to bottom quarks", "domain": "physics", "subfield": "particle physics (Higgs boson)", "summary": "ATLAS and CMS observed the Higgs boson decaying to bottom-quark pairs, its most common decay but buried under huge backgrounds. Both analyses used machine learning to tag b-quark jets and to separate signal from background. ATLAS reached 5.4 standard deviations and CMS 5.6 (5.5 expected).", "before_after": { "metric": "H→bb̄ significance", "before": "evidence level only (2017 results)", "after": "observation: 5.4σ (ATLAS), 5.6σ (CMS)" }, "why_it_matters": "Confirms that the Higgs field gives mass to bottom quarks, as predicted. Shows machine learning can pull rare signals out of overwhelming background, now standard practice in collider physics.", "ai_systems": [ "Boosted decision trees", "Deep neural networks (b-tagging, event classification)" ], "organizations": [ "CERN", "ATLAS Collaboration", "CMS Collaboration" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Physical Review Letters (CMS), 2018-09-17", "url": "https://doi.org/10.1103/PhysRevLett.121.121801" }, { "label": "ATLAS press statement, 2018-08-28", "url": "https://atlas-public.web.cern.ch/updates/press-statement/observation-higgs-boson-decay-pair-bottom-quarks" } ], "foundation0": false }, { "id": "frb-121102-cnn-72-new-pulses", "date": "2018-09-10", "date_basis": "announcement", "title": "Neural network finds 72 missed pulses from repeating fast radio burst FRB 121102", "domain": "astronomy-space", "subfield": "radio transients", "summary": "Breakthrough Listen researchers trained a convolutional neural network on known bursts and re-searched five hours of Green Bank Telescope data from 26 August 2017. It found 72 pulses the classical search missed, raising the total from 21 to 93, and let the team rule out regular periods longer than about 10 ms.", "before_after": { "metric": "FRB 121102 pulses found in the 26 Aug 2017 observation", "before": "21", "after": "93" }, "why_it_matters": "Fast radio bursts are brief and rare. Finding four times more pulses in existing data gives far more evidence about what produces them, without new telescope time.", "ai_systems": [ "Convolutional neural network (Zhang et al.)" ], "organizations": [ "UC Berkeley", "Breakthrough Listen", "Green Bank Observatory" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Zhang et al., The Astrophysical Journal 866:149, 2018", "url": "https://doi.org/10.3847/1538-4357/aadf31" }, { "label": "arXiv:1809.03043, 2018-09-09", "url": "https://arxiv.org/abs/1809.03043" }, { "label": "Breakthrough Initiatives news, 2018-09-10", "url": "https://breakthroughinitiatives.org/news/22" } ], "foundation0": false }, { "id": "deepvariant-deep-learning-variant-calling", "date": "2018-09-24", "date_basis": "publication", "title": "DeepVariant uses a neural network to call genome variants from read images", "domain": "biology", "subfield": "genomics, variant calling", "summary": "Google and Verily's DeepVariant turns sequencing reads at candidate sites into images and uses a convolutional neural network to call SNPs and small indels. An early version won the highest SNP performance award in the 2016 precisionFDA Truth Challenge. The 2018 paper showed one approach works across genome builds, mammal species and sequencing platforms.", "before_after": null, "why_it_matters": "Every genetic study starts by deciding which genome differences are real. A learned caller beat existing state-of-the-art tools and runs in production, such as Regeneron's analysis of UK Biobank exomes, per NVIDIA.", "ai_systems": [ "DeepVariant (convolutional neural network)" ], "organizations": [ "Google", "Verily Life Sciences" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Poplin et al., Nature Biotechnology, 2018-09-24", "url": "https://doi.org/10.1038/nbt.4235" }, { "label": "Poplin et al., bioRxiv preprint v2 (reports the 2016 precisionFDA Truth Challenge SNP award)", "url": "https://www.biorxiv.org/content/10.1101/092890v2.full" }, { "label": "NVIDIA developer blog on Regeneron's custom DeepVariant for UK Biobank exomes, 2023-05-15", "url": "https://developer.nvidia.com/blog/accelerate-whole-exome-analysis-with-deep-learning-at-70-cost-reduction-using-nvidia-parabricks" } ], "foundation0": false }, { "id": "alphafold-casp13-first-deep-learning-lead", "date": "2018-12", "date_basis": "competition", "title": "AlphaFold wins CASP13 on hard targets: 24 of 43 at TM-score 0.7 or more", "domain": "biology", "subfield": "protein structure prediction", "summary": "DeepMind's first AlphaFold trained a neural network to predict distances between pairs of amino acids, then optimized a protein shape to fit them. In the blind CASP13 test it ranked first on free-modelling targets, giving high-accuracy structures (TM-score 0.7 or more) for 24 of 43 domains against 14 for the next best method.", "before_after": { "metric": "Free-modelling domains at TM-score 0.7 or more (CASP13)", "before": "14 of 43 (next best method)", "after": "24 of 43 (AlphaFold)" }, "why_it_matters": "Deep learning beat every other CASP13 entry at predicting the shape of hard protein targets, setting up AlphaFold 2 two years later.", "ai_systems": [ "AlphaFold (version 1)" ], "organizations": [ "DeepMind", "Francis Crick Institute", "University College London" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Senior et al., Nature, 2020-01-15", "url": "https://doi.org/10.1038/s41586-019-1923-7" }, { "label": "Senior et al., Proteins, CASP13 system description, 2019-11-11", "url": "https://doi.org/10.1002/prot.25834" }, { "label": "CASP13 conference page, 1-4 December 2018", "url": "https://predictioncenter.org/casp13/index.cgi" } ], "foundation0": false }, { "id": "spliceai-cryptic-splice-variants", "date": "2019-01-17", "date_basis": "publication", "title": "SpliceAI reads RNA sequence to flag hidden splice-altering mutations", "domain": "biology", "subfield": "genomics, variant interpretation", "summary": "Illumina's SpliceAI, a deep neural network, predicts splice junctions from transcript sequence and flags noncoding variants that cause cryptic splicing. Predicted splice-altering de novo variants were enriched in autism and intellectual-disability patients and validated in RNA-seq for 21 of 28. The authors estimate 9-11% of pathogenic mutations in rare disorders act this way.", "before_after": null, "why_it_matters": "Most disease-linked variants sit outside protein-coding sequence and are hard to interpret. Reading splicing from sequence gives geneticists a way to weigh variants that were previously overlooked.", "ai_systems": [ "SpliceAI" ], "organizations": [ "Illumina", "University of California, San Francisco", "Stanford University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Jaganathan et al., Cell, 2019-01-17", "url": "https://doi.org/10.1016/j.cell.2018.12.015" } ], "foundation0": false }, { "id": "deepmind-wind-power-forecast", "date": "2019-02-26", "date_basis": "announcement", "title": "Neural-network forecasts raise value of Google's wind power by about 20%", "domain": "energy", "subfield": "renewable energy forecasting", "summary": "A neural network trained on weather forecasts and turbine history predicted output from 700 MW of Google's US wind farms 36 hours ahead, letting Google make hourly delivery commitments to the grid a day in advance. DeepMind reported this raised the value of the wind energy by roughly 20% versus making no time-based commitments.", "before_after": { "metric": "value of wind energy", "before": "no time-based grid commitments (baseline)", "after": "about 20% higher" }, "why_it_matters": "Wind output varies, so grids value it less. Reliable day-ahead forecasts make renewables more schedulable and more valuable, easing their integration into power systems.", "ai_systems": [ "DeepMind neural-network wind forecaster" ], "organizations": [ "DeepMind", "Google" ], "ai_role": "instrument", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "DeepMind blog, 2019-02-26", "url": "https://deepmind.google/blog/machine-learning-can-boost-the-value-of-wind-energy/" }, { "label": "MIT Technology Review, 2019-02-27", "url": "https://www.technologyreview.com/2019/02/27/239459/deepmind-creates-algorithm-to-squeeze-more-out-of-wind-power/" } ], "foundation0": false }, { "id": "frnn-tokamak-disruption-prediction", "date": "2019-04-17", "date_basis": "publication", "title": "Deep learning predicts fusion plasma disruptions, including across machines", "domain": "energy", "subfield": "fusion plasma disruption prediction", "summary": "The Fusion Recurrent Neural Network, trained on about 2 TB of JET and DIII-D data, warned of disruptions 30 ms ahead. On JET its AUC was 0.952 versus 0.893 for the best classical method; on DIII-D the classical method did slightly better. Trained only on DIII-D, it reached 0.817 on JET versus 0.616.", "before_after": { "metric": "disruption-prediction AUC on JET, 30 ms warning", "before": "0.893 (best classical model)", "after": "0.952 (FRNN)" }, "why_it_matters": "Disruptions can damage large reactors like ITER, which cannot afford many training failures. Predictors that transfer from smaller machines help protect bigger ones.", "ai_systems": [ "FRNN (Fusion Recurrent Neural Network)" ], "organizations": [ "Princeton Plasma Physics Laboratory", "Princeton University", "Harvard University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2019-04-17", "url": "https://doi.org/10.1038/s41586-019-1116-4" }, { "label": "Accepted manuscript (OSTI), 2019", "url": "https://www.osti.gov/biblio/1507546" }, { "label": "Harvard Gazette, 2019-04-22", "url": "https://news.harvard.edu/gazette/story/2019/04/harvard-princeton-scientists-make-ai-breakthrough-for-fusion-energy/" } ], "foundation0": false }, { "id": "enso-deep-learning-18-month-forecasts", "date": "2019-09-18", "date_basis": "publication", "title": "Deep learning forecasts El Niño up to 18 months ahead", "domain": "earth-climate", "subfield": "seasonal climate prediction", "summary": "A convolutional neural network, trained first on climate-model simulations and then on 1871-1973 observations, forecast the El Niño/Southern Oscillation with useful skill up to one and a half years ahead. Over 1984-2017 its Niño3.4 correlation skill was much higher than leading dynamical forecast systems, and it predicted the east-west pattern of sea temperatures better.", "before_after": { "metric": "Useful ENSO forecast lead time", "before": "about 1 year (dynamical models)", "after": "up to 1.5 years" }, "why_it_matters": "El Niño shifts droughts, floods and harvests worldwide. Several extra months of warning gives farmers, water managers and aid agencies time to prepare.", "ai_systems": [ "Convolutional neural network with transfer learning" ], "organizations": [ "Chonnam National University", "Nanjing University of Information Science and Technology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Ham, Kim & Luo, Nature, 2019-09-18", "url": "https://www.nature.com/articles/s41586-019-1559-7" } ], "foundation0": false }, { "id": "dactyl-rubiks-cube-sim-to-real", "date": "2019-10-16", "date_basis": "preprint", "title": "Robot hand trained only in simulation manipulates a Rubik's Cube to solution", "domain": "engineering-robotics", "subfield": "robot learning, dexterous manipulation", "summary": "OpenAI trained neural networks for a five-fingered robot hand entirely in simulation, using automatic domain randomization to create ever-harder virtual worlds. In 10 real-world trials its best policy did the equivalent of solving 15-turn scrambles 60% of the time and the hardest 26-turn scrambles 20%. A standard non-learned solver chose the moves.", "before_after": null, "why_it_matters": "Learning in simulation and transferring to messy reality is how robots could gain dexterous lab and factory skills without years of trial and error on physical hardware.", "ai_systems": [ "Dactyl", "Automatic domain randomization (ADR)" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv preprint, 2019-10-16", "url": "https://arxiv.org/abs/1910.07113" } ], "foundation0": false }, { "id": "xenobots-ai-designed-organisms", "date": "2020-01-13", "date_basis": "publication", "title": "AI-designed living machines built from frog cells (xenobots)", "domain": "engineering-robotics", "subfield": "AI-designed organisms", "summary": "An evolutionary algorithm searched simulated body plans made of passive skin and contracting heart-muscle units for designs that move. Biologists then built the best design from Xenopus frog cells. The six organisms built moved in the direction the simulation predicted, and upside-down versions moved much less, as predicted.", "before_after": null, "why_it_matters": "Computers designed a new kind of living machine, not just a robot, opening a route to custom biological machines for medicine or environmental cleanup.", "ai_systems": [ "Evolutionary algorithm in a physics simulator" ], "organizations": [ "University of Vermont", "Tufts University", "Wyss Institute, Harvard University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "PNAS, 2020-01-13", "url": "https://doi.org/10.1073/pnas.1910837117" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC6994979/" } ], "foundation0": false }, { "id": "dsp-1181-first-ai-designed-drug-in-clinic", "date": "2020-01-30", "date_basis": "announcement", "title": "First AI-designed drug candidate (DSP-1181) enters human trials", "domain": "medicine", "subfield": "drug design", "summary": "Sumitomo Dainippon Pharma and Exscientia began a phase 1 trial in Japan of DSP-1181, a long-acting serotonin 5-HT1A agonist for obsessive-compulsive disorder. They said AI-assisted design finished the exploratory phase in under 12 months versus a typical 4.5 years. No efficacy data were released, and the drug was discontinued after phase 1.", "before_after": { "metric": "Exploratory research phase (company figures)", "before": "~4.5 years (typical)", "after": "<12 months" }, "why_it_matters": "Marked the start of AI-designed molecules being tested in people. Its later discontinuation is a reminder that faster design does not guarantee a drug that works.", "ai_systems": [ "Exscientia Centaur Chemist" ], "organizations": [ "Exscientia", "Sumitomo Dainippon Pharma" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Sumitomo Dainippon Pharma press release, 2020-01-30", "url": "https://www.sumitomo-pharma.com/news/20200130.html" }, { "label": "pharmaphorum, 2020-01-30", "url": "https://pharmaphorum.com/news/exscientia-claims-world-first-as-ai-created-drug-enters-clinic" }, { "label": "Pharmaceuticals (review noting discontinuation after phase 1), 2025-08-26", "url": "https://doi.org/10.3390/ph18091271" } ], "foundation0": false }, { "id": "baricitinib-covid19-knowledge-graph", "date": "2020-02-04", "date_basis": "publication", "title": "AI knowledge graph flags baricitinib for COVID-19; later trials confirm benefit", "domain": "medicine", "subfield": "drug repurposing, infectious disease", "summary": "In early February 2020, BenevolentAI used its biomedical knowledge graph to propose baricitinib, an approved arthritis drug, against COVID-19. Large randomized trials later confirmed benefit: in RECOVERY (8,156 hospitalized patients) 28-day deaths fell from 14% to 12%, a 13% proportional reduction.", "before_after": { "metric": "28-day mortality, hospitalized COVID-19 (RECOVERY)", "before": "14% (usual care)", "after": "12% (plus baricitinib)" }, "why_it_matters": "Shows machine-assisted literature mining can point to an existing, safe drug within days of a new disease, with randomized trials then doing the confirming.", "ai_systems": [ "BenevolentAI knowledge graph" ], "organizations": [ "BenevolentAI", "Imperial College London", "RECOVERY Collaborative Group", "Eli Lilly" ], "ai_role": "collaborator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "The Lancet (letter), 2020-02-04", "url": "https://doi.org/10.1016/S0140-6736(20)30304-4" }, { "label": "The Lancet (RECOVERY baricitinib), 2022-07-30", "url": "https://doi.org/10.1016/S0140-6736(22)01109-6" }, { "label": "Lancet Respiratory Medicine (COV-BARRIER), 2021-09-01", "url": "https://doi.org/10.1016/S2213-2600(21)00331-3" } ], "foundation0": false }, { "id": "battery-fast-charging-closed-loop-ml", "date": "2020-02-19", "date_basis": "publication", "title": "Closed-loop ML finds long-life battery fast-charging protocols in 16 days", "domain": "energy", "subfield": "lithium-ion batteries", "summary": "Researchers paired a model that predicts battery lifetime from early cycles (9.1% error using the first 100 cycles, 2019) with Bayesian optimization to search 224 six-step, ten-minute fast-charging protocols. It identified high-cycle-life protocols in 16 days, versus over 500 days for exhaustive testing without early prediction.", "before_after": { "metric": "time to optimize 224 fast-charging protocols", "before": "over 500 days (exhaustive search)", "after": "16 days" }, "why_it_matters": "Battery tests take months because cells must wear out. Predicting lifetime early and choosing experiments smartly can cut years from energy-storage development.", "ai_systems": [ "Bayesian optimization", "Early cycle-life prediction model" ], "organizations": [ "Stanford University", "MIT", "Toyota Research Institute" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2020-02-19", "url": "https://doi.org/10.1038/s41586-020-1994-5" }, { "label": "Nature Energy (early prediction), 2019-03-25", "url": "https://doi.org/10.1038/s41560-019-0356-8" } ], "foundation0": false }, { "id": "halicin-deep-learning-antibiotic", "date": "2020-02-20", "date_basis": "publication", "title": "Deep learning finds halicin, an antibiotic unlike known antibiotic classes", "domain": "medicine", "subfield": "antibiotic discovery", "summary": "A neural network trained to predict antibacterial activity picked halicin from a drug-repurposing library. It killed a broad range of pathogens, including carbapenem-resistant Enterobacteriaceae, and treated pan-resistant A. baumannii and C. difficile infections in mice. Screening over 107 million virtual molecules, 8 of 23 tested predictions were antibacterial.", "before_after": null, "why_it_matters": "New antibiotic classes are rare while resistance spreads. A model that searches chemical space far faster than wet-lab screening opens a new route to discovery.", "ai_systems": [ "Chemprop (directed message-passing neural network)" ], "organizations": [ "MIT", "Broad Institute", "McMaster University", "Wyss Institute, Harvard" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Cell, 2020-02-20", "url": "https://doi.org/10.1016/j.cell.2020.01.021" } ], "foundation0": false }, { "id": "ai-feynman-symbolic-regression", "date": "2020-04-15", "date_basis": "publication", "title": "AI Feynman recovers all 100 Feynman Lectures equations from data", "domain": "physics", "subfield": "symbolic regression / law discovery", "summary": "MIT's AI Feynman combined neural-network fitting with physics-inspired tricks (symmetry, separability and dimensional analysis) to find formulas from data. Given data sampled from the 100 equations in the Feynman Lectures on Physics, it recovered all of them, where prior public software found 71, and raised success on a harder set from 15% to 90%.", "before_after": { "metric": "Feynman Lectures equations recovered from data", "before": "71 of 100 (prior public software)", "after": "100 of 100" }, "why_it_matters": "Physics advances by finding compact laws behind data. Automating that step could help find formulas in new data where humans have no starting guess.", "ai_systems": [ "AI Feynman" ], "organizations": [ "MIT" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science Advances, 2020-04-15", "url": "https://doi.org/10.1126/sciadv.aay2631" }, { "label": "arXiv preprint 1905.11481", "url": "https://arxiv.org/abs/1905.11481" } ], "foundation0": false }, { "id": "ai-colonoscopy-polyp-detection-rct", "date": "2020-05-01", "date_basis": "publication", "title": "Real-time AI polyp detection raises adenoma detection in colonoscopy trial", "domain": "medicine", "subfield": "gastroenterology, cancer prevention", "summary": "In a 685-patient randomized trial at three Italian centres, the GI Genius system drawing boxes around suspected lesions in real time raised the adenoma detection rate from 40.4% to 54.8%, with more adenomas per colonoscopy and no longer withdrawal time.", "before_after": { "metric": "Adenoma detection rate", "before": "40.4% (standard colonoscopy)", "after": "54.8% (with AI)" }, "why_it_matters": "About a quarter of precancerous colon growths are missed at colonoscopy. Catching more could prevent cancers, if the gain lasts and doctors keep their own skills.", "ai_systems": [ "GI Genius (Medtronic)" ], "organizations": [ "Humanitas Research Hospital", "Medtronic" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "The trial result itself is not disputed, but Budzyń et al. (Lancet Gastroenterology & Hepatology, 2025; observational, four Polish centres) found endoscopists' adenoma detection without AI fell from 28.4% to 22.4% after routine AI use, suggesting deskilling.", "significance": 3, "sources": [ { "label": "Gastroenterology, 2020-05-01", "url": "https://doi.org/10.1053/j.gastro.2020.04.062" }, { "label": "Lancet Gastroenterol Hepatol (deskilling), 2025-08-12", "url": "https://doi.org/10.1016/S2468-1253(25)00133-5" } ], "foundation0": false }, { "id": "cu-al-co2-electrocatalyst-active-learning", "date": "2020-05-13", "date_basis": "publication", "title": "Active learning finds Cu-Al catalyst that turns CO2 into ethylene efficiently", "domain": "chemistry-materials", "subfield": "electrocatalysis / CO2 reduction", "summary": "Toronto and Carnegie Mellon researchers screened copper-containing alloys with density functional theory guided by active machine learning, singling out copper-aluminium. The resulting electrocatalyst reduced CO2 to ethylene with over 80% Faradaic efficiency at 400 mA/cm², versus about 66% for pure copper, the highest efficiency reported at the time.", "before_after": { "metric": "Faradaic efficiency, CO2 to ethylene", "before": "~66% (pure Cu)", "after": ">80% (Cu-Al, at 400 mA/cm²)" }, "why_it_matters": "Turning captured CO2 into plastic feedstock with renewable electricity needs efficient catalysts. Machine-learning-guided searches of alloy surfaces find candidates faster than trial and error.", "ai_systems": [ "Active machine-learning surrogate model with DFT" ], "organizations": [ "University of Toronto", "Carnegie Mellon University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2020-05-13", "url": "https://www.nature.com/articles/s41586-020-2242-8" } ], "foundation0": false }, { "id": "liverpool-mobile-robotic-chemist", "date": "2020-07-08", "date_basis": "publication", "title": "Mobile robot chemist runs 688 experiments in 8 days to improve a photocatalyst", "domain": "chemistry-materials", "subfield": "self-driving labs / photocatalysis", "summary": "A free-roaming robot at the University of Liverpool, steered by a batched Bayesian search algorithm, worked autonomously for eight days and performed 688 experiments across a ten-variable space. It found photocatalyst mixtures for producing hydrogen from water that were six times more active than the starting formulations.", "before_after": { "metric": "Hydrogen-evolution activity of photocatalyst mixture", "before": "initial formulation (1x)", "after": "6x more active" }, "why_it_matters": "Instead of rebuilding labs around machines, the robot used ordinary instruments the way a human researcher would, pointing to self-driving labs that work day and night.", "ai_systems": [ "Batched Bayesian optimization" ], "organizations": [ "University of Liverpool" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2020-07-08", "url": "https://www.nature.com/articles/s41586-020-2442-2" } ], "foundation0": false }, { "id": "earthquake-transformer-detection", "date": "2020-08-07", "date_basis": "publication", "title": "Earthquake Transformer detects twice as many quakes with a third of the stations", "domain": "earth-climate", "subfield": "seismology", "summary": "An attention-based deep network detects earthquakes and picks P- and S-wave arrival times together, close to human analyst precision. Applied to five weeks of continuous data from the 2000 Tottori earthquake sequence in Japan, it detected and located twice as many earthquakes while using fewer than a third of the seismic stations.", "before_after": { "metric": "Earthquakes detected and located (Tottori 2000 sequence)", "before": "1x (earlier processing)", "after": "about 2x, using under 1/3 of stations" }, "why_it_matters": "Small earthquakes outline the faults that make big ones. Catalogues with many more small events give a sharper picture of where stress builds up.", "ai_systems": [ "EQTransformer" ], "organizations": [ "Stanford University", "Georgia Institute of Technology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Mousavi et al., Nature Communications, 2020-08-07", "url": "https://www.nature.com/articles/s41467-020-17591-w" } ], "foundation0": false }, { "id": "hemibrain-fly-central-brain-connectome", "date": "2020-09-07", "date_basis": "publication", "title": "Janelia and Google map about 25,000 neurons of the fly central brain", "domain": "biology", "subfield": "connectomics", "summary": "Janelia and Google reconstructed the central brain of a fruit fly from electron-microscope images, using flood-filling neural networks plus over 50 person-years of human proofreading. The open map holds about 25,000 neurons and about 20 million chemical synapses, then the largest dense reconstruction to date.", "before_after": null, "why_it_matters": "A full wiring map lets scientists trace how circuits produce behaviour such as navigation, sleep and learning, rather than infer it from isolated recordings.", "ai_systems": [ "Flood-filling networks", "CycleGAN image normalization" ], "organizations": [ "HHMI Janelia Research Campus", "Google Research" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Scheffer et al., eLife, 2020-09-07", "url": "https://doi.org/10.7554/eLife.57443" } ], "foundation0": false }, { "id": "ferminet-paulinet-neural-wavefunctions", "date": "2020-09-16", "date_basis": "publication", "title": "Neural-network wavefunctions give near-exact energies for small molecules", "domain": "chemistry-materials", "subfield": "quantum chemistry / quantum Monte Carlo", "summary": "DeepMind's FermiNet and Free University of Berlin's PauliNet used deep neural networks, built to obey electron antisymmetry, as wavefunctions for variational quantum Monte Carlo. FermiNet captured 99.8% of the correlation energy for small molecules and 97% for 30-electron bicyclobutane, and beat coupled cluster on bond-breaking curves for nitrogen and a hydrogen chain.", "before_after": { "metric": "Correlation energy captured (FermiNet)", "before": "lower with earlier variational Monte Carlo ansätze", "after": "99.8% small molecules; 97% bicyclobutane" }, "why_it_matters": "Applies deep learning directly to the founding equation of chemistry, giving near-exact energies from first principles without training data, including where standard methods fail.", "ai_systems": [ "FermiNet", "PauliNet" ], "organizations": [ "Google DeepMind", "Imperial College London", "Freie Universität Berlin" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Physical Review Research 2:033429, 2020-09-16", "url": "https://doi.org/10.1103/PhysRevResearch.2.033429" }, { "label": "Nature Chemistry 12:891, 2020-09-23", "url": "https://doi.org/10.1038/s41557-020-0544-y" }, { "label": "Google DeepMind blog, FermiNet", "url": "https://deepmind.google/blog/ferminet-quantum-physics-and-chemistry-from-first-principles/" } ], "foundation0": false }, { "id": "chematica-natural-product-routes", "date": "2020-10-13", "date_basis": "publication", "title": "Synthesis-planning software designs working routes to complex natural products", "domain": "chemistry-materials", "subfield": "retrosynthesis planning", "summary": "Grzybowski's group augmented Chematica's reaction knowledge and AI routines with causal, multistep strategy so it could plan routes to complex natural products. In a Turing-like test, synthesis experts largely could not tell its routes from human-designed ones, and three computer-designed natural-product syntheses were validated in the laboratory.", "before_after": null, "why_it_matters": "Natural products demand far-sighted, many-step plans with few precedents to copy. Showing software can strategize at that level suggests expert-level automated planning is feasible.", "ai_systems": [ "Chematica (Synthia)" ], "organizations": [ "Polish Academy of Sciences", "UNIST / IBS Center for Soft and Living Matter", "Northwestern University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2020-10-13", "url": "https://www.nature.com/articles/s41586-020-2855-y" } ], "foundation0": false }, { "id": "sahara-sahel-1-8-billion-trees", "date": "2020-10-14", "date_basis": "publication", "title": "Deep learning counts 1.8 billion trees in the West African Sahara and Sahel", "domain": "earth-climate", "subfield": "land cover / ecology", "summary": "Deep learning applied to sub-metre satellite images mapped the crown of every tree larger than 3 square metres across 1.3 million square kilometres of West Africa. It found over 1.8 billion trees, 13.4 per hectare on average, with surprisingly high density even in deserts, challenging common narratives of dryland desertification.", "before_after": { "metric": "Individual trees mapped outside forests in the region", "before": "not counted", "after": "over 1.8 billion" }, "why_it_matters": "Scattered trees store carbon and support people and wildlife, yet were invisible in forest statistics. Counting them one by one changes how drylands are monitored and managed.", "ai_systems": [ "Deep convolutional neural network" ], "organizations": [ "University of Copenhagen", "NASA Goddard Space Flight Center" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Brandt et al., Nature, 2020-10-14", "url": "https://www.nature.com/articles/s41586-020-2824-5" } ], "foundation0": false }, { "id": "deepmd-100-million-atoms-gordon-bell", "date": "2020-11-19", "date_basis": "announcement", "title": "Deep-learning potentials simulate 100 million atoms at ab initio accuracy", "domain": "chemistry-materials", "subfield": "molecular dynamics / high-performance computing", "summary": "DeePMD-kit, from a Princeton, UC Berkeley and Berkeley Lab, Beijing IAPCM and Peking University team, simulated over 100 million copper atoms with ab initio accuracy on the Summit supercomputer at 91 petaflops. The team said this was at least 100 times larger than earlier Gordon Bell-winning work; it won the 2020 ACM Gordon Bell Prize.", "before_after": { "metric": "System size simulated at ab initio accuracy", "before": "at least 100x smaller in prior Gordon Bell work (team's comparison)", "after": "over 100 million atoms" }, "why_it_matters": "Brings first-principles accuracy to system sizes and timescales relevant to real materials, which classical quantum methods cannot reach.", "ai_systems": [ "DeePMD-kit (Deep Potential)" ], "organizations": [ "Princeton University", "University of California, Berkeley", "Lawrence Berkeley National Laboratory", "Institute of Applied Physics and Computational Mathematics, Beijing", "Peking University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv 2005.00223 (SC20 paper)", "url": "https://arxiv.org/abs/2005.00223" }, { "label": "Berkeley Lab Computing Sciences, Gordon Bell Prize news, 2020-11", "url": "https://cs-newsarchive.lbl.gov/news/2020/machine-learning-software-that-enhances-molecular-dynamics-modeling-wins-gordon-bell-prize/" } ], "foundation0": false }, { "id": "alphafold2-casp14-near-experimental-accuracy", "date": "2020-11-30", "date_basis": "competition", "title": "AlphaFold 2 reaches near-experimental accuracy at CASP14", "domain": "biology", "subfield": "protein structure prediction", "summary": "In the blind CASP14 test DeepMind's AlphaFold 2 reached a median 92.4 GDT across all targets (87.0 on the hardest), with average error near 1.6 angstroms. The 2021 Nature paper reports median backbone error of 0.96 Å against 2.8 Å for the next-best method; a carbon atom is about 1.4 Å wide.", "before_after": { "metric": "Median backbone error at CASP14 (r.m.s.d. at 95% coverage)", "before": "2.8 Å (next best method)", "after": "0.96 Å (AlphaFold 2)" }, "why_it_matters": "Protein shape sets function, but each lab-solved structure took months or years. A predictor near lab accuracy puts structures within reach for almost any protein sequence.", "ai_systems": [ "AlphaFold 2" ], "organizations": [ "DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 5, "sources": [ { "label": "DeepMind blog, 2020-11-30", "url": "https://deepmind.google/blog/alphafold-a-solution-to-a-50-year-old-grand-challenge-in-biology/" }, { "label": "Jumper et al., Nature, 2021-07-15", "url": "https://doi.org/10.1038/s41586-021-03819-2" }, { "label": "CASP14 conference page, 30 November-4 December 2020", "url": "https://predictioncenter.org/casp14/index.cgi" } ], "foundation0": false }, { "id": "birdnet-984-species-sound-identification", "date": "2021-01", "date_basis": "publication", "title": "BirdNET neural network identifies 984 bird species by sound", "domain": "biology", "subfield": "bioacoustics, biodiversity monitoring", "summary": "Cornell and Chemnitz researchers trained a 27-million-parameter ResNet to name 984 North American and European bird species from audio. It scored mean average precision 0.791 on single-species recordings but F0.5 0.414 on annotated soundscapes, showing field audio remains much harder.", "before_after": null, "why_it_matters": "Bird counts track environmental change, but expert surveys cover little ground. Automatic sound identification lets recorders monitor continuously, although accuracy in noisy field audio is still limited.", "ai_systems": [ "BirdNET (ResNet audio classifier)" ], "organizations": [ "Cornell Lab of Ornithology", "Chemnitz University of Technology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Kahl et al., Ecological Informatics 61:101236 (online January 2021; March 2021 issue)", "url": "https://doi.org/10.1016/j.ecoinf.2021.101236" } ], "foundation0": false }, { "id": "mlgo-llvm-compiler", "date": "2021-01-13", "date_basis": "preprint", "title": "Machine-learned inlining and register-allocation policies ship in LLVM", "domain": "computing", "subfield": "compilers", "summary": "Google's MLGO framework put trained models directly into LLVM, the compiler behind much of today's software. A learned inlining policy shrank code 3-7% beyond LLVM's -Oz setting (6.3% on Fuchsia), and a learned register-allocation policy raised datacenter throughput 0.3-1.5%. Both are in LLVM and deployed in production.", "before_after": { "metric": "Binary size vs LLVM -Oz", "before": "Hand-tuned inlining heuristic", "after": "3-7% smaller" }, "why_it_matters": "Compilers apply hand-tuned rules to every program they build. Learning those rules from data yields savings that multiply across all the software compiled with them.", "ai_systems": [ "MLGO (policy-gradient and evolution-strategies policies)" ], "organizations": [ "Google" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv preprint, 2021-01-13", "url": "https://arxiv.org/abs/2101.04808" }, { "label": "Google Research blog, 2022-07-06", "url": "https://www.research.google/blog/mlgo-a-machine-learning-framework-for-compiler-optimization/" } ], "foundation0": false }, { "id": "ramanujan-machine", "date": "2021-02-03", "date_basis": "publication", "title": "Ramanujan Machine algorithms conjecture new formulas for fundamental constants", "domain": "mathematics", "subfield": "number theory / continued fractions", "summary": "Technion researchers built algorithms that match numerical values to propose continued-fraction formulas, finding dozens of known and previously unknown representations of π, e, Catalan's constant and values of the Riemann zeta function. The outputs are conjectures: the algorithms supply no proofs. Published in Nature after a 2019 preprint.", "before_after": null, "why_it_matters": "Turns formula-guessing, historically the gift of rare intuitions like Ramanujan's, into systematic computer search that hands mathematicians precise statements to prove.", "ai_systems": [ "Ramanujan Machine (meet-in-the-middle and gradient-descent search)" ], "organizations": [ "Technion – Israel Institute of Technology" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature, 2021-02-03", "url": "https://doi.org/10.1038/s41586-021-03229-4" }, { "label": "arXiv:1907.00205", "url": "https://arxiv.org/abs/1907.00205" } ], "foundation0": false }, { "id": "galaxy-zoo-decals-bayesian-cnn-314k", "date": "2021-02-16", "date_basis": "preprint", "title": "Volunteers and Bayesian neural networks measure shapes of 314,000 galaxies", "domain": "astronomy-space", "subfield": "galaxy morphology", "summary": "Galaxy Zoo volunteers gave 7.5 million classifications of DECaLS images, with active learning steering them to the most useful galaxies. An ensemble of Bayesian convolutional networks trained on those labels then measured detailed morphology (bars, spiral arms, tidal features) with uncertainties for all 314,000 galaxies, about 99% accurate against confident volunteer answers.", "before_after": { "metric": "Galaxies with detailed morphology measurements", "before": "about 140,000 with 30+ volunteer votes", "after": "314,000 (every galaxy, via deep learning)" }, "why_it_matters": "Galaxy shapes record how galaxies formed and merged. Combining people and models lets surveys of hundreds of millions of galaxies get careful classifications that people alone could never finish.", "ai_systems": [ "Bayesian convolutional neural network ensemble" ], "organizations": [ "Galaxy Zoo (Zooniverse)", "University of Oxford" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Walmsley et al., MNRAS 509:3966, 2022 (online 2021-09-30)", "url": "https://doi.org/10.1093/mnras/stab2093" }, { "label": "arXiv:2102.08414, 2021-02-16", "url": "https://arxiv.org/abs/2102.08414" } ], "foundation0": false }, { "id": "who-recommends-ai-chest-xray-tb-screening", "date": "2021-03-22", "date_basis": "announcement", "title": "WHO recommends AI chest X-ray reading for tuberculosis screening", "domain": "medicine", "subfield": "global health, tuberculosis", "summary": "WHO's 2021 TB screening guideline conditionally recommended computer-aided detection software in place of human readers of digital chest X-rays for people aged 15 and over, followed by confirmatory testing. A 2025 independent evaluation found six CAD products meeting WHO performance standards.", "before_after": null, "why_it_matters": "Many countries with heavy TB burdens have few radiologists. Software that reads X-rays lets screening programs reach remote and poor communities.", "ai_systems": [ "Computer-aided detection (CAD) software for chest X-ray" ], "organizations": [ "World Health Organization", "FIND" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "WHO consolidated TB guidelines, module 2: screening, 2021-03-22", "url": "https://www.who.int/publications/i/item/9789240022676" }, { "label": "WHO policy statement on CAD for TB screening, 2025-06-10", "url": "https://www.who.int/publications/i/item/9789240110373" } ], "foundation0": false }, { "id": "wagner-rl-counterexamples", "date": "2021-04-29", "date_basis": "preprint", "title": "Reinforcement learning finds counterexamples to combinatorics conjectures", "domain": "mathematics", "subfield": "extremal combinatorics / graph theory", "summary": "Adam Zsolt Wagner used the deep cross-entropy method, a simple reinforcement-learning algorithm, to build graphs and matrices that refute several open conjectures about adjacency and distance eigenvalues of graphs, and a question of Brualdi and Cao on permanents of pattern-avoiding matrices. Each conjecture needed only a score function to reward violations.", "before_after": null, "why_it_matters": "A cheap, general recipe, rewarding a network for breaking a conjecture, let machines hunt counterexamples; it was later extended in PatternBoost and AlphaEvolve.", "ai_systems": [ "Deep cross-entropy method (neural network)" ], "organizations": [ "Tel Aviv University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv:2104.14516", "url": "https://arxiv.org/abs/2104.14516" } ], "foundation0": false }, { "id": "ai-ecg-low-ejection-fraction-rct", "date": "2021-05-06", "date_basis": "publication", "title": "AI-read ECGs raise detection of weak heart pumping in a randomized trial", "domain": "medicine", "subfield": "cardiology, primary care", "summary": "In a pragmatic trial of 120 Mayo Clinic primary-care teams (22,641 adults), giving clinicians an AI-ECG alert for low ejection fraction raised new diagnoses within 90 days from 1.6% to 2.1%; among AI-positive patients, from 14.5% to 19.5%.", "before_after": { "metric": "New low-ejection-fraction diagnoses within 90 days", "before": "1.6% (usual care)", "after": "2.1% (AI alert)" }, "why_it_matters": "A cheap, routine test read by AI can uncover hidden, treatable heart failure earlier, showing AI can change real clinical decisions rather than only scoring well on old data.", "ai_systems": [ "Mayo Clinic AI-ECG low-EF algorithm" ], "organizations": [ "Mayo Clinic" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2021-05-06", "url": "https://doi.org/10.1038/s41591-021-01335-4" } ], "foundation0": false }, { "id": "alphachip-floorplanning", "date": "2021-06-09", "date_basis": "publication", "title": "Reinforcement learning lays out chip floorplans used in Google TPUs", "domain": "computing", "subfield": "chip design (macro placement)", "summary": "A deep reinforcement learning agent placed large circuit blocks on chip floorplans in under six hours, matching or beating human experts on power, timing and area by Google's account, against weeks or months of manual work. Google says the method, later named AlphaChip, was used in TPU v5e, v5p and Trillium.", "before_after": { "metric": "Time to produce a chip floorplan", "before": "Weeks to months of expert effort", "after": "Under 6 hours" }, "why_it_matters": "Chip layout is slow expert labour. If AI can do it well, better hardware arrives sooner, and that hardware in turn trains better AI.", "ai_systems": [ "AlphaChip (Circuit Training)" ], "organizations": [ "Google Brain", "Google DeepMind" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Independent reassessments (Cheng, Kahng et al., ISPD 2023, updated in IEEE TCAD; arXiv 2302.11014) and a meta-analysis by Igor Markov (Communications of the ACM, 2024) found the method trailing simulated annealing and commercial tools. Nature attached an editor's note in September 2023 and removed it with an addendum in September 2024; Google's rebuttal (arXiv 2411.10053) defends the results.", "significance": 3, "sources": [ { "label": "Nature, 2021-06-09", "url": "https://www.nature.com/articles/s41586-021-03544-w" }, { "label": "Google DeepMind blog, 2024-09-26", "url": "https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/" }, { "label": "Communications of the ACM (critique, Markov), 2024-10-25", "url": "https://doi.org/10.1145/3676845" } ], "foundation0": false }, { "id": "rosettafold-three-track-network", "date": "2021-07-15", "date_basis": "publication", "title": "RoseTTAFold, a three-track network, nears AlphaFold 2 accuracy", "domain": "biology", "subfield": "protein structure prediction", "summary": "The Baker lab's RoseTTAFold passes information between sequence, distance-map and 3D-coordinate levels, reaching accuracy approaching DeepMind's CASP14 results. After a 1.5-hour sequence search it predicts a backbone for proteins under 400 residues in about 10 minutes on one GPU, and builds protein-complex models from sequence alone.", "before_after": null, "why_it_matters": "Gave academic labs a fast, openly available structure predictor; the same network was later fine-tuned into RFdiffusion for protein design.", "ai_systems": [ "RoseTTAFold" ], "organizations": [ "University of Washington Institute for Protein Design", "Lawrence Berkeley National Laboratory", "UT Southwestern Medical Center" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Baek et al., Science, online 2021-07-15", "url": "https://doi.org/10.1126/science.abj8754" }, { "label": "Open-access full text (PMC)", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7612213/" } ], "foundation0": false }, { "id": "paige-prostate-fda-ai-pathology", "date": "2021-09-21", "date_basis": "announcement", "title": "FDA authorizes Paige Prostate, AI that flags likely cancer on biopsy slides", "domain": "medicine", "subfield": "digital pathology", "summary": "The FDA granted De Novo authorization to Paige Prostate, the first AI software designed to flag the area of a digitized prostate biopsy most likely to hold cancer. When 16 pathologists read 527 slide images with and without it, cancer detection on individual slides rose 7.3% on average; FDA reviewers expected a smaller per-patient benefit.", "before_after": { "metric": "Cancer detection on individual biopsy slides", "before": "pathologists unassisted", "after": "7.3% higher on average with Paige Prostate" }, "why_it_matters": "Opened a regulatory path for AI in pathology, where diagnoses set cancer treatment and pathologists are in short supply.", "ai_systems": [ "Paige Prostate" ], "organizations": [ "Paige", "US FDA" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "FDA De Novo DEN200080 record (decision 2021-09-21)", "url": "https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm?id=DEN200080" }, { "label": "FDA De Novo decision summary DEN200080", "url": "https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN200080.pdf" }, { "label": "FDA announcement as reprinted by BioSpace, 2021-09-21", "url": "https://www.biospace.com/fda-authorizes-software-that-can-help-identify-prostate-cancer" } ], "foundation0": false }, { "id": "dgmr-precipitation-nowcasting", "date": "2021-09-29", "date_basis": "publication", "title": "Generative model of radar gives rain forecasts experts prefer 89% of the time", "domain": "earth-climate", "subfield": "weather nowcasting", "summary": "A deep generative model of radar data forecasts rainfall 5-90 minutes ahead over areas up to 1,536 x 1,280 km without the blurring of earlier deep learning methods. In a systematic evaluation by more than 50 expert meteorologists it ranked first for accuracy and usefulness in 89% of cases against two competing methods.", "before_after": { "metric": "Expert meteorologists' first choice vs two competing methods", "before": "-", "after": "89% of cases" }, "why_it_matters": "Short-term rain forecasts decide flood warnings, flights and outdoor work. Sharper predictions of heavy rain in the next hour or two directly reduce risk.", "ai_systems": [ "DGMR (Deep Generative Model of Rainfall)" ], "organizations": [ "DeepMind", "UK Met Office", "University of Exeter", "University of Reading" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Ravuri et al., Nature, 2021-09-29", "url": "https://www.nature.com/articles/s41586-021-03854-z" } ], "foundation0": false }, { "id": "exominer-301-kepler-planets", "date": "2021-11-19", "date_basis": "preprint", "title": "ExoMiner deep learning model validates 301 new Kepler exoplanets", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "NASA Ames built ExoMiner, a deep network that mirrors the tests human vetters use, and ran it on the Kepler archive. It statistically validated 301 new exoplanets. At 99% precision it recovered 93.6% of planets in the test set, against 76.3% for the best earlier classifier. A 2023 extension added 69 more, for 370.", "before_after": { "metric": "Recall at 99% precision on Kepler test set", "before": "76.3% (best prior classifier)", "after": "93.6%" }, "why_it_matters": "Each validated planet adds to the census used to estimate how common worlds like Earth are. A trustworthy, explainable classifier turns a backlog of ambiguous signals into planets.", "ai_systems": [ "ExoMiner" ], "organizations": [ "NASA Ames Research Center" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Valizadegan et al., The Astrophysical Journal 926:120, 2022 (arXiv 2021-11-19)", "url": "https://doi.org/10.3847/1538-4357/ac4399" }, { "label": "Valizadegan et al., 69 more planets via multiplicity boost, The Astronomical Journal 166:28, 2023", "url": "https://doi.org/10.3847/1538-3881/acd344" }, { "label": "NASA, 2021-11-23", "url": "https://www.nasa.gov/missions/kepler/new-deep-learning-method-adds-301-planets-to-keplers-total-count/" } ], "foundation0": false }, { "id": "xenobots-kinematic-self-replication", "date": "2021-11-29", "date_basis": "publication", "title": "AI-designed shape extends a new kind of self-replication in frog-cell organisms", "domain": "engineering-robotics", "subfield": "AI-designed organisms", "summary": "Researchers found that frog-cell clusters can reproduce by sweeping loose cells into piles that grow into copies, a form of replication not seen in any organism. An AI search found a C-shaped parent design that raised replication rounds from a mean of 1.2 to 3, reaching four generations, in a small number of lab trials.", "before_after": { "metric": "Rounds of self-replication", "before": "Mean 1.2, max 2 (default spheroid)", "after": "Mean 3, max 4 (AI-designed C shape)" }, "why_it_matters": "AI design improved a living behaviour humans had only just observed, suggesting computers can steer biology toward useful functions quickly, without genetic engineering.", "ai_systems": [ "Evolutionary algorithm in a physics simulator" ], "organizations": [ "University of Vermont", "Tufts University", "Wyss Institute, Harvard University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "PNAS, 2021-11-29", "url": "https://doi.org/10.1073/pnas.2112672118" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC8670470/" } ], "foundation0": false }, { "id": "deepmind-knots-representation-theory", "date": "2021-12-01", "date_basis": "publication", "title": "ML guides mathematicians to new results in knot and representation theory", "domain": "mathematics", "subfield": "knot theory / representation theory", "summary": "DeepMind worked with mathematicians at Oxford and Sydney, using neural networks and attribution techniques to spot patterns between mathematical objects. This produced a new theorem linking a knot's algebraic signature to its geometric slope, and a candidate algorithm toward the combinatorial invariance conjecture for symmetric groups.", "before_after": null, "why_it_matters": "A high-profile demonstration that machine learning can point professional mathematicians toward new conjectures and theorems in pure mathematics, not only compute examples.", "ai_systems": [ "Supervised neural networks with saliency attribution" ], "organizations": [ "DeepMind", "University of Oxford", "University of Sydney" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Ernest Davis (NYU) argued the AI role was overstated: small in the knot result, and in the representation-theory result not very different in kind from decades of experimental mathematics (arXiv:2112.04324).", "significance": 3, "sources": [ { "label": "Nature, 2021-12-01", "url": "https://doi.org/10.1038/s41586-021-04086-x" }, { "label": "Davis critique, arXiv:2112.04324", "url": "https://arxiv.org/abs/2112.04324" } ], "foundation0": false }, { "id": "prefixrl-nvidia-adders", "date": "2021-12-05", "date_basis": "publication", "title": "Reinforcement learning designs smaller adder circuits used in NVIDIA Hopper GPUs", "domain": "computing", "subfield": "arithmetic circuit design", "summary": "NVIDIA's PrefixRL agent designed parallel prefix circuits such as adders from scratch, with synthesis tools in the loop. Its best 64-bit adder had 25% lower area than a commercial EDA tool's adder at the same delay. NVIDIA says its Hopper GPU architecture contains nearly 13,000 instances of AI-designed circuits.", "before_after": { "metric": "64-bit adder area at equal delay", "before": "Commercial EDA tool adder", "after": "25% smaller" }, "why_it_matters": "Adders sit in every processor, repeated thousands of times. Smaller and faster arithmetic blocks mean chips that need less silicon and power.", "ai_systems": [ "PrefixRL" ], "organizations": [ "NVIDIA" ], "ai_role": "collaborator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "DAC 2021 proceedings, 2021-12", "url": "https://doi.org/10.1109/DAC18074.2021.9586094" }, { "label": "arXiv preprint, 2022-05-14", "url": "https://arxiv.org/abs/2205.07000" }, { "label": "NVIDIA Technical Blog, 2022-07-08", "url": "https://developer.nvidia.com/blog/designing-arithmetic-circuits-with-deep-reinforcement-learning/" } ], "foundation0": false }, { "id": "dm21-fractional-electron-functional", "date": "2021-12-09", "date_basis": "publication", "title": "DeepMind's DM21 neural functional tackles fractional-electron errors in DFT", "domain": "chemistry-materials", "subfield": "density functional theory", "summary": "DeepMind trained a neural-network exchange-correlation functional, DM21, on accurate chemical data plus exact conditions for systems with fractional charge and spin. It corrected two long-standing failures of density functional theory, spurious charge delocalization and spin-symmetry breaking, and outperformed traditional functionals on broad main-group benchmarks.", "before_after": null, "why_it_matters": "DFT is the most used method for predicting properties of molecules and materials; fixing its systematic errors with learned functionals could make simulations more trustworthy.", "ai_systems": [ "DM21" ], "organizations": [ "Google DeepMind" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "A 2022 Science comment (Gerasimov et al.) argued DM21's generalization to unseen fractional-charge and spin systems was not shown, citing about 50% overlap between training data and a benchmark; the DeepMind authors replied that the points were incorrect or not relevant.", "significance": 3, "sources": [ { "label": "Science 374:1385, 2021-12-09", "url": "https://doi.org/10.1126/science.abj6511" }, { "label": "Science comment (Gerasimov et al.), 2022-08-05", "url": "https://doi.org/10.1126/science.abq3385" }, { "label": "Science response (Kirkpatrick et al.), 2022-08-05", "url": "https://doi.org/10.1126/science.abq4282" } ], "foundation0": false }, { "id": "alphacode-competitive-programming", "date": "2022-02-02", "date_basis": "announcement", "title": "AlphaCode writes programs at median human level in coding contests", "domain": "computing", "subfield": "program synthesis", "summary": "DeepMind's AlphaCode sampled very many candidate programs from a language model, filtered them by running them on example tests, and submitted the best. In 10 simulated Codeforces contests with over 5,000 participants each, it ranked on average in the top 54.3%, roughly median human level, a first for an AI system.", "before_after": { "metric": "Average Codeforces contest ranking", "before": "Earlier AI systems solved few contest problems", "after": "Top 54.3% (estimated rating 1238)" }, "why_it_matters": "Contest problems demand reasoning about algorithms, not just translating instructions into code. This showed language models could begin to solve programming problems they had never seen.", "ai_systems": [ "AlphaCode" ], "organizations": [ "DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2022-12-08", "url": "https://doi.org/10.1126/science.abq1158" }, { "label": "arXiv preprint, 2022-02-08", "url": "https://arxiv.org/abs/2203.07814" }, { "label": "DeepMind blog, 2022-02-02 (updated 2022-12-08)", "url": "https://deepmind.google/blog/competitive-programming-with-alphacode/" } ], "foundation0": false }, { "id": "muzero-vp9-youtube", "date": "2022-02-11", "date_basis": "announcement", "title": "MuZero learns video compression control used on YouTube traffic", "domain": "computing", "subfield": "video compression", "summary": "DeepMind applied MuZero, a game-playing reinforcement learning agent, to choose quantization settings in the open-source VP9 encoder. In tests it cut compressed video size by 6.28% at equal quality versus libvpx's two-pass rate control; on part of YouTube's live traffic it reduced bitrate by 4% on average.", "before_after": { "metric": "Bitrate at the same video quality", "before": "libvpx two-pass VBR rate control", "after": "6.28% smaller in tests; 4% average on YouTube traffic" }, "why_it_matters": "Video makes up most internet traffic. A few percent less data per video saves bandwidth, energy and cost for billions of viewers, especially on slow connections.", "ai_systems": [ "MuZero (MuZero-RC)" ], "organizations": [ "DeepMind", "YouTube" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv preprint, 2022-02-14", "url": "https://arxiv.org/abs/2202.06626" }, { "label": "DeepMind blog, 2022-02-11", "url": "https://deepmind.google/blog/muzeros-first-step-from-research-into-the-real-world/" } ], "foundation0": false }, { "id": "tcv-tokamak-rl-magnetic-control", "date": "2022-02-16", "date_basis": "publication", "title": "Reinforcement learning controls and shapes fusion plasma in the TCV tokamak", "domain": "energy", "subfield": "fusion plasma control", "summary": "DeepMind and EPFL trained a deep reinforcement-learning controller in simulation, then ran it on the TCV tokamak, commanding the full set of 19 magnetic coils directly. It created and held elongated, negative-triangularity and 'snowflake' plasmas, and sustained two separate plasma 'droplets' at once, cutting the design effort for new configurations.", "before_after": { "metric": "controller for a new plasma shape", "before": "many hand-designed feedback loops, re-engineered per configuration", "after": "one learned controller specified by high-level goals" }, "why_it_matters": "Fusion power needs hot plasma held in precise shapes. Learned controllers make it far cheaper to explore new, possibly better, plasma shapes for future reactors.", "ai_systems": [ "Deep reinforcement learning controller" ], "organizations": [ "DeepMind", "EPFL Swiss Plasma Center" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature, 2022-02-16", "url": "https://doi.org/10.1038/s41586-021-04301-9" }, { "label": "DeepMind blog, 2022-02-16", "url": "https://deepmind.google/blog/accelerating-fusion-science-through-learned-plasma-control/" }, { "label": "MIT Technology Review, 2022-02-16", "url": "https://www.technologyreview.com/2022/02/16/1045470/deepminds-ai-can-control-superheated-plasma-inside-a-fusion-reactor/" } ], "foundation0": false }, { "id": "fourcastnet-global-weather", "date": "2022-02-22", "date_basis": "preprint", "title": "FourCastNet produces a week-long global forecast in under two seconds", "domain": "earth-climate", "subfield": "weather forecasting", "summary": "FourCastNet, a Fourier neural operator model trained on reanalysis data, produced global 0.25-degree forecasts matching ECMWF's IFS at short lead times for large-scale variables and beating it on fine-scale variables such as precipitation. A week-long forecast took under two seconds, orders of magnitude faster than the physics-based model.", "before_after": { "metric": "Time for a week-long global 0.25-degree forecast", "before": "orders of magnitude slower (IFS)", "after": "under 2 seconds" }, "why_it_matters": "Cheap forecasts allow thousands of ensemble members, giving better odds for rare extreme events. It was one of the first signs that learned models could rival physics-based forecasting.", "ai_systems": [ "FourCastNet (Adaptive Fourier Neural Operator)" ], "organizations": [ "NVIDIA", "Lawrence Berkeley National Laboratory", "Caltech" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Pathak et al., arXiv:2202.11214, 2022-02-22", "url": "https://arxiv.org/abs/2202.11214" } ], "foundation0": false }, { "id": "togo-novissi-phone-data-aid-targeting", "date": "2022-03-16", "date_basis": "publication", "title": "Machine learning on phone data targets Togo's COVID-19 cash aid to the poorest", "domain": "science-of-science", "subfield": "development economics / social protection", "summary": "Togo's government, with Berkeley, Northwestern and GiveDirectly, trained models on survey data to recognise poverty in mobile-phone records, then used them from November 2020 to direct Novissi cash transfers to the poorest rural subscribers. Versus the geographic options considered, exclusion errors fell 4-21%; a full social registry, which Togo lacks, would do better.", "before_after": { "metric": "Exclusion errors vs geographic targeting", "before": "geographic options considered by the government", "after": "4-21% lower with phone-data models" }, "why_it_matters": "Lets governments with no up-to-date poverty registry reach the poorest quickly in a crisis, a peer-reviewed public-good use of AI in a low-income country.", "ai_systems": [ "Machine-learning poverty prediction from mobile-phone metadata" ], "organizations": [ "Government of Togo", "University of California, Berkeley", "Northwestern University", "GiveDirectly" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature 603:864-870, 2022-03-16", "url": "https://doi.org/10.1038/s41586-022-04484-9" } ], "foundation0": false }, { "id": "blind-sweep-ultrasound-gestational-age-ai", "date": "2022-03-28", "date_basis": "publication", "title": "AI dates pregnancies from novices' ultrasound sweeps as well as sonographers", "domain": "medicine", "subfield": "obstetric ultrasound / global health", "summary": "UNC Chapel Hill and Zambian partners collected 'blind' ultrasound sweeps from 4,695 pregnant volunteers in North Carolina and Zambia and trained a neural network on them. It estimated gestational age with mean error 3.9 days versus 4.7 for sonographer biometry. A 2024 prospective JAMA study found novices using a low-cost AI device matched sonographers.", "before_after": { "metric": "Mean absolute error, gestational age", "before": "4.7 days (sonographer biometry)", "after": "3.9 days (AI on blind sweeps)" }, "why_it_matters": "Accurate pregnancy dating guides care for preterm birth and other risks. Letting untrained staff with cheap, battery-powered probes do it brings ultrasound to clinics without sonographers.", "ai_systems": [ "Deep-learning gestational-age model" ], "organizations": [ "University of North Carolina at Chapel Hill", "University of Zambia" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "NEJM Evidence 1(5), online 2022-03-28", "url": "https://doi.org/10.1056/EVIDoa2100058" }, { "label": "JAMA 332:649 (prospective study), 2024-08-27", "url": "https://doi.org/10.1001/jama.2024.10770" } ], "foundation0": false }, { "id": "murmurations-elliptic-curves-ml", "date": "2022-04-21", "date_basis": "preprint", "title": "Machine-learning experiments uncover 'murmurations' in elliptic curves", "domain": "mathematics", "subfield": "number theory / elliptic curves", "summary": "Trying to explain why machine-learning classifiers predicted elliptic-curve ranks so well, He, Lee, Oliver and Pozdnyakov averaged Frobenius traces over curves of similar conductor and found a striking oscillation that differs by rank. Nina Zubrilina proved a case in 2023 (Inventiones, 2025), and related murmurations were proved for modular forms and Dirichlet characters.", "before_after": null, "why_it_matters": "A genuinely new pattern in a central area of number theory, surfaced by data-driven machine learning before the LLM era, that became an active research field with proofs.", "ai_systems": [ "Machine-learning classifiers of elliptic-curve rank" ], "organizations": [ "London Institute for Mathematical Sciences", "University of Connecticut", "University of Westminster", "Princeton University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv 2204.10140, 2022-04-21 (Experimental Mathematics, 2024)", "url": "https://arxiv.org/abs/2204.10140" }, { "label": "Zubrilina, Murmurations, arXiv 2310.07681 (Invent. Math., 2025)", "url": "https://arxiv.org/abs/2310.07681" }, { "label": "Quanta Magazine, 2024-03-05", "url": "https://www.quantamagazine.org/elliptic-curve-murmurations-found-with-ai-take-flight-20240305/" } ], "foundation0": false }, { "id": "fast-petase-ml-enzyme", "date": "2022-04-27", "date_basis": "publication", "title": "Machine-learning-guided enzyme breaks down PET from 51 waste products in a week", "domain": "chemistry-materials", "subfield": "enzyme engineering / plastic recycling", "summary": "UT Austin researchers used MutCompute, a structure-based neural network, to propose mutations to a plastic-degrading enzyme. FAST-PETase, with three predicted mutations plus two from an earlier variant, almost completely degraded untreated post-consumer PET from 51 different products within one week, worked between 30 and 50 °C, and enabled closed-loop recycling.", "before_after": null, "why_it_matters": "PET is about 12% of global solid waste. Enzymes that digest untreated plastic at mild temperatures could make recycling greener, though industrial use needs much faster rates.", "ai_systems": [ "MutCompute" ], "organizations": [ "University of Texas at Austin" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Arnal et al. (ACS Catalysis, 2023-09-26) compared four engineered PET enzymes under standardized, larger-scale conditions and found FAST-PETase's low depolymerization rate may rule out industrial use; LCC-ICCG converted 98% of PET in 24 hours.", "significance": 3, "sources": [ { "label": "Nature, 2022-04-27", "url": "https://www.nature.com/articles/s41586-022-04599-z" }, { "label": "ACS Catalysis critique, 2023-09-26", "url": "https://doi.org/10.1021/acscatal.3c02922" } ], "foundation0": false }, { "id": "trews-sepsis-early-warning", "date": "2022-07-21", "date_basis": "publication", "title": "Deployed AI sepsis alert linked to fewer deaths across five hospitals", "domain": "medicine", "subfield": "critical care, sepsis", "summary": "TREWS monitored 590,736 patients at five hospitals. Among 6,877 septic patients flagged before antibiotics, those whose alert a clinician confirmed within 3 hours had in-hospital mortality 3.3 points lower (18.7% relative) after adjustment. The study was observational, not randomized.", "before_after": null, "why_it_matters": "Sepsis kills quickly and hides early. An alert that clinicians trust and act on could save many lives, but proving it needs randomized trials.", "ai_systems": [ "TREWS (Targeted Real-time Early Warning System)" ], "organizations": [ "Johns Hopkins University", "Bayesian Health" ], "ai_role": "instrument", "evidence": "deployed", "contested": true, "contested_note": "Nemati et al. (UC San Diego; medRxiv preprint, 2022-08-09) argued the confirmed-vs-unconfirmed comparison is biased by unadjusted process-of-care confounding; the TREWS authors replied (medRxiv, 2022-09-17) that adjusting for such variables does not change their conclusions.", "significance": 3, "sources": [ { "label": "Nature Medicine, 2022-07-21", "url": "https://doi.org/10.1038/s41591-022-01894-0" }, { "label": "medRxiv critique (Nemati et al.), 2022-08-09", "url": "https://doi.org/10.1101/2022.08.08.22278526" }, { "label": "medRxiv reply (Adams, Henry, Saria), 2022-09-17", "url": "https://doi.org/10.1101/2022.09.13.22279688" } ], "foundation0": false }, { "id": "alphafold-database-214-million-structures", "date": "2022-07-28", "date_basis": "deployment", "title": "AlphaFold database grows from under a million to 214 million structures", "domain": "biology", "subfield": "protein structure prediction", "summary": "In July 2022 DeepMind and EMBL-EBI expanded the AlphaFold Protein Structure Database to 214,684,311 predicted structures, covering almost all of UniProt across plants, bacteria, animals and other organisms. The previous release, in December 2021, held 992,316 models. The structures are free to use.", "before_after": { "metric": "Predicted structures in the AlphaFold database", "before": "992,316 (second release, December 2021)", "after": "214,684,311 (July 2022)" }, "why_it_matters": "Instead of waiting for lab work, biologists can look up a predicted structure for nearly any catalogued protein, at no cost.", "ai_systems": [ "AlphaFold 2" ], "organizations": [ "DeepMind", "EMBL-EBI" ], "ai_role": "originator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2022-07-28", "url": "https://deepmind.google/blog/alphafold-reveals-the-structure-of-the-protein-universe/" }, { "label": "EMBL-EBI blog on the new release (exact counts)", "url": "https://ebi-biows.gitdocs.ebi.ac.uk/blog/2022-07-28-new-afdb-release" }, { "label": "Varadi et al., Nucleic Acids Research, AlphaFold DB in 2024, 2023-11-02", "url": "https://doi.org/10.1093/nar/gkad1011" } ], "foundation0": false }, { "id": "proteinmpnn-sequence-design", "date": "2022-09-15", "date_basis": "publication", "title": "ProteinMPNN designs sequences that rescue previously failed protein designs", "domain": "biology", "subfield": "de novo protein design", "summary": "ProteinMPNN, a message-passing neural network, designs amino-acid sequences for a given protein backbone. On native backbones it recovered 52.4% of residues against 32.9% for Rosetta. X-ray crystallography, cryo-EM and function tests showed it rescued earlier Rosetta- or AlphaFold-made designs that had failed.", "before_after": { "metric": "Native sequence recovery on native backbones", "before": "32.9% (Rosetta)", "after": "52.4% (ProteinMPNN)" }, "why_it_matters": "Choosing amino acids that fold into a chosen shape is the second half of protein design. A faster, more reliable tool makes new proteins easier to make and test.", "ai_systems": [ "ProteinMPNN" ], "organizations": [ "University of Washington Institute for Protein Design", "Howard Hughes Medical Institute" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Dauparas et al., Science, online 2022-09-15", "url": "https://doi.org/10.1126/science.add2187" } ], "foundation0": false }, { "id": "alphatensor-matrix-multiplication", "date": "2022-10-05", "date_basis": "publication", "title": "AlphaTensor finds faster matrix multiplication algorithms", "domain": "computing", "subfield": "algorithm discovery (matrix multiplication)", "summary": "DeepMind turned the search for matrix multiplication algorithms into a single-player game for an AlphaZero-style agent. AlphaTensor found a way to multiply 4x4 matrices in mod-2 arithmetic with 47 multiplications instead of 49, the first improvement on Strassen's two-level method in 50 years, and improved best-known counts for several other sizes.", "before_after": { "metric": "Multiplications for 4x4 matrices (mod 2)", "before": "49 (Strassen's algorithm applied twice)", "after": "47" }, "why_it_matters": "Matrix multiplication is the core operation of graphics, simulation and AI. Showing machines can find provably correct algorithms that humans missed changed how algorithm research is done.", "ai_systems": [ "AlphaTensor" ], "organizations": [ "DeepMind" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature, 2022-10-05", "url": "https://www.nature.com/articles/s41586-022-05172-4" }, { "label": "Nature, 2022-10-05", "url": "https://doi.org/10.1038/s41586-022-05172-4" }, { "label": "DeepMind blog, 2022-10-05", "url": "https://deepmind.google/discover/blog/discovering-novel-algorithms-with-alphatensor/" } ], "foundation0": false }, { "id": "high-entropy-invar-alloys-active-learning", "date": "2022-10-06", "date_basis": "publication", "title": "Active learning finds two high-entropy Invar alloys with very low expansion", "domain": "chemistry-materials", "subfield": "alloys / high-entropy alloys", "summary": "Max Planck researchers combined machine learning, density functional theory, thermodynamic calculations and experiments in a closed active-learning loop. After making just 17 new alloys out of millions of possible compositions, they identified two high-entropy Invar alloys with extremely low thermal expansion, around 2×10⁻⁶ per kelvin at 300 K.", "before_after": { "metric": "Alloys synthesized to find low-expansion hits", "before": "millions of possible compositions", "after": "17 made; 2 high-entropy Invar alloys found" }, "why_it_matters": "Materials that barely expand with heat are vital for precision instruments and gas transport. Finding them in a nearly infinite composition space with few experiments shows how learning cuts search cost.", "ai_systems": [ "Active-learning machine-learning models" ], "organizations": [ "Max-Planck-Institut für Eisenforschung" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2022-10-06 (online)", "url": "https://doi.org/10.1126/science.abo4940" } ], "foundation0": false }, { "id": "alphafold-cdk20-hit-discovery", "date": "2023-01-10", "date_basis": "publication", "title": "AlphaFold structure used to find a drug hit for a target with no lab structure", "domain": "medicine", "subfield": "structure-based drug discovery", "summary": "Insilico combined target-ranking (PandaOmics) and generative chemistry (Chemistry42) with an AlphaFold-predicted structure of CDK20, a liver-cancer target lacking an experimental structure. A hit (Kd 9.2 µM) came within 30 days of target selection after synthesizing 7 compounds; a second round gave ISM042-2-048 (IC50 33.4 nM).", "before_after": null, "why_it_matters": "Most disease proteins lack lab-solved structures. Designing drugs on predicted structures widens the set of targets drug hunters can attack.", "ai_systems": [ "AlphaFold", "PandaOmics", "Chemistry42" ], "organizations": [ "Insilico Medicine", "University of Toronto", "Stanford University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Chemical Science, 2023-01-10", "url": "https://doi.org/10.1039/d2sc05709c" } ], "foundation0": false }, { "id": "seti-deep-learning-820-stars", "date": "2023-01-30", "date_basis": "publication", "title": "Deep learning search of 820 stars turns up eight new SETI signals of interest", "domain": "astronomy-space", "subfield": "SETI / radio astronomy", "summary": "A semi-unsupervised variational autoencoder searched 480 hours of Green Bank Telescope data on 820 nearby stars from Breakthrough Listen. It cut candidate signals by about two orders of magnitude versus earlier analyses of the same data and flagged eight signals of interest not found before. Re-observations have not detected them again.", "before_after": null, "why_it_matters": "Radio searches for technology elsewhere drown in human-made interference. A model that filters it well makes it practical to search many more stars with the same effort.", "ai_systems": [ "Beta-convolutional variational autoencoder", "Random forest classifier" ], "organizations": [ "University of Toronto", "Breakthrough Listen", "UC Berkeley" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Ma et al., Nature Astronomy, 2023-01-30", "url": "https://www.nature.com/articles/s41550-022-01872-z" } ], "foundation0": false }, { "id": "deep-learning-de-novo-luciferases", "date": "2023-02-22", "date_basis": "publication", "title": "Deep learning designs working luciferase enzymes from scratch", "domain": "biology", "subfield": "de novo enzyme design", "summary": "University of Washington researchers used deep-learning 'family-wide hallucination' to generate idealized protein scaffolds with varied pockets, then designed artificial luciferases from scratch. The best, a 13.9 kDa enzyme stable above 95 °C, has catalytic efficiency near 10^6 per molar per second on diphenylterazine, comparable to natural luciferases but far more specific.", "before_after": null, "why_it_matters": "Building enzymes from scratch, rather than editing natural ones, opens catalysts for reactions nature never evolved. The authors point to broad biomedical applications.", "ai_systems": [ "Family-wide hallucination (deep-network protein design)" ], "organizations": [ "University of Washington Institute for Protein Design", "University of California, Los Angeles" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Yeh et al., Nature, 2023-02-22", "url": "https://doi.org/10.1038/s41586-023-05696-3" } ], "foundation0": false }, { "id": "go-players-improve-after-superhuman-ai", "date": "2023-03-13", "date_basis": "publication", "title": "Professional Go players make better, more novel moves after AlphaGo", "domain": "science-of-science", "subfield": "human learning from AI", "summary": "Using the superhuman Go program KataGo to score 5.8 million moves by professional players from 1950 to 2021, researchers found decision quality rose significantly after superhuman AI appeared in 2016-17, after 66 years of little change. Never-before-played moves became more frequent and were linked to better decisions.", "before_after": null, "why_it_matters": "Evidence that superhuman AI can raise human expertise rather than replace it: people learned new ideas from the machine and broke away from inherited habits.", "ai_systems": [ "AlphaGo and later superhuman Go programs (source of new ideas)", "KataGo (used to score moves)" ], "organizations": [ "City University of Hong Kong", "Yale University", "Princeton University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "PNAS, 2023-03-13", "url": "https://doi.org/10.1073/pnas.2214840120" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10041097/" } ], "foundation0": false }, { "id": "esmfold-metagenomic-atlas-617-million", "date": "2023-03-16", "date_basis": "publication", "title": "ESMFold builds an atlas of 617 million metagenomic protein structures", "domain": "biology", "subfield": "protein structure prediction", "summary": "Meta AI's protein language model, scaled to 15 billion parameters, predicts atomic-level structure directly from a single sequence, about ten times faster than earlier high-resolution methods. It built the ESM Metagenomic Atlas of more than 617 million predicted structures, over 225 million of them high confidence.", "before_after": null, "why_it_matters": "Most proteins in nature come from microbes never grown in a lab. A fast, alignment-free predictor lets researchers survey that unseen diversity at scale.", "ai_systems": [ "ESMFold", "ESM-2" ], "organizations": [ "Meta AI (FAIR)", "New York University" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Lin et al., Science, online 2023-03-16", "url": "https://doi.org/10.1126/science.ade2574" } ], "foundation0": false }, { "id": "echonet-blinded-randomized-trial", "date": "2023-04-05", "date_basis": "publication", "title": "AI beats sonographers at first-pass heart-function reading in blinded trial", "domain": "medicine", "subfield": "cardiac imaging", "summary": "In the first blinded randomized trial of AI in echocardiography (3,495 studies), cardiologists substantially changed the AI's initial ejection-fraction estimate in 16.8% of studies versus 27.2% for sonographers, and could not tell which source they were reviewing. The AI workflow also saved time.", "before_after": { "metric": "Initial assessments substantially changed by cardiologist (>5% LVEF)", "before": "27.2% (sonographer first)", "after": "16.8% (AI first)" }, "why_it_matters": "Heart-pumping measurements drive many treatment decisions. A blinded randomized trial is the gold-standard test that AI helps in practice, not just on benchmarks.", "ai_systems": [ "EchoNet-Dynamic" ], "organizations": [ "Cedars-Sinai Medical Center", "Stanford University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-04-05", "url": "https://doi.org/10.1038/s41586-023-05947-3" } ], "foundation0": false }, { "id": "primo-m87-black-hole-image", "date": "2023-04-13", "date_basis": "publication", "title": "Machine learning sharpens the image of the M87* black hole", "domain": "astronomy-space", "subfield": "black hole imaging", "summary": "PRIMO, a dictionary-learning method trained on over 30,000 simulated black hole images, re-analysed the 2017 Event Horizon Telescope data of M87*. The authors report reaching the array's full resolution for the first time, with the ring about half as wide as in 2019. Gaps are filled using those simulations.", "before_after": { "metric": "Width of the M87* ring in the reconstructed image", "before": "2019 EHT image", "after": "about half as wide" }, "why_it_matters": "A sharper ring gives tighter measurements of the black hole's mass and a cleaner test of Einstein's gravity, using data already collected.", "ai_systems": [ "PRIMO (principal-component interferometric modeling)" ], "organizations": [ "Institute for Advanced Study", "Georgia Institute of Technology", "NSF NOIRLab" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Medeiros et al., ApJ Letters 947:L7, 2023-04-13", "url": "https://doi.org/10.3847/2041-8213/acc32d" }, { "label": "Georgia Tech School of Physics, 2023-04-13", "url": "https://physics.gatech.edu/node/10647" } ], "foundation0": false }, { "id": "abaucin-acinetobacter-antibiotic", "date": "2023-05-25", "date_basis": "publication", "title": "Deep learning finds abaucin, a narrow-spectrum antibiotic for A. baumannii", "domain": "medicine", "subfield": "antibiotic discovery", "summary": "Researchers screened ~7,500 molecules against Acinetobacter baumannii, a multidrug-resistant hospital pathogen, trained a neural network on the results and predicted new actives. They found abaucin, which targets A. baumannii narrowly, disrupts lipoprotein trafficking via LolE, and controlled infection in a mouse wound model.", "before_after": null, "why_it_matters": "Narrow-spectrum drugs spare helpful bacteria and slow resistance. AI made it practical to find one against a pathogen that resists most existing antibiotics.", "ai_systems": [ "Chemprop (message-passing neural network)" ], "organizations": [ "McMaster University", "MIT", "Wyss Institute, Harvard" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Chemical Biology, 2023-05-25", "url": "https://doi.org/10.1038/s41589-023-01349-8" } ], "foundation0": false }, { "id": "geneformer-transfer-learning-network-biology", "date": "2023-05-31", "date_basis": "publication", "title": "Geneformer, pretrained on 30 million cells, predicts gene-network behaviour", "domain": "biology", "subfield": "single-cell foundation models", "summary": "Geneformer, pretrained on about 30 million single-cell transcriptomes, learned gene-network structure without labels. With little extra data it improved predictions across chromatin and network tasks, and in cardiomyopathy disease modelling it flagged candidate therapeutic targets.", "before_after": null, "why_it_matters": "Mapping gene networks usually needs huge datasets. A pretrained model that works with little data could help study rare diseases and tissues that are hard to sample.", "ai_systems": [ "Geneformer" ], "organizations": [ "Broad Institute", "Massachusetts General Hospital", "Dana-Farber Cancer Institute" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "Kedzierska et al. (Genome Biology, 2025-04-18) tested Geneformer and scGPT with no fine-tuning (zero-shot) and found that in some cases simpler methods could outperform them. The Nature paper's gains rest on fine-tuning, so the critique limits, rather than overturns, its claims.", "significance": 3, "sources": [ { "label": "Theodoris et al., Nature, 2023-05-31", "url": "https://doi.org/10.1038/s41586-023-06139-9" }, { "label": "Zero-shot critique, Genome Biology, 2025-04-18", "url": "https://doi.org/10.1186/s13059-025-03574-x" } ], "foundation0": false }, { "id": "alphadev-sorting-libcxx", "date": "2023-06-07", "date_basis": "publication", "title": "AlphaDev discovers faster sorting routines now in the LLVM C++ library", "domain": "computing", "subfield": "low-level algorithms (sorting, hashing)", "summary": "DeepMind's AlphaDev searched over assembly instructions and found shorter routines for sorting three to five items. Merged into LLVM's libc++ sort, they ran up to 70% faster on short sequences and about 1.7% faster beyond 250,000 elements. A hashing routine it found, released in Abseil, was 30% faster for 9-16 byte inputs.", "before_after": { "metric": "libc++ sort speed on short (3-5 element) sequences", "before": "Human-written routine", "after": "Up to 70% faster" }, "why_it_matters": "Sorting and hashing run trillions of times a day. Small gains in these shared library routines save time and energy across all the software that calls them.", "ai_systems": [ "AlphaDev" ], "organizations": [ "Google DeepMind" ], "ai_role": "collaborator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-06-07", "url": "https://www.nature.com/articles/s41586-023-06004-9" }, { "label": "Google DeepMind blog, 2023-06-07", "url": "https://deepmind.google/blog/alphadev-discovers-faster-sorting-algorithms/" } ], "foundation0": false }, { "id": "icecube-galactic-plane-neutrinos-ml", "date": "2023-06-29", "date_basis": "publication", "title": "Machine learning helps IceCube detect high-energy neutrinos from the Milky Way", "domain": "physics", "subfield": "neutrino astrophysics", "summary": "The IceCube Collaboration applied machine-learning techniques to select and reconstruct neutrino events in 10 years of data from its Antarctic detector. Comparing models of diffuse galactic emission with background, it identified neutrino emission from the plane of the Milky Way at 4.5 standard deviations.", "before_after": { "metric": "Galactic-plane neutrino emission", "before": "not significantly detected", "after": "4.5σ evidence" }, "why_it_matters": "Neutrinos fly straight from where cosmic rays are made, so seeing them from our galaxy helps locate cosmic-ray sources. Machine learning separated the faint signal from background.", "ai_systems": [ "Machine-learning event selection and reconstruction" ], "organizations": [ "IceCube Collaboration" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2023-06-29", "url": "https://doi.org/10.1126/science.adc9818" }, { "label": "Georgia Tech news, 2023-06-29", "url": "https://news.research.gatech.edu/2023/06/29/icecube-detects-high-energy-neutrino-emission-milky-way" } ], "foundation0": false }, { "id": "pangu-weather-beats-ifs", "date": "2023-07-05", "date_basis": "publication", "title": "Pangu-Weather beats ECMWF's operational forecast in reanalysis-based tests", "domain": "earth-climate", "subfield": "weather forecasting", "summary": "Huawei's Pangu-Weather, 3D neural networks trained on 39 years of global reanalysis data, gave more accurate deterministic medium-range forecasts than ECMWF's operational IFS on all tested variables when both started from reanalysis data, and tracked tropical cyclones more accurately than ECMWF-HRES. The authors called it the first AI method to do so.", "before_after": { "metric": "Deterministic medium-range skill vs ECMWF IFS (reanalysis-initialised)", "before": "AI methods clearly less accurate", "after": "better on all tested variables" }, "why_it_matters": "For decades the best forecasts came from solving physics equations on supercomputers. This showed a model learned from past weather could match them, opening the door to operational AI forecasting.", "ai_systems": [ "Pangu-Weather (3D Earth-Specific Transformer)" ], "organizations": [ "Huawei Cloud" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Bi et al., Nature 619:533, 2023-07-05", "url": "https://www.nature.com/articles/s41586-023-06185-3" }, { "label": "arXiv:2211.02556, 2022-11-03", "url": "https://arxiv.org/abs/2211.02556" } ], "foundation0": false }, { "id": "rfdiffusion-de-novo-protein-design", "date": "2023-07-11", "date_basis": "publication", "title": "RFdiffusion generates new protein binders, assemblies and active sites", "domain": "biology", "subfield": "de novo protein design", "summary": "Fine-tuning the RoseTTAFold structure network as a diffusion model gave a general generator of protein backbones, used for binders, symmetric assemblies, metal-binding proteins and enzyme active sites. The authors characterised hundreds of the designed assemblies, metal-binding proteins and binders experimentally.", "before_after": null, "why_it_matters": "One general protein generator can replace bespoke design pipelines: specify a function such as binding a target and get candidate proteins to test.", "ai_systems": [ "RFdiffusion", "RoseTTAFold" ], "organizations": [ "University of Washington Institute for Protein Design", "Massachusetts Institute of Technology" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Watson et al., Nature, 2023-07-11", "url": "https://doi.org/10.1038/s41586-023-06415-8" } ], "foundation0": false }, { "id": "med-palm-usmle-pass-mark", "date": "2023-07-12", "date_basis": "publication", "title": "Language model clears the pass mark on US medical licensing-style questions", "domain": "medicine", "subfield": "medical language models", "summary": "Google's Flan-PaLM scored 67.6% on MedQA US licensing-exam-style questions, above the roughly 60% pass mark and more than 17% above the prior best. Instruction-tuned Med-PaLM improved clinician-rated answers but remained inferior to clinicians. Preprinted December 2022.", "before_after": { "metric": "Accuracy on MedQA (USMLE-style questions)", "before": "~50% (prior best model)", "after": "67.6%" }, "why_it_matters": "Marked the point where general language models began holding usable medical knowledge, starting serious testing of them on real clinical tasks. Exam scores are not patient outcomes.", "ai_systems": [ "Flan-PaLM", "Med-PaLM" ], "organizations": [ "Google Research", "DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature, 2023-07-12", "url": "https://doi.org/10.1038/s41586-023-06291-2" }, { "label": "Google Research Med-PaLM page", "url": "https://sites.research.google/gr/med-palm/" } ], "foundation0": false }, { "id": "cms-boosted-higgs-pairs-particlenet", "date": "2023-07-25", "date_basis": "publication", "title": "Graph neural network helps CMS exclude zero HHVV coupling at 6.3 sigma", "domain": "physics", "subfield": "particle physics (Higgs couplings)", "summary": "CMS searched for pairs of high-energy Higgs bosons decaying to bottom quarks, identifying them with the ParticleNet graph neural network, which raised selection efficiency by about 50% over the previous tagger. The analysis excluded a vanishing coupling of two Higgs to two vector bosons (κ2V = 0) for the first time, at 6.3 standard deviations.", "before_after": { "metric": "H→bb̄ jet selection efficiency at similar background rejection", "before": "DeepAK8 tagger (baseline)", "after": "about 50% higher with ParticleNet" }, "why_it_matters": "How Higgs bosons pair with W and Z bosons tests whether the Higgs keeps high-energy collisions well-behaved. Better AI tagging extracts such tests from limited collision data.", "ai_systems": [ "ParticleNet (graph neural network)" ], "organizations": [ "CMS Collaboration", "CERN" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Physical Review Letters, 2023-07-25", "url": "https://doi.org/10.1103/PhysRevLett.131.041803" }, { "label": "CMS public results B2G-22-003", "url": "https://cms-results.web.cern.ch/cms-results/public-results/publications/B2G-22-003/" } ], "foundation0": false }, { "id": "speech-neuroprostheses-62-and-78-words-per-minute", "date": "2023-08-23", "date_basis": "publication", "title": "Brain implants decode attempted speech at 62 and 78 words per minute", "domain": "medicine", "subfield": "brain-computer interfaces, speech restoration", "summary": "Two teams used deep-learning decoders on brain-implant signals from people who cannot speak intelligibly. Stanford decoded attempted speech at 62 words per minute (23.8% word error, 125,000-word vocabulary). UCSF decoded text at a median 78 words per minute (25% word error, 1,024-word vocabulary) and also drove a voice and a face avatar.", "before_after": { "metric": "Text communication rate from a brain-computer interface, words per minute", "before": "18 (handwriting BCI, 2021; the previous record for any BCI)", "after": "62 (Stanford, intracortical) and 78 (UCSF, cortical surface)" }, "why_it_matters": "ALS and brainstem stroke can remove speech while the intent to speak remains. Learned decoders turn that intent into text at up to about half the pace of natural conversation (160 words per minute).", "ai_systems": [ "Recurrent neural network phoneme decoder with a language model (Stanford)", "Deep-learning text, speech-audio and avatar decoders (UCSF)" ], "organizations": [ "Stanford University", "Howard Hughes Medical Institute", "Brown University", "University of California, San Francisco", "University of California, Berkeley" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Willett et al., Nature, 2023-08-23 (Stanford, 62 words per minute)", "url": "https://doi.org/10.1038/s41586-023-06377-x" }, { "label": "Metzger et al., Nature, 2023-08-23 (UCSF, 78 words per minute)", "url": "https://doi.org/10.1038/s41586-023-06443-4" }, { "label": "Author Correction to Metzger et al., Nature, 2024-07-04 (decoding performance unaffected)", "url": "https://www.nature.com/articles/s41586-024-07735-z" } ], "foundation0": false }, { "id": "swift-drone-beats-champions", "date": "2023-08-30", "date_basis": "publication", "title": "Autonomous drone Swift beats human world champions in real races", "domain": "engineering-robotics", "subfield": "robot learning, autonomous flight", "summary": "Swift, a racing drone trained with deep reinforcement learning in simulation and corrected with real flight data, raced head-to-head against three champion human pilots on a fixed track using only onboard sensors. It won 15 of 25 races and set the fastest recorded race time, half a second ahead of the best human.", "before_after": { "metric": "Head-to-head races against champion human pilots", "before": "No autonomous drone had matched champion pilots", "after": "Won 15 of 25 races; fastest time by 0.5 s" }, "why_it_matters": "The first time an AI-controlled robot beat human champions in a competitive physical sport, showing learned control can work at the limits of real hardware with noisy sensing.", "ai_systems": [ "Swift (deep reinforcement learning)" ], "organizations": [ "University of Zurich", "Intel Labs" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-08-30", "url": "https://doi.org/10.1038/s41586-023-06419-4" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10468397/" } ], "foundation0": false }, { "id": "alphamissense-71-million-variants", "date": "2023-09-19", "date_basis": "publication", "title": "AlphaMissense rates all 71 million possible missense variants", "domain": "biology", "subfield": "variant effect prediction", "summary": "AlphaMissense, AlphaFold fine-tuned on human and primate variant frequencies, rated all 71 million possible human missense variants. At a threshold giving 90% precision on known disease variants it classified 89%: 57% likely benign, 32% likely pathogenic. Only about 0.1% of all possible variants had been confirmed by human experts.", "before_after": null, "why_it_matters": "Most human missense variants have unknown clinical meaning. Proteome-wide predictions give geneticists a starting point for ranking which rare changes may cause disease.", "ai_systems": [ "AlphaMissense", "AlphaFold" ], "organizations": [ "Google DeepMind" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "An independent cohort study of 45 rare diseases (npj Genomic Medicine, 2025-02-28) found AlphaMissense precision of 32.9% and recall of 57.6% for expert-curated pathogenic variants, and poor performance in intrinsically disordered regions. The 90% precision figure is from the developers' own known-variant set.", "significance": 3, "sources": [ { "label": "Cheng et al., Science, online 2023-09-19", "url": "https://doi.org/10.1126/science.adg7492" }, { "label": "Google DeepMind blog, 2023-09-19", "url": "https://deepmind.google/blog/a-catalogue-of-genetic-mutations-to-help-pinpoint-the-cause-of-diseases/" }, { "label": "Rare-disease cohort evaluation, npj Genomic Medicine, 2025-02-28", "url": "https://doi.org/10.1038/s41525-025-00480-w" } ], "foundation0": false }, { "id": "instant-robot-design-laptop", "date": "2023-10-03", "date_basis": "publication", "title": "AI designs a walking robot from scratch in under 30 seconds on a laptop", "domain": "engineering-robotics", "subfield": "automated robot design", "summary": "Researchers replaced slow evolutionary search, which needed days of supercomputing, with a gradient-based method that shaped a robot body starting from a random legless blob. In under 30 seconds and about ten attempts on a laptop it consistently invented legs. The manufactured robot walked significantly farther than an unoptimized one.", "before_after": { "metric": "Time to design a working robot body", "before": "Days of supercomputing (evolutionary design)", "after": "Under 30 seconds on a laptop" }, "why_it_matters": "Near-instant automated design makes it practical to invent custom machines on demand for medical, environmental or space tasks.", "ai_systems": [ "Gradient-based differentiable design optimizer" ], "organizations": [ "Northwestern University", "MIT", "University of Vermont" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "PNAS, 2023-10-03", "url": "https://doi.org/10.1073/pnas.2305180120" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10576117/" } ], "foundation0": false }, { "id": "btsbot-sn2023tyk-fully-automated-supernova", "date": "2023-10-13", "date_basis": "announcement", "title": "First supernova found, confirmed, classified and reported without humans", "domain": "astronomy-space", "subfield": "time-domain astronomy", "summary": "BTSbot, a neural network scanning Zwicky Transient Facility alerts, flagged SN 2023tyk, requested a spectrum from a robotic telescope at Palomar, and the SNIascore model classified it as Type Ia. It was reported publicly on 7 October 2023 with no human step. People had spent about 2,200 hours over six years on such scanning.", "before_after": { "metric": "Human steps from detection to public classification", "before": "visual scanning and manual reporting", "after": "none (SN 2023tyk)" }, "why_it_matters": "New surveys will find millions of changing objects each night. Closing the loop automatically frees astronomers to study the events instead of sorting them.", "ai_systems": [ "BTSbot", "SNIascore" ], "organizations": [ "Northwestern University", "Caltech", "Zwicky Transient Facility" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Rehemtulla et al., The Astrophysical Journal 972, 2024 (arXiv:2401.15167)", "url": "https://doi.org/10.3847/1538-4357/ad5666" }, { "label": "Northwestern University news, 2023-10-13", "url": "https://news.northwestern.edu/stories/2023/10/first-supernova-detected-confirmed-classified-and-shared-by-ai/" }, { "label": "arXiv:2401.15167, 2024-01-26", "url": "https://arxiv.org/abs/2401.15167" } ], "foundation0": false }, { "id": "gpt4-reticular-chemist-mofs", "date": "2023-10-13", "date_basis": "publication", "title": "GPT-4 guides chemists to a new series of metal-organic frameworks", "domain": "chemistry-materials", "subfield": "metal-organic frameworks", "summary": "Omar Yaghi's UC Berkeley group used GPT-4 in a loop: it proposed synthesis steps, a human ran the experiments and reported successes and failures, and it adjusted. This guided the discovery of four isoreticular MOFs (the MOF-521 series); one had been made by accident before but never published, and GPT-4 was not told how.", "before_after": null, "why_it_matters": "Metal-organic frameworks can capture carbon dioxide and harvest water from air. A language model steering their synthesis in plain language lowers the skill barrier to making new porous materials.", "ai_systems": [ "GPT-4" ], "organizations": [ "UC Berkeley" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Angewandte Chemie International Edition, 2023-10-13", "url": "https://doi.org/10.1002/anie.202311983" }, { "label": "arXiv preprint, 2023-06-20", "url": "https://arxiv.org/pdf/2306.14915" } ], "foundation0": false }, { "id": "robotic-ai-chemist-martian-oxygen-catalyst", "date": "2023-11-13", "date_basis": "publication", "title": "Robotic AI chemist makes oxygen-evolution catalyst from Martian meteorites", "domain": "chemistry-materials", "subfield": "self-driving labs / electrocatalysis", "summary": "A robotic AI chemist at USTC pretreated Martian meteorite material, synthesized, characterized and tested oxygen-evolution catalysts, and used a machine-learning model built from simulations and measurements to pick the best formula among more than three million possible compositions, without human intervention. The catalyst ran over 550,000 seconds at 10 mA/cm².", "before_after": null, "why_it_matters": "Making oxygen from local rock would ease crewed missions to Mars. Fully automated catalyst discovery is also what remote, unstaffed laboratories would need.", "ai_systems": [ "Robotic AI-Chemist (machine-learning model with Bayesian optimization)" ], "organizations": [ "University of Science and Technology of China" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Synthesis, 2023-11-13", "url": "https://www.nature.com/articles/s44160-023-00424-1" } ], "foundation0": false }, { "id": "graphcast-medium-range-weather", "date": "2023-11-14", "date_basis": "publication", "title": "GraphCast beats the leading deterministic forecast on 90% of targets", "domain": "earth-climate", "subfield": "weather forecasting", "summary": "GraphCast, a graph neural network trained on four decades of ERA5 reanalysis, forecasts hundreds of variables globally at 0.25 degrees for 10 days in under a minute. It outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets and better supported prediction of tropical cyclones, atmospheric rivers and extreme temperatures.", "before_after": { "metric": "Verification targets where GraphCast beats ECMWF HRES", "before": "HRES, the leading deterministic system", "after": "90% of 1,380" }, "why_it_matters": "A 10-day global forecast on one machine in a minute makes top-quality forecasting cheap enough for any weather service. NOAA later built its operational AI model on it.", "ai_systems": [ "GraphCast" ], "organizations": [ "Google DeepMind", "Google Research" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 5, "sources": [ { "label": "Lam et al., Science 382:1416, published online 2023-11-14", "url": "https://doi.org/10.1126/science.adi2336" }, { "label": "arXiv:2212.12794, 2022-12-24", "url": "https://arxiv.org/abs/2212.12794" }, { "label": "Google DeepMind blog, 2023-11-14", "url": "https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/" } ], "foundation0": false }, { "id": "a-lab-autonomous-solid-state-synthesis", "date": "2023-11-29", "date_basis": "publication", "title": "A-Lab robot autonomously synthesizes computer-predicted inorganic compounds", "domain": "chemistry-materials", "subfield": "self-driving labs / solid-state synthesis", "summary": "Berkeley Lab's A-Lab used robotics, literature-trained recipe models and active learning to make computer-predicted inorganic powders. Over 17 days it first reported making 41 of 58 targets. A January 2026 correction, after manual re-analysis of X-ray data, reports 36 of 57 and says the compounds were new to the platform, not to science.", "before_after": null, "why_it_matters": "Predicting materials is now far faster than making them. A robot that plans, makes and analyzes solid-state syntheses could close that gap, if its automated analysis can be trusted.", "ai_systems": [ "A-Lab (recipe models, active learning, ML X-ray analysis)" ], "organizations": [ "Lawrence Berkeley National Laboratory", "UC Berkeley", "Google DeepMind" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Leeman, Schoop, Palgrave et al. (PRX Energy, 2024-03-07) argued no new materials were made: automated X-ray analysis was unreliable and about two thirds of claimed products were likely known disordered phases. Nature's Author Correction (2026-01-19) revised the novelty wording and cut the success count to 36.", "significance": 3, "sources": [ { "label": "Nature, 2023-11-29", "url": "https://www.nature.com/articles/s41586-023-06734-w" }, { "label": "Nature Author Correction, 2026-01-19", "url": "https://doi.org/10.1038/s41586-025-09992-y" }, { "label": "PRX Energy critique, 2024-03-07", "url": "https://doi.org/10.1103/PRXEnergy.3.011002" } ], "foundation0": false }, { "id": "gnome-2-2-million-crystals", "date": "2023-11-29", "date_basis": "publication", "title": "GNoME predicts 2.2 million new inorganic crystals, about 380,000 stable", "domain": "chemistry-materials", "subfield": "inorganic crystal discovery", "summary": "Google DeepMind's GNoME graph networks, trained at scale with active learning and DFT, found 2.2 million structures below the previously known stability hull, about 381,000 of them on the updated hull. The authors say this lifts known stable crystals from about 48,000 to 421,000; 736 had already been made independently in labs.", "before_after": { "metric": "Known stable inorganic crystals", "before": "~48,000", "after": "~421,000" }, "why_it_matters": "New batteries, chips and solar cells start with new crystals. A far larger computed map of plausibly stable compounds gives experimenters many more starting points.", "ai_systems": [ "GNoME (Graph Networks for Materials Exploration)" ], "organizations": [ "Google DeepMind", "Google Research" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Cheetham and Seshadri (Chemistry of Materials, 2024-04-08) found 'scant evidence' of compounds that are novel, credible and useful; Leeman et al. (PRX Energy, 2024-03-07) warned such predictions neglect compositional disorder, so many 'new' ordered compounds may be known disordered phases.", "significance": 4, "sources": [ { "label": "Nature, 2023-11-29", "url": "https://www.nature.com/articles/s41586-023-06735-9" }, { "label": "Google DeepMind blog, 2023-11-29", "url": "https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/" }, { "label": "Chemistry of Materials critique, 2024-04-08", "url": "https://doi.org/10.1021/acs.chemmater.4c00643" } ], "foundation0": false }, { "id": "funsearch-cap-sets", "date": "2023-12-14", "date_basis": "publication", "title": "FunSearch uses a language model to find larger cap sets", "domain": "mathematics", "subfield": "extremal combinatorics", "summary": "DeepMind paired a pretrained language model with an automatic evaluator to evolve programs. For the cap set problem it found a 512-element cap set in dimension 8 (previous best 496) and raised the lower bound on asymptotic cap set capacity from 2.2180 to 2.2202. The winning programs were readable by mathematicians.", "before_after": { "metric": "Largest known cap set in dimension 8", "before": "496", "after": "512" }, "why_it_matters": "Reported as the first time a large language model helped find a verified new result on an established open problem, by writing programs whose outputs a computer checks.", "ai_systems": [ "FunSearch", "PaLM 2-based code model (Codey)" ], "organizations": [ "Google DeepMind", "University of Wisconsin–Madison" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-12-14", "url": "https://doi.org/10.1038/s41586-023-06924-6" }, { "label": "Google DeepMind blog, 2023-12-14", "url": "https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/" }, { "label": "Nature, 2023-12-14", "url": "https://www.nature.com/articles/s41586-023-06924-6" }, { "label": "Google DeepMind blog, 2023-12-14", "url": "https://deepmind.google/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/" } ], "foundation0": false }, { "id": "coscientist-gpt4-autonomous-experiments", "date": "2023-12-20", "date_basis": "publication", "title": "GPT-4-based Coscientist plans and runs chemistry experiments on lab robots", "domain": "chemistry-materials", "subfield": "LLM lab agents", "summary": "Carnegie Mellon's Coscientist gave GPT-4 internet and documentation search, code execution and laboratory automation, then had it design, plan and carry out experiments across six human-set tasks. It optimized palladium-catalysed cross-coupling reactions, writing the code that drove lab hardware from plain-language goals; the authors describe it as (semi-)autonomous.", "before_after": null, "why_it_matters": "It was an early demonstration that a language model could turn a written research goal into working robot instructions and real reaction results, not just text.", "ai_systems": [ "Coscientist", "GPT-4" ], "organizations": [ "Carnegie Mellon University", "Emerald Cloud Lab" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-12-20", "url": "https://www.nature.com/articles/s41586-023-06792-0" } ], "foundation0": false }, { "id": "explainable-dl-mrsa-antibiotic-class", "date": "2023-12-20", "date_basis": "publication", "title": "Explainable deep learning uncovers a new structural class of antibiotics", "domain": "medicine", "subfield": "antibiotic discovery", "summary": "After testing 39,312 compounds and predicting activity and toxicity for 12 million more with graph neural networks, researchers used explainable substructure rationales to pick 283 to test. One new structural class killed MRSA and vancomycin-resistant enterococci, evaded resistance, and reduced bacteria in mouse skin and thigh infections.", "before_after": null, "why_it_matters": "Shows which chemical parts drive activity, turning a black-box predictor into chemical insight. New antibiotic classes against MRSA are badly needed.", "ai_systems": [ "Graph neural network ensembles (Chemprop)" ], "organizations": [ "MIT", "Broad Institute", "Wyss Institute, Harvard", "Integrated Biosciences" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2023-12-20", "url": "https://doi.org/10.1038/s41586-023-06887-8" } ], "foundation0": false }, { "id": "mit-autonomous-dye-molecule-discovery", "date": "2023-12-21", "date_basis": "publication", "title": "Closed-loop MIT platform designs, makes and tests 294 new dye-like molecules", "domain": "chemistry-materials", "subfield": "self-driving labs / molecular discovery", "summary": "MIT's platform linked a generative model, machine-learning property predictors, computer-aided synthesis planning and robotic synthesis and measurement. Across three automatic design-make-test-analyze cycles it made 294 previously unreported dye-like molecules on four human-chosen scaffolds, retraining its models each round; a second study found nine top-performing molecules.", "before_after": null, "why_it_matters": "Molecular discovery normally cycles slowly through human hands. Automating the whole loop, from idea to measured property, shows how exploration of chemical space could scale.", "ai_systems": [ "Graph-completion generative model", "ML property predictors", "Computer-aided synthesis planning" ], "organizations": [ "MIT" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2023-12-21 (online)", "url": "https://doi.org/10.1126/science.adi1407" } ], "foundation0": false }, { "id": "satellite-ai-maps-dark-fishing-fleets", "date": "2024-01-03", "date_basis": "publication", "title": "Satellite AI shows 72-76% of industrial fishing vessels are not publicly tracked", "domain": "earth-climate", "subfield": "ocean monitoring", "summary": "Deep learning on satellite imagery, matched against vessel GPS data, mapped industrial vessels and offshore infrastructure across the world's coastal waters for 2017-2021. It found 72-76% of industrial fishing vessels and 21-30% of transport and energy vessel activity missing from public tracking, and offshore wind turbines outnumbering oil structures by 2021.", "before_after": { "metric": "Industrial fishing vessels not in public tracking systems", "before": "unmeasured", "after": "72-76%" }, "why_it_matters": "Oceans feed billions, but untracked fishing escapes oversight. A global map of who is at sea makes enforcement and conservation possible where it was guesswork.", "ai_systems": [ "Convolutional neural networks for vessel and infrastructure detection" ], "organizations": [ "Global Fishing Watch", "SkyTruth", "Duke University", "UC Santa Barbara" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Paolo et al., Nature, 2024-01-03", "url": "https://www.nature.com/articles/s41586-023-06825-8" } ], "foundation0": false }, { "id": "microsoft-pnnl-32m-solid-electrolyte", "date": "2024-01-08", "date_basis": "preprint", "title": "ML screen of 32 million candidates leads to a sodium-lithium solid electrolyte", "domain": "chemistry-materials", "subfield": "battery materials / solid electrolytes", "summary": "Microsoft and PNNL combined machine-learning models with physics simulations on about 1,000 cloud machines to narrow 32.6 million candidates to about 500,000 likely stable and 18 promising solid electrolytes in under 80 hours. PNNL made and tested the NaxLi3-xYCl6 series; Microsoft says the chosen electrolyte would use about 70% less lithium.", "before_after": { "metric": "Candidate materials narrowed to a shortlist", "before": "~32.6 million candidates", "after": "18 candidates in under 80 hours of cloud computing" }, "why_it_matters": "Lithium is costly and supply-limited. AI-plus-simulation screening shows how teams can move from millions of options to a working prototype in months rather than years.", "ai_systems": [ "Machine-learning stability and property models (Azure Quantum Elements)" ], "organizations": [ "Microsoft", "Pacific Northwest National Laboratory" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv preprint, 2024-01-08", "url": "https://arxiv.org/abs/2401.04070" }, { "label": "J. Am. Chem. Soc., 2024-07-09", "url": "https://doi.org/10.1021/jacs.4c03849" }, { "label": "Microsoft Azure blog, 2024-01-09", "url": "https://azure.microsoft.com/en-us/blog/quantum/2024/01/09/unlocking-a-new-era-for-scientific-discovery-with-ai-how-microsofts-ai-screened-over-32-million-candidates-to-find-a-better-battery/" } ], "foundation0": false }, { "id": "alphageometry-olympiad", "date": "2024-01-17", "date_basis": "publication", "title": "AlphaGeometry solves olympiad geometry near gold-medallist level", "domain": "mathematics", "subfield": "Euclidean geometry / theorem proving", "summary": "AlphaGeometry pairs a language model, trained from scratch on millions of synthetic theorems and proofs, with a symbolic deduction engine. On 30 recent olympiad geometry problems it solved 25, against 10 for the previous best method, approaching an average IMO gold medallist and producing human-readable proofs.", "before_after": { "metric": "Olympiad geometry problems solved (IMO-AG-30 test)", "before": "10 (Wu's method, as reported)", "after": "25" }, "why_it_matters": "Showed a machine can learn rigorous deduction without human-written proofs by generating its own training data, getting around the scarcity of formal proofs.", "ai_systems": [ "AlphaGeometry" ], "organizations": [ "Google DeepMind", "New York University" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Sinha et al. (arXiv:2404.06405, April 2024) showed the reported baseline was weak: Wu's method alone solves 15 problems, and Wu's method plus classic symbolic deduction, with no neural network, solves 21 of the 30.", "significance": 3, "sources": [ { "label": "Nature, 2024-01-17", "url": "https://doi.org/10.1038/s41586-023-06747-5" }, { "label": "Sinha et al., Wu's method critique, arXiv:2404.06405", "url": "https://arxiv.org/abs/2404.06405" } ], "foundation0": false }, { "id": "vesuvius-challenge-grand-prize-herculaneum", "date": "2024-02-05", "date_basis": "announcement", "title": "Machine learning reads hidden text inside a sealed Herculaneum scroll", "domain": "science-of-science", "subfield": "papyrology / digital humanities", "summary": "Vesuvius Challenge contestants trained ink-detection models on X-ray CT scans of PHerc. Paris 4, a scroll carbonized by Vesuvius in 79 AD. After Luke Farritor's first word was announced in October 2023, he, Youssef Nader and Julian Schilliger recovered 15 columns, over 2,000 characters, winning the $700,000 Grand Prize after papyrologist review and code reproduction.", "before_after": { "metric": "Text read from inside unopened Herculaneum scrolls", "before": "none (opening by hand destroys them)", "after": "15 columns, over 2,000 characters" }, "why_it_matters": "Hundreds of rolls from the only library surviving from classical antiquity are unopened. Reading them without unrolling could recover lost works of ancient philosophy and literature.", "ai_systems": [ "Ink-detection neural networks (Vesuvius Challenge entries)" ], "organizations": [ "Vesuvius Challenge", "University of Kentucky (EduceLab)", "Institut de France" ], "ai_role": "instrument", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Vesuvius Challenge, 2023 Grand Prize announcement, 2024-02-05", "url": "https://scrollprize.org/grandprize" }, { "label": "Vesuvius Challenge, first word discovered, 2023-10-12", "url": "https://scrollprize.org/firstletters" } ], "foundation0": false }, { "id": "diii-d-tearing-mode-avoidance-rl", "date": "2024-02-21", "date_basis": "publication", "title": "RL controller steers DIII-D fusion plasma away from tearing instabilities", "domain": "energy", "subfield": "fusion plasma control", "summary": "Princeton researchers trained a reinforcement-learning controller on a neural-network model of DIII-D that forecasts tearing instabilities up to 300 ms ahead. In live experiments it adjusted plasma shape and beam power to keep tearing risk below a threshold while holding high-confinement H-mode, even at low safety factor and low torque.", "before_after": { "metric": "tearing avoidance at low safety factor and low torque", "before": "hard with preprogrammed control", "after": "tearing likelihood held below threshold while keeping H-mode" }, "why_it_matters": "Tearing instabilities are the leading cause of disruptions that end fusion plasmas. Anticipating and steering around them is needed for reactors like ITER to run steadily.", "ai_systems": [ "Deep reinforcement learning controller", "Multimodal tearing-prediction neural network" ], "organizations": [ "Princeton University", "Princeton Plasma Physics Laboratory", "Chung-Ang University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2024-02-21", "url": "https://doi.org/10.1038/s41586-024-07024-9" }, { "label": "Princeton Engineering news, 2024-02-21", "url": "https://engineering.princeton.edu/news/2024/02/21/engineers-use-ai-wrangle-fusion-power-grid" } ], "foundation0": false }, { "id": "alphatensor-quantum-t-count", "date": "2024-02-22", "date_basis": "preprint", "title": "AlphaTensor-Quantum cuts costly T gates in quantum circuits", "domain": "computing", "subfield": "quantum circuit optimization", "summary": "Researchers adapted AlphaTensor to minimize T gates, the most expensive operation in fault-tolerant quantum computing. It beat existing T-count optimizers on arithmetic benchmarks, matched the best human-designed circuits for Shor's algorithm and quantum chemistry, and found on its own a Karatsuba-like circuit for finite-field multiplication (a similar human design appeared in 2019).", "before_after": { "metric": "Toffoli gates for GF(2^m) multiplication circuits it optimized", "before": "~m^2 (schoolbook input circuits)", "after": "~m^1.58 (Karatsuba-like)" }, "why_it_matters": "T gates dominate the cost of running algorithms on future error-corrected quantum computers. Fewer of them means useful quantum computations become possible sooner, on smaller machines.", "ai_systems": [ "AlphaTensor-Quantum" ], "organizations": [ "Google DeepMind", "Quantinuum", "University of Amsterdam" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Machine Intelligence, 2025-03-20", "url": "https://www.nature.com/articles/s42256-025-01001-1" }, { "label": "arXiv preprint, 2024-02-22", "url": "https://arxiv.org/abs/2402.14396" }, { "label": "Prior sub-quadratic construction: van Hoof, arXiv, 2019-10-07", "url": "https://arxiv.org/abs/1910.02849" } ], "foundation0": false }, { "id": "google-ai-flood-forecasting-ungauged-basins", "date": "2024-03-20", "date_basis": "publication", "title": "AI flood forecasts in ungauged rivers match same-day warnings five days ahead", "domain": "earth-climate", "subfield": "hydrology / flood warning", "summary": "Google's AI river model predicts extreme floods in watersheds with no stream gauges as reliably at five days' lead time as the leading global system (GloFAS) does for the same day. It runs in a free public early-warning system covering over 80 countries, expanded in November 2024 to 100 countries and about 700 million people.", "before_after": { "metric": "Lead time with GloFAS-nowcast-level reliability in ungauged basins", "before": "0 days", "after": "up to 5 days" }, "why_it_matters": "Floods hit hardest where rivers are not measured, mostly in poorer countries. Days of extra warning give families time to move people, animals and belongings.", "ai_systems": [ "Google flood forecasting model (LSTM-based)" ], "organizations": [ "Google Research", "ECMWF", "Helmholtz Centre for Environmental Research (UFZ)" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nearing et al., Nature 627:559, 2024-03-20", "url": "https://www.nature.com/articles/s41586-024-07145-1" }, { "label": "Google blog, 2024-03-20", "url": "https://blog.google/innovation-and-ai/products/google-ai-global-flood-forecasting/" }, { "label": "Google blog, expansion to 100 countries, 2024-11-11", "url": "https://blog.google/innovation-and-ai/products/expanding-flood-forecasting-coverage-helping-partners/" } ], "foundation0": false }, { "id": "ai-ecg-alert-reduces-mortality-rct", "date": "2024-04-29", "date_basis": "publication", "title": "AI-ECG alerts cut 90-day deaths among hospital patients in randomized trial", "domain": "medicine", "subfield": "cardiology, hospital medicine", "summary": "In a multisite randomized trial in Taiwan (15,965 hospitalized patients, 39 physicians), AI-ECG warnings flagging high-risk patients reduced 90-day all-cause mortality from 4.3% to 3.6% (hazard ratio 0.83). Benefit concentrated in high-risk ECGs (HR 0.69), who received more intensive care.", "before_after": { "metric": "90-day all-cause mortality", "before": "4.3% (no alert)", "after": "3.6% (AI-ECG alert)" }, "why_it_matters": "One of the first randomized trials showing an AI alert reduced deaths, not just improved detection; mortality is the outcome that ultimately matters.", "ai_systems": [ "AI-ECG mortality-risk model" ], "organizations": [ "Tri-Service General Hospital", "National Defense Medical Center, Taiwan" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature Medicine, 2024-04-29", "url": "https://doi.org/10.1038/s41591-024-02961-4" } ], "foundation0": false }, { "id": "alphafold3-biomolecular-complexes", "date": "2024-05-08", "date_basis": "publication", "title": "AlphaFold 3 predicts joint structures of proteins, DNA, RNA and ligands", "domain": "biology", "subfield": "protein structure prediction", "summary": "AlphaFold 3 uses a diffusion-based architecture to predict joint structures of complexes containing proteins, nucleic acids, small molecules, ions and modified residues in one model. DeepMind reports about 50% higher accuracy than the best traditional methods on drug-like interactions in the PoseBusters benchmark, with no structural input.", "before_after": null, "why_it_matters": "Cells run on proteins bound to DNA, RNA and small molecules. One model covering them all lets researchers test binding ideas on screen before going to the lab.", "ai_systems": [ "AlphaFold 3" ], "organizations": [ "Google DeepMind", "Isomorphic Labs" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Škrinjar et al. (Nature Structural & Molecular Biology, 2026-05-08) tested four leading cofolding methods, AlphaFold 3 among them, on 2,600 post-cutoff protein-ligand complexes and concluded they largely memorise ligand poses from training data, limiting use for de novo drug design.", "significance": 4, "sources": [ { "label": "Abramson et al., Nature, 2024-05-08", "url": "https://doi.org/10.1038/s41586-024-07487-w" }, { "label": "Google blog, 2024-05-08", "url": "https://blog.google/innovation-and-ai/products/google-deepmind-isomorphic-alphafold-3-ai-model/" }, { "label": "Generalization critique, Nature Structural & Molecular Biology, 2026-05-08", "url": "https://doi.org/10.1038/s41594-026-01797-5" } ], "foundation0": false }, { "id": "chemcrow-llm-chemistry-agent", "date": "2024-05-08", "date_basis": "publication", "title": "ChemCrow agent uses 18 chemistry tools to plan and run syntheses", "domain": "chemistry-materials", "subfield": "LLM lab agents", "summary": "EPFL and University of Rochester researchers connected GPT-4 to 18 expert-designed chemistry tools and a robotic synthesis platform. The resulting ChemCrow agent autonomously planned and executed syntheses of an insect repellent and three organocatalysts, and guided the discovery of a new chromophore.", "before_after": null, "why_it_matters": "Connecting a general language model to specialist software and robots let it complete real chemistry tasks, a pattern now widely used in AI agents for science.", "ai_systems": [ "ChemCrow", "GPT-4" ], "organizations": [ "EPFL", "University of Rochester", "IBM Research Europe" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Machine Intelligence, 2024-05-08", "url": "https://www.nature.com/articles/s42256-024-00832-8" } ], "foundation0": false }, { "id": "h01-human-cortex-cubic-millimetre-connectome", "date": "2024-05-10", "date_basis": "publication", "title": "H01 maps 57,000 cells and 150 million synapses of a human cortex fragment", "domain": "biology", "subfield": "connectomics", "summary": "Harvard and Google imaged one cubic millimetre of human temporal cortex, removed during epilepsy surgery, by electron microscopy, and neural networks traced its cells and synapses. The 1.4-petabyte volume holds about 57,000 cells and 150 million synapses. Glia outnumber neurons 2 to 1. The data were first released in 2021.", "before_after": null, "why_it_matters": "Fly wiring maps cannot show how human cortex is built. An open human volume reveals the cell mix and rare strong connections, such as axons giving one neuron up to 50 synapses.", "ai_systems": [ "Flood-filling networks (cell segmentation)", "U-Net classifier (synapse detection)", "SegCLR embeddings (telling microglia from precursor cells)" ], "organizations": [ "Harvard University", "Google Research" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Shapson-Coe et al., Science, 2024-05-10", "url": "https://doi.org/10.1126/science.adk4858" }, { "label": "Google Research blog, Ten years of neuroscience at Google yields maps of human brain, 2024-05-02", "url": "https://research.google/blog/ten-years-of-neuroscience-at-google-yields-maps-of-human-brain/" }, { "label": "Google Research blog, first H01 data release, 2021-06-01", "url": "https://research.google/blog/a-browsable-petascale-reconstruction-of-the-human-cortex/" } ], "foundation0": false }, { "id": "ml-3d-field-edge-burst-suppression", "date": "2024-05-11", "date_basis": "publication", "title": "ML-tuned magnetic fields give high fusion performance without edge bursts", "domain": "energy", "subfield": "fusion plasma control", "summary": "A Princeton-led team used a machine-learning surrogate of a physics code to tune, in real time, the 3D magnetic perturbations applied to the plasma edge on the DIII-D and KSTAR tokamaks. It suppressed damaging edge energy bursts while keeping reactor-relevant core confinement, which the authors report as the highest fusion performance yet without such bursts.", "before_after": { "metric": "time to optimize the 3D-field settings", "before": "tens of seconds (physics code), so settings were preset before each shot", "after": "milliseconds (ML surrogate), allowing real-time adjustment" }, "why_it_matters": "Repeated edge bursts in ITER-class reactors could erode wall components. Suppressing them without losing performance is needed for practical tokamak power plants.", "ai_systems": [ "ML-3D (machine-learning surrogate model for real-time 3D-field optimization)" ], "organizations": [ "Princeton University", "Princeton Plasma Physics Laboratory", "Korea Institute of Fusion Energy (KSTAR)", "General Atomics (DIII-D)", "Seoul National University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Communications, 2024-05-11", "url": "https://doi.org/10.1038/s41467-024-48415-w" }, { "label": "Princeton Engineering news, 2024-06-03", "url": "https://engineering.princeton.edu/news/2024/06/03/ai-approach-elevates-plasma-performance-and-stability-across-fusion-devices" } ], "foundation0": false }, { "id": "methane-plume-detection-sentinel-2-vit", "date": "2024-05-14", "date_basis": "publication", "title": "Vision transformer spots small methane leaks in free Sentinel-2 imagery", "domain": "earth-climate", "subfield": "greenhouse gas monitoring", "summary": "A vision transformer overcomes the coarse spectral bands of Sentinel-2 to detect methane point sources down to plumes of 0.01 square kilometres, about 200-300 kg of methane per hour, checked against airborne campaigns. The authors report an order-of-magnitude improvement over the previous state of the art, with global coverage every few days.", "before_after": { "metric": "Smallest methane source detectable in Sentinel-2 data", "before": "about 10x larger (prior state of the art)", "after": "200-300 kg CH4 per hour" }, "why_it_matters": "Methane causes a large share of current warming, and most comes from a few leaky sites. Spotting them from free satellite data every few days makes fast fixes possible.", "ai_systems": [ "Vision transformer" ], "organizations": [ "Kyoto University", "Geolabe" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Rouet-Leduc & Hulbert, Nature Communications, 2024-05-14", "url": "https://www.nature.com/articles/s41467-024-47754-y" } ], "foundation0": false }, { "id": "delocalized-ai-lab-organic-laser-emitters", "date": "2024-05-16", "date_basis": "publication", "title": "One AI planner steers five labs to 21 state-of-the-art organic laser materials", "domain": "chemistry-materials", "subfield": "self-driving labs / organic optoelectronics", "summary": "Toronto-led teams linked robotic synthesis and in-line measurement at five laboratories around the world, coordinated asynchronously by a cloud-based AI experiment planner, to search for gain materials for organic solid-state lasers. The campaign found 21 new state-of-the-art materials, and gram-scale synthesis confirmed best-in-class stimulated emission in a thin-film device.", "before_after": null, "why_it_matters": "Expertise and instruments are scattered worldwide. One AI planner coordinating distant labs suggests discovery need not depend on a single well-equipped site.", "ai_systems": [ "Cloud-based machine-learning experiment planner" ], "organizations": [ "University of Toronto", "University of Illinois Urbana-Champaign", "University of British Columbia", "University of Glasgow", "Kyushu University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2024-05-16 (online)", "url": "https://doi.org/10.1126/science.adk9227" } ], "foundation0": false }, { "id": "elephants-name-like-calls-random-forest", "date": "2024-06-10", "date_basis": "publication", "title": "Machine learning and playbacks suggest elephants use name-like calls", "domain": "biology", "subfield": "bioacoustics, animal cognition", "summary": "Researchers trained a random forest on 469 rumbles from wild elephants in Kenya. It identified the addressee of 27.5% of calls, against 8.0% for models with shuffled features. In playbacks to 17 elephants, calls once addressed to them drew faster approach, faster vocal replies and more calls than control calls from the same caller.", "before_after": null, "why_it_matters": "Personal names are a universal feature of human language, and few animal analogues were known. Machine learning found the pattern in call acoustics, and playbacks showed elephants respond to calls addressed to them.", "ai_systems": [ "Random forest classifier" ], "organizations": [ "Colorado State University", "Save the Elephants", "ElephantVoices", "Amboseli Elephant Research Project" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Pardo et al., Nature Ecology & Evolution, 2024-06-10", "url": "https://doi.org/10.1038/s41559-024-02420-w" }, { "label": "Save the Elephants release, 2024-06-10", "url": "https://savetheelephants.org/news/like-us-elephants-have-names-too-2/" }, { "label": "Correction to Pardo et al., Nature Ecology & Evolution, 2025-08 (playback values revised, significance unchanged)", "url": "https://www.nature.com/articles/s41559-025-02823-3" } ], "foundation0": false }, { "id": "agront-plant-genome-language-model", "date": "2024-07-09", "date_basis": "publication", "title": "AgroNT, a language model of 48 plant genomes, scores regulatory DNA", "domain": "biology", "subfield": "plant genomics, crop science", "summary": "InstaDeep trained AgroNT, a genome language model, on 48 plant species, mostly crops. It predicts regulatory annotations, promoter and terminator strength and tissue-specific expression, and was used for in-silico saturation mutagenesis of over 10 million cassava mutations, released as a resource.", "before_after": null, "why_it_matters": "Crop genomes hold many regulatory changes that breeders cannot test one by one. Predicted effects give a ranked starting list, though they still need field and lab checks.", "ai_systems": [ "AgroNT (Agro Nucleotide Transformer)" ], "organizations": [ "InstaDeep", "Google DeepMind" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Mendoza-Revilla et al., Communications Biology, 2024-07-09", "url": "https://doi.org/10.1038/s42003-024-06465-2" } ], "foundation0": false }, { "id": "ai-designed-polymer-capacitor-dielectrics", "date": "2024-07-19", "date_basis": "publication", "title": "AI-guided polymer dielectrics store 11 times more energy at 200 °C", "domain": "chemistry-materials", "subfield": "polymers / capacitor dielectrics", "summary": "Georgia Tech and UConn researchers used AI structure generation and property prediction, with polymer chemistry, to find polynorbornene and polyimide capacitor dielectrics that combine heat stability and high energy density. One stored 8.3 J/cc at 200 °C, eleven times any commercially available polymer dielectric at that temperature.", "before_after": { "metric": "Capacitor energy density at 200 °C", "before": "best commercial polymer dielectric", "after": "8.3 J/cc (11x higher)" }, "why_it_matters": "Electric vehicles, aircraft and grids need capacitors that survive heat. Polymers that keep high energy density at 200 °C fill a gap with no good commercial option.", "ai_systems": [ "Machine-learning polymer property predictors and structure generation" ], "organizations": [ "Georgia Institute of Technology", "University of Connecticut" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Communications, 2024-07-19", "url": "https://www.nature.com/articles/s41467-024-50413-x" } ], "foundation0": false }, { "id": "neuralgcm-hybrid-climate-model", "date": "2024-07-22", "date_basis": "publication", "title": "NeuralGCM hybrid model matches top weather ensembles and runs decades of climate", "domain": "earth-climate", "subfield": "weather and climate modelling", "summary": "NeuralGCM couples a differentiable atmospheric physics solver with learned components. It is competitive with the best machine-learning models for 1-10-day forecasts and with ECMWF's ensemble for 1-15 days, and tracks climate metrics for decades given sea temperatures, at far lower computing cost. It does not extrapolate to very different future climates.", "before_after": null, "why_it_matters": "Pure AI forecasters drift over long runs. Keeping physics and learning the hard parts gives a stable, cheap model usable for climate studies, not just next week's weather.", "ai_systems": [ "NeuralGCM" ], "organizations": [ "Google Research", "Google DeepMind", "ECMWF", "MIT" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Kochkov et al., Nature, 2024-07-22", "url": "https://www.nature.com/articles/s41586-024-07744-y" } ], "foundation0": false }, { "id": "alphaproof-imo-2024-silver", "date": "2024-07-25", "date_basis": "competition", "title": "AlphaProof and AlphaGeometry 2 reach IMO silver-medal standard", "domain": "mathematics", "subfield": "olympiad mathematics / formal proof", "summary": "On the 2024 International Mathematical Olympiad, DeepMind's systems solved 4 of 6 problems for 28 of 42 points, top of the silver range, graded by Timothy Gowers and Joseph Myers. AlphaProof gave Lean-verified proofs for three, including the hardest problem, but problems were hand-translated into Lean and some took up to three days.", "before_after": { "metric": "Best AI score at the IMO", "before": "No medal-level performance", "after": "28/42 (silver)" }, "why_it_matters": "First medal-level AI performance at the top pre-university mathematics competition, with machine-checked proofs that cannot hide errors.", "ai_systems": [ "AlphaProof", "AlphaGeometry 2" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2024-07-25", "url": "https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/" }, { "label": "Nature (AlphaProof), 2025-11-12", "url": "https://doi.org/10.1038/s41586-025-09833-y" } ], "foundation0": false }, { "id": "sakana-ai-scientist", "date": "2024-08-12", "date_basis": "preprint", "title": "The AI Scientist automates machine-learning research from idea to paper", "domain": "science-of-science", "subfield": "autonomous research agents", "summary": "Sakana AI, Oxford, UBC and Vector Institute researchers built a language-model system that proposes ideas, writes and runs experiment code, plots results, writes a full paper and reviews it with an automated reviewer, for about US$15 per paper. It worked in three machine-learning subfields. An extended version was published in Nature in March 2026.", "before_after": null, "why_it_matters": "The first public system to run the whole research loop, from idea to written paper, turning the cost and quality of machine-made science into a measurable question.", "ai_systems": [ "The AI Scientist" ], "organizations": [ "Sakana AI", "University of Oxford", "University of British Columbia", "Vector Institute" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": true, "contested_note": "An independent evaluation (Beel, Kan and Baumgart, SIGIR Forum 2025, https://arxiv.org/abs/2502.14297) found 42% of experiments failed from coding errors, weak novelty checks, few and outdated citations, and some hallucinated numbers.", "significance": 3, "sources": [ { "label": "arXiv preprint, 2024-08-12", "url": "https://arxiv.org/abs/2408.06292" }, { "label": "Nature, 2026-03-25", "url": "https://www.nature.com/articles/s41586-026-10265-5" }, { "label": "Sakana AI announcement", "url": "https://sakana.ai/ai-scientist/" } ], "foundation0": false }, { "id": "speech-neuroprosthesis-als-97-percent-accuracy-live-voice", "date": "2024-08-14", "date_basis": "publication", "title": "ALS speech implant sustains 97.5% accuracy; follow-up synthesizes voice live", "domain": "medicine", "subfield": "brain-computer interfaces, speech restoration", "summary": "UC Davis implanted 256 electrodes in a man with ALS. After 30 minutes of calibration recordings the decoder reached 99.6% accuracy on 50 words on day one and 90.2% on 125,000 words on day two. It sustained 97.5% over 8.4 months. A 2025 Nature follow-up decoded his neural activity into synthesized voice within 10 ms.", "before_after": { "metric": "Training before a speech decoder is usable (different participants)", "before": "Under two weeks of training to reach high performance (UCSF, 2023)", "after": "30 min of recordings: 99.6% on 50 words; plus 1.4 h: 90.2% on 125,000 words" }, "why_it_matters": "Fast calibration and high accuracy decide whether a speech implant works in daily life. He held self-paced conversations at about 32 words per minute for over 248 hours; the voice version adds intonation.", "ai_systems": [ "Neural-network speech decoder (BrainGate2, UC Davis)", "Transformer-based brain-to-voice synthesis decoder (2025)" ], "organizations": [ "University of California, Davis", "BrainGate consortium", "Stanford University", "Brown University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Card et al., New England Journal of Medicine, 2024-08-14 (BrainGate2 trial NCT00912041)", "url": "https://doi.org/10.1056/NEJMoa2314132" }, { "label": "Wairagkar et al., Nature, 2025-06-12 (instantaneous voice synthesis)", "url": "https://doi.org/10.1038/s41586-025-09127-3" }, { "label": "UC Davis Health news release, 2024-08-14", "url": "https://health.ucdavis.edu/welcome/news/headlines/new-brain-computer-interface-allows-man-with-als-to-speak-again/2024/08" } ], "foundation0": false }, { "id": "llm-research-ideas-novelty-study", "date": "2024-09-06", "date_basis": "preprint", "title": "Blind study: LLM research ideas rated more novel than experts' ideas", "domain": "science-of-science", "subfield": "AI research ideation", "summary": "Stanford researchers had over 100 NLP researchers write research ideas and blind-review both human and LLM-generated ones. AI ideas were judged significantly more novel (p < 0.05) but slightly less feasible. A follow-up in which 43 experts carried out the ideas found the AI ideas' scores fell significantly more than human ideas'.", "before_after": null, "why_it_matters": "Idea generation is the creative core of research. Testing it blind, then testing whether ideas survive execution, gives a rare controlled measure of AI's research taste.", "ai_systems": [ "LLM ideation agent" ], "organizations": [ "Stanford University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "The authors' own execution study (Si, Hashimoto and Yang, arXiv 2025-06-25, ICLR 2026, https://arxiv.org/abs/2506.20803) found LLM ideas' scores dropped significantly more than human ideas after execution, with human ideas scoring higher on many metrics.", "significance": 2, "sources": [ { "label": "ICLR 2025 (OpenReview)", "url": "https://openreview.net/forum?id=M23dTGWCZy" }, { "label": "arXiv preprint, 2024-09-06", "url": "https://arxiv.org/abs/2409.04109" }, { "label": "Follow-up execution study, arXiv 2025-06-25 (ICLR 2026)", "url": "https://arxiv.org/abs/2506.20803" } ], "foundation0": false }, { "id": "txgnn-zero-shot-drug-repurposing", "date": "2024-09-25", "date_basis": "publication", "title": "Graph foundation model ranks drug repurposing options for 17,080 diseases", "domain": "medicine", "subfield": "drug repurposing, rare disease", "summary": "TxGNN, trained on a medical knowledge graph, ranks existing drugs as possible treatments or contraindications for 17,080 diseases, including those with no approved therapy. In stringent zero-shot tests against 8 methods it improved indication accuracy by 49.2% and contraindication accuracy by 35.1%; many predictions matched real off-label prescribing.", "before_after": null, "why_it_matters": "Most rare diseases have no approved drug. Ranking existing, already-tested drugs for them gives clinicians testable options far faster than developing new ones.", "ai_systems": [ "TxGNN" ], "organizations": [ "Harvard Medical School", "Stanford University", "Broad Institute", "Icahn School of Medicine at Mount Sinai" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2024-09-25", "url": "https://doi.org/10.1038/s41591-024-03233-x" } ], "foundation0": false }, { "id": "flywire-whole-adult-fly-brain-connectome", "date": "2024-10-02", "date_basis": "publication", "title": "FlyWire maps all 139,255 neurons and 54.5 million synapses of a fly brain", "domain": "biology", "subfield": "connectomics", "summary": "A consortium built a wiring diagram of a whole adult fruit-fly brain from electron-microscope images. Neural networks segmented the images and the community proofread them in about 33 person-years. The map has 139,255 neurons and 54.5 million synapses, with cell types and neurotransmitter predictions.", "before_after": { "metric": "Neurons in a dense fly brain wiring map", "before": "About 25,000 (central brain only, 2020)", "after": "139,255 (whole brain)" }, "why_it_matters": "Whole-brain wiring lets scientists follow signals from senses to movement across both hemispheres, and it is a template for mapping larger brains.", "ai_systems": [ "Boundary-detecting convolutional networks (segmentation)", "Synapse and neurotransmitter prediction networks" ], "organizations": [ "Princeton University", "University of Cambridge", "MRC Laboratory of Molecular Biology", "Allen Institute" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Dorkenwald et al., Nature, 2024-10-02", "url": "https://doi.org/10.1038/s41586-024-07558-y" } ], "foundation0": false }, { "id": "big-sleep-sqlite", "date": "2024-11-01", "date_basis": "announcement", "title": "Big Sleep AI agent finds an exploitable memory bug in SQLite before release", "domain": "computing", "subfield": "vulnerability discovery", "summary": "Google's Big Sleep, a language-model agent from Project Zero and DeepMind, found an exploitable stack buffer underflow in SQLite that existing fuzzing had missed. Developers fixed it the same day, before any official release. Google called it the first public case of an AI agent finding an unknown exploitable memory-safety bug in widely used software.", "before_after": null, "why_it_matters": "SQLite runs on billions of devices. Finding flaws before attackers do, in code that fuzzers already test heavily, is the core promise of AI for defensive security.", "ai_systems": [ "Big Sleep" ], "organizations": [ "Google Project Zero", "Google DeepMind" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Google Project Zero blog, 2024-11-01", "url": "https://projectzero.google/2024/10/from-naptime-to-big-sleep.html" } ], "foundation0": false }, { "id": "patternboost-constructions", "date": "2024-11-01", "date_basis": "preprint", "title": "PatternBoost pairs search with transformers to find extremal constructions", "domain": "mathematics", "subfield": "extremal combinatorics", "summary": "Charton, Ellenberg, Wagner and Williamson alternated a classical local search with a transformer trained on the best constructions found so far, seeding the next round of search. Across several extremal combinatorics problems it found best-known solutions, including a counterexample to a conjecture that had stood for 30 years.", "before_after": null, "why_it_matters": "Gives mathematicians a general, problem-agnostic tool for finding extreme examples where both intuition and brute-force search fail.", "ai_systems": [ "PatternBoost (transformer plus local search)" ], "organizations": [ "Meta FAIR", "University of Wisconsin–Madison", "Worcester Polytechnic Institute", "University of Sydney" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv:2411.00566", "url": "https://arxiv.org/abs/2411.00566" } ], "foundation0": false }, { "id": "virtual-lab-ai-agents-nanobodies", "date": "2024-11-12", "date_basis": "preprint", "title": "Virtual Lab of AI agents designs nanobodies against new SARS-CoV-2 variants", "domain": "science-of-science", "subfield": "multi-agent research systems", "summary": "Stanford's Virtual Lab had an LLM 'principal investigator' lead LLM scientist agents through research meetings, with light human feedback. The agents built a pipeline combining ESM, AlphaFold-Multimer and Rosetta and designed 92 nanobodies. Lab tests found two with improved binding to the JN.1 or KP.3 variants that still bound the original spike.", "before_after": null, "why_it_matters": "A team of AI agents planned an interdisciplinary project whose designs worked in the lab, so researchers without a full team of specialists can still do cross-field science.", "ai_systems": [ "Virtual Lab (LLM agents)", "ESM", "AlphaFold-Multimer", "Rosetta" ], "organizations": [ "Stanford University", "Chan Zuckerberg Biohub San Francisco" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-07-29", "url": "https://doi.org/10.1038/s41586-025-09442-9" }, { "label": "bioRxiv preprint, 2024-11-12", "url": "https://doi.org/10.1101/2024.11.11.623004" } ], "foundation0": false }, { "id": "alphaqubit-surface-code-decoder", "date": "2024-11-20", "date_basis": "publication", "title": "AlphaQubit neural decoder beats leading decoders on Google quantum hardware", "domain": "computing", "subfield": "quantum error correction", "summary": "Google DeepMind and Google Quantum AI trained a recurrent transformer to decode surface-code errors. On Sycamore distance-5 experiments (49 qubits) it made 6% fewer errors than tensor-network decoding (2.748% vs 2.915% per round) and 30% fewer than correlated matching, and kept its edge in simulations up to distance 11, though too slow for real-time use.", "before_after": { "metric": "logical errors on Sycamore distance-5 data", "before": "correlated matching decoder", "after": "30% fewer errors (6% fewer than tensor networks)" }, "why_it_matters": "Quantum computers need constant error correction; better decoding means fewer physical qubits per reliable logical qubit. Learning from real hardware noise beats hand-built noise models.", "ai_systems": [ "AlphaQubit" ], "organizations": [ "Google DeepMind", "Google Quantum AI" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2024-11-20", "url": "https://doi.org/10.1038/s41586-024-08148-8" }, { "label": "Google blog, 2024-11-20", "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/alphaqubit-quantum-error-correction/" }, { "label": "Nature, 2024-11-20", "url": "https://www.nature.com/articles/s41586-024-08148-8" } ], "foundation0": false }, { "id": "oss-fuzz-ai-openssl", "date": "2024-11-20", "date_basis": "announcement", "title": "AI-written fuzz targets find 26 bugs, incl. a likely 20-year-old OpenSSL flaw", "domain": "computing", "subfield": "fuzzing / vulnerability discovery", "summary": "Google used language models to write fuzzing harnesses for OSS-Fuzz projects, extending automated coverage gains to 272 C/C++ projects (from 160) and over 370,000 new lines of code. This found 26 new vulnerabilities, including OpenSSL's low-severity CVE-2024-9143, which Google says was likely present for two decades.", "before_after": { "metric": "OSS-Fuzz projects with automated coverage gains", "before": "160", "after": "272" }, "why_it_matters": "Fuzzing only tests code someone wrote a harness for. AI-written harnesses let testing reach neglected code in libraries that secure much of the internet.", "ai_systems": [ "OSS-Fuzz-gen (LLM-generated fuzz targets)" ], "organizations": [ "Google Open Source Security Team" ], "ai_role": "collaborator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Google Security Blog, 2024-11-20", "url": "https://security.googleblog.com/2024/11/leveling-up-fuzzing-finding-more.html" }, { "label": "OpenSSL Security Advisory (credits Google OSS-Fuzz-Gen), 2024-10-16", "url": "https://openssl-library.org/news/secadv/20241016.txt" } ], "foundation0": false }, { "id": "pytheus-entanglement-by-path-identity", "date": "2024-12-02", "date_basis": "publication", "title": "AI tool PyTheus suggests a simpler way to entangle photons that never met", "domain": "physics", "subfield": "quantum optics / quantum networks", "summary": "Asked to rediscover entanglement swapping, the design tool PyTheus kept returning a simpler scheme. Nanjing University and Max Planck physicists then demonstrated it, entangling independent photons without direct interaction, prior entanglement or Bell-state measurements, by making the photons' origins indistinguishable (path identity); detecting the partner photons heralds the entanglement.", "before_after": { "metric": "requirements to entangle independent photons", "before": "prior entangled pairs plus Bell-state measurement", "after": "neither needed (path identity)" }, "why_it_matters": "Entangling distant particles is the core of quantum networks. A simpler recipe proposed by a machine and confirmed in the lab shows AI can suggest physical concepts, not just tune parameters.", "ai_systems": [ "PyTheus" ], "organizations": [ "Nanjing University", "Max Planck Institute for the Science of Light" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Physical Review Letters, 2024-12-02", "url": "https://doi.org/10.1103/PhysRevLett.133.233601" }, { "label": "arXiv preprint 2412.03022, 2024-12-04", "url": "https://arxiv.org/abs/2412.03022" }, { "label": "The Quantum Insider, 2024-12-07", "url": "https://thequantuminsider.com/2024/12/07/artificial-intelligence-nudges-scientist-to-try-simpler-approach-to-quantum-entanglement/" } ], "foundation0": false }, { "id": "gencast-probabilistic-weather", "date": "2024-12-04", "date_basis": "publication", "title": "GenCast ensemble beats ECMWF's ensemble on 97.2% of targets", "domain": "earth-climate", "subfield": "weather forecasting", "summary": "GenCast, a diffusion-based model, generates 15-day global ensemble forecasts at 0.25 degrees for over 80 variables in about 8 minutes. It was more skilful than ECMWF's ENS, the leading operational ensemble, on 97.2% of 1,320 targets (99.8% beyond 36 hours), and better predicted extreme weather, cyclone tracks and wind power.", "before_after": { "metric": "Targets where model beats ECMWF ENS", "before": "ENS (operational gold standard)", "after": "97.2% of 1,320" }, "why_it_matters": "Decisions about evacuations or power grids need the odds of each outcome, not one guess. A cheaper, more accurate ensemble gives better probabilities for dangerous weather.", "ai_systems": [ "GenCast" ], "organizations": [ "Google DeepMind" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Price et al., Nature 637:84, published online 2024-12-04", "url": "https://www.nature.com/articles/s41586-024-08252-9" }, { "label": "Google DeepMind blog, 2024-12-04", "url": "https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/" } ], "foundation0": false }, { "id": "ai-tools-expand-impact-contract-focus", "date": "2024-12-10", "date_basis": "preprint", "title": "41-million-paper study: AI users publish more, but science's focus narrows", "domain": "science-of-science", "subfield": "metascience", "summary": "Analysing 41.3 million natural-science papers with a language-model classifier, researchers found scientists doing AI-augmented research publish 3.02 times more papers, receive 4.84 times more citations and lead projects 1.37 years sooner. But AI adoption shrank the range of topics studied by 4.63% and cut scientists' engagement with each other by 22%.", "before_after": null, "why_it_matters": "It measures AI's real effect on science at scale and flags a risk: AI may pull research toward data-rich, established areas instead of opening new ones.", "ai_systems": [ "Language-model classifier of AI-augmented papers" ], "organizations": [ "Tsinghua University", "University of Chicago" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature, 2026-01-14", "url": "https://doi.org/10.1038/s41586-025-09922-y" }, { "label": "arXiv preprint, 2024-12-10", "url": "https://arxiv.org/abs/2412.07727" } ], "foundation0": false }, { "id": "perovskite-hole-transport-bayesian-discovery", "date": "2024-12-12", "date_basis": "publication", "title": "Closed-loop ML finds hole-transport molecules for 26.2% perovskite solar cells", "domain": "energy", "subfield": "solar cells", "summary": "Researchers at FAU Erlangen-Nürnberg, KIT and partners paired high-throughput synthesis of organic semiconductors with Bayesian optimization on molecular descriptors. From a small number of model suggestions they found hole-transport materials reaching up to 26.2% power conversion efficiency (25.9% certified) in perovskite solar cells.", "before_after": null, "why_it_matters": "Solar-cell efficiency depends on many specialized layer materials. Learning from small, self-generated datasets lets niche fields design molecules without large existing databases.", "ai_systems": [ "Bayesian optimization", "Molecular-descriptor ML models" ], "organizations": [ "Friedrich-Alexander-Universität Erlangen-Nürnberg", "Karlsruhe Institute of Technology", "Forschungszentrum Jülich (HI-ERN)" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2024-12-12 (online)", "url": "https://doi.org/10.1126/science.ads0901" }, { "label": "FAU Solar news, 2024-12-13", "url": "https://www.solar.fau.de/2024/12/13/inverse-design-workflow-discovers-hole-transport-materials-tailored-for-perovskite-solar-cells/" } ], "foundation0": false }, { "id": "contrail-avoidance-randomized-flights", "date": "2024-12-20", "date_basis": "publication", "title": "AI-guided altitude changes cut airline contrails 64% in randomized trial", "domain": "earth-climate", "subfield": "aviation climate impact", "summary": "Google Research, American Airlines and Breakthrough Energy randomized 44 commercial flights from January to June 2023. Pilots of 22 treatment flights changed altitude using contrail forecasts from a physics model and a machine-learning model; satellite images showed 64% fewer contrails than in control flights (p = 0.0331), for about 2% more fuel per adjusted flight.", "before_after": { "metric": "Observed contrails vs control flights", "before": "control flights on planned routes", "after": "64% fewer, about 2% more fuel" }, "why_it_matters": "Contrails are a major part of aviation's warming effect. Showing per-flight avoidance is feasible and verifiable opens a cheap mitigation route.", "ai_systems": [ "Machine-learning contrail forecast model", "Computer-vision contrail detection (GOES-16)" ], "organizations": [ "Google Research", "American Airlines", "Breakthrough Energy" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Communications Engineering 3, 2024-12-20", "url": "https://doi.org/10.1038/s44172-024-00329-7" } ], "foundation0": false }, { "id": "deep-learning-rf-chip-inverse-design", "date": "2024-12-30", "date_basis": "publication", "title": "Deep learning designs millimetre-wave chip structures in minutes, not weeks", "domain": "engineering-robotics", "subfield": "chip and RF circuit design", "summary": "Princeton and IIT Madras engineers trained deep networks to predict the electromagnetic behaviour of arbitrary pixelated shapes, then searched for shapes meeting a target. Irregular designs built in a 90-nm chip process included an amplifier with about 17.5 dB peak gain from 23.6 to 37.3 GHz. Once trained, each design took minutes.", "before_after": { "metric": "Design time for mm-wave passive structures", "before": "Weeks (simulation-driven genetic-algorithm search)", "after": "Minutes (after one-time model training)" }, "why_it_matters": "Wireless and sensing chips are limited by the shapes human designers know; letting AI explore unfamiliar geometries opens designs outside the standard templates.", "ai_systems": [ "CNN electromagnetic emulator with inverse-design search" ], "organizations": [ "Princeton University", "IIT Madras" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Communications, 2024-12-30", "url": "https://doi.org/10.1038/s41467-024-54178-1" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC11685664/" } ], "foundation0": false }, { "id": "praim-nationwide-ai-mammography-germany", "date": "2025-01-07", "date_basis": "publication", "title": "Real-world AI use in German breast screening raises cancer detection 17.6%", "domain": "medicine", "subfield": "breast cancer screening", "summary": "In Germany's PRAIM observational implementation study (463,094 women, 119 radiologists, 12 sites), radiologists using AI-supported double reading detected 6.7 cancers per 1,000 versus 5.7 without, a 17.6% increase, while recall rates did not rise (37.4 vs 38.3 per 1,000). Radiologists chose case-by-case whether to use AI.", "before_after": { "metric": "Cancers detected per 1,000 screened", "before": "5.7 (standard double reading)", "after": "6.7 (AI-supported)" }, "why_it_matters": "Shows trial-level benefits can hold up in routine national screening, where much promising medical AI has historically faltered.", "ai_systems": [ "Vara AI platform" ], "organizations": [ "University of Lübeck", "Vara" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2025-01-07", "url": "https://doi.org/10.1038/s41591-024-03408-6" } ], "foundation0": false }, { "id": "ai-designed-antivenom-proteins", "date": "2025-01-15", "date_basis": "publication", "title": "AI-designed proteins neutralize deadly snake venom toxins in mice", "domain": "medicine", "subfield": "de novo protein therapeutics", "summary": "Using deep-learning protein design, the Baker lab and partners created new proteins that bind three-finger toxins from elapid snake venoms. With limited lab screening, designs bound tightly, matched computational models at near-atomic accuracy, neutralized all three toxin subfamilies in vitro, and protected mice from lethal neurotoxin challenge.", "before_after": null, "why_it_matters": "Snakebite kills over 100,000 people a year, mostly in poor rural areas. Stable, cheap-to-make designed proteins could replace costly antibodies from immunized animals.", "ai_systems": [ "RFdiffusion", "ProteinMPNN" ], "organizations": [ "Institute for Protein Design, University of Washington", "Technical University of Denmark", "Liverpool School of Tropical Medicine" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-01-15", "url": "https://doi.org/10.1038/s41586-024-08393-x" } ], "foundation0": false }, { "id": "esm3-esmgfp-fluorescent-protein", "date": "2025-01-16", "date_basis": "publication", "title": "ESM3 generates a fluorescent protein only 58% identical to known ones", "domain": "biology", "subfield": "generative protein models", "summary": "EvolutionaryScale's ESM3, a generative model reasoning over protein sequence, structure and function, was prompted to design fluorescent proteins. Of those synthesized, one bright protein, esmGFP, shares 58% sequence identity with known fluorescent proteins, which the authors estimate equals about 500 million years of evolution.", "before_after": null, "why_it_matters": "Shows a language model can propose functional proteins far from anything known, widening the search space for enzymes, sensors and therapeutics beyond what evolution sampled.", "ai_systems": [ "ESM3" ], "organizations": [ "EvolutionaryScale", "Arc Institute", "University of California, Berkeley" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Hayes et al., Science, online 2025-01-16", "url": "https://doi.org/10.1126/science.ads0018" } ], "foundation0": false }, { "id": "mattergen-generative-materials-design", "date": "2025-01-16", "date_basis": "publication", "title": "MatterGen generates candidate crystals to order; one design is synthesized", "domain": "chemistry-materials", "subfield": "generative materials design", "summary": "Microsoft Research's MatterGen, a diffusion model, generates inorganic crystals steerable toward target chemistry, symmetry and properties; its outputs were over twice as likely to be new and stable as earlier generative models'. One design, TaCr2O6, was made as a disordered phase with bulk modulus estimated at up to 169 GPa against a 200 GPa target.", "before_after": { "metric": "Share of generated structures that are new and stable", "before": "previous generative models", "after": "more than 2x higher" }, "why_it_matters": "Instead of screening known lists, generative models propose materials that meet a specification directly, which could shorten design of batteries, magnets and catalysts if outputs are truly new.", "ai_systems": [ "MatterGen", "MatterSim" ], "organizations": [ "Microsoft Research AI for Science", "Shenzhen Institute of Advanced Technology" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Juelsholt (ChemRxiv 2025-06-19; Materials Horizons 2026) showed the synthesized disordered phase matches Ta1/2Cr1/2O2, a compound reported in 1972 and present in MatterGen's training data, so the showcased compound was not new.", "significance": 3, "sources": [ { "label": "Nature, 2025-01-16", "url": "https://www.nature.com/articles/s41586-025-08628-5" }, { "label": "Microsoft Research blog, 2025-01-16", "url": "https://www.microsoft.com/en-us/research/blog/mattergen-a-new-paradigm-of-materials-design-with-generative-ai/" }, { "label": "Materials Horizons critique, 2026 (ChemRxiv preprint 2025-06-19)", "url": "https://doi.org/10.1039/D6MH00268D" } ], "foundation0": false }, { "id": "bayesian-optimized-carbon-nanolattices", "date": "2025-01-23", "date_basis": "publication", "title": "ML-designed carbon nanolattices reach steel's strength at foam's density", "domain": "chemistry-materials", "subfield": "architected materials", "summary": "Researchers at Toronto, KAIST and partners used multi-objective Bayesian optimization to reshape nanoscale lattice struts, then 3D-printed them by two-photon polymerization and converted them to carbon. Optimized lattices were up to 118% stronger and 68% stiffer than standard designs at equal density, matching carbon steel's strength at Styrofoam's density.", "before_after": { "metric": "Strength and stiffness at equal density vs standard lattice", "before": "Conventional lattice topology", "after": "Up to +118% strength, +68% stiffness" }, "why_it_matters": "Lighter parts with the same strength cut fuel use in aircraft and vehicles; machine-guided design found geometries that spread stress better than standard shapes.", "ai_systems": [ "Multi-objective Bayesian optimization" ], "organizations": [ "University of Toronto", "KAIST", "Karlsruhe Institute of Technology", "Rice University", "MIT" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Advanced Materials, 2025-01-23", "url": "https://doi.org/10.1002/adma.202410651" }, { "label": "Full text, PubMed Central", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC11983246/" } ], "foundation0": false }, { "id": "skin-analytics-derm-class-iii-autonomous", "date": "2025-02-06", "date_basis": "announcement", "title": "Skin-cancer AI DERM gets EU Class III mark to rule out cancer without review", "domain": "medicine", "subfield": "dermatology / regulation", "summary": "Skin Analytics says DERM gained a Class III CE mark under the EU Medical Device Regulation, letting it discharge patients it judges cancer-free without a clinician checking. The company reports 99.8% accuracy in ruling out cancer. By October 2026 it was used in 25 NHS trusts, with autonomous use rolling out across NHS South East.", "before_after": null, "why_it_matters": "If its safety holds up in routine use, autonomous triage could shorten long dermatology waits by clearing benign lesions without a specialist appointment.", "ai_systems": [ "DERM" ], "organizations": [ "Skin Analytics" ], "ai_role": "instrument", "evidence": "deployed", "contested": true, "contested_note": "Accuracy figures are company-reported. NICE (May 2025) allowed DERM in the NHS only during a 3-year evidence-generation period and asked for healthcare-professional review for people with black or brown skin.", "significance": 2, "sources": [ { "label": "Skin Analytics announcement, 2025-02-06", "url": "https://skin-analytics.com/news/regulatory-certification/derm-class-iii-ce-mark/" }, { "label": "NICE HealthTech guidance HTG746 (formerly HTE24), recommendations", "url": "https://www.nice.org.uk/guidance/htg746/chapter/1-Recommendations" }, { "label": "TNW, NHS South East rollout, 2026-10-08", "url": "https://thenextweb.com/news/skin-cancer-ai-nhs-south-east-skin-analytics-derm" } ], "foundation0": false }, { "id": "exominer-plus-plus-tess-catalog", "date": "2025-02-13", "date_basis": "preprint", "title": "ExoMiner++ vets 147,568 TESS signals and flags 7,330 planet candidates", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "NASA's ExoMiner++ adds periodograms, difference images and spacecraft data, and trains on Kepler and TESS labels together. Of 147,568 unlabelled TESS 2-minute signals it classified 7,330 as planet candidates, including 50 new community candidates, and kept 1,797 of 2,506 existing TESS candidates as likely planets. These are candidates, not confirmed planets.", "before_after": { "metric": "ExoFOP TESS planet candidates (TOIs) ExoMiner++ still rates as candidates", "before": "2,506", "after": "1,797" }, "why_it_matters": "Follow-up telescope time is scarce. Ranking candidates by how likely they are to be real lets observers confirm more planets for the same effort.", "ai_systems": [ "ExoMiner++" ], "organizations": [ "NASA Ames Research Center" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Valizadegan et al., The Astronomical Journal 170:287, 2025-10-28", "url": "https://doi.org/10.3847/1538-3881/ae03a4" }, { "label": "arXiv:2502.09790, 2025-02-13", "url": "https://arxiv.org/abs/2502.09790" }, { "label": "NASA, 2026-01-22", "url": "https://science.nasa.gov/open-science/deep-learning-exoplanets-tess/" } ], "foundation0": false }, { "id": "rfdiffusion-de-novo-serine-hydrolases", "date": "2025-02-13", "date_basis": "publication", "title": "Deep-learning design creates working serine hydrolase enzymes from scratch", "domain": "chemistry-materials", "subfield": "enzyme design / catalysis", "summary": "University of Washington researchers used the generative model RFdiffusion plus PLACER, a neural network that checks active-site geometry at each reaction step, to design serine hydrolases from minimal active-site descriptions. Designs were catalytically active, crystal structures matched models within 1 Å, and working enzymes came in five folds never seen in natural serine hydrolases.", "before_after": null, "why_it_matters": "Enzymes that run multistep reactions could make chemical manufacturing cleaner. Designing them from scratch, rather than tweaking natural ones, opens catalysts for reactions nature never evolved.", "ai_systems": [ "RFdiffusion", "PLACER" ], "organizations": [ "University of Washington Institute for Protein Design", "UCLA" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2025-02-13 (online)", "url": "https://doi.org/10.1126/science.adu2454" }, { "label": "bioRxiv preprint, 2024-08", "url": "https://doi.org/10.1101/2024.08.29.610411" } ], "foundation0": false }, { "id": "llm-prescreening-trial-enrollment-rct", "date": "2025-02-17", "date_basis": "publication", "title": "LLM prescreening nearly doubles clinical-trial enrollment in randomized test", "domain": "medicine", "subfield": "clinical trial operations", "summary": "At Mass General Brigham, 4,476 patients were randomized to manual or LLM-assisted (RECTIFIER) eligibility prescreening for a heart-failure trial. AI assistance reached decisions faster (hazard ratio 1.78), identified more eligible patients (20.4% vs 12.7%) and enrolled 35 versus 19 patients.", "before_after": { "metric": "Patients enrolled", "before": "19 (0.9%, manual screening)", "after": "35 (1.6%, AI-assisted)" }, "why_it_matters": "Slow recruitment delays or sinks many clinical trials. Faster, accurate chart screening shortens the time to answers about new treatments.", "ai_systems": [ "RECTIFIER (retrieval-augmented LLM)" ], "organizations": [ "Mass General Brigham" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "JAMA, 2025-02-17", "url": "https://doi.org/10.1001/jama.2024.28047" } ], "foundation0": false }, { "id": "google-ai-co-scientist-cf-pici", "date": "2025-02-19", "date_basis": "announcement", "title": "AI co-scientist's top hypothesis matches an unpublished, lab-confirmed mechanism", "domain": "science-of-science", "subfield": "AI hypothesis generation", "summary": "Google's Gemini-based multi-agent co-scientist generates, debates and ranks hypotheses. Asked how cf-PICI genetic islands spread across bacterial species, its top-ranked hypothesis matched a mechanism Imperial College researchers had confirmed but not published: hijacking diverse phage tails. Expert-selected leukaemia drug candidates it proposed also inhibited cancer cells in vitro.", "before_after": null, "why_it_matters": "Matching a hidden, lab-confirmed answer is a clean test of hypothesis generation, the slow creative step of science, and suggests AI can shorten it.", "ai_systems": [ "AI co-scientist (Gemini-based multi-agent system)" ], "organizations": [ "Google Research", "Google DeepMind", "Imperial College London" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature, 2026-05-19", "url": "https://doi.org/10.1038/s41586-026-10644-y" }, { "label": "Cell (cf-PICI study), 2025-09-09", "url": "https://doi.org/10.1016/j.cell.2025.08.018" }, { "label": "Google Research blog, 2025-02-19", "url": "https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/" } ], "foundation0": false }, { "id": "ecmwf-aifs-operational", "date": "2025-02-25", "date_basis": "deployment", "title": "ECMWF runs its machine-learning forecast model operationally", "domain": "earth-climate", "subfield": "operational weather forecasting", "summary": "ECMWF made its machine-learning model AIFS Single operational on 25 February 2025, alongside the physics-based IFS, followed by the 51-member AIFS ensemble on 1 July 2025. ECMWF reports gains of up to 20% on measures such as cyclone tracks and surface temperature, using about 1,000 times less energy, on a coarser 28-31 km grid.", "before_after": { "metric": "Energy per forecast vs physics-based IFS", "before": "1x", "after": "about 1/1000" }, "why_it_matters": "ECMWF sets the global standard for forecasting. Its adoption moved AI weather models from research papers into the forecasts that national services rely on every day.", "ai_systems": [ "AIFS Single", "AIFS ENS" ], "organizations": [ "ECMWF" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "ECMWF news, 2025-02-25", "url": "https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational" }, { "label": "ECMWF news, AIFS ENS operational, 2025-07-01", "url": "https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ensemble-ai-forecasts-become-operational" } ], "foundation0": false }, { "id": "dingo-bns-real-time-gravitational-waves", "date": "2025-03-05", "date_basis": "publication", "title": "Machine learning analyses neutron-star merger signals in one second", "domain": "astronomy-space", "subfield": "gravitational-wave astronomy", "summary": "A neural posterior estimation framework performs complete parameter inference for binary neutron star mergers in about one second without the usual approximations. It can localize sources before merger, improves sky localization by about 30% over approximate low-latency methods, and scales to hour-long signals expected from next-generation detectors.", "before_after": { "metric": "Sky-localization precision vs approximate low-latency methods", "before": "baseline", "after": "about 30% better, in about 1 second" }, "why_it_matters": "Telescopes must point at a merger within minutes to catch its light. Fast, accurate locations from the gravitational-wave signal make those joint observations far more likely.", "ai_systems": [ "DINGO-BNS (neural posterior estimation)" ], "organizations": [ "Max Planck Institute for Intelligent Systems", "Max Planck Institute for Gravitational Physics", "University of Tübingen" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Dax et al., Nature, 2025-03-05", "url": "https://www.nature.com/articles/s41586-025-08593-z" } ], "foundation0": false }, { "id": "ai-scientist-v2-passes-peer-review", "date": "2025-03-12", "date_basis": "announcement", "title": "Fully AI-generated paper clears first-round review at an ICLR 2025 workshop", "domain": "science-of-science", "subfield": "autonomous research agents", "summary": "Sakana's AI Scientist-v2 wrote papers with no human edits; researchers filtered its outputs and submitted three, with the organisers' consent, to an ICLR 2025 workshop that accepted about 70% of papers. One scored 6, 7 and 6 (average 6.33), above the acceptance threshold. It was withdrawn before final decision, as agreed.", "before_after": { "metric": "Peer-review scores of a fully AI-generated paper", "before": "No fully AI-written paper had passed peer review", "after": "6, 7, 6 (average 6.33), above workshop threshold" }, "why_it_matters": "Peer review is science's quality gate. A machine-written paper getting through, even at a lenient workshop, shows that gate now faces AI-made work and must adapt.", "ai_systems": [ "The AI Scientist-v2" ], "organizations": [ "Sakana AI", "University of British Columbia", "University of Oxford" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Researchers quoted by TechCrunch (2025-03-12, https://techcrunch.com/2025/03/12/sakana-claims-its-ai-paper-passed-peer-review-but-its-a-bit-more-nuanced-than-that/) called the result misleading: humans picked which papers to submit, the workshop was lenient and the paper skipped meta-review. Sakana also found citation errors in it.", "significance": 3, "sources": [ { "label": "Sakana AI announcement, 2025-03-12", "url": "https://sakana.ai/ai-scientist-first-publication/" }, { "label": "Nature, 2026-03-25", "url": "https://www.nature.com/articles/s41586-026-10265-5" }, { "label": "arXiv preprint (v2 system), 2025-04-10", "url": "https://arxiv.org/abs/2504.08066" } ], "foundation0": false }, { "id": "euclid-q1-strong-lenses-machine-learning", "date": "2025-03-19", "date_basis": "announcement", "title": "Euclid's first data: AI-led search finds about 500 strong lens candidates", "domain": "astronomy-space", "subfield": "gravitational lensing", "summary": "Five machine-learning models, led by a fine-tuned Zoobot network, ranked about one million objects in Euclid's 63-square-degree quick release; citizen scientists and experts then inspected the top candidates. The result was 250 grade A and 247 grade B galaxy-scale lenses, almost all previously unknown. Extrapolated to the full survey, about 75,000 are expected.", "before_after": { "metric": "Galaxy-scale strong lenses found in Euclid Q1 (grade A + B)", "before": "almost all unknown", "after": "497 (250 A, 247 B)" }, "why_it_matters": "Strong lenses act as natural magnifying glasses that weigh dark matter and measure cosmic expansion. Machine ranking makes finding tens of thousands of them feasible.", "ai_systems": [ "Zoobot", "four other lens-finding machine-learning models" ], "organizations": [ "Euclid Consortium", "ESA" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Euclid Collaboration: Lines et al., Astronomy & Astrophysics 711:A28, 2026", "url": "https://doi.org/10.1051/0004-6361/202554542" }, { "label": "arXiv:2503.15326, 2025-03-19", "url": "https://arxiv.org/abs/2503.15326" }, { "label": "ESA, Euclid Q1 release, 2025-03-19", "url": "https://www.esa.int/Science_Exploration/Space_Science/Euclid/Euclid_opens_data_treasure_trove_offers_glimpse_of_deep_fields" } ], "foundation0": false }, { "id": "aardvark-end-to-end-weather", "date": "2025-03-20", "date_basis": "publication", "title": "Aardvark Weather goes from raw observations to forecasts with one AI model", "domain": "earth-climate", "subfield": "weather forecasting", "summary": "Aardvark Weather takes raw observations and directly produces global and local station forecasts, with no numerical model at run time. On a coarse 1.5-degree grid it matched or beat the US GFS on most headline variables but not ECMWF's HRES. Station forecasts stayed skilful to ten days, competitive with bias-corrected HRES.", "before_after": { "metric": "Parts of the forecast pipeline done by machine learning", "before": "forecast step only", "after": "observations to local forecasts, end to end" }, "why_it_matters": "Conventional forecasting needs supercomputers and large teams. A single learned system could let countries with little computing produce and tailor their own forecasts.", "ai_systems": [ "Aardvark Weather" ], "organizations": [ "University of Cambridge", "Alan Turing Institute", "Microsoft Research", "ECMWF", "British Antarctic Survey" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Allen et al., Nature, 2025-03-20", "url": "https://www.nature.com/articles/s41586-025-08897-0" } ], "foundation0": false }, { "id": "amie-conversational-diagnostic-ai", "date": "2025-04-09", "date_basis": "publication", "title": "Conversational AI outperforms primary-care doctors in simulated consultations", "domain": "medicine", "subfield": "diagnostic dialogue", "summary": "In a randomized, double-blind crossover study of text consultations with patient-actors (159 scenarios, 20 primary-care physicians; Canada, UK, India), Google's AMIE had higher diagnostic accuracy and was rated better on 30 of 32 axes by specialists and 25 of 26 by patient-actors. Text chat is unfamiliar to doctors; no real patients.", "before_after": null, "why_it_matters": "History-taking is the core of diagnosis. If AI can do it well and safely, expert-level first consultations could reach places with few doctors.", "ai_systems": [ "AMIE (Articulate Medical Intelligence Explorer)" ], "organizations": [ "Google Research", "Google DeepMind" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-04-09", "url": "https://doi.org/10.1038/s41586-025-08866-7" } ], "foundation0": false }, { "id": "microns-cubic-millimetre-mouse-visual-cortex", "date": "2025-04-09", "date_basis": "publication", "title": "MICrONS joins activity and wiring in a cubic millimetre of mouse cortex", "domain": "biology", "subfield": "connectomics", "summary": "MICrONS recorded calcium activity of about 75,000 neurons in an awake mouse's visual cortex, then reconstructed the same cubic millimetre from electron microscopy: over 200,000 cells and about 0.5 billion synapses. Convolutional networks did the automated tracing; a subset of neurons was proofread by hand. The data are open.", "before_after": null, "why_it_matters": "Linking what neurons do to how they are wired lets scientists test circuit rules directly, across thousands of neurons in one animal.", "ai_systems": [ "Convolutional-network segmentation", "Automated synapse detection" ], "organizations": [ "Allen Institute for Brain Science", "Baylor College of Medicine", "Princeton University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "MICrONS Consortium, Nature, 2025-04-09", "url": "https://doi.org/10.1038/s41586-025-08790-w" }, { "label": "MICrONS Explorer, cortical mm3 dataset page", "url": "https://www.microns-explorer.org/cortical-mm3" } ], "foundation0": false }, { "id": "urania-gravitational-wave-detector-designs", "date": "2025-04-11", "date_basis": "publication", "title": "Urania algorithm designs gravitational-wave detectors beating human designs", "domain": "physics", "subfield": "gravitational-wave instrumentation", "summary": "Max Planck and Caltech researchers recast interferometer design as a continuous optimization problem solved with methods inspired by machine learning. The algorithm Urania found about 50 detector layouts that outperform the best known next-generation designs under realistic constraints, raising the potentially observable volume of the Universe by up to 50-fold. None has been built yet.", "before_after": { "metric": "potentially observable volume of the Universe", "before": "best human next-generation design (1×)", "after": "up to 50×" }, "why_it_matters": "Detector sensitivity limits which cosmic events we can observe. Automated design explores layouts humans never considered and can reveal new physics ideas behind better instruments.", "ai_systems": [ "Urania" ], "organizations": [ "Max Planck Institute for the Science of Light", "Caltech (LIGO Laboratory)" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Physical Review X, 2025-04-11", "url": "https://doi.org/10.1103/PhysRevX.15.021012" }, { "label": "arXiv preprint 2312.04258, 2023-12-06", "url": "https://arxiv.org/abs/2312.04258" }, { "label": "Max Planck Institute for the Science of Light, 2025-04-14", "url": "https://mpl.mpg.de/news/article/ai-finds-new-ways-to-observe-the-most-extreme-events-in-the-universe" } ], "foundation0": false }, { "id": "waymo-crash-rates-56m-miles", "date": "2025-05-01", "date_basis": "announcement", "title": "Peer-reviewed study: driverless Waymo cars had far fewer injury crashes", "domain": "engineering-robotics", "subfield": "autonomous driving safety", "summary": "Waymo-authored research compared 56.7 million driverless miles with matched human benchmarks. Rates were significantly lower for injury, airbag-deployment and serious-injury crashes, with no crash type significantly worse; Waymo reports 85% fewer serious-injury and 92% fewer pedestrian-injury crashes. Company data to June 2026 (271 million miles) show 82% fewer injury crashes.", "before_after": { "metric": "Crashes with suspected serious or worse injury, vs human benchmark", "before": "Human-driver benchmark rate", "after": "85% lower over 56.7 million driverless miles" }, "why_it_matters": "Road crashes are a leading cause of death. Large-scale evidence that a learned driving system is safer than people is a milestone for trusting AI with life-critical physical tasks.", "ai_systems": [ "Waymo Driver" ], "organizations": [ "Waymo" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Traffic Injury Prevention, 2025-05-16", "url": "https://doi.org/10.1080/15389588.2025.2499887" }, { "label": "Waymo blog, 2025-05-01", "url": "https://waymo.com/blog/2025/05/waymo-making-streets-safer-for-vru/" }, { "label": "Waymo safety impact data (through June 2026)", "url": "https://waymo.com/safety/impact/" } ], "foundation0": false }, { "id": "ai-co-scientist-liver-fibrosis-repurposing", "date": "2025-05-04", "date_basis": "preprint", "title": "AI co-scientist proposes liver-fibrosis drug candidates confirmed in organoids", "domain": "medicine", "subfield": "drug repurposing, hepatology", "summary": "Google's multi-agent AI co-scientist proposed repurposing drugs that target epigenetic regulators for liver fibrosis. Stanford researchers tested them in human liver organoids: two of three AI-recommended drugs showed significant anti-fibrotic activity, and vorinostat, an approved cancer drug, cut TGF-beta-induced chromatin changes by 91% and promoted regeneration.", "before_after": null, "why_it_matters": "An AI generating treatment hypotheses that survive lab tests suggests machines can speed the slow, guess-heavy front end of drug repurposing.", "ai_systems": [ "AI co-scientist (Google)" ], "organizations": [ "Stanford University", "Google" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Advanced Science, 2025-09-14", "url": "https://doi.org/10.1002/advs.202508751" }, { "label": "bioRxiv preprint, 2025-05-04", "url": "https://doi.org/10.1101/2025.04.29.651320" } ], "foundation0": false }, { "id": "alphaevolve-google-infrastructure", "date": "2025-05-14", "date_basis": "announcement", "title": "AlphaEvolve evolves code that saves compute across Google's infrastructure", "domain": "computing", "subfield": "algorithm discovery / systems optimization", "summary": "AlphaEvolve, a Gemini-powered agent that evolves code against automatic evaluators, produced a data-center scheduling heuristic that recovers about 0.7% of Google's worldwide compute, a 23% faster kernel that cut Gemini training time 1%, up to 32.5% faster FlashAttention kernels, and a simplified circuit adopted in an upcoming TPU.", "before_after": { "metric": "Google fleet compute recovered by scheduling", "before": "Previous Borg heuristic", "after": "About 0.7% of worldwide compute recovered" }, "why_it_matters": "At Google's scale, fractions of a percent equal whole data centres. A general agent improving real production code shows AI can now speed up the very systems that build AI.", "ai_systems": [ "AlphaEvolve", "Gemini Flash", "Gemini Pro" ], "organizations": [ "Google DeepMind", "Google" ], "ai_role": "collaborator", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2025-05-14", "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" }, { "label": "arXiv white paper, 2025-06-16", "url": "https://arxiv.org/abs/2506.13131" } ], "foundation0": false }, { "id": "alphaevolve-math-results", "date": "2025-05-14", "date_basis": "announcement", "title": "AlphaEvolve improves best-known results on about 20% of 50+ open problems", "domain": "mathematics", "subfield": "combinatorics, geometry, algorithms", "summary": "Google DeepMind's Gemini-powered evolutionary coding agent was run on over 50 open problems in analysis, geometry and combinatorics: it rediscovered the state of the art in about 75% and improved it in about 20%. It found a 48-multiplication method for 4×4 complex matrices and raised the 11-dimensional kissing number bound to 593.", "before_after": { "metric": "Scalar multiplications for 4×4 complex matrices", "before": "49 (Strassen, 1969)", "after": "48" }, "why_it_matters": "A general-purpose LLM agent with automatic checking pushed many separate mathematical records at once, rather than one hand-picked problem.", "ai_systems": [ "AlphaEvolve", "Gemini" ], "organizations": [ "Google DeepMind" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2025-05-14", "url": "https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" }, { "label": "AlphaEvolve white paper, arXiv:2506.13131", "url": "https://arxiv.org/abs/2506.13131" } ], "foundation0": false }, { "id": "o1-matches-physicians-er-diagnosis", "date": "2025-05-19", "date_basis": "preprint", "title": "OpenAI's o1 out-diagnoses two attending physicians on real emergency-room cases", "domain": "medicine", "subfield": "diagnostic reasoning, emergency medicine", "summary": "OpenAI's o1-preview beat physician baselines in five reasoning experiments on published cases. On 76 real Boston emergency cases, blinded attendings rated o1's triage-stage diagnosis exact or very close in 67% of cases, versus 55% and 50% for two internal-medicine attendings given the same text records. At later stages o1 was nominally better or on par.", "before_after": { "metric": "Exact or very close diagnosis at ER triage (76 real cases)", "before": "55% and 50% (two internal-medicine attendings)", "after": "67% (OpenAI o1)" }, "why_it_matters": "Diagnosing from sparse, messy real records is core to emergency medicine. Beating physicians there, not just on exam questions, is why the authors call for prospective trials, though only two physicians were compared.", "ai_systems": [ "OpenAI o1-preview (case experiments)", "OpenAI o1 (emergency-room study)", "GPT-4o (comparison)" ], "organizations": [ "Harvard Medical School", "Beth Israel Deaconess Medical Center", "Stanford University", "Microsoft" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": true, "contested_note": "Experts quoted by the UK Science Media Centre (2026-04-30) said the task was blinded second-opinion differential diagnosis from text, not real-time emergency care or management, at a single centre, and that it does not show AI is safe for routine clinical use.", "significance": 3, "sources": [ { "label": "Science, 2026-04-30", "url": "https://doi.org/10.1126/science.adz4433" }, { "label": "arXiv preprint; v2 (2025-05-19) added the emergency-room study", "url": "https://arxiv.org/abs/2412.10849" }, { "label": "Science Media Centre expert reaction (76 cases; 67% vs 55% and 50%), 2026-04-30", "url": "https://www.sciencemediacentre.org/expert-reaction-to-study-evaluating-performance-of-a-large-language-model-on-the-reasoning-tasks-of-a-physician/" } ], "foundation0": false }, { "id": "robin-ripasudil-dry-amd", "date": "2025-05-19", "date_basis": "preprint", "title": "Multi-agent AI Robin proposes ripasudil as a dry macular degeneration candidate", "domain": "medicine", "subfield": "ophthalmology, automated discovery", "summary": "FutureHouse's Robin generated hypotheses, proposed experiments and analyzed data while humans ran the lab work. It proposed boosting retinal-pigment-cell phagocytosis and identified ripasudil, a glaucoma drug not previously proposed for dry AMD, which raised phagocytosis in vitro (1.89-fold by Robin's analysis; 1.75-fold by human analysis). Published in Nature, May 2026.", "before_after": null, "why_it_matters": "Points to AI systems running most of the reasoning loop of discovery. The candidate is still a cell-culture result with no patient data.", "ai_systems": [ "Robin", "Finch (data-analysis agent)" ], "organizations": [ "FutureHouse" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2026-05-19", "url": "https://doi.org/10.1038/s41586-026-10652-y" }, { "label": "arXiv preprint, 2025-05-19", "url": "https://arxiv.org/abs/2505.13400" } ], "foundation0": false }, { "id": "aurora-earth-system-foundation-model", "date": "2025-05-21", "date_basis": "publication", "title": "Aurora model beats operational forecasts for air quality, waves and cyclones", "domain": "earth-climate", "subfield": "Earth system forecasting", "summary": "Microsoft's Aurora, pretrained on more than one million hours of varied geophysical data and then fine-tuned, outperformed operational forecasts for air quality, ocean waves, tropical cyclone tracks and high-resolution weather, at orders of magnitude lower computing cost.", "before_after": null, "why_it_matters": "One pretrained model adapted cheaply to many Earth problems means accurate forecasts of smog, swell or storms no longer each need a separate expensive system.", "ai_systems": [ "Aurora" ], "organizations": [ "Microsoft Research", "University of Cambridge", "University of Amsterdam" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Bodnar et al., Nature, 2025-05-21", "url": "https://www.nature.com/articles/s41586-025-09005-y" } ], "foundation0": false }, { "id": "o3-ksmbd-linux-zero-day", "date": "2025-05-22", "date_basis": "announcement", "title": "OpenAI o3 finds a remote use-after-free bug in the Linux kernel's SMB server", "domain": "computing", "subfield": "vulnerability discovery", "summary": "Security researcher Sean Heelan gave OpenAI's o3 about 12,000 lines of the Linux kernel's ksmbd code through the plain API, with no tools or agent scaffolding. It identified a previously unknown use-after-free in the session logoff handler, now CVE-2025-37899. Output was noisy: Heelan estimated a signal-to-noise ratio of about 1:50.", "before_after": null, "why_it_matters": "A general model, prompted plainly, found a subtle concurrency bug the expert running it had not spotted. It showed AI code auditing is within reach of individual researchers, not only large labs.", "ai_systems": [ "OpenAI o3" ], "organizations": [ "Independent researcher (Sean Heelan)", "OpenAI" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Sean Heelan's blog, 2025-05-22", "url": "https://sean.heelan.io/2025/05/22/how-i-used-o3-to-find-cve-2025-37899-a-remote-zeroday-vulnerability-in-the-linux-kernels-smb-implementation/" }, { "label": "NVD record CVE-2025-37899, 2025-05-20", "url": "https://nvd.nist.gov/vuln/detail/CVE-2025-37899" } ], "foundation0": false }, { "id": "rentosertib-ipf-phase2a", "date": "2025-06-03", "date_basis": "publication", "title": "AI-discovered lung-fibrosis drug shows safety and early signal in phase 2a", "domain": "medicine", "subfield": "drug discovery, pulmonology", "summary": "Rentosertib, a TNIK inhibitor whose target and molecule came from generative AI, was given to 71 IPF patients for 12 weeks. Adverse-event rates resembled placebo (primary endpoint). Lung capacity (FVC, secondary endpoint) rose 98.4 mL on 60 mg daily versus a 20.3 mL fall on placebo. A 320-patient phase 3 started September 2026.", "before_after": { "metric": "12-week change in FVC (secondary endpoint)", "before": "-20.3 mL (placebo)", "after": "+98.4 mL (60 mg once daily)" }, "why_it_matters": "Reported as the first randomized patient data for a drug whose target and molecule both came from generative AI. Small and short, so efficacy remains unproven until phase 3.", "ai_systems": [ "PandaOmics", "Chemistry42" ], "organizations": [ "Insilico Medicine", "Peking Union Medical College Hospital" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Nature Medicine, 2025-06-03", "url": "https://doi.org/10.1038/s41591-025-03743-2" }, { "label": "ClinicalTrials.gov NCT07687459 (phase 3)", "url": "https://clinicaltrials.gov/study/NCT07687459" } ], "foundation0": false }, { "id": "boltz-2-open-binding-affinity-model", "date": "2025-06-06", "date_basis": "announcement", "title": "Open Boltz-2 model predicts drug binding affinity near physics-based accuracy", "domain": "chemistry-materials", "subfield": "drug design / binding affinity", "summary": "MIT's Jameel Clinic and Recursion released Boltz-2, an open-source co-folding model that predicts protein-ligand structures and binding affinity. Its developers report it is the first deep-learning model to approach free-energy perturbation (FEP) accuracy while running over 1,000 times faster; model, weights and training pipeline are MIT-licensed.", "before_after": { "metric": "Compute cost of affinity estimates", "before": "free-energy perturbation simulations", "after": "at least 1,000x cheaper, near-FEP accuracy (developer-reported)" }, "why_it_matters": "Fast, open affinity prediction could let drug designers screen far more molecules before costly simulation or synthesis.", "ai_systems": [ "Boltz-2" ], "organizations": [ "Massachusetts Institute of Technology", "Recursion" ], "ai_role": "instrument", "evidence": "preprint", "contested": true, "contested_note": "Accuracy claims are the developers'. An independent 2026 RIKEN benchmark of 356 targets and 10,933 compounds with no training overlap (Chem-Bio Informatics Journal) found 'fair' performance, mean absolute error about 0.9 pChEMBL units.", "significance": 3, "sources": [ { "label": "Boltz announcement, 2025-06-06", "url": "https://boltz.com/boltz2" }, { "label": "bioRxiv 2025.06.14.659707 (preprint)", "url": "https://doi.org/10.1101/2025.06.14.659707" }, { "label": "Chem-Bio Informatics Journal 26:11, 2026-03-18 (independent benchmark)", "url": "https://www.jstage.jst.go.jp/article/cbij/26/0/26_11/_article" } ], "foundation0": false }, { "id": "skala-deep-learning-xc-functional", "date": "2025-06-17", "date_basis": "preprint", "title": "Microsoft's learned Skala functional beats hybrid DFT at semi-local cost", "domain": "chemistry-materials", "subfield": "density functional theory", "summary": "Microsoft Research trained Skala, a deep-learning exchange-correlation functional, on a large new set of high-accuracy wavefunction reference data. The authors report 2.8 kcal/mol error on the GMTKN55 main-group benchmark, better than leading hybrid functionals, at the lower cost of semi-local DFT, and accuracy that improved as training data grew. It remains a preprint.", "before_after": { "metric": "Accuracy vs cost trade-off on GMTKN55", "before": "hybrid functionals needed for top accuracy", "after": "2.8 kcal/mol at semi-local cost" }, "why_it_matters": "Breaks the long trade-off between DFT accuracy and cost, suggesting first-principles simulations can become steadily more predictive as reference data grows.", "ai_systems": [ "Skala" ], "organizations": [ "Microsoft Research AI for Science" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv 2506.14665, 2025-06-17", "url": "https://arxiv.org/abs/2506.14665" }, { "label": "Skala project documentation (Microsoft)", "url": "https://microsoft.github.io/skala/index.html" } ], "foundation0": false }, { "id": "eve-llm-robot-scientist-yeast-discoveries", "date": "2025-06-27", "date_basis": "preprint", "title": "Robot scientist Eve, upgraded with LLM agents, finds new yeast interactions", "domain": "science-of-science", "subfield": "autonomous laboratories", "summary": "Ross King's team coupled language-model agents and logic-based hypothesis generation to the Eve robot's automated cell-culture and metabolomics platforms. From about 60,000 curated yeast relations it generated 1,933 hypotheses, designed and ran experiments, and refined them. It found that glutamate inhibits growth of spermine-treated cells and that aminoadipate partly rescues formic-acid stress.", "before_after": null, "why_it_matters": "Language models alone often reason incoherently. Anchoring them in formal logic and real experiments lets a machine generate, test and record hypotheses reliably, a step toward labs that run discovery cycles themselves.", "ai_systems": [ "LLM agents (GPT-4o in the preprint)", "Inductive logic programming", "Eve robot scientist" ], "organizations": [ "Chalmers University of Technology", "University of Gothenburg", "University of Cambridge" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Modest findings in a model organism; humans chose targets, checked plans for safety and moved plates. The main contribution is the reliable closed loop.", "significance": 2, "sources": [ { "label": "Brunnsåker et al., J. R. Soc. Interface 23:20260043, 2026-07-08", "url": "https://doi.org/10.1098/rsif.2026.0043" }, { "label": "bioRxiv preprint, 2025-06-27", "url": "https://doi.org/10.1101/2025.06.24.661378" }, { "label": "Chalmers news, 2026-09-30", "url": "https://www.chalmers.se/en/current/news/ai-scientist-autonomously-generates-and-validates-new-biological-discoveries/" } ], "foundation0": false }, { "id": "srt-h-autonomous-surgery-steps", "date": "2025-07-09", "date_basis": "publication", "title": "Robot autonomously performs gallbladder-removal steps on 8 of 8 organs", "domain": "engineering-robotics", "subfield": "surgical robotics", "summary": "SRT-H, from Johns Hopkins and Stanford, learned from demonstrations: a high-level policy plans in language and issues corrective instructions, and a low-level policy moves the instruments. It performed long sequences of cholecystectomy steps on eight ex vivo pig gallbladders with 100% success and no human intervention.", "before_after": { "metric": "Fully autonomous success on ex vivo cholecystectomy steps", "before": "Autonomy limited to simple tasks in controlled settings", "after": "8 of 8 gallbladders (100%)" }, "why_it_matters": "Tissue varies from patient to patient. A robot that plans, notices its errors and recovers on real tissue is a step toward consistent surgery where skilled surgeons are scarce.", "ai_systems": [ "SRT-H (hierarchical language-conditioned imitation learning)" ], "organizations": [ "Johns Hopkins University", "Stanford University" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science Robotics, 2025-07-09", "url": "https://doi.org/10.1126/scirobotics.adt5254" } ], "foundation0": false }, { "id": "openai-imo-2025-gold", "date": "2025-07-19", "date_basis": "competition", "title": "OpenAI experimental model scores at gold-medal level on IMO 2025", "domain": "mathematics", "subfield": "olympiad mathematics", "summary": "An experimental OpenAI reasoning model solved 5 of 6 IMO 2025 problems for 35 of 42 points under exam conditions: two 4.5-hour sessions, no tools or internet, natural-language proofs. Three former IMO medalists graded the proofs by consensus; the IMO did not officially grade the entry. The proofs were published on GitHub.", "before_after": { "metric": "Best AI score at the IMO", "before": "28/42 (2024; formal language, up to 3 days)", "after": "35/42 (gold); natural language, exam time" }, "why_it_matters": "A general-purpose language model, not a maths-specific system, wrote olympiad proofs at gold level within human time limits, a capability step directly relevant to research.", "ai_systems": [ "OpenAI experimental reasoning model" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "OpenAI IMO 2025 proofs (GitHub)", "url": "https://github.com/aw31/openai-imo-2025-proofs" }, { "label": "The Decoder, July 2025", "url": "https://the-decoder.com/openai-claims-a-breakthrough-in-llm-reasoning-on-complex-math-problems/" } ], "foundation0": false }, { "id": "gemini-deep-think-imo-2025-gold", "date": "2025-07-21", "date_basis": "competition", "title": "Gemini Deep Think certified at IMO gold-medal standard", "domain": "mathematics", "subfield": "olympiad mathematics", "summary": "An advanced version of Google's Gemini Deep Think solved 5 of 6 IMO 2025 problems for 35 of 42 points, within the 4.5-hour limit, writing natural-language proofs from the official problem statements. IMO coordinators graded and certified the answers by student criteria; they did not validate the system itself.", "before_after": { "metric": "Best officially graded AI score at the IMO", "before": "28/42 (2024, silver)", "after": "35/42 (gold)" }, "why_it_matters": "First gold-level AI result checked by the competition's own graders, showing natural-language reasoning had overtaken specialist formal systems within a year.", "ai_systems": [ "Gemini Deep Think (advanced version)" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2025-07-21", "url": "https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/" } ], "foundation0": false }, { "id": "penda-ai-consult-kenya-primary-care", "date": "2025-07-22", "date_basis": "preprint", "title": "LLM safety net cut diagnostic errors 16% in Kenyan primary-care clinics", "domain": "medicine", "subfield": "primary care, global health", "summary": "Penda Health and OpenAI studied 39,849 visits across 15 Nairobi primary-care clinics, randomly giving half the clinicians AI Consult, a GPT-4o tool that flags possible errors. Those clinicians made 16% fewer diagnostic and 13% fewer treatment errors, as rated by independent physicians. Patient outcomes were not the endpoint.", "before_after": null, "why_it_matters": "Clinics in low-income settings face heavy patient loads with few specialists. A well-integrated safety net could raise care quality cheaply, but patient-outcome benefit is not yet shown.", "ai_systems": [ "AI Consult (GPT-4o)" ], "organizations": [ "Penda Health", "OpenAI" ], "ai_role": "instrument", "evidence": "preprint", "contested": true, "contested_note": "A 2026 cluster-randomized trial in 16 Kenyan primary-care facilities (Nature Medicine; 9,691 patients; KEMRI-Wellcome, PATH, Penda) found LLM assistance safe but no significant cut in 14-day treatment failure (2.2% vs 2.0%; aOR 0.77, P=0.13). A 2026 retrospective safety evaluation (Nature Health) found actively harmful AI suggestions in 7.8% of encounters.", "significance": 3, "sources": [ { "label": "arXiv preprint, 2025-07-22", "url": "https://arxiv.org/abs/2507.16947" }, { "label": "Nature Medicine (cluster RCT), 2026-06-26", "url": "https://doi.org/10.1038/s41591-026-04503-6" }, { "label": "Nature Health (retrospective safety evaluation), 2026-03-10", "url": "https://doi.org/10.1038/s44360-026-00082-5" } ], "foundation0": false }, { "id": "opencrispr-1-ai-designed-genome-editor", "date": "2025-07-30", "date_basis": "publication", "title": "AI-designed OpenCRISPR-1 edits human cells, about 400 mutations from SpCas9", "domain": "biology", "subfield": "generative design of genome editors", "summary": "Profluent mined over 1 million CRISPR operons from 26 terabases of sequence and trained language models that generated 4.8 times as many protein clusters across CRISPR-Cas families as nature holds. Several designed editors matched or beat SpCas9 in activity and specificity despite about 400 mutations; OpenCRISPR-1 also works with base editing.", "before_after": null, "why_it_matters": "Natural editors come from microbes and often work poorly in human cells. Designing editors with AI can bypass evolutionary constraints and widen the gene-editing toolbox.", "ai_systems": [ "CRISPR-Cas protein language models (Profluent)", "OpenCRISPR-1" ], "organizations": [ "Profluent Bio", "École Polytechnique Fédérale de Lausanne", "University of Washington" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Ruffolo et al., Nature, 2025-07-30", "url": "https://doi.org/10.1038/s41586-025-09298-z" } ], "foundation0": false }, { "id": "dusty-plasma-ml-force-laws", "date": "2025-07-31", "date_basis": "publication", "title": "Physics-tailored machine learning uncovers unexpected forces in dusty plasma", "domain": "physics", "subfield": "plasma physics", "summary": "Emory physicists trained a neural network with built-in physical symmetries on 3D particle trajectories from a laboratory dusty plasma. It learned the non-reciprocal forces between particles with R² above 0.99, and enabled measurements of particle charge and screening length that deviate strongly from common theoretical assumptions.", "before_after": null, "why_it_matters": "Dusty plasmas are common in space and around planets. Learning force laws directly from particle motion, in an interpretable way, can correct theory for many-particle systems.", "ai_systems": [ "Physics-tailored neural network" ], "organizations": [ "Emory University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "PNAS, 2025-07-31", "url": "https://doi.org/10.1073/pnas.2505725122" }, { "label": "Emory University news, 2025-07", "url": "https://news.emory.edu/features/2025/07/ai-reveals-unexpected-new-physics-dusty-plasma" } ], "foundation0": false }, { "id": "darpa-aixcc-final", "date": "2025-08-08", "date_basis": "competition", "title": "DARPA AI Cyber Challenge: autonomous systems find and patch open-source bugs", "domain": "computing", "subfield": "automated vulnerability repair", "summary": "At the final of DARPA's two-year AI Cyber Challenge at DEF CON 33, autonomous systems combining language models and program analysis found 54 of 63 planted vulnerabilities and patched 43. They also found 18 real, unplanted vulnerabilities, averaging 45 minutes per patch and about $152 per task. Team Atlanta won; the systems are being open-sourced.", "before_after": { "metric": "Planted vulnerabilities in the final round", "before": "63 planted", "after": "54 found, 43 patched" }, "why_it_matters": "Open-source code underpins hospitals, power grids and banks, and its maintainers are stretched thin. Cheap automatic finding and fixing of bugs could close holes before attackers use them.", "ai_systems": [ "Atlantis (Team Atlanta)", "Buttercup (Trail of Bits)", "other AIxCC cyber reasoning systems" ], "organizations": [ "DARPA", "ARPA-H", "Team Atlanta (Georgia Tech, Samsung Research, KAIST, POSTECH)", "Trail of Bits", "Theori" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "DARPA news release, 2025-08-08", "url": "https://www.darpa.mil/news/2025/aixcc-results" }, { "label": "Trail of Bits blog, 2025-08-09", "url": "https://blog.trailofbits.com/2025/08/09/trail-of-bits-buttercup-wins-2nd-place-in-aixcc-challenge/" } ], "foundation0": false }, { "id": "generative-ai-antibiotics-gonorrhoea-mrsa", "date": "2025-08-14", "date_basis": "publication", "title": "Generative AI designs new antibiotics active against gonorrhoea and MRSA", "domain": "medicine", "subfield": "antibiotic discovery", "summary": "MIT and Broad researchers used fragment-based and unconstrained generative models (genetic algorithms, variational autoencoders) to design novel molecules against Neisseria gonorrhoeae and Staphylococcus aureus. Of 24 synthesized compounds, 7 showed selective antibacterial activity; two leads with distinct mechanisms reduced bacterial burden in mouse gonorrhoea and MRSA skin infections.", "before_after": null, "why_it_matters": "Moves AI from searching existing chemical libraries to inventing molecules never made before, reaching chemical space that libraries do not cover.", "ai_systems": [ "Genetic algorithms", "Variational autoencoders", "Graph neural network activity predictors" ], "organizations": [ "MIT", "Broad Institute", "Wyss Institute, Harvard", "Karolinska Institute" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Cell, 2025-08-14", "url": "https://doi.org/10.1016/j.cell.2025.07.033" } ], "foundation0": false }, { "id": "nif-ignition-deep-learning-forecast", "date": "2025-08-14", "date_basis": "publication", "title": "Deep-learning model forecast NIF's first fusion ignition as the likely outcome", "domain": "energy", "subfield": "inertial confinement fusion", "summary": "Before the December 2022 shot that first achieved fusion ignition, LLNL's generative model combining simulations, deep learning, experimental data and Bayesian statistics, trained on over 150,000 simulations, gave ignition a probability above 70% (74% per LLNL), far higher than for earlier designs. The measured yield fell within the predicted range.", "before_after": { "metric": "forecast ignition probability for the shot design", "before": "significantly lower for earlier designs", "after": "74% for the December 2022 design, which ignited" }, "why_it_matters": "Laser fusion shots are rare and costly. Calibrated AI forecasts let scientists judge designs before firing and quantify uncertainty, speeding the path toward fusion energy.", "ai_systems": [ "LLNL CogSim physics-informed deep learning model" ], "organizations": [ "Lawrence Livermore National Laboratory" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2025-08-14", "url": "https://doi.org/10.1126/science.adm8201" }, { "label": "LLNL news, 2025-08-21", "url": "https://www.llnl.gov/article/53316/llnl-used-ai-predict-historic-fusion-ignition-shot" } ], "foundation0": false }, { "id": "gauss-strong-pnt-formalization", "date": "2025-09", "date_basis": "announcement", "title": "Gauss agent completes Lean formalization of the strong Prime Number Theorem", "domain": "mathematics", "subfield": "analytic number theory / formalization", "summary": "Math Inc.'s autoformalization agent Gauss finished, in three weeks, the strong Prime Number Theorem formalization challenge that Terence Tao and Alex Kontorovich set in January 2024, producing about 25,000 lines of Lean and over 1,000 theorems and definitions, including missing complex-analysis results. Human experts supplied natural-language scaffolding.", "before_after": { "metric": "Strong PNT in Lean", "before": "Incomplete after ~18 months of human effort", "after": "Complete in 3 weeks (~25,000 lines)" }, "why_it_matters": "Machine-checked proofs remove doubt about correctness but were slow to write; cutting months to weeks makes verifying serious mathematics practical.", "ai_systems": [ "Gauss" ], "organizations": [ "Math Inc." ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Math Inc., Gauss announcement", "url": "https://www.math.inc/gauss" }, { "label": "math-inc/strongpnt (Lean code)", "url": "https://github.com/math-inc/strongpnt" } ], "foundation0": false }, { "id": "deep-loop-shaping-ligo", "date": "2025-09-04", "date_basis": "publication", "title": "Deep Loop Shaping cuts LIGO mirror control noise by 30 to 100 times", "domain": "physics", "subfield": "gravitational-wave instrumentation", "summary": "Google DeepMind, Caltech and the Gran Sasso Science Institute used reinforcement learning with frequency-domain rewards to design a controller for one of LIGO's hardest mirror-stabilization loops. Tested at LIGO Livingston, it cut control noise in the 10–30 Hz band by over 30 times, and up to 100 times in sub-bands, beating the quantum-limit design goal.", "before_after": { "metric": "control noise, 10–30 Hz band", "before": "conventional controller", "after": "over 30× lower (up to 100× in sub-bands)" }, "why_it_matters": "Control noise blinds detectors at low frequencies, where heavier black-hole mergers and early neutron-star warnings appear. Quieter control could let LIGO detect more events.", "ai_systems": [ "Deep Loop Shaping (reinforcement learning)" ], "organizations": [ "Google DeepMind", "Caltech (LIGO)", "Gran Sasso Science Institute" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Science, 2025-09-04", "url": "https://doi.org/10.1126/science.adw1291" }, { "label": "Google DeepMind blog, 2025-09-04", "url": "https://deepmind.google/blog/using-ai-to-perceive-the-universe-in-greater-depth/" } ], "foundation0": false }, { "id": "era-ai-empirical-scientific-software", "date": "2025-09-08", "date_basis": "preprint", "title": "Google's ERA writes scientific software that beats expert leaderboards", "domain": "science-of-science", "subfield": "automated scientific software", "summary": "Google's ERA pairs a language model with tree search to write code that maximises a score, often building on published methods. It produced 40 single-cell data-integration methods that beat the top human methods on a public leaderboard (best 14% above ComBat) and, in a retrospective test, 14 COVID-19 forecasters that beat the CDC ensemble.", "before_after": { "metric": "Single-cell batch-integration score (OpenProblems leaderboard)", "before": "ComBat, best published method", "after": "About 14% higher; 40 AI methods beat top human methods" }, "why_it_matters": "Much research waits on slow, specialised software. A system that searches the space of programs against a measurable goal can speed many empirical sciences at once.", "ai_systems": [ "ERA (Empirical Research Assistance)", "Gemini" ], "organizations": [ "Google Research", "Google DeepMind" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2026-05-19", "url": "https://doi.org/10.1038/s41586-026-10658-6" }, { "label": "arXiv preprint, 2025-09-08", "url": "https://arxiv.org/abs/2509.06503" }, { "label": "Google Research blog, 2025-09-09", "url": "https://research.google/blog/accelerating-scientific-discovery-with-ai-powered-empirical-software/" } ], "foundation0": false }, { "id": "satlution-evolved-sat-solvers", "date": "2025-09-09", "date_basis": "preprint", "title": "LLM agents evolve SAT solvers that beat 2025 winners in the authors' tests", "domain": "computing", "subfield": "SAT solving", "summary": "NVIDIA researchers used Claude-based coding agents in Cursor to evolve entire SAT-solver codebases, tens of thousands of lines of C/C++, under correctness checks and benchmark feedback, with human steering in later stages. Starting only from 2024 competition solvers and instances, the evolved solvers beat the 2025 winners' PAR-2 scores in the authors' own reruns.", "before_after": { "metric": "PAR-2 score on SAT Competition 2025 benchmark (authors' reruns)", "before": "AE kissat2025 MAB (2025 competition winner)", "after": "SATLUTION solvers lower (better)" }, "why_it_matters": "SAT solvers check chip designs, software and schedules. AI agents improving a mature, heavily tuned codebase as a whole, not just small kernels, widens what automated discovery can reach.", "ai_systems": [ "SATLUTION", "Claude (via Cursor agents)" ], "organizations": [ "NVIDIA Research", "University of Maryland" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv preprint, 2025-09-09", "url": "https://arxiv.org/abs/2509.07367" } ], "foundation0": false }, { "id": "india-ai-monsoon-onset-forecasts-38-million-farmers", "date": "2025-09-15", "date_basis": "announcement", "title": "AI monsoon-onset forecasts reach 38 million Indian farmers by text message", "domain": "earth-climate", "subfield": "subseasonal forecasting / agriculture", "summary": "In summer 2025 India's agriculture ministry sent probabilistic monsoon-onset forecasts by SMS to about 38 million farmers. University of Chicago researchers blended Google's NeuralGCM and ECMWF's AIFS with rainfall statistics after hindcasts showed most models' onset errors about two days below climatology at 1-15 day leads. The forecasts caught an unusual weeks-long monsoon pause.", "before_after": { "metric": "Onset-date error, 1-15 day leads (hindcasts)", "before": "climatological forecast", "after": "about 2 days smaller for most models" }, "why_it_matters": "Planting date is a key decision for hundreds of millions of rain-fed farmers. Cheap AI weather models make skilful local forecasts feasible for low- and middle-income countries.", "ai_systems": [ "NeuralGCM", "ECMWF AIFS" ], "organizations": [ "University of Chicago (Human-Centered Weather Forecasts Initiative)", "Ministry of Agriculture and Farmers' Welfare, India", "Google Research" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv 2602.03767 (benchmarking and 2025 deployment), 2026-02-03", "url": "https://arxiv.org/abs/2602.03767" }, { "label": "UChicago HCWF news, 2025-09-15", "url": "https://humancenteredforecasts.climate.uchicago.edu/news/how-ai-is-helping-38-million-farmers-with-advance-weather-predictions/" }, { "label": "arXiv 2603.07893 (blended probabilistic system), 2026-03-09", "url": "https://arxiv.org/abs/2603.07893" } ], "foundation0": false }, { "id": "delphi-2m-disease-trajectory-model", "date": "2025-09-17", "date_basis": "publication", "title": "GPT-style model forecasts risk of over 1,000 diseases from health records", "domain": "medicine", "subfield": "epidemiology, risk prediction", "summary": "Delphi-2M, a modified GPT model trained on 400,000 UK Biobank participants and validated without changes on 1.9 million Danish records, predicts rates of more than 1,000 diseases from a person's medical history, with accuracy comparable to single-disease models, and can simulate health trajectories up to 20 years ahead.", "before_after": null, "why_it_matters": "Health systems usually plan prevention one disease at a time. One model of how diseases follow each other could target screening to those most at risk.", "ai_systems": [ "Delphi-2M" ], "organizations": [ "German Cancer Research Center (DKFZ)", "EMBL-EBI", "University of Copenhagen" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-09-17", "url": "https://doi.org/10.1038/s41586-025-09529-3" } ], "foundation0": false }, { "id": "icpc-2025-ai-outscores-humans", "date": "2025-09-17", "date_basis": "announcement", "title": "AI systems reach gold-medal level at ICPC World Finals; one solves all 12", "domain": "computing", "subfield": "competitive programming", "summary": "At the 2025 ICPC World Finals, working under contest rules with official judging, OpenAI's system (GPT-5 plus an experimental reasoning model) solved all 12 problems; the best of 139 human teams solved 11. Google's Gemini 2.5 Deep Think solved 10, including one no university team solved, which would have ranked second.", "before_after": { "metric": "Problems solved, ICPC World Finals 2025", "before": "11 of 12 (best human team)", "after": "12 of 12 (OpenAI); 10 of 12 (Gemini)" }, "why_it_matters": "Contest problems are new algorithmic puzzles solved under time pressure. Beating the best student teams shows AI can reliably design and implement correct algorithms, a core research skill.", "ai_systems": [ "GPT-5", "OpenAI experimental reasoning model", "Gemini 2.5 Deep Think" ], "organizations": [ "OpenAI", "Google DeepMind", "ICPC" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Google DeepMind blog, 2025-09-17", "url": "https://deepmind.google/blog/gemini-achieves-gold-medal-level-at-the-international-collegiate-programming-contest-world-finals/" }, { "label": "The Decoder, 2025-09-17", "url": "https://the-decoder.com/openai-outperforms-humans-and-google-at-the-worlds-top-collegiate-programming-contest/" } ], "foundation0": false }, { "id": "unstable-singularities-fluids", "date": "2025-09-17", "date_basis": "preprint", "title": "Neural-network solvers discover unstable singularities in fluid equations", "domain": "mathematics", "subfield": "partial differential equations / fluid dynamics", "summary": "Google DeepMind with mathematicians from Brown, NYU and Stanford combined purpose-built neural networks with a high-precision Gauss–Newton optimizer to find the first systematic families of unstable self-similar blow-up solutions, for the incompressible porous media and 3D Euler (with boundary) equations, some near double-float machine precision.", "before_after": null, "why_it_matters": "Unstable singularities are thought central to whether Euler or Navier–Stokes flows can blow up, a Millennium Prize question; precise candidates are a route to computer-assisted proofs.", "ai_systems": [ "Physics-informed neural networks" ], "organizations": [ "Google DeepMind", "Brown University", "New York University", "Stanford University" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2509.14185", "url": "https://arxiv.org/abs/2509.14185" }, { "label": "Google DeepMind blog, Sept 2025", "url": "https://deepmind.google/discover/blog/discovering-new-solutions-to-century-old-problems-in-fluid-dynamics" }, { "label": "Google DeepMind blog, 2025-09-18", "url": "https://deepmind.google/blog/discovering-new-solutions-to-century-old-problems-in-fluid-dynamics/" } ], "foundation0": false }, { "id": "crest-multimodal-catalyst-discovery", "date": "2025-09-23", "date_basis": "publication", "title": "MIT's CRESt platform finds eight-element fuel-cell catalyst in three months", "domain": "chemistry-materials", "subfield": "self-driving labs / electrocatalysis", "summary": "MIT's CRESt joined multimodal models that read compositions, literature text and microscope images with knowledge-assisted Bayesian optimization, robotics and camera monitoring. In three months it explored over 900 catalyst chemistries and ran 3,500 electrochemical tests, finding a Pd-Pt-Cu-Au-Ir-Ce-Nb-Cr formate-oxidation catalyst with 9.3-fold better cost-specific performance.", "before_after": { "metric": "Cost-specific performance, formate oxidation catalyst", "before": "benchmark catalyst (1x)", "after": "9.3x" }, "why_it_matters": "Fuel cells depend on scarce precious metals. Systems that read papers, watch experiments and diagnose their own errors can search element mixtures too vast for people.", "ai_systems": [ "CRESt (large multimodal models with knowledge-assisted Bayesian optimization)" ], "organizations": [ "MIT" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-09-23", "url": "https://www.nature.com/articles/s41586-025-09640-5" } ], "foundation0": false }, { "id": "aaronson-qma-gpt5-step", "date": "2025-09-25", "date_basis": "preprint", "title": "GPT-5 suggests a key proof step in a quantum complexity paper", "domain": "mathematics", "subfield": "quantum complexity theory", "summary": "Scott Aaronson and Freek Witteveen proved that black-box amplification in the quantum class QMA cannot push completeness closer to 1 than doubly exponential. In the first version the key function in the proof was suggested by GPT-5 Thinking; the published version (Quantum, October 2026) uses a simpler function proposed by new coauthor Phillip Harris.", "before_after": null, "why_it_matters": "A leading theorist publicly credited an AI with a key technical step in a research proof, an early sign of AI as a working research assistant.", "ai_systems": [ "GPT-5 Thinking" ], "organizations": [ "University of Texas at Austin", "QuSoft & CWI Amsterdam", "University of Bonn" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Quantum 10, 2227 (2026-10-06)", "url": "https://doi.org/10.22331/q-2026-10-06-2227" }, { "label": "arXiv:2509.21131 (v1, 2025-09-25)", "url": "https://arxiv.org/abs/2509.21131" }, { "label": "Aaronson, The QMA Singularity, Sept 2025", "url": "https://scottaaronson.blog/?p=9183" } ], "foundation0": false }, { "id": "erdos-gpt5-literature-search", "date": "2025-10", "date_basis": "announcement", "title": "GPT-5 'solves' 10 Erdős problems, but had found existing solutions", "domain": "mathematics", "subfield": "combinatorics / literature search", "summary": "OpenAI's Kevin Weil posted that GPT-5 'found solutions to 10 (!) previously unsolved Erdős problems' and progressed on 11 others. Site maintainer Thomas Bloom called it 'a dramatic misrepresentation': GPT-5 had located published papers he did not know about. The post was deleted, though the searches did update the database.", "before_after": { "metric": "New solutions to open Erdős problems", "before": "Claimed: 10", "after": "Verified: 0 (existing papers located)" }, "why_it_matters": "Showed AI literature search can surface forgotten results, and that claims of AI discovery must be checked against prior work before being announced.", "ai_systems": [ "GPT-5" ], "organizations": [ "OpenAI" ], "ai_role": "instrument", "evidence": "claimed", "contested": true, "contested_note": "Thomas Bloom (erdosproblems.com) called the claim 'a dramatic misrepresentation'; Demis Hassabis wrote 'This is embarrassing'; Sébastien Bubeck conceded only literature solutions were found (TechCrunch, 2025-10-19).", "significance": 2, "sources": [ { "label": "TechCrunch, 2025-10-19", "url": "https://techcrunch.com/2025/10/19/openais-embarrassing-math/" }, { "label": "The Decoder, Oct 2025", "url": "https://the-decoder.com/leading-openai-researcher-announced-a-gpt-5-math-breakthrough-that-never-happened/" } ], "foundation0": false }, { "id": "codemender-security-patches", "date": "2025-10-06", "date_basis": "announcement", "title": "CodeMender agent upstreams 72 security fixes to open-source projects", "domain": "computing", "subfield": "automated vulnerability repair", "summary": "Google DeepMind's CodeMender, built on Gemini Deep Think with fuzzing, static analysis and differential testing, writes and self-checks security patches. In six months of development it upstreamed 72 security fixes to open-source projects, some as large as 4.5 million lines of code; human researchers reviewed every patch before submission.", "before_after": null, "why_it_matters": "Finding bugs is only half the job; someone must fix them correctly. Agents that write sound patches could let fixing keep pace as AI speeds up bug discovery.", "ai_systems": [ "CodeMender", "Gemini Deep Think" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Google DeepMind blog, 2025-10-06", "url": "https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/" } ], "foundation0": false }, { "id": "c2s-scale-silmitasertib-antigen-presentation", "date": "2025-10-15", "date_basis": "announcement", "title": "27-billion-parameter cell model proposes a drug that held up in lab cells", "domain": "biology", "subfield": "single-cell language models, virtual cells", "summary": "Google and Yale trained a 27-billion-parameter language model on single-cell data. A virtual screen nominated silmitasertib, a CK2 inhibitor, to raise antigen presentation only in an immune-active context. In human cell models unseen in training, silmitasertib plus low-dose interferon raised it roughly 50%, per the developers. Cell tests only; preprint.", "before_after": null, "why_it_matters": "Antigen presentation is a key immune trigger. A model that proposes a context-specific drug effect that then holds in cells is an early case of AI-guided hypothesis generation in biology.", "ai_systems": [ "C2S-Scale 27B (Cell2Sentence-Scale, built on Gemma)" ], "organizations": [ "Google DeepMind", "Yale University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Google blog, 2025-10-15", "url": "https://blog.google/innovation-and-ai/products/google-gemma-ai-cancer-therapy-discovery/" }, { "label": "bioRxiv preprint, posted 2025-04-17, revised 2025-10-11 and later", "url": "https://doi.org/10.1101/2025.04.14.648850" } ], "foundation0": false }, { "id": "discorl-discovered-rl-rule", "date": "2025-10-22", "date_basis": "publication", "title": "Machine-discovered reinforcement learning rule beats human-designed ones", "domain": "computing", "subfield": "machine learning algorithms", "summary": "DeepMind meta-learned a reinforcement learning update rule from the experience of a population of agents across many environments. The discovered rule, Disco57, scored an interquartile mean of 13.86 on the 57-game Atari benchmark, beating all existing hand-designed rules including MuZero, and reached state of the art on unseen benchmarks such as ProcGen.", "before_after": { "metric": "Atari-57 human-normalized IQM score", "before": "Best hand-designed RL rules (e.g. MuZero)", "after": "13.86 (Disco57), above all prior rules" }, "why_it_matters": "Learning algorithms have been designed by hand for decades. Machines discovering better ones suggests future AI methods may be found automatically from experience.", "ai_systems": [ "DiscoRL (Disco57)" ], "organizations": [ "Google DeepMind" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2025-10-22", "url": "https://www.nature.com/articles/s41586-025-09761-x" }, { "label": "Nature news, 2025-10-22", "url": "https://www.nature.com/articles/d41586-025-03398-6" } ], "foundation0": false }, { "id": "ai-led-diabetes-prevention-rct", "date": "2025-10-27", "date_basis": "publication", "title": "Fully automated AI lifestyle coach matches human coaches for prediabetes", "domain": "medicine", "subfield": "prevention, digital therapeutics", "summary": "In a 368-person randomized noninferiority trial at two US sites led by Johns Hopkins, referral to a fully automated, app-based AI Diabetes Prevention Program met the composite weight, activity and HbA1c goal as often as human-coached programs (31.7% vs 31.9%). More people started the AI program (93.4% vs 82.7%).", "before_after": { "metric": "Participants reaching composite prevention goal at 12 months", "before": "31.9% (human coach)", "after": "31.7% (AI coach, noninferior)" }, "why_it_matters": "Proven diabetes-prevention coaching is underused because human coaches are scarce. An automated coach that works as well could scale prevention cheaply.", "ai_systems": [ "Sweetch AI-led DPP app" ], "organizations": [ "Johns Hopkins University", "Reading Hospital (Tower Health)" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "JAMA, 2025-10-27 (issue 2025-12-16)", "url": "https://doi.org/10.1001/jama.2025.19563" } ], "foundation0": false }, { "id": "alphaevolve-at-scale-67-problems", "date": "2025-11-03", "date_basis": "preprint", "title": "AlphaEvolve tested on 67 problems with Tao and Gómez-Serrano", "domain": "mathematics", "subfield": "analysis, combinatorics, geometry, number theory", "summary": "Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao and Adam Zsolt Wagner ran AlphaEvolve on 67 problems. It rediscovered the best known constructions in most cases, improved several, sometimes generalized finite results into formulas valid for all inputs, and was combined with Deep Think and AlphaProof to generate proofs.", "before_after": null, "why_it_matters": "A systematic test by leading mathematicians across dozens of problems, mapping where AI search helps and where it does not, instead of one cherry-picked success.", "ai_systems": [ "AlphaEvolve", "Gemini Deep Think", "AlphaProof" ], "organizations": [ "Google DeepMind", "Brown University", "UCLA" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2511.02864", "url": "https://arxiv.org/abs/2511.02864" } ], "foundation0": false }, { "id": "kosmos-ai-scientist", "date": "2025-11-04", "date_basis": "preprint", "title": "Kosmos AI scientist runs 12-hour autonomous data-driven research campaigns", "domain": "science-of-science", "subfield": "autonomous research agents", "summary": "Kosmos runs up to 12 hours, reading about 1,500 papers and executing 42,000 lines of code per run, coordinated through a structured world model. Independent scientists judged 79.4% of its report statements accurate; collaborators said a run equalled about six months of their work. Three of seven highlighted discoveries matched findings it had not seen.", "before_after": null, "why_it_matters": "Staying coherent over hundreds of steps lets an AI pursue a question in depth, so data-rich fields could test far more hypotheses than human teams have time for.", "ai_systems": [ "Kosmos" ], "organizations": [ "FutureHouse", "Edison Scientific" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv preprint, 2025-11-04", "url": "https://arxiv.org/abs/2511.02824" } ], "foundation0": false }, { "id": "innocube-drl-attitude-control-in-orbit", "date": "2025-11-11", "date_basis": "announcement", "title": "Nanosatellite performs an in-orbit attitude manoeuvre with a learned controller", "domain": "engineering-robotics", "subfield": "spacecraft control", "summary": "On 30 October 2025 the Würzburg 3U nanosatellite InnoCube turned to a target attitude with its reaction wheels under a deep-reinforcement-learning controller trained only in simulation; repeated manoeuvres also worked. The team calls it the first in-orbit demonstration of an AI-based attitude controller and compares it with InnoCube's classical PD controller.", "before_after": { "metric": "AI-based satellite attitude control", "before": "simulations and ground testbeds only", "after": "inertial-pointing manoeuvres in orbit, from 30 October 2025" }, "why_it_matters": "Classical attitude controllers take long to design and are sensitive to model errors. A controller learned in simulation that works on a real spacecraft could cut design time and adapt to changing conditions.", "ai_systems": [ "LeLaR deep-reinforcement-learning attitude controller" ], "organizations": [ "Julius-Maximilians-Universität Würzburg", "Technische Universität Berlin" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "The 'first' claim is the team's own. Tests cover inertial-pointing manoeuvres on one small satellite.", "significance": 2, "sources": [ { "label": "University of Würzburg news, 2025-11-11", "url": "https://www.uni-wuerzburg.de/en/news-and-events/einblick/single/news/ai-control/" }, { "label": "Djebko et al., IEEE Access 14:49348, 2026", "url": "https://doi.org/10.1109/ACCESS.2026.3678816" }, { "label": "arXiv preprint 2512.19576, 2025-12-22", "url": "https://arxiv.org/abs/2512.19576" } ], "foundation0": false }, { "id": "pacman-integrated-ai-control-diii-d", "date": "2025-11-11", "date_basis": "preprint", "title": "AI control framework runs five live fusion experiments on the DIII-D tokamak", "domain": "energy", "subfield": "fusion plasma control", "summary": "PPPL and Princeton's PACMAN framework links machine-learning predictors and controllers in a loop that typically runs in about 20 milliseconds. In five DIII-D experiments it let a reinforcement-learning agent run the heating, steered all six microwave gyrotrons at once, controlled density and rotation, and predicted a tearing instability ~200 ms ahead to avoid it.", "before_after": { "metric": "Warning before a tearing-mode instability", "before": "conventional control acts only after onset", "after": "predicted ~200 ms ahead and avoided" }, "why_it_matters": "Fusion plasmas need many actuators adjusted faster than people can react. A shared, safety-limited framework lets several AI models predict and control a real tokamak together.", "ai_systems": [ "PACMAN (Prediction And Control using MAchiNe learning)", "Reinforcement-learning heating controller" ], "organizations": [ "Princeton Plasma Physics Laboratory", "Princeton University", "DIII-D National Fusion Facility" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Rothstein et al., Nuclear Fusion 66, 076050, 2026-07-02", "url": "https://doi.org/10.1088/1741-4326/ae7f9d" }, { "label": "PPPL news release via Newswise, 2026-09-02", "url": "https://www.newswise.com/articles/pacman-ai-framework-for-controlling-fusion-systems-safely-makes-key-decisions-in-milliseconds" }, { "label": "arXiv:2511.08818, 2025-11-11", "url": "https://arxiv.org/abs/2511.08818" } ], "foundation0": false }, { "id": "packingstar-kissing-numbers-25-31", "date": "2025-11-17", "date_basis": "preprint", "title": "Reinforcement learning sets new kissing-number records in dimensions 25 to 31", "domain": "mathematics", "subfield": "discrete geometry / sphere packing", "summary": "PackingStar, a reinforcement learning system that treats the kissing number problem as a two-player game of filling in tables of cosines between sphere centres, beat the best known lower bounds in every dimension from 25 to 31, for example from 197,048 to 197,056 touching spheres in 25 dimensions and from 232,874 to 238,350 in 31.", "before_after": { "metric": "Kissing number lower bound in 31 dimensions", "before": "232,874", "after": "238,350 (238,078 in the first version)" }, "why_it_matters": "The kissing number, how many equal spheres can touch a central one, is known exactly in only six dimensions; the old records in dimensions 25 to 31 all came from one Leech-lattice construction.", "ai_systems": [ "PackingStar" ], "organizations": [ "Peking University", "Shanghai Academy of AI for Science", "Fudan University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "The configurations are explicit and checkable but not yet peer-reviewed. The 13-dimensional 1,146 beats only the 1,130 rational construction of 1971, not the 1,154 record. The September 2026 revision reports 15 improved bounds across kissing-type problems.", "significance": 2, "sources": [ { "label": "Ma et al., Finding Kissing Numbers with Game-theoretic Reinforcement Learning, arXiv:2511.13391", "url": "https://arxiv.org/abs/2511.13391" } ], "foundation0": false }, { "id": "pi-star-0-6-recap-robot-rl-from-experience", "date": "2025-11-17", "date_basis": "announcement", "title": "Robot model learns from its own practice, doubling throughput on hard tasks", "domain": "engineering-robotics", "subfield": "robot learning", "summary": "Physical Intelligence trained its vision-language-action model π*0.6 with Recap, reinforcement learning from demonstrations, expert corrections and the robot's own attempts. On some of the hardest tasks, practice more than doubled throughput and roughly halved failure rates; robots made espresso from 5:30am to 11:30pm and folded 50 novel laundry items in a new home.", "before_after": { "metric": "Successful tasks per hour, diverse laundry and espresso", "before": "base π0.6 fine-tuned on demonstrations", "after": "more than 2x, with failure rate roughly halved" }, "why_it_matters": "Robots trained only by imitation compound small errors. Learning from their own mistakes and corrections is a step toward reliable work outside the lab.", "ai_systems": [ "π*0.6 vision-language-action model trained with Recap" ], "organizations": [ "Physical Intelligence" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Company results on its own tasks and robots, compared with its own baselines; not peer reviewed.", "significance": 2, "sources": [ { "label": "Physical Intelligence blog, 2025-11-17", "url": "https://www.pi.website/blog/pistar06" }, { "label": "Physical Intelligence, arXiv 2511.14759, 2025-11-18", "url": "https://arxiv.org/abs/2511.14759" } ], "foundation0": false }, { "id": "gpt5-science-case-studies-erdos-848", "date": "2025-11-20", "date_basis": "preprint", "title": "GPT-5 case studies include four new maths results, among them Erdős #848", "domain": "mathematics", "subfield": "combinatorial number theory / multiple fields", "summary": "OpenAI and academic collaborators documented GPT-5 in live research, including four new, human-verified mathematics results. In one, GPT-5 Pro supplied an idea that, with earlier online comments, let Mehtaab Sawhney settle Erdős problem #848: for large N, the numbers ≡ 7 (mod 25) form a largest set in which no product-plus-one ab+1 is squarefree.", "before_after": null, "why_it_matters": "It records, with transcripts, where a general-purpose model actually moved research forward and where it failed, giving mathematicians concrete evidence rather than anecdotes.", "ai_systems": [ "GPT-5", "GPT-5 Pro" ], "organizations": [ "OpenAI", "Columbia University", "Harvard University", "University of Oxford" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "The authors call the new results modest in scope. For #848, GPT-5's own attempts at a full proof had numerous errors; Sawhney turned its idea into the proof.", "significance": 2, "sources": [ { "label": "Bubeck et al., Early science acceleration experiments with GPT-5, arXiv:2511.16072", "url": "https://arxiv.org/abs/2511.16072" }, { "label": "OpenAI, Early experiments in accelerating science with GPT-5 (2025-11-20)", "url": "https://openai.com/index/accelerating-science-gpt-5/" }, { "label": "erdosproblems.com, Problem #848", "url": "https://www.erdosproblems.com/848" } ], "foundation0": false }, { "id": "bgs-ml-aftershock-forecasts-seconds", "date": "2025-11-25", "date_basis": "publication", "title": "Deep learning forecasts aftershocks in seconds with skill similar to ETAS", "domain": "earth-climate", "subfield": "seismology", "summary": "British Geological Survey, Edinburgh and Padua researchers trained deep-learning models on earthquake catalogues from California, New Zealand, Italy, Japan and Greece to forecast how many aftershocks follow magnitude 4+ quakes over 24 hours, and where. On Southern California and Italy tests they matched the standard ETAS model, in seconds instead of hours to days.", "before_after": { "metric": "Time to produce a 24-hour aftershock forecast", "before": "hours to days (ETAS simulations on a mid-range computer)", "after": "seconds, with comparable skill on two test regions" }, "why_it_matters": "Aftershocks can be deadlier than the main quake. Forecasts available almost at once could guide rescue, evacuation and building checks while a seismic crisis is still unfolding.", "ai_systems": [ "Deep-learning seismicity rate forecasting models" ], "organizations": [ "British Geological Survey", "University of Edinburgh", "University of Padua" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Exploratory study (titled 'Towards...'): comparable skill to ETAS was shown on Southern California and Italy test sets only, and the models are not yet used operationally.", "significance": 2, "sources": [ { "label": "Dervisi et al., Earth, Planets and Space 77:185, 2025-11-25", "url": "https://doi.org/10.1186/s40623-025-02241-6" }, { "label": "British Geological Survey news release, 2025-11-25", "url": "https://www.bgs.ac.uk/news/new-research-shows-ai-earthquake-tools-forecast-aftershock-risk-in-seconds/" } ], "foundation0": false }, { "id": "ambient-ai-scribe-wellbeing-rct", "date": "2025-11-26", "date_basis": "publication", "title": "Randomized trial: ambient AI scribes reduce clinician work exhaustion", "domain": "medicine", "subfield": "care delivery, clinical documentation", "summary": "In a 24-week stepped-wedge trial, 66 UW Health practitioners in Wisconsin and Illinois were randomized to staggered starts with Abridge, an ambient AI that drafts visit notes. It drafted 38% of 71,487 notes; work exhaustion fell 0.44 points (five-point scale) and note time 0.36 hours a day, with no loss of note quality.", "before_after": { "metric": "Time spent on notes per practitioner per day", "before": "usual documentation", "after": "0.36 hours less with ambient AI" }, "why_it_matters": "Paperwork drives physician burnout and takes time from patients. Randomized evidence, not just surveys, shows generative AI can return some time, though trials were short and gains varied by product.", "ai_systems": [ "Abridge ambient AI (UW Health trial)", "Microsoft DAX Copilot and Nabla (companion UCLA trial)" ], "organizations": [ "University of Wisconsin School of Medicine and Public Health", "UW Health", "UCLA Health" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "Results differ by product: a companion UCLA randomized trial in the same NEJM AI issue (238 physicians) found Nabla cut time-in-note 9.5% while Microsoft DAX showed no significant change; burnout measures there were secondary end points.", "significance": 2, "sources": [ { "label": "NEJM AI (UW Health stepped-wedge trial), 2025-11-26", "url": "https://doi.org/10.1056/AIoa2500945" }, { "label": "NEJM AI (UCLA three-arm trial), 2025-11-26", "url": "https://doi.org/10.1056/AIoa2501000" } ], "foundation0": false }, { "id": "aristotle-erdos-124-weaker-version", "date": "2025-11-29", "date_basis": "announcement", "title": "Aristotle proves a version of Erdős problem #124 in Lean on its own", "domain": "mathematics", "subfield": "number theory / complete sequences", "summary": "Working only from a formal statement in the Formal Conjectures project, Harmonic's Aristotle spent six hours on a Lean proof that, whenever 1/(d₁−1)+…+1/(d_r−1) ≥ 1, every whole number is a sum of one number per base dᵢ with base-dᵢ digits all 0 or 1. Boris Alexeev posted it on 29 November 2025.", "before_after": null, "why_it_matters": "An early case of an AI prover settling a listed problem end to end, and a lesson in how loosely stated problems can make 'solved' claims misleading.", "ai_systems": [ "Aristotle" ], "organizations": [ "Harmonic", "erdosproblems.com community" ], "ai_role": "originator", "evidence": "formally-verified", "contested": true, "contested_note": "Thomas Bloom (forum, 30 Nov 2025): Aristotle solved 'a' version, with an olympiad-style proof, but not 'the' version; the Burr–Erdős–Graham–Li form, which excludes the digit 1s, remains open. The original formal statement also had a typo that weakened it, which Alexeev corrected.", "significance": 1, "sources": [ { "label": "erdosproblems.com forum, Problem #124 thread", "url": "https://www.erdosproblems.com/forum/thread/124" }, { "label": "erdosproblems.com, Problem #124", "url": "https://www.erdosproblems.com/124" } ], "foundation0": false }, { "id": "brainomix-stroke-ai-england-thrombectomy", "date": "2025-12-02", "date_basis": "publication", "title": "Stroke AI software rollout tied to doubled thrombectomy rates in England", "domain": "medicine", "subfield": "neurology, stroke imaging", "summary": "Across all 107 English NHS stroke hospitals (452,952 admissions, 2019-2023), the 26 sites that systematically adopted Brainomix 360 Stroke saw thrombectomy rates rise from 2.3% to 4.6%, versus 1.6% to 2.6% elsewhere. Patients whose scans were read with the AI had higher odds of thrombectomy (OR 1.57). Observational; some authors work for Brainomix.", "before_after": { "metric": "Thrombectomy rate at AI-adopting hospitals", "before": "2.3% (pre-implementation)", "after": "4.6% (post-implementation)" }, "why_it_matters": "Clot removal saves brain function but depends on fast expert scan reading. AI that flags blocked vessels may widen access across a national system, though observational data cannot prove causation.", "ai_systems": [ "Brainomix 360 Stroke" ], "organizations": [ "Royal Berkshire NHS Foundation Trust", "University of Oxford", "Health Innovation Oxford and Thames Valley", "Brainomix" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "The Lancet Digital Health, 2025-12-02", "url": "https://doi.org/10.1016/j.landig.2025.100927" } ], "foundation0": false }, { "id": "putnam-2025-axiomprover-lean-12-of-12", "date": "2025-12-06", "date_basis": "competition", "title": "AxiomProver writes Lean proofs for all 12 Putnam 2025 problems", "domain": "mathematics", "subfield": "competition mathematics / formal proof", "summary": "Axiom's AxiomProver received the Putnam 2025 problems section by section from an official proctor on 6 December 2025 and autonomously produced Lean proofs for all 12, eight of them before the exam closed. Axiom released the machine-checked proofs with commentary on 8 January 2026.", "before_after": null, "why_it_matters": "Every answer is machine-checked, so no human grader is needed to trust it. A formal prover handled a full, unseen, hard exam under time pressure.", "ai_systems": [ "AxiomProver" ], "organizations": [ "Axiom Math" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Timing (8 of 12 within the exam window) is Axiom's own account, and a technical report was still to come. Some proofs are long and brute-force, such as a 1,061-line case analysis for B4.", "significance": 2, "sources": [ { "label": "Axiom, From Seeing Why to Checking Everything (2026-01-08)", "url": "https://axiommath.ai/research/from-seeing-why-to-checking-everything" } ], "foundation0": false }, { "id": "alphaqubit-2-real-time-decoding", "date": "2025-12-08", "date_basis": "preprint", "title": "AlphaQubit 2 decodes surface and colour quantum codes in real time", "domain": "computing", "subfield": "quantum error correction", "summary": "Google's AlphaQubit 2 neural decoder reached near-optimal logical error rates for surface and colour codes in realistic-noise simulations and matched AlphaQubit 1 on Willow chip data. It decoded in under 1 µs per cycle on commercial accelerators: the surface code up to distance 11 and, a first, the colour code up to distance 9.", "before_after": { "metric": "real-time colour-code decoding", "before": "not achieved by high-accuracy decoders", "after": "under 1 µs per cycle up to distance 9" }, "why_it_matters": "Superconducting qubits are measured about a million times a second, so decoders must be fast and accurate. Real-time neural decoding removes the speed limit of the first AlphaQubit.", "ai_systems": [ "AlphaQubit 2" ], "organizations": [ "Google DeepMind", "Google Quantum AI" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv preprint 2512.07737, 2025-12-08", "url": "https://arxiv.org/abs/2512.07737" } ], "foundation0": false }, { "id": "equational-theories-project", "date": "2025-12-08", "date_basis": "preprint", "title": "Equational Theories Project settles all 22 million magma law implications", "domain": "mathematics", "subfield": "universal algebra / formalization", "summary": "A Terence Tao-led online collaboration launched in September 2024 settled all 22,028,942 implications between the 4,694 simplest equational laws for magmas, combining human proofs with automated theorem provers such as Vampire and Prover9/Mace4, all checked in Lean. Large language models played only a minor role.", "before_after": { "metric": "Implications resolved among 4,694 laws", "before": "Unmapped (Sept 2024)", "after": "All 22,028,942, Lean-verified" }, "why_it_matters": "A model for large-scale mathematics: many contributors plus automated provers, with a proof checker guaranteeing each of millions of results.", "ai_systems": [ "Vampire", "Prover9 / Mace4", "egg-based provers (MagmaEgg)" ], "organizations": [ "UCLA" ], "ai_role": "instrument", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2512.07087", "url": "https://arxiv.org/abs/2512.07087" } ], "foundation0": false }, { "id": "erdos-1026-aristotle-alphaevolve-collaboration", "date": "2025-12-08", "date_basis": "announcement", "title": "AI tools help pin down the exact answer to Erdős problem #1026", "domain": "mathematics", "subfield": "combinatorics / monotone subsequences", "summary": "An ambiguous 1971 Erdős question on monotone subsequence sums was made precise and answered online. Harmonic's Aristotle proved the square case in Lean on 7 December 2025, AlphaEvolve data suggested a formula for every length, and an AI literature search found a 2024 packing result that completed it. Terence Tao wrote up the full proof.", "before_after": null, "why_it_matters": "A documented case of AI tools fitting into open collaborative mathematics, with formal proving, numerical exploration and literature search each checked by people.", "ai_systems": [ "Aristotle", "AlphaEvolve", "AI deep research tools" ], "organizations": [ "Harmonic", "Google DeepMind", "UCLA", "erdosproblems.com community" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "The square case Aristotle proved was already in Tidor, Wang and Yang (2016), whom erdosproblems.com credits with the first proof; Tao traces its packing idea to Seidenberg (1959). The final assembly relies on cited results and was not formalized.", "significance": 1, "sources": [ { "label": "Terence Tao, The story of Erdős problem #1026 (2025-12-08)", "url": "https://terrytao.wordpress.com/2025/12/08/the-story-of-erdos-problem-126/" }, { "label": "erdosproblems.com, Problem #1026", "url": "https://www.erdosproblems.com/1026" } ], "foundation0": false }, { "id": "fda-qualifies-aim-nash-drug-development-tool", "date": "2025-12-08", "date_basis": "announcement", "title": "FDA qualifies its first AI drug-development tool, for liver biopsy scoring", "domain": "medicine", "subfield": "drug development, digital pathology", "summary": "The FDA qualified AIM-NASH, PathAI's cloud AI that scores liver biopsies for fat, inflammation, ballooning and scarring, for MASH drug trials. Pathologists accept or reject every AI score. Validation showed AI-assisted reads agreed with expert consensus as well as individual pathologists did. Europe's medicines agency had qualified the tool earlier in 2025.", "before_after": null, "why_it_matters": "MASH drug trials hinge on biopsy scores that pathologists often disagree on. A regulator-qualified AI reader can make trial end points more consistent and cheaper, and sets a precedent for AI in drug development.", "ai_systems": [ "AIM-NASH (renamed AIM-MASH AI Assist)" ], "organizations": [ "PathAI", "US Food and Drug Administration" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "FDA announcement, 2025-12-08", "url": "https://www.fda.gov/drugs/drug-safety-and-availability/fda-qualifies-first-ai-drug-development-tool-will-be-used-mash-clinical-trials" }, { "label": "Nature Medicine clinical validation, 2024-11-04", "url": "https://doi.org/10.1038/s41591-024-03301-2" }, { "label": "PathAI press release (notes prior EMA qualification), 2025-12-09", "url": "https://www.pathai.com/news/pathais-aim-mash-ai-assist-becomes-first-ai-powered-pathology-tool-to-receive-fda-qualification-for-mash-clinical-trials" } ], "foundation0": false }, { "id": "rec-4881-fap-polyp-reduction", "date": "2025-12-08", "date_basis": "announcement", "title": "Drug flagged by an AI cell-imaging screen shrinks polyps in hereditary disease", "domain": "medicine", "subfield": "drug discovery, gastroenterology", "summary": "Recursion reported phase 1b/2 TUPELO data for REC-4881, a MEK1/2 inhibitor its AI-driven cell-imaging screen flagged for familial adenomatous polyposis. Median polyp burden fell 43% after 12 weeks on drug (12 patients) and stood 53% below baseline at week 25, 12 weeks after stopping (11 patients). Grade 3 side effects affected 15.8% of 19 patients.", "before_after": null, "why_it_matters": "Familial adenomatous polyposis has no approved drug and often ends in colon removal. Small, open-label, company-reported data suggest AI phenotypic screens can repurpose existing molecules for rare diseases.", "ai_systems": [ "Recursion OS phenomics platform (early version)" ], "organizations": [ "Recursion" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Recursion press release, 2025-12-08", "url": "https://ir.recursion.com/news-releases/news-release-details/positive-phase-1b2-results-ongoing-rec-4881-tupelo-trial-0" }, { "label": "Release mirror (AAP/GlobeNewswire), 2025-12-08", "url": "https://www.aap.com.au/aapreleases/globenewswire9598312/" } ], "foundation0": false }, { "id": "gpt-5-2-learning-curve-monotonicity", "date": "2025-12-11", "date_basis": "preprint", "title": "GPT-5.2 Pro proves Gaussian maximum likelihood improves with more data", "domain": "mathematics", "subfield": "statistical learning theory", "summary": "Mark Sellke and Steven Yin report that GPT-5.2 Pro variants derived every result in their paper, the first nontrivial proofs that the maximum likelihood estimator's average log-loss never worsens with more data, for Gaussians with unknown covariance and Gamma variables with unknown scale. The Gaussian case was flagged open; humans only prompted and checked.", "before_after": null, "why_it_matters": "Learning theorists had shown that extra data can hurt some learners; settling the textbook Gaussian case shows when the basic intuition holds, and the proofs came from the model.", "ai_systems": [ "GPT-5.2 Pro" ], "organizations": [ "OpenAI", "Harvard University" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Not yet peer-reviewed; OpenAI says external experts reviewed the proofs.", "significance": 2, "sources": [ { "label": "Sellke and Yin, On Learning-Curve Monotonicity for Maximum Likelihood Estimators, arXiv:2512.10220", "url": "https://arxiv.org/abs/2512.10220" }, { "label": "OpenAI, Advancing science and math with GPT-5.2 (2025-12-11)", "url": "https://openai.com/index/gpt-5-2-for-science-and-math/" } ], "foundation0": false }, { "id": "leap71-noyron-20kn-methalox-engines", "date": "2025-12-11", "date_basis": "announcement", "title": "Two computer-designed 20 kN methane rocket engines hot-fired within three weeks", "domain": "engineering-robotics", "subfield": "rocket propulsion design", "summary": "LEAP 71 says its Noyron computational engineering model autonomously designed a bell-nozzle engine and a full-scale aerospike, each 20 kN, burning methane and liquid oxygen, and that specification to first flame took under three weeks. The bell engine ran above 93% combustion efficiency; the aerospike reached its 50 bar chamber pressure in a single burn.", "before_after": null, "why_it_matters": "Engine design is slow, iterative expert work. Generating working hardware directly from a specification could shorten development cycles for propulsion and other complex machines.", "ai_systems": [ "Noyron (Large Computational Engineering Model)" ], "organizations": [ "LEAP 71", "Aconity3D" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "Company announcement only. Noyron encodes physics and engineering rules rather than learning from data, and the aerospike ran once because of start-up transient problems.", "significance": 2, "sources": [ { "label": "LEAP 71 announcement, 2025-12-11", "url": "https://leap71.com/2025/12/11/leap-71-hot-fires-two-orbital-class-methalox-engines-designed-autonomously-by-noyron/" } ], "foundation0": false }, { "id": "ale-agent-wins-atcoder-heuristic-contest", "date": "2025-12-14", "date_basis": "competition", "title": "AI agent places first ahead of 804 humans in an AtCoder Heuristic Contest", "domain": "computing", "subfield": "algorithm engineering / optimization", "summary": "Sakana AI's ALE-Agent, a system making about 4,800 calls to GPT-5.2 and Gemini 3 Pro, entered the four-hour AtCoder Heuristic Contest 058 on 14 December 2025 by special permission, unrated, and took first place among 804 human entrants with an approach that beat the problem setter's intended solution. It cost about US$1,300.", "before_after": null, "why_it_matters": "Heuristic contests reward open-ended optimisation, close to real engineering problems like scheduling and routing, rather than puzzles with known answers.", "ai_systems": [ "ALE-Agent", "GPT-5.2", "Gemini 3 Pro" ], "organizations": [ "Sakana AI" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Sakana AI announcement, 2026-01-05", "url": "https://sakana.ai/ahc058/" }, { "label": "AtCoder contest history for ALE-Agent's account (rank 1 in AHC058)", "url": "https://atcoder.jp/users/fishylene/history?contestType=heuristic" }, { "label": "AtCoder Heuristic Contest 058 page", "url": "https://atcoder.jp/contests/ahc058" } ], "foundation0": false }, { "id": "stoc-2026-gemini-pre-submission-feedback", "date": "2025-12-15", "date_basis": "announcement", "title": "Gemini tool checks STOC 2026 papers for proof errors before submission", "domain": "science-of-science", "subfield": "peer review / proof checking", "summary": "STOC 2026 offered authors optional, private feedback from a Gemini Deep Think tool tuned for mathematical rigour. Over 80% of papers submitted by the experiment's end opted in. More than 100 authors answered a survey and 97% of those quoted found it helpful; one said it caught a bug that made their proof entirely incorrect.", "before_after": null, "why_it_matters": "Proofs in theory papers are long, and reviewers rarely check every step. An automated rigour check before submission catches errors when they are cheapest to fix.", "ai_systems": [ "Paper Assistant Tool (PAT), built on an advanced Gemini 2.5 Deep Think" ], "organizations": [ "Google Research", "ACM STOC 2026" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "Results come from Google's own survey of consenting respondents, with no count of errors found. Reviews went to authors only, not the program committee (STOC 2026 programme page).", "significance": 2, "sources": [ { "label": "Google Research blog, 2025-12-15 (updated 2026-05-18)", "url": "https://research.google/blog/gemini-provides-automated-feedback-for-theoretical-computer-scientists-at-stoc-2026/" }, { "label": "STOC 2026, automated pre-submission feedback programme", "url": "https://acm-stoc.org/stoc2026/stoc2026-LLM_feedback.html" } ], "foundation0": false }, { "id": "hubble-archive-anomalymatch-1300-anomalies", "date": "2025-12-16", "date_basis": "publication", "title": "AI search of Hubble's archive finds over 1,300 anomalies, 800 undocumented", "domain": "astronomy-space", "subfield": "galaxies / archival anomaly search", "summary": "ESA's David O'Ryan and Pablo Gómez ran AnomalyMatch, a semi-supervised, active-learning neural network, over 99.6 million image cutouts from the Hubble Legacy Archive in about two and a half days. Inspecting its top-ranked sources, they confirmed more than 1,300 anomalies, over 800 previously undocumented, including 86 new candidate lenses and 18 jellyfish galaxies.", "before_after": { "metric": "Systematic anomaly search of the whole Hubble Legacy Archive", "before": "never done; anomalies found by hand or by chance", "after": "~100 million cutouts searched in ~2.5 days" }, "why_it_matters": "Rare objects such as lenses and merging galaxies test ideas about dark matter and galaxy growth. Fast AI triage makes it practical to mine old archives and coming surveys like Euclid and Rubin.", "ai_systems": [ "AnomalyMatch (semi-supervised, active-learning neural network)" ], "organizations": [ "European Space Agency (ESAC)" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "O'Ryan & Gómez, Astronomy & Astrophysics 704, A227, 2025-12-16", "url": "https://doi.org/10.1051/0004-6361/202555512" }, { "label": "ESA/Hubble science release heic2603, 2026-01-27", "url": "https://esahubble.org/news/heic2603/" } ], "foundation0": false }, { "id": "kagome-superconductors-ml-guided-yru3b2-luru3b2", "date": "2025-12-16", "date_basis": "preprint", "title": "Machine-learning screen leads to two new kagome superconductors", "domain": "chemistry-materials", "subfield": "superconducting materials", "summary": "The SuperC consortium used machine-learning-accelerated high-throughput screening plus first-principles calculations to pick kagome-lattice compounds. Rice University then made YRu3B2 and LuRu3B2 and measured bulk superconductivity, with critical temperatures of 0.81 K and 0.95 K and nearly 100% superconducting volume fractions.", "before_after": { "metric": "Known superconductivity in YRu3B2 and LuRu3B2", "before": "none reported before December 2025", "after": "bulk superconductors, Tc 0.81 K and 0.95 K" }, "why_it_matters": "Superconductor discovery is still mostly trial and error. A screen that cheaply narrows candidates before synthesis could scale the search, though these temperatures are far from practical.", "ai_systems": [ "Machine-learning high-throughput screening model" ], "organizations": [ "SuperC consortium (Aalto University)", "Rice University", "Donostia International Physics Center", "Princeton University", "Ruhr University Bochum" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Gaggl et al. (arXiv:2512.09314, 2025-12-10; Physica Scripta, 2026) independently reported bulk superconductivity in YRu3B2 (Tc 0.7 K) six days earlier, without machine learning.", "significance": 2, "sources": [ { "label": "Albu Mustaf et al., Physical Review Research 8, 023308, 2026-06-17", "url": "https://doi.org/10.1103/lpqj-7hyg" }, { "label": "arXiv:2512.16945, 2025-12-16", "url": "https://arxiv.org/abs/2512.16945" }, { "label": "Aalto University news release, 2026-06-29", "url": "https://www.aalto.fi/en/news/researchers-identify-new-superconductors-unlocking-process-that-could-yield-thousands-more" } ], "foundation0": false }, { "id": "ai-ecg-cirrhosis-detection-cluster-rct", "date": "2025-12-17", "date_basis": "publication", "title": "AI-read ECGs double new diagnoses of advanced liver disease in primary care", "domain": "medicine", "subfield": "hepatology, primary care screening", "summary": "Mayo Clinic cluster-randomized 98 primary-care teams; in half, clinicians were alerted when a model reading routine ECGs flagged likely advanced chronic liver disease. Among 15,596 adults, new diagnoses of advanced fibrosis or cirrhosis within 180 days were 1.0% with alerts versus 0.5% without (odds ratio 2.09). Clinician follow-up of alerts was uneven.", "before_after": { "metric": "New advanced chronic liver disease diagnoses within 180 days", "before": "0.5% (usual care)", "after": "1.0% (AI-ECG alerts)" }, "why_it_matters": "Cirrhosis is often found only after complications. Using a cheap test that patients already get to flag hidden liver disease could bring earlier treatment at almost no extra cost.", "ai_systems": [ "ECG-based machine learning model for advanced chronic liver disease" ], "organizations": [ "Mayo Clinic" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2025-12-17", "url": "https://doi.org/10.1038/s41591-025-04058-y" } ], "foundation0": false }, { "id": "noaa-operational-ai-global-models", "date": "2025-12-17", "date_basis": "deployment", "title": "NOAA puts AI global forecast models and a hybrid AI-physics ensemble into use", "domain": "earth-climate", "subfield": "operational weather forecasting", "summary": "NOAA deployed three AI-based global systems built from a fine-tuned GraphCast: AIGFS, whose 16-day forecast uses 0.3% of the GFS's computing; a 31-member AI ensemble with early skill gains of 18-24 hours; and a 62-member hybrid AI-physics ensemble that beat both on most metrics. NOAA notes weaker hurricane intensity forecasts in version 1.", "before_after": { "metric": "Computing for a 16-day global forecast", "before": "GFS (100%)", "after": "AIGFS (0.3%)" }, "why_it_matters": "The US national weather service now runs AI forecasts operationally, and blending AI and physics ensembles points to a way of keeping the strengths of both.", "ai_systems": [ "AIGFS", "AIGEFS", "HGEFS", "GraphCast (base model)" ], "organizations": [ "NOAA", "Google DeepMind" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "NOAA news release, 2025-12-17", "url": "https://www.noaa.gov/news-release/noaa-deploys-new-generation-of-ai-driven-global-weather-models" } ], "foundation0": false }, { "id": "llm-adoption-scientific-production-science", "date": "2025-12-18", "date_basis": "publication", "title": "Scientists using LLMs post 24-89% more papers; polished prose loses its signal", "domain": "science-of-science", "subfield": "metascience", "summary": "Cornell and Berkeley researchers analysed 2.1 million preprints, 28,000 peer-review reports and 246 million document accesses. Scientists who began drafting with large language models raised paper output by 23.7% to 89.3%, depending on field and background. LLM use reversed the link between complex writing and paper quality, and adopters cited more diverse, younger work.", "before_after": { "metric": "Paper output after adopting LLMs for drafting", "before": "pre-adoption baseline", "after": "+23.7% to +89.3%, depending on field and author background" }, "why_it_matters": "Journals, funders and hiring committees use output and writing quality as signals of merit. Measuring how AI changes both shows those signals need rethinking.", "ai_systems": [ "Large language models (ChatGPT and others), detected by text analysis" ], "organizations": [ "Cornell University", "University of California, Berkeley" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "LLM use is inferred from text, and the associations are observational, not causal.", "significance": 2, "sources": [ { "label": "Kusumegi et al., Science 390:1240, 2025-12-18", "url": "https://doi.org/10.1126/science.adw3000" }, { "label": "Author version, arXiv 2601.13187", "url": "https://arxiv.org/abs/2601.13187" } ], "foundation0": false }, { "id": "masai-ai-mammography-rct", "date": "2026-01", "date_basis": "publication", "title": "MASAI trial: AI-supported mammography finds more cancers, misses no more", "domain": "medicine", "subfield": "breast cancer screening", "summary": "Sweden's MASAI trial randomized 105,934 women to AI-supported screening (Transpara triaging scans to single or double reading) or standard double reading. AI found 29% more cancers (6.4 vs 5.0 per 1,000) with 44% less reading; interval cancers were non-inferior (1.55 vs 1.76 per 1,000) and sensitivity rose from 73.8% to 80.5%.", "before_after": { "metric": "Screening sensitivity", "before": "73.8% (standard double reading)", "after": "80.5% (AI-supported)" }, "why_it_matters": "Finding more cancers only helps if no more are missed between screens. A large randomized trial showing both, with far less radiologist time, is strong evidence for real screening programs.", "ai_systems": [ "Transpara 1.7.0 (ScreenPoint Medical)" ], "organizations": [ "Lund University", "Skåne University Hospital", "Radboud University Medical Centre" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "The Lancet (interval cancers, primary endpoint), 2026-01-31 issue", "url": "https://doi.org/10.1016/S0140-6736(25)02464-X" }, { "label": "Lancet Digital Health (screening performance), 2025-02-03", "url": "https://doi.org/10.1016/S2589-7500(24)00267-X" }, { "label": "Lancet Oncology (interim safety analysis), 2023-08", "url": "https://doi.org/10.1016/S1470-2045(23)00298-X" } ], "foundation0": false }, { "id": "gendsa-v2-low-dose-angiography-rct", "date": "2026-01-02", "date_basis": "publication", "title": "Generative AI cuts X-ray dose by two-thirds in angiography-guided procedures", "domain": "medicine", "subfield": "interventional radiology, imaging", "summary": "GenDSA-V2, trained on over 5 million images from 46,829 patients at 70 centres, generates angiography frames so fewer X-ray exposures are needed. In a blinded randomized trial of 1,068 patients undergoing procedures for brain aneurysms, lung cancer or liver cancer, mean air kerma fell from 457.4 to 151.3 mGy, with similar complications (7.5% vs 8.1%).", "before_after": { "metric": "Mean air kerma radiation dose per procedure", "before": "457.4 mGy (standard protocol)", "after": "151.3 mGy (GenDSA-V2)" }, "why_it_matters": "Digital subtraction angiography guides over 100,000 procedures a day, exposing patients and staff to radiation. A prospectively validated low-dose method could cut cumulative exposure, especially for interventional teams.", "ai_systems": [ "GenDSA-V2" ], "organizations": [ "Union Hospital, Tongji Medical College, Huazhong University of Science and Technology", "Wuhan University", "Neusoft Medical Systems" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2026-01-02", "url": "https://doi.org/10.1038/s41591-025-04042-6" } ], "foundation0": false }, { "id": "schwartz-claude-c-parameter-resummation", "date": "2026-01-05", "date_basis": "preprint", "title": "Claude carries out a particle-theory calculation under a physicist's supervision", "domain": "physics", "subfield": "theoretical particle physics (QCD)", "summary": "Harvard's Matthew Schwartz steered Claude Opus 4.5, by text prompts only, to resum the Sudakov shoulder of the C-parameter in electron-positron collisions at NLL+NLO. Claude did the calculations, Monte Carlo checks and writing; Schwartz caught errors, including a wrong factorization formula, and validated the result. He estimates two weeks instead of about a year.", "before_after": { "metric": "Time for a second-year-graduate-level theory project (author's estimate)", "before": "about a year", "after": "two weeks (110 drafts, 36M tokens)" }, "why_it_matters": "It shows a language model can carry out a real, checkable theoretical physics calculation when an expert steers it, which could free physicists for harder questions.", "ai_systems": [ "Claude Opus 4.5" ], "organizations": [ "Harvard University", "NSF Institute for Artificial Intelligence and Fundamental Interactions" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Not peer reviewed. Schwartz reports Claude was sloppy, adjusted parameters to make plots match, and first used a wrong factorization formula; he says expert checking of every step was essential.", "significance": 3, "sources": [ { "label": "Schwartz, arXiv:2601.02484, 2026-01-05", "url": "https://arxiv.org/abs/2601.02484" }, { "label": "Anthropic guest post by M. Schwartz, 2026-03-23", "url": "https://www.anthropic.com/research/vibe-physics" } ], "foundation0": false }, { "id": "erdos-728-autonomous", "date": "2026-01-06", "date_basis": "announcement", "title": "AI resolves Erdős problem #728 with a Lean proof, essentially autonomously", "domain": "mathematics", "subfield": "number theory / factorial divisibility", "summary": "Erdős problem #728 on factorial divisibility was vaguely stated; once a precise version was agreed, Kevin Barreto ran GPT-5.2 Pro and Harmonic's Aristotle, which produced a Lean proof by 6 January 2026. It was described as the first Erdős problem resolved essentially autonomously by AI; Pomerance's 2014 paper uses a similar method.", "before_after": null, "why_it_matters": "Marked the shift from AI rediscovering known results to AI producing new, machine-checked solutions to problems collected from one of the most prolific mathematicians.", "ai_systems": [ "GPT-5.2 Pro", "Aristotle" ], "organizations": [ "OpenAI", "Harmonic" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Sothanaphan, writeup of Aristotle's proof, arXiv:2601.07421", "url": "https://arxiv.org/abs/2601.07421" }, { "label": "erdosproblems.com forum, #728", "url": "https://www.erdosproblems.com/forum/thread/728?order=oldest" }, { "label": "Tao et al., AI contributions to Erdős problems (wiki)", "url": "https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems" } ], "foundation0": false }, { "id": "als-agentic-ai-accelerator-experiments", "date": "2026-01-16", "date_basis": "publication", "title": "LLM agent runs multistage experiments on Berkeley's Advanced Light Source", "domain": "physics", "subfield": "accelerator physics / lab automation", "summary": "Berkeley Lab's Accelerator Assistant, an agent driven by large language models, turns plain-language requests into executable plans: retrieving archived data, finding control channels, writing scripts, operating the accelerator within safety limits and analyzing results. On the Advanced Light Source it cut experiment preparation about 100-fold versus an expert's manual scripting.", "before_after": { "metric": "time to prepare an experiment workflow", "before": "several hours (manual expert scripting)", "after": "a few minutes from one prompt (about 100× faster)" }, "why_it_matters": "Big facilities lose scarce beam time to setup work. Agents that safely turn intent into machine operations could let scientists run more, and more complex, experiments.", "ai_systems": [ "Accelerator Assistant (LLM agent built on the Osprey framework)" ], "organizations": [ "Lawrence Berkeley National Laboratory" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Physical Review Research, 2026-01-16", "url": "https://doi.org/10.1103/jtqy-9jz1" }, { "label": "arXiv preprint 2509.17255, 2025-09-21", "url": "https://arxiv.org/abs/2509.17255" }, { "label": "Berkeley Lab ATAP news, 2026", "url": "https://atap.lbl.gov/news/agentic-ai-enables-advanced-light-source-experiments/" } ], "foundation0": false }, { "id": "coffies-ai-active-region-emergence-forecast", "date": "2026-01-19", "date_basis": "preprint", "title": "Transformer model warns of emerging sunspot regions hours before they surface", "domain": "astronomy-space", "subfield": "heliophysics / space weather", "summary": "NASA's COFFIES team (NJIT, Princeton, NASA Ames) trained a sliding-window transformer on Solar Dynamics Observatory data from 46 active regions, looking for subtle acoustic-power and magnetic changes from flux rising below the surface. Forecasting up to 12 hours ahead, the published model caught emergence onset on average 9.24 hours early.", "before_after": { "metric": "Error (RMSE) of continuum-intensity forecasts at active-region emergence", "before": "LSTM baseline", "after": "10.6% lower; onset flagged 9.24 h early on average" }, "why_it_matters": "Active regions launch the flares and coronal mass ejections that threaten satellites, power grids and astronauts. Warning hours before they appear would extend space-weather forecasts.", "ai_systems": [ "EarlyDetect sliding-window transformer" ], "organizations": [ "NASA COFFIES DRIVE Science Center", "New Jersey Institute of Technology", "Princeton University", "NASA Ames Research Center" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Retrospective test on 46 active regions, not an operational forecast; lead times vary more than with the LSTM baseline. The January 2026 preprint reported a 4.73 h average lead.", "significance": 2, "sources": [ { "label": "Tirona et al., JGR: Machine Learning and Computation, 2026-08-14", "url": "https://doi.org/10.1029/2025JH001207" }, { "label": "NASA Science, COFFIES, 2026-08-14", "url": "https://science.nasa.gov/science-research/heliophysics/nasas-coffies-uses-ai-to-predict-storm-causing-active-regions-on-sun/" }, { "label": "arXiv:2601.13144, 2026-01-19", "url": "https://arxiv.org/abs/2601.13144" } ], "foundation0": false }, { "id": "prea-llm-chatbot-specialist-referral-rct", "date": "2026-01-19", "date_basis": "publication", "title": "LLM chatbot takes patient histories before specialist visits in randomized trial", "domain": "medicine", "subfield": "care delivery, outpatient referral", "summary": "PreA, a GPT-4o mini chatbot co-designed with patients and clinicians, took histories and drafted referral reports before specialist visits at two hospitals in western China. In a trial randomizing 2,069 patients across 24 specialties, it cut consultation time 28.7% (4.41 to 3.14 minutes) and raised physician-rated care coordination; unassisted use matched staff-assisted use.", "before_after": { "metric": "Specialist consultation time", "before": "4.41 min (no chatbot)", "after": "3.14 min (PreA used alone)" }, "why_it_matters": "Specialists in crowded, under-resourced hospitals get minutes per patient. An AI that gathers the history and drafts the referral returns that time to examination and decisions.", "ai_systems": [ "PreA (built on OpenAI GPT-4o mini)" ], "organizations": [ "Chinese Academy of Medical Sciences & Peking Union Medical College", "First Affiliated Hospital of Guilin Medical University", "Affiliated Hospital of Gansu Medical College", "Peking University", "Tencent (WeChat AI)" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Medicine, 2026-01-19", "url": "https://doi.org/10.1038/s41591-025-04176-7" } ], "foundation0": false }, { "id": "mars-multi-agent-robotic-perovskite-lab", "date": "2026-01-21", "date_basis": "publication", "title": "19 LLM agents and lab robots run closed-loop perovskite materials research", "domain": "chemistry-materials", "subfield": "autonomous laboratories", "summary": "MARS, from the Shenzhen Institute of Advanced Technology, coordinates 19 language-model agents and 16 domain tools in orchestrator, scientist, engineer, executor and analyst groups that drive robotic experiments. It optimised perovskite nanocrystal synthesis within 10 iterations and designed a water-stable 'core-shell-corona' perovskite composite in 3.5 hours.", "before_after": null, "why_it_matters": "Most self-driving labs tune one fixed recipe. Agents that also plan, pick tools and interpret results move automation toward labs that carry whole research projects.", "ai_systems": [ "MARS multi-agent and robot system (19 LLM agents, 16 tools)" ], "organizations": [ "Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences", "National University of Singapore", "City University of Hong Kong" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Two demonstrations within one materials family, set up by the authors; details here rest on the abstract-level institute summary.", "significance": 2, "sources": [ { "label": "Shi et al., Matter 9(2):102577, online 2026-01-21", "url": "https://doi.org/10.1016/j.matt.2025.102577" }, { "label": "Chinese Academy of Sciences news, 2026-01-22", "url": "https://english.cas.cn/newsroom/research-news/202602/t20260210_1150376.shtml" } ], "foundation0": false }, { "id": "aisle-openssl-12-cves", "date": "2026-01-27", "date_basis": "announcement", "title": "AISLE's AI analyzer credited with all 12 flaws in OpenSSL's Jan 2026 fixes", "domain": "computing", "subfield": "vulnerability discovery", "summary": "OpenSSL's 27 January 2026 security advisory fixed 12 vulnerabilities, all reported by AISLE researchers (one also found independently by another researcher), who say their autonomous AI analyzer found them. One was high severity (a stack overflow in CMS message parsing), some date back to 1998, and AISLE's fixes were adopted for 5 of the 12.", "before_after": null, "why_it_matters": "OpenSSL secures much of the internet's encrypted traffic and is heavily audited. An AI system finding every flaw in a release shows how far automated auditing has come.", "ai_systems": [ "AISLE autonomous analyzer" ], "organizations": [ "AISLE", "OpenSSL Foundation" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "OpenSSL Security Advisory, 2026-01-27", "url": "https://openssl-library.org/news/secadv/20260127.txt" }, { "label": "AISLE blog, January 2026", "url": "https://aisle.com/blog/aisle-discovered-12-out-of-12-openssl-vulnerabilities" } ], "foundation0": false }, { "id": "alphagenome-regulatory-variant-effects", "date": "2026-01-28", "date_basis": "publication", "title": "AlphaGenome predicts regulatory effects of DNA variants from 1 million bases", "domain": "biology", "subfield": "regulatory genomics", "summary": "AlphaGenome reads 1 million bases of DNA and predicts thousands of functional genomic tracks at up to single-base resolution, including gene expression, chromatin accessibility, transcription-factor binding and splicing. Trained on human and mouse genomes, it matches or beats the strongest external models in 25 of 26 variant-effect evaluations.", "before_after": null, "why_it_matters": "Most disease-linked DNA changes lie outside genes, in regulatory regions. One model scoring a variant across many layers of regulation helps researchers prioritize which changes matter.", "ai_systems": [ "AlphaGenome" ], "organizations": [ "Google DeepMind" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Avsec et al., Nature, 2026-01-28", "url": "https://doi.org/10.1038/s41586-025-10014-0" }, { "label": "Google DeepMind blog, 2025-06-25 (preprint announcement; updated for Nature)", "url": "https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/" } ], "foundation0": false }, { "id": "perseverance-ai-planned-mars-drives", "date": "2026-01-30", "date_basis": "announcement", "title": "Perseverance drives Mars routes planned by generative AI for the first time", "domain": "astronomy-space", "subfield": "planetary rover operations", "summary": "NASA JPL used Anthropic's Claude vision-language models to analyse orbital images and terrain-slope data and write waypoints for Perseverance, a job human planners have done for 28 years. After checks on JPL's digital twin covering over 500,000 telemetry variables, the rover drove 210 m on 8 December 2025 and 246 m on 10 December.", "before_after": { "metric": "Who plans Mars rover routes", "before": "human rover planners only", "after": "AI-generated waypoints; 210 m and 246 m driven" }, "why_it_matters": "Signal delays rule out joystick driving on Mars. Letting AI plan safe routes could mean longer drives, less operator workload and more science per day on distant worlds.", "ai_systems": [ "Claude (Anthropic vision-language models)" ], "organizations": [ "NASA Jet Propulsion Laboratory", "Anthropic" ], "ai_role": "originator", "evidence": "deployed", "contested": false, "contested_note": "Two demonstration drives. Rover drivers made minor changes using ground-level images Claude had not seen (Anthropic). Rovers already had onboard autonomous navigation; the first is generative-AI route planning.", "significance": 3, "sources": [ { "label": "NASA JPL news, 2026-01-30", "url": "https://www.jpl.nasa.gov/news/nasas-perseverance-rover-completes-first-ai-planned-drive-on-mars/" }, { "label": "Anthropic feature, Claude on Mars, 2026-01-30", "url": "https://www.anthropic.com/features/claude-on-mars" } ], "foundation0": false }, { "id": "fel-conjecture-axiomprover", "date": "2026-02-03", "date_basis": "preprint", "title": "AxiomProver proves Fel's conjecture with a formal Lean proof", "domain": "mathematics", "subfield": "commutative algebra / numerical semigroups", "summary": "Axiom Math's AxiomProver, given a natural-language statement, produced a complete Lean/Mathlib proof of Fel's conjecture, an explicit formula for syzygy invariants of numerical semigroups. Axiom says no hints, helper lemmas or manual patches were added after kickoff; humans wrote the statement, definitions and the accompanying paper.", "before_after": null, "why_it_matters": "An open conjecture went from statement to machine-checked proof without human mathematical input, so correctness rests on the Lean checker rather than trust in the AI.", "ai_systems": [ "AxiomProver" ], "organizations": [ "Axiom Math" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "arXiv:2602.03716", "url": "https://arxiv.org/abs/2602.03716" }, { "label": "Axiom, 2026-02-05", "url": "https://axiommath.ai/research/proof-of-concept" } ], "foundation0": false }, { "id": "gemini-deep-think-tcs-case-studies", "date": "2026-02-03", "date_basis": "preprint", "title": "Researchers log Gemini Deep Think helping settle open problems in theory", "domain": "science-of-science", "subfield": "human-AI research collaboration", "summary": "Thirty-six researchers led from Google Research describe more than 15 cases where Gemini Deep Think helped resolve open questions, mostly in theoretical computer science, plus economics, optimisation and physics. Examples: a three-item counterexample refuting a 2015 online submodular welfare conjecture, and a fatal flaw it found in a cryptography preprint, confirmed by its authors.", "before_after": null, "why_it_matters": "A documented record of what a model contributed, and which working habits made it useful, lets other researchers judge the claims and reuse the techniques.", "ai_systems": [ "Gemini Deep Think and advanced variants" ], "organizations": [ "Google Research", "Carnegie Mellon University", "EPFL", "Bar-Ilan University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Section authors vouch only for their own sections; some results are partial, and most are not peer reviewed.", "significance": 2, "sources": [ { "label": "Woodruff et al., arXiv 2602.03837, 2026-02-03", "url": "https://arxiv.org/abs/2602.03837" } ], "foundation0": false }, { "id": "claude-agent-team-c-compiler", "date": "2026-02-05", "date_basis": "announcement", "title": "Sixteen Claude agents write a C compiler that builds the Linux kernel", "domain": "computing", "subfield": "compilers / autonomous software engineering", "summary": "Nicholas Carlini ran 16 Claude Opus 4.6 agents in parallel, mostly unsupervised, over nearly 2,000 sessions and about $20,000 of API use. Without internet access they wrote a 100,000-line Rust C compiler that builds a bootable Linux 6.9 on x86, ARM and RISC-V and passes about 99% of most compiler test suites.", "before_after": null, "why_it_matters": "A compiler is a large, tightly specified system. Agents building one from scratch show autonomous teams can deliver big, testable software; output slower than unoptimized GCC shows the limits.", "ai_systems": [ "Claude Opus 4.6" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "Anthropic lists limits: it calls GCC for 16-bit real-mode boot code, uses GCC's assembler and linker, and emits code slower than GCC with optimizations off. Chris Lattner (Modular, 2026-02-18) called it a competent textbook implementation tuned to pass tests rather than build general abstractions.", "significance": 2, "sources": [ { "label": "Anthropic Engineering, Building a C compiler with a team of parallel Claudes (2026-02-05)", "url": "https://www.anthropic.com/engineering/building-c-compiler" }, { "label": "Chris Lattner, The Claude C Compiler: what it reveals about the future of software (2026-02-18)", "url": "https://www.modular.com/blog/the-claude-c-compiler-what-it-reveals-about-the-future-of-software" } ], "foundation0": false }, { "id": "claude-opus-4-6-500-open-source-vulnerabilities", "date": "2026-02-05", "date_basis": "announcement", "title": "Claude Opus 4.6 finds 500+ high-severity bugs in heavily fuzzed open-source code", "domain": "computing", "subfield": "vulnerability discovery / open-source security", "summary": "Anthropic's red team ran Claude Opus 4.6 in a virtual machine with standard tools against open-source projects and validated more than 500 high-severity vulnerabilities, some missed for decades despite years of fuzzing. Examples in Ghostscript, OpenSC and the CGIF library are now patched by maintainers; human researchers validated each finding and wrote the first patches.", "before_after": null, "why_it_matters": "Fuzzers miss bugs that need reasoning about code history or an algorithm's logic; finding them at scale helps under-resourced maintainers fix flaws before attackers find them.", "ai_systems": [ "Claude Opus 4.6" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "The 500+ count is Anthropic's own. Most reports were still in disclosure and the work focused on memory-corruption bugs, so the total cannot yet be checked independently.", "significance": 3, "sources": [ { "label": "Anthropic Frontier Red Team, Evaluating and mitigating the growing risk of LLM-discovered 0-days (2026-02-05)", "url": "https://www.anthropic.com/research/zero-days" } ], "foundation0": false }, { "id": "gpt5-ginkgo-autonomous-lab-cell-free-protein", "date": "2026-02-05", "date_basis": "preprint", "title": "GPT-5 runs 36,000 lab experiments and cuts cell-free protein cost 40%", "domain": "science-of-science", "subfield": "autonomous laboratories", "summary": "GPT-5 designed, analysed and refined experiments run by Ginkgo Bioworks' robotic cloud lab, testing more than 36,000 cell-free protein synthesis recipes on over 580 plates in six rounds. Humans mainly prepared and loaded reagents and plates. It made the benchmark protein sfGFP at $422 per gram in reaction components, 40% below the previous best, $698.", "before_after": { "metric": "Specific cost of sfGFP by cell-free synthesis (20 µL reactions, 2 mL tubes)", "before": "$698 per gram (Olsen et al., 2025)", "after": "$422 per gram, with titre up 27%" }, "why_it_matters": "Cell-free protein synthesis makes proteins without living cells but is costly. A model that closes the loop with robots shows how labs could search recipe spaces too large for people.", "ai_systems": [ "GPT-5" ], "organizations": [ "OpenAI", "Ginkgo Bioworks" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Not peer reviewed; one protein in one cell-free system. The cost counts reaction components only, not manufacturing, and R&D World notes yields depend strongly on oxygenation and reaction geometry.", "significance": 3, "sources": [ { "label": "Smith et al., bioRxiv, 2026-02-05", "url": "https://doi.org/10.64898/2026.02.05.703998" }, { "label": "Preprint PDF (OpenAI)", "url": "https://cdn.openai.com/pdf/5a12a3bc-96b7-4e07-9386-db6ee5bb2ed9/using-a-gpt-5-driven-autonomous-lab-to-optimize-the-cost-and-titer-of-cell-free-protein-synthesis.pdf" }, { "label": "R&D World, 2026-02-05", "url": "https://www.rdworldonline.com/openais-gpt-5-autonomously-ran-36000-protein-synthesis-experiments-in-ginkgo-bioworks-cloud-lab/" } ], "foundation0": false }, { "id": "aletheia-autonomous-research", "date": "2026-02-10", "date_basis": "preprint", "title": "Aletheia agent writes a research paper without human intervention", "domain": "mathematics", "subfield": "arithmetic geometry / combinatorics", "summary": "Google DeepMind's Aletheia, built on an advanced Gemini Deep Think, generates, checks and revises proofs in natural language. It wrote a paper on 'eigenweights' in arithmetic geometry with no human intervention. Run on 700 'open' Erdős problems, 137 of 200 gradable answers were flawed; of 13 meaningful ones, 4 were new and 9 already known.", "before_after": null, "why_it_matters": "Moves AI from competition problems into the real research workflow of finding literature, proving and writing, with explicit accounting of what was autonomous.", "ai_systems": [ "Aletheia", "Gemini Deep Think" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2602.10177", "url": "https://arxiv.org/abs/2602.10177" }, { "label": "Erdős case study, arXiv:2601.22401", "url": "https://arxiv.org/abs/2601.22401" } ], "foundation0": false }, { "id": "gpt-5-2-single-minus-gluon-amplitudes", "date": "2026-02-12", "date_basis": "preprint", "title": "GPT-5.2 conjectures formula showing single-minus gluon amplitudes are nonzero", "domain": "physics", "subfield": "theoretical particle physics (scattering amplitudes)", "summary": "A preprint by Guevara, Lupsasca, Skinner, Strominger and Weil shows tree-level gluon amplitudes with one negative-helicity gluon, usually presumed zero, are nonzero in special 'half-collinear' kinematics. Per the paper, GPT-5.2 Pro first conjectured the key all-n formula, a new internal OpenAI model proved it, and the authors checked it by hand.", "before_after": { "metric": "single-minus tree-level gluon amplitudes", "before": "presumed to vanish", "after": "nonzero closed-form formula in half-collinear kinematics" }, "why_it_matters": "Scattering amplitudes are the core predictions of particle theory. A language model supplied the key conjecture and proof, though the result covers a narrow, unphysical-signature regime.", "ai_systems": [ "GPT-5.2 Pro", "Internal OpenAI model (a scaffolded GPT-5.2, per OpenAI)" ], "organizations": [ "OpenAI", "Institute for Advanced Study", "Vanderbilt University", "University of Cambridge", "Harvard University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": true, "contested_note": "Not yet peer reviewed. D. Louapre (Hugging Face blog, 2026-02-19) calls the AI conjecture a real contribution but stresses the narrow scope (measure-zero kinematics in (2,2) signature), says the physics framing was human work, and says the AI workflow is under-documented.", "significance": 3, "sources": [ { "label": "arXiv preprint 2602.12176, 2026-02-12", "url": "https://arxiv.org/abs/2602.12176" }, { "label": "OpenAI blog, February 2026", "url": "https://openai.com/index/new-result-theoretical-physics/" }, { "label": "Hugging Face blog (critique), 2026-02-19", "url": "https://huggingface.co/blog/dlouapre/gpt-single-minus-gluons" } ], "foundation0": false }, { "id": "ai-replication-political-science-384-studies", "date": "2026-02-17", "date_basis": "preprint", "title": "AI workflow re-runs 384 political science papers to measure reproducibility", "domain": "science-of-science", "subfield": "reproducibility", "summary": "Yiqing Xu (Stanford) and Leo Yang Yang (HKBU) built a Claude Code-coordinated workflow that fetches replication files, rebuilds environments, runs code and matches published regression tables, for about $2,000 in API costs. Across 384 sampled papers from three top journals, fully or largely reproducible papers rose from 20.8% before data-archiving rules to 82.5% after.", "before_after": { "metric": "Share of sampled papers fully or largely reproducible", "before": "20.8% before DA-RT archiving rules", "after": "82.5% after (92.1% when replication files were accessible)" }, "why_it_matters": "Checking published results by hand is too slow to do at scale. Automated replication turns reproducibility from a spot check into something a whole field can measure.", "ai_systems": [ "Claude Code (adaptive coordinator)", "Claude Opus 4.7 (table extraction)" ], "organizations": [ "Stanford University", "Hong Kong Baptist University" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Xu and Yang, arXiv 2602.16733 (v3, 2026-06-01)", "url": "https://arxiv.org/abs/2602.16733" } ], "foundation0": false }, { "id": "virtual-biotech-agents-trial-outcomes", "date": "2026-02-23", "date_basis": "preprint", "title": "Over 37,000 AI agents mine 55,984 trials for what makes drug targets succeed", "domain": "medicine", "subfield": "drug development, clinical trials", "summary": "Stanford's Virtual Biotech, a team of AI agents organized like a drug company, annotated outcomes of 55,984 clinical trials and linked drug targets to single-cell data. Drugs aimed at cell-type-specific genes were 48% more likely to reach market and had 32% fewer adverse events, an association. It also proposed a lung cancer strategy targeting B7-H3.", "before_after": null, "why_it_matters": "Most drugs fail in trials, often for reasons visible only across thousands of studies. Agent teams that can read and structure the whole trial record could steer target choice earlier and more cheaply.", "ai_systems": [ "The Virtual Biotech (multi-agent LLM system)" ], "organizations": [ "Stanford University", "PHD Biosciences" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "bioRxiv preprint, 2026-02-23", "url": "https://doi.org/10.64898/2026.02.23.707551" }, { "label": "Science, 2026-09-17", "url": "https://doi.org/10.1126/science.aeg6779" } ], "foundation0": false }, { "id": "knuth-claudes-cycles", "date": "2026-02-28", "date_basis": "announcement", "title": "Claude finds the construction for Knuth's Hamiltonian cycle problem", "domain": "mathematics", "subfield": "graph theory / combinatorics", "summary": "Donald Knuth reported that Claude Opus 4.6, prompted and coached by Filip Stappers, found in about an hour a rule splitting the arcs of a digraph on m³ vertices into three Hamiltonian cycles for every odd m. Knuth wrote the proof, which Kim Morrison checked in Lean; GPT models later settled even m.", "before_after": { "metric": "Values of m with a known general decomposition", "before": "m = 3 proved; 4 ≤ m ≤ 16 found by computer search", "after": "All odd m (Claude's rule, proved); even m ≥ 8 soon after via GPT-5.3-codex and GPT-5.4 Pro" }, "why_it_matters": "One of computing's most respected figures, previously sceptical of generative AI, documented an AI solving a problem he had worked on for weeks.", "ai_systems": [ "Claude Opus 4.6" ], "organizations": [ "Stanford University", "Anthropic" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Knuth, Claude's Cycles (2026-02-28, rev. 2026-04-14)", "url": "https://cs.stanford.edu/~knuth/papers/claude-cycles.pdf" } ], "foundation0": false }, { "id": "lean-zip-ai-verified-deflate", "date": "2026-02-28", "date_basis": "announcement", "title": "AI agents write and formally verify a zlib-compatible compressor in Lean", "domain": "computing", "subfield": "formal verification of software", "summary": "With loose human supervision, Claude-based coding agents wrote a zlib-format DEFLATE compressor and decompressor in the Lean language and proved, checked by Lean's kernel, that decompressing compressed data returns the original for every input and compression level (given a size limit). The proof spans over 1,100 theorems in about 32,000 lines.", "before_after": null, "why_it_matters": "As AI writes more of the world's code, people need guarantees it is correct. Machine-checked proofs let anyone confirm the result without trusting the AI that produced it.", "ai_systems": [ "Claude" ], "organizations": [ "Lean FRO" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Leonardo de Moura's blog, 2026-02-28", "url": "https://leodemoura.github.io/blog/2026-2-28-when-ai-writes-the-worlds-software-who-verifies-it/" }, { "label": "lean-zip repository (GitHub)", "url": "https://github.com/kim-em/lean-zip" } ], "foundation0": false }, { "id": "design-conductor-agent-risc-v-cpu", "date": "2026-03", "date_basis": "preprint", "title": "AI agent designs a Linux-capable RISC-V CPU from a 219-word spec in 12 hours", "domain": "engineering-robotics", "subfield": "chip design", "summary": "Verkor's Design Conductor agent turned a 219-word requirements document into a verified, tape-out-ready layout of a Linux-capable RISC-V CPU in 12 hours, unaided. It met timing at 1.48 GHz in the ASAP7 7 nm academic kit and scored 3,261 on CoreMark, about a 2011 Intel Celeron. A May follow-up built an LLM accelerator.", "before_after": { "metric": "Time from specification to verified CPU layout", "before": "months to years for human teams", "after": "12 hours, fully autonomous" }, "why_it_matters": "Chip design takes large expert teams years. Agents that carry a specification all the way to a verified layout could make custom hardware cheap enough for many more uses.", "ai_systems": [ "Design Conductor (agent built on frontier LLMs)" ], "organizations": [ "Verkor" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Company report. No chip was fabricated: verification was in simulation and layout used the predictive ASAP7 kit (IEEE Spectrum). The core is far simpler than leading-edge CPUs.", "significance": 2, "sources": [ { "label": "Verkor, arXiv 2603.08716 (submitted 2026-02-06)", "url": "https://arxiv.org/abs/2603.08716" }, { "label": "IEEE Spectrum, 2026-04-22", "url": "https://spectrum.ieee.org/ai-chip-design" }, { "label": "Design Conductor 2.0, arXiv 2605.05170, 2026-05-06", "url": "https://arxiv.org/abs/2605.05170" } ], "foundation0": false }, { "id": "sphere-packing-formalization-gauss", "date": "2026-03-02", "date_basis": "announcement", "title": "Viazovska's sphere-packing proofs in dimensions 8 and 24 formally verified", "domain": "mathematics", "subfield": "discrete geometry / formalization", "summary": "Math Inc.'s Gauss joined a Lean project begun by Sidharth Hariharan and Maryna Viazovska, proving all remaining 8-dimensional results in five days (announced 23 February 2026), then formalized the 24-dimensional proof in about two weeks from the original paper alone. The roughly 200,000-line release also fixed a typo in the published paper.", "before_after": { "metric": "Lean formalization of optimal sphere packing", "before": "~70,000 lines, incomplete", "after": "Complete for dimensions 8 and 24 (~200,000 lines)" }, "why_it_matters": "First machine-checked verification of Fields Medal-winning work; the team had estimated six more months for dimension 8 alone with existing tools.", "ai_systems": [ "Gauss" ], "organizations": [ "Math Inc.", "EPFL" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Math Inc., sphere packing", "url": "https://www.math.inc/sphere-packing" }, { "label": "IEEE Spectrum, 2026-03-02", "url": "https://spectrum.ieee.org/ai-proof-verification" }, { "label": "EPFL news, 2026-03-11", "url": "https://actu.epfl.ch/news/prof-viazovska-s-proofs-of-sphere-packing-formaliz/" } ], "foundation0": false }, { "id": "ml-warming-pattern-emergent-constraint", "date": "2026-03-03", "date_basis": "publication", "title": "Machine learning on warming patterns cuts future-warming uncertainty by 70%", "domain": "earth-climate", "subfield": "climate projections", "summary": "Applied to large climate-model ensembles, machine learning found which grid cells' 1971-2020 warming trends best predict future global warming, mainly tropical and polar ones. Using the observed pattern cut projection error variance by about 70%, versus 48% with the global-mean trend, putting the chance of passing 2 °C by mid-century under SSP3-7.0 near 80%.", "before_after": { "metric": "Reduction in error variance of future global warming projections", "before": "about 48% (global-mean warming trend constraint)", "after": "about 70% (ML spatial-pattern constraint)" }, "why_it_matters": "Policy hinges on how fast the world warms. Narrower projections tied to observed regional warming tell planners how soon Paris Agreement thresholds are likely to be crossed.", "ai_systems": [ "Machine-learning emergent-constraint models" ], "organizations": [ "East China Normal University", "Nanjing University of Information Science and Technology", "Pacific Climate Impacts Consortium, University of Victoria" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Li et al., Nature Communications 17, 3366, 2026-03-03", "url": "https://doi.org/10.1038/s41467-026-70205-9" } ], "foundation0": false }, { "id": "evo-2-genome-language-model", "date": "2026-03-04", "date_basis": "publication", "title": "Evo 2 DNA model scores genetic variants without task-specific training", "domain": "biology", "subfield": "genome language models", "summary": "Evo 2, from Arc Institute, NVIDIA and others, is a 40-billion-parameter DNA model trained on about 9 trillion bases with a 1-million-base context. It scores genetic variants, including BRCA1 variants, without task-specific training. Zero-shot human gene-essentiality prediction is modest (AUROC 0.66), within the range of simple conservation scores.", "before_after": null, "why_it_matters": "A DNA model needing no per-task training could help read mutation effects in any species, though its own gene-essentiality result shows clear limits.", "ai_systems": [ "Evo 2" ], "organizations": [ "Arc Institute", "NVIDIA", "Stanford University", "University of California, Berkeley" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Brixi et al., Nature, 2026-03-04", "url": "https://doi.org/10.1038/s41586-026-10176-5" }, { "label": "bioRxiv preprint, posted February 2025", "url": "https://doi.org/10.1101/2025.02.18.638918" } ], "foundation0": false }, { "id": "gpt-5-2-single-minus-graviton-amplitudes", "date": "2026-03-04", "date_basis": "preprint", "title": "GPT-5.2 Pro helps show single-minus graviton amplitudes are also nonzero", "domain": "physics", "subfield": "theoretical physics (gravity scattering amplitudes)", "summary": "Three weeks after their gluon result, Guevara, Lupsasca, Skinner, Strominger and Weil showed that tree-level graviton amplitudes with one negative-helicity graviton, often presumed zero, are nonzero in special 'half-collinear' kinematics, deriving and solving a recursion for any number of gravitons. The paper says GPT-5.2 Pro and an internal OpenAI model played a significant role throughout.", "before_after": { "metric": "Single-minus tree-level graviton amplitudes", "before": "often presumed to vanish", "after": "nonzero; recursion solved for any number of gravitons" }, "why_it_matters": "Gravity amplitudes are harder than gluon ones. Carrying the AI-assisted method over to gravitons suggests it generalizes, though the result holds only in a special kinematic regime.", "ai_systems": [ "GPT-5.2 Pro", "Internal OpenAI model" ], "organizations": [ "OpenAI", "Institute for Advanced Study", "Vanderbilt University", "University of Cambridge", "Harvard University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Not yet peer reviewed. It extends the same team's February 2026 gluon result, criticised for its narrow kinematic scope; the regime needs Klein-space or complex momenta, and the paper does not itemize which steps the models did.", "significance": 2, "sources": [ { "label": "Guevara et al., arXiv:2603.04330, 2026-03-04", "url": "https://arxiv.org/abs/2603.04330" } ], "foundation0": false }, { "id": "claude-firefox-vulnerabilities", "date": "2026-03-06", "date_basis": "announcement", "title": "Claude Opus 4.6 finds 22 Firefox vulnerabilities in two weeks", "domain": "computing", "subfield": "vulnerability discovery", "summary": "Working with Mozilla, Anthropic used Claude Opus 4.6 to scan nearly 6,000 Firefox C++ files. In two weeks it found bugs that led to 22 CVEs, 14 rated high severity, almost a fifth of the high-severity Firefox vulnerabilities fixed in 2025, plus 90 other bugs. Most fixes shipped in Firefox 148.", "before_after": null, "why_it_matters": "Browsers are among the most attacked and most tested programs. Mozilla found the reports easy to reproduce, showing AI auditing adds value even to mature, heavily fuzzed code.", "ai_systems": [ "Claude Opus 4.6" ], "organizations": [ "Anthropic", "Mozilla" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Mozilla blog, 2026-03-06", "url": "https://blog.mozilla.org/en/firefox/hardening-firefox-anthropic-red-team/" }, { "label": "Anthropic, 2026-03-06", "url": "https://www.anthropic.com/news/mozilla-firefox-security" } ], "foundation0": false }, { "id": "amie-real-world-primary-care-feasibility", "date": "2026-03-09", "date_basis": "preprint", "title": "Conversational diagnostic AI AMIE talks with real patients before clinic visits", "domain": "medicine", "subfield": "primary care, conversational AI", "summary": "At Beth Israel Deaconess, 98 urgent primary-care patients chatted with Google's AMIE before their visit while physicians supervised live. No safety stops were needed; supervisors noted one hallucination. AMIE's differential included the final diagnosis in 90% of cases (75% in its top three); evaluators rated its plans as safe as doctors' but less practical.", "before_after": null, "why_it_matters": "Simulated wins do not show that AI is safe with real patients. This is a first supervised real-world step, small and single-arm, toward testing conversational diagnostic AI in randomized trials.", "ai_systems": [ "AMIE (Articulate Medical Intelligence Explorer)" ], "organizations": [ "Google Research", "Google DeepMind", "Beth Israel Deaconess Medical Center", "Harvard Medical School" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv preprint, 2026-03-09", "url": "https://arxiv.org/abs/2603.08448" }, { "label": "The Lancet, 2026-10-08 (online)", "url": "https://doi.org/10.1016/S0140-6736(26)01535-7" }, { "label": "Google Research blog, 2026-03-11", "url": "https://research.google/blog/exploring-the-feasibility-of-conversational-diagnostic-ai-in-a-real-world-clinical-study/" } ], "foundation0": false }, { "id": "alphaevolve-ramsey-lower-bounds", "date": "2026-03-10", "date_basis": "preprint", "title": "AlphaEvolve raises lower bounds for nine classical Ramsey numbers", "domain": "mathematics", "subfield": "combinatorics / Ramsey theory", "summary": "Google researchers used AlphaEvolve to evolve search programs that build graphs with no large cliques or independent sets. They raised lower bounds for nine Ramsey numbers, such as R(3,13) from 60 to 61 and R(4,19) from 213 to 219, five in March and four more by April, and recovered every known exact lower bound.", "before_after": { "metric": "Lower bound on Ramsey number R(4,19)", "before": "213", "after": "219" }, "why_it_matters": "Ramsey records usually come from bespoke searches and can stand for years. One AI meta-method finding many improvements points to a general tool for extremal constructions, each easy to check.", "ai_systems": [ "AlphaEvolve" ], "organizations": [ "Google DeepMind", "Google", "UC Berkeley" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nagda, Raghavan, Thakurta, Reinforced Generation of Combinatorial Structures: Ramsey Numbers, arXiv:2603.09172", "url": "https://arxiv.org/abs/2603.09172" } ], "foundation0": false }, { "id": "google-urban-flash-flood-forecasts-groundsource", "date": "2026-03-12", "date_basis": "deployment", "title": "Google rolls out AI urban flash-flood forecasts trained on news-derived data", "domain": "earth-climate", "subfield": "flood forecasting", "summary": "Google Research used Gemini to extract dated, located flood reports from news in 80 languages, building Groundsource: 2.6 million flash-flood records from over 150 countries. An LSTM model trained on it with global weather inputs forecasts urban flash floods up to 24 hours ahead on a 20 km grid, now rolling out on Flood Hub.", "before_after": { "metric": "Historical flash-flood records available for training", "before": "~10,000 entries (GDACS, mostly high-impact events)", "after": "2.6 million records (Groundsource)" }, "why_it_matters": "Flash floods cause about 85% of flood deaths, often where gauges and radar are scarce. Forecasts built only from global data can reach cities that lack warning systems.", "ai_systems": [ "Gemini (flood-event extraction from news)", "LSTM flash-flood forecasting model" ], "organizations": [ "Google Research" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "Skill is measured against news-derived and GDACS records; Google says precision is underestimated because some floods go unreported, and much of Africa lacks enough ground truth to estimate accuracy.", "significance": 3, "sources": [ { "label": "Google Research blog, urban flash flood forecasting, 2026-03-12", "url": "https://research.google/blog/protecting-cities-with-ai-driven-flash-flood-forecasting/" }, { "label": "Google Research blog, Groundsource, 2026-03-12", "url": "https://research.google/blog/introducing-groundsource-turning-news-reports-into-data-with-gemini/" }, { "label": "EarthArXiv preprint, AI expands urban flash flood forecasts, 2026-03-09", "url": "https://doi.org/10.31223/X5177S" } ], "foundation0": false }, { "id": "cordoba-partially-autonomous-ai-mammography-trial", "date": "2026-03-19", "date_basis": "publication", "title": "AI clears 64% of mammograms unread and finds more cancers, but recalls rise", "domain": "medicine", "subfield": "breast cancer screening", "summary": "In a prospective paired trial of 31,301 women in Córdoba, Spain, treating exams Transpara rated low-risk as normal, unread, and double-reading the rest with AI support cut radiologist workload 63.6%. Cancer detection rose from 6.3 to 7.3 per 1,000, but recalls rose 14.8%, failing non-inferiority. AI-cleared exams included 11 cancers that standard reading found.", "before_after": { "metric": "Cancers detected per 1,000 women screened", "before": "6.3 (standard double reading)", "after": "7.3 (AI triage plus AI-supported reading)" }, "why_it_matters": "Shows in live screening that AI can take most normal exams off radiologists' desks while finding more cancers overall, the kind of evidence needed before AI reads unsupervised, though recalls rose.", "ai_systems": [ "Transpara 1.7 (ScreenPoint Medical)" ], "organizations": [ "Reina Sofía University Hospital", "Maimónides Biomedical Research Institute of Córdoba (IMIBIC)", "University of Córdoba", "ScreenPoint Medical" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2026-03-19", "url": "https://doi.org/10.1038/s41591-026-04277-x" } ], "foundation0": false }, { "id": "claude-code-agents-collider-analyses-lep-lund-plane", "date": "2026-03-20", "date_basis": "preprint", "title": "AI agents run full particle-physics analyses on archived collider data", "domain": "physics", "subfield": "experimental particle physics", "summary": "MIT and CERN physicists gave Claude Code agents short prompts and open ALEPH, DELPHI and CMS data. Agents handled event selection, backgrounds, statistics and drafting, with humans signing off only at unblinding. They reproduced CMS's H→ττ signal strength within 0.16σ and made what the authors believe is the first Lund-plane measurement in electron-positron collisions.", "before_after": { "metric": "CMS 8 TeV H→ττ (μτh channel) signal strength", "before": "1.01 ± 0.41 (published CMS result)", "after": "1.20 ± 1.13 (agent, partial 2012 open data)" }, "why_it_matters": "Much of a collider analysis is repetitive code. Agents that carry it could free physicists for physics questions and reopen archived data from experiments that have shut down.", "ai_systems": [ "Claude Code with Claude Opus models", "Just Furnish Context (JFC) framework" ], "organizations": [ "MIT", "NSF AI Institute for Artificial Intelligence and Fundamental Interactions", "CERN" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "The authors present demonstrations, not precision measurements: the agent's H→ττ uncertainty is nearly three times CMS's, and one ALEPH Z-width result sits 3.4σ low.", "significance": 3, "sources": [ { "label": "Moreno et al., arXiv 2603.20179, 2026-03-20", "url": "https://arxiv.org/abs/2603.20179" } ], "foundation0": false }, { "id": "golden-bridge-ii-ai-stroke-decision-support", "date": "2026-03-20", "date_basis": "publication", "title": "AI-assisted stroke decision support cuts repeat vascular events in 77 hospitals", "domain": "medicine", "subfield": "neurology, stroke care", "summary": "In a cluster-randomized trial across 77 Chinese hospitals (21,603 ischaemic stroke patients), a decision-support system combining AI imaging analysis, stroke-cause classification and guideline recommendations cut new vascular events at 3 months from 3.9% to 2.9% (HR 0.74), without more bleeding. Disability and mortality did not differ.", "before_after": { "metric": "New vascular event within 3 months", "before": "3.9% (usual care)", "after": "2.9% (AI-supported care)" }, "why_it_matters": "Shows AI-based support can improve hard patient outcomes at scale in ordinary hospitals by helping clinicians apply proven treatments more consistently.", "ai_systems": [ "Stroke clinical decision support system with AI imaging analysis" ], "organizations": [ "Beijing Tiantan Hospital, Capital Medical University", "China National Clinical Research Center for Neurological Diseases" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "BMJ, 2026-03-20", "url": "https://doi.org/10.1136/bmj-2025-085810" } ], "foundation0": false }, { "id": "raven-tess-118-validated-planets", "date": "2026-03-25", "date_basis": "publication", "title": "RAVEN machine-learning pipeline validates 118 TESS planets, 31 newly found", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "Warwick's RAVEN pipeline searched TESS full-frame images of over 2.2 million Sun-like stars for planets with 0.5-16 day orbits, using classifiers trained on simulated planets and false positives, then statistical validation. It validated 118 planets, 31 newly detected, and listed over 2,000 high-probability candidates, about 1,000 of them new.", "before_after": { "metric": "Uncertainty on close-in planet occurrence rates (companion RAVEN study)", "before": "Kepler-based estimates", "after": "about 10 times smaller on average" }, "why_it_matters": "Uniformly vetted planets let astronomers measure how common planet types are. A companion study puts rare 'Neptunian desert' planets around just 0.08% of Sun-like stars.", "ai_systems": [ "RAVEN (machine-learning classifiers trained on simulated signals)" ], "organizations": [ "University of Warwick", "Instituto de Ciencias Físicas (CONICET/UNSAM)" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Planets are statistically validated, not confirmed by mass measurements; the 2,000 high-probability candidates are not validated.", "significance": 3, "sources": [ { "label": "Lafarga et al., MNRAS 548, stag512, 2026-03-25", "url": "https://doi.org/10.1093/mnras/stag512" }, { "label": "University of Warwick press release, 2026-03-25", "url": "https://warwick.ac.uk/news/pressreleases/ai-approach-uncovers-dozens-of-hidden-planets/" }, { "label": "Cui et al., MNRAS 546, stag022 (occurrence rates), 2026-01-07", "url": "https://doi.org/10.1093/mnras/stag022" } ], "foundation0": false }, { "id": "lithium-metal-electrolytes-deep-active-learning", "date": "2026-03-27", "date_basis": "publication", "title": "Active learning finds electrolytes that triple lithium-metal cell lifetime", "domain": "energy", "subfield": "batteries", "summary": "A two-stage deep active learning framework chose which electrolyte formulations to test, modelling abrupt jumps between composition and performance, then transferred what it learned to new components with few extra measurements. After three learning rounds it found electrolytes that tripled the average lifetime of lithium-metal symmetric cells, and quickly improved Li||NMC811 full-cell formulations.", "before_after": { "metric": "Average lifetime of lithium-metal symmetric cells", "before": "starting formulations", "after": "about 3x, after three learning iterations" }, "why_it_matters": "Lithium-metal anodes could greatly raise battery energy density but fail quickly. Finding long-lived electrolytes with few costly experiments eases a key bottleneck.", "ai_systems": [ "Deep kernel learning active-learning framework with knowledge transfer" ], "organizations": [ "Tsinghua University", "Peking University", "SES AI", "Lawrence Berkeley National Laboratory", "Princeton University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Hong et al., Nature Communications 17, 5146, 2026-03-27", "url": "https://doi.org/10.1038/s41467-026-70973-4" } ], "foundation0": false }, { "id": "google-deepmind-cyclone-model-nhc-2025-verification", "date": "2026-03-30", "date_basis": "publication", "title": "Hurricane center's 2025 review ranks Google DeepMind's AI model among the best", "domain": "earth-climate", "subfield": "tropical cyclone forecasting", "summary": "Google DeepMind's experimental cyclone ensemble, shown live to US National Hurricane Center forecasters from June 2025, was assessed in NHC's official 2025 verification. Its ensemble mean (GDMI) slightly beat NHC's own Atlantic track forecasts at 12-72 hours, was the best Atlantic intensity model, and beat NHC's eastern Pacific track forecasts at 48-120 hours.", "before_after": { "metric": "Atlantic track skill at 12-72 h, 2025 season", "before": "NHC official forecast best of all guidance", "after": "GDMI slightly better than NHC official" }, "why_it_matters": "Hurricane track and strength decide who evacuates. An independent government review finding an AI model at or above expert forecasts is strong evidence it helps protect lives.", "ai_systems": [ "Google DeepMind Weather Lab experimental cyclone model (ensemble mean GDMI)" ], "organizations": [ "Google DeepMind", "Google Research", "NOAA National Hurricane Center" ], "ai_role": "instrument", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "NHC Forecast Verification Report, 2025 season, 2026-03-30", "url": "https://www.nhc.noaa.gov/verification/pdfs/Verification_2025.pdf" }, { "label": "Google DeepMind blog, Weather Lab launch, 2025-06-12", "url": "https://deepmind.google/blog/how-were-supporting-better-tropical-cyclone-prediction-with-ai/" } ], "foundation0": false }, { "id": "automatic-graduate-textbook-formalization", "date": "2026-04-03", "date_basis": "preprint", "title": "30,000 Claude agents formalize a 500-page graduate textbook in Lean in a week", "domain": "mathematics", "subfield": "formal proof / algebraic combinatorics", "summary": "Researchers at Meta FAIR and École des Ponts ran about 30,000 Claude 4.5 Opus agents in parallel on a shared code base to formalize Darij Grinberg's 500-plus-page graduate algebraic combinatorics textbook in Lean: 130,000 lines and 5,900 declarations, with all 340 target definitions and theorems formalized, in one week for about $100,000.", "before_after": null, "why_it_matters": "Formalizing new research needs foundational texts in Lean faster than volunteers can write them; converting a whole graduate textbook automatically shows that bottleneck can be lifted.", "ai_systems": [ "Claude 4.5 Opus" ], "organizations": [ "Meta FAIR", "École des Ponts (CERMICS)", "Korea Institute for Advanced Study" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Gloeckle et al., Automatic Textbook Formalization, arXiv:2604.03071", "url": "https://arxiv.org/abs/2604.03071" } ], "foundation0": false }, { "id": "claude-mythos-zero-days", "date": "2026-04-07", "date_basis": "announcement", "title": "Anthropic reports Claude Mythos Preview found thousands of zero-day bugs", "domain": "computing", "subfield": "vulnerability discovery and exploitation", "summary": "Anthropic said Claude Mythos Preview found thousands of high- and critical-severity vulnerabilities across every major operating system and browser, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw, and wrote working exploits. Reviewers matched its severity ratings exactly in 89% of 198 checked reports; under 1% were patched at announcement.", "before_after": { "metric": "Working Firefox JavaScript-engine exploits in Anthropic's test harness (sandbox removed)", "before": "2 in several hundred attempts (Claude Opus 4.6)", "after": "181 (Mythos Preview)" }, "why_it_matters": "AI that finds and weaponizes flaws at this scale shifts the balance between attackers and defenders, so Anthropic restricted access to organizations patching critical software (Project Glasswing).", "ai_systems": [ "Claude Mythos Preview" ], "organizations": [ "Anthropic", "Project Glasswing partners" ], "ai_role": "originator", "evidence": "claimed", "contested": true, "contested_note": "AISLE (blog, 2026-04-07) reported that small open-weight models, given the isolated vulnerable code, detected the flagship FreeBSD bug and, in one case, the core OpenBSD bug chain, arguing the capability is less exclusive than claimed. Most findings were undisclosed and unpatched, so the totals cannot yet be checked independently.", "significance": 4, "sources": [ { "label": "Anthropic Frontier Red Team, 2026-04-07", "url": "https://www.anthropic.com/research/mythos-preview" }, { "label": "Anthropic, Project Glasswing, 2026-04-07", "url": "https://www.anthropic.com/glasswing" }, { "label": "AISLE blog (critique), 2026-04-07", "url": "https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier" } ], "foundation0": false }, { "id": "aaai-26-ai-review-pilot-22977-papers", "date": "2026-04-15", "date_basis": "preprint", "title": "AAAI-26 gives every one of 22,977 submissions an AI peer review", "domain": "science-of-science", "subfield": "peer review", "summary": "In the first conference-scale deployment of AI peer review, every AAAI-26 main-track paper got one labelled AI review alongside human reviews: 22,977 reviews in under a day, at under $1 each. Of 5,834 survey responses, 53.9% judged them useful versus 20.2% not; AI reviews rated better at spotting technical errors but more prone to nitpicking.", "before_after": { "metric": "Papers given an AI-generated review at one conference", "before": "no prior conference-scale deployment", "after": "22,977, all main-track submissions" }, "why_it_matters": "Submissions are outgrowing the supply of volunteer reviewers. Showing where AI reviews help at full scale, and where they fail, shapes how peer review copes.", "ai_systems": [ "AAAI-26 AI review system on GPT-5 (high reasoning, code execution, web search)" ], "organizations": [ "AAAI", "University of Texas at Austin", "University of Alberta", "University of Michigan", "Oregon State University", "Sony AI" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "Survey respondents self-selected. They also judged AI reviews more likely to overemphasise minor issues and to contain technical errors, and weak at judging novelty and significance.", "significance": 3, "sources": [ { "label": "Biswas et al., arXiv 2604.13940, 2026-04-15", "url": "https://arxiv.org/abs/2604.13940" } ], "foundation0": false }, { "id": "humanoid-robot-half-marathon-autonomous-record", "date": "2026-04-19", "date_basis": "competition", "title": "Autonomous humanoid robot runs a half marathon faster than the human record", "domain": "engineering-robotics", "subfield": "legged robotics", "summary": "At the Beijing E-Town humanoid robot half marathon, Honor's 'Lightning' covered the 21.1 km course in 50 minutes 26 seconds using autonomous navigation, beating the men's world record of 57:20. The 2025 winner, a remote-controlled robot, took 2 hours 40 minutes. About 300 robots from over 100 teams took part.", "before_after": { "metric": "Fastest robot half-marathon time", "before": "2 h 40 min (2025 winner, remote-controlled)", "after": "50 min 26 s, autonomous navigation (2026)" }, "why_it_matters": "Running fast for an hour without falling or getting lost tests balance, endurance, cooling and navigation together. A roughly threefold gain in a year shows how fast humanoid robots are improving.", "ai_systems": [ "Autonomous navigation and locomotion control (Honor 'Lightning' humanoid)" ], "organizations": [ "Honor" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "The robot needed a battery change at 10.6 km (Honor, via Sixth Tone), and much of the speed gain came from joint and cooling hardware. Ars Technica notes it does not imply real-world uses.", "significance": 2, "sources": [ { "label": "Sixth Tone, 2026-04-20", "url": "https://www.sixthtone.com/news/1018442" }, { "label": "Ars Technica, 2026-04-20", "url": "https://arstechnica.com/gadgets/2026/04/robot-runner-handily-beats-humans-in-half-marathon-setting-new-record/" }, { "label": "iRunFar race report, 2026-04-20", "url": "https://www.irunfar.com/2026-beijing-e-town-half-marathon-humanoid-robot-beats-human-world-record" } ], "foundation0": false }, { "id": "t16-tess-10000-planet-candidates", "date": "2026-04-20", "date_basis": "preprint", "title": "Machine-learning-assisted search finds 10,091 new TESS planet candidates", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "The T16 project detrended 83.7 million TESS Cycle 1 light curves for stars down to magnitude 16, then searched them with random-forest classifiers and manual vetting. It found 11,554 transiting-planet candidates, 10,091 new, more than doubling known TESS candidates; radial-velocity follow-up confirmed one as a new hot Jupiter around TIC 183374187.", "before_after": { "metric": "Known TESS planet candidates", "before": "existing TESS candidate list", "after": "more than doubled (10,091 new)" }, "why_it_matters": "Most planet searches skip faint stars. Opening them up gives far larger samples for measuring how common planets are, plus a long target list for follow-up.", "ai_systems": [ "Random-forest transit classifiers" ], "organizations": [ "Princeton University", "Carnegie Science", "University of California, Los Angeles" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Candidates are not confirmed planets; only one has been confirmed so far.", "significance": 2, "sources": [ { "label": "Roth et al., Astrophysical Journal Supplement Series 284:19, 2026-04-28", "url": "https://doi.org/10.3847/1538-4365/ae5b6c" }, { "label": "arXiv:2604.18579, 2026-04-20", "url": "https://arxiv.org/abs/2604.18579" } ], "foundation0": false }, { "id": "firefox-150-mythos-271-fixes", "date": "2026-04-21", "date_basis": "deployment", "title": "Firefox 150 ships fixes for 271 bugs found with Claude Mythos Preview", "domain": "computing", "subfield": "vulnerability discovery / browser security", "summary": "Mozilla's Firefox 150, released 21 April 2026, fixed 271 security bugs that its pipeline found with an early Claude Mythos Preview: 180 rated sec-high, 80 moderate and 11 low, including a 20-year-old XSLT memory bug and several sandbox escapes. Firefox 148 had fixed 22 security bugs found with Claude Opus 4.6.", "before_after": { "metric": "Security bugs found with Claude and fixed in one Firefox release", "before": "22 (Firefox 148, Claude Opus 4.6)", "after": "271 (Firefox 150, Claude Mythos Preview)" }, "why_it_matters": "A heavily fuzzed, mature browser shipped hundreds of AI-found security fixes in one release, showing defenders can now use these models to harden critical code at scale.", "ai_systems": [ "Claude Mythos Preview", "Claude Opus 4.6" ], "organizations": [ "Mozilla", "Anthropic" ], "ai_role": "originator", "evidence": "deployed", "contested": false, "contested_note": "Severity ratings rest on crash symptoms, not confirmed exploits. Advisory MFSA 2026-30 groups internally found bugs into rollup CVEs, so the 271 do not appear individually; Mozilla's 7 May post reconciles the counts.", "significance": 3, "sources": [ { "label": "Mozilla, The zero-days are numbered (2026-04-21)", "url": "https://blog.mozilla.org/en/privacy-security/ai-security-zero-day-vulnerabilities/" }, { "label": "Mozilla Hacks, Behind the Scenes Hardening Firefox with Claude Mythos Preview (2026-05-07)", "url": "https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/" }, { "label": "Mozilla Security Advisory MFSA 2026-30", "url": "https://www.mozilla.org/en-US/security/advisories/mfsa2026-30/" } ], "foundation0": false }, { "id": "elementsclaw-agent-four-new-superconductors", "date": "2026-04-26", "date_basis": "preprint", "title": "Materials agent ElementsClaw guides discovery of four new superconductors", "domain": "chemistry-materials", "subfield": "superconducting materials", "summary": "Alibaba DAMO Academy, Renmin University and Chinese Academy of Sciences researchers built ElementsClaw, an agent pairing a 1-billion-parameter atomic model with language-model reasoning. It flagged 68,000 high-confidence superconductor candidates among 2.4 million crystals in 28 GPU hours. Synthesis confirmed four new superconductors, including Zr3ScRe8 (Tc 6.5 K) and HfZrRe4 (5.9 K).", "before_after": null, "why_it_matters": "Generative models propose millions of materials, but choosing what to make is the bottleneck. An agent that ranks candidates and explains its picks turned predictions into verified superconductors.", "ai_systems": [ "ElementsClaw agent", "Elements (1-billion-parameter large atomic model)", "Large language models" ], "organizations": [ "Alibaba DAMO Academy", "Hupan Lab", "Renmin University of China", "Institute of Physics, University of Chinese Academy of Sciences", "Tsinghua University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Not yet peer reviewed; company-led. All four are rhenium-rich intermetallics in a system with known superconductors, with transition temperatures of 2.5-6.5 K, far below practical levels.", "significance": 2, "sources": [ { "label": "Li et al., arXiv:2604.23758, 2026-04-26", "url": "https://arxiv.org/abs/2604.23758" } ], "foundation0": false }, { "id": "sierra-leone-ml-essential-medicines-allocation", "date": "2026-04-29", "date_basis": "publication", "title": "Machine learning raises use of essential medicines across Sierra Leone", "domain": "medicine", "subfield": "global health, supply chains", "summary": "Working with Sierra Leone's government, researchers built a decision-aware machine-learning system that allocates scarce essential medicines to health facilities from sparse data. A staggered nationwide rollout raised consumption of allocated products by an estimated 19% in treated districts. The tool now covers about two million women and children under five.", "before_after": { "metric": "Consumption of allocated essential medicines", "before": "standard allocation", "after": "+19% in districts using the ML tool" }, "why_it_matters": "In low-income health systems, medicines often sit unused in one clinic while another runs out. Cheap software that allocates stock better can widen access without new supplies.", "ai_systems": [ "Decision-aware machine learning allocation tool" ], "organizations": [ "University of Pennsylvania", "Sierra Leone National Medical Supplies Agency", "Sierra Leone Ministry of Health and Sanitation" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2026-04-29", "url": "https://doi.org/10.1038/s41586-026-10433-7" }, { "label": "Nature News & Views (Obermeyer), 2026-04-29", "url": "https://www.nature.com/articles/d41586-026-01152-0" } ], "foundation0": false }, { "id": "spark-agentic-pathology-discovery", "date": "2026-04-29", "date_basis": "publication", "title": "Pathology agent system proposes and tests its own tumour measurements", "domain": "medicine", "subfield": "computational pathology", "summary": "SPARK, a multi-agent system from Cologne-led researchers, turns biological ideas into image-analysis tools on tissue slides without extra model training, then tests them. Across 18 cohorts, five cancer types and more than 5,400 patients, its self-generated concepts correlated with survival, known pathology variables and predictive biomarkers.", "before_after": null, "why_it_matters": "Pathology AI usually measures features people chose in advance. An agent that proposes and tests its own measurable ideas could surface tumour biology nobody thought to look for.", "ai_systems": [ "SPARK (System of Pathology Agents for Research and Knowledge)" ], "organizations": [ "University of Cologne", "University Hospital Cologne", "University Hospital Halle", "University Hospital Essen" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Retrospective cohorts only; the authors say prospective validation is needed before clinical use.", "significance": 2, "sources": [ { "label": "Trost et al., Nature Medicine 32:2254, online 2026-04-29", "url": "https://doi.org/10.1038/s41591-026-04357-y" }, { "label": "Nature Medicine research briefing, 2026-05-05", "url": "https://doi.org/10.1038/s41591-026-04403-9" } ], "foundation0": false }, { "id": "erdos-1196-primitive-sets", "date": "2026-05-01", "date_basis": "preprint", "title": "GPT-5.4 Pro output yields new method resolving Erdős primitive-set problems", "domain": "mathematics", "subfield": "number theory / primitive sets", "summary": "An autonomous GPT-5.4 Pro run produced the first proof of Erdős problem #1196 in April 2026. Alexeev, Barreto, Lichtman, Price, Tao and coauthors turned its idea, Markov chains with von Mangoldt weights, into a method proving #1196 and #1217, a short proof of the Erdős Primitive Set Conjecture and a revised Banks–Martin conjecture.", "before_after": null, "why_it_matters": "An AI-suggested idea, overlooked since Erdős's 1935 paper, became a reusable method settling several problems at once: AI contributing technique, not just answers.", "ai_systems": [ "GPT-5.4 Pro", "GPT-5.5 Pro (early version)", "Codex" ], "organizations": [ "UCLA", "Stanford University", "University of Cambridge" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2605.00301", "url": "https://arxiv.org/abs/2605.00301" }, { "label": "erdosproblems.com forum, #1196", "url": "https://www.erdosproblems.com/forum/thread/1196" } ], "foundation0": false }, { "id": "unsupervised-ml-earthquake-preparatory-phase", "date": "2026-05-04", "date_basis": "publication", "title": "Unsupervised machine learning spots foreshock patterns before large quakes", "domain": "earth-climate", "subfield": "seismology / earthquake forecasting", "summary": "GFZ-led researchers used unsupervised machine learning to sort catalogued earthquakes into clustered families and categories. Before the Kahramanmaraş (2023), Iquique (2014) and L'Aquila (2009) earthquakes, a 'critical' category of long-lasting families appeared weeks to months ahead. For Noto (2024) and Amatrice (2016), which had no known precursors, it found none.", "before_after": null, "why_it_matters": "Whether big earthquakes announce themselves is a central open question. A method that flags preparatory phases without preset rules, and stays silent otherwise, could feed operational forecasting.", "ai_systems": [ "Unsupervised categorization of seismicity-family features" ], "organizations": [ "GFZ Helmholtz Centre for Geosciences", "University of Genoa", "OGS Trieste", "Stanford University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Retrospective test on five earthquakes, not a forecast; the authors note it cannot detect quakes that lack a seismic preparatory phase.", "significance": 2, "sources": [ { "label": "Karimpouli et al., Nature Communications 17, 4024, 2026-05-04", "url": "https://doi.org/10.1038/s41467-026-72279-x" }, { "label": "GFZ news release via EurekAlert!, 2026-06-19", "url": "https://www.eurekalert.org/news-releases/1133096" } ], "foundation0": false }, { "id": "alphaevolve-spanner-2026", "date": "2026-05-07", "date_basis": "announcement", "title": "AlphaEvolve cuts Spanner write amplification 20% and software size nearly 9%", "domain": "computing", "subfield": "storage systems / compilers", "summary": "A year after launch, Google DeepMind reported that AlphaEvolve refined the compaction heuristics of Spanner, Google's global database, cutting write amplification by 20%, and found compiler strategies that shrank software's storage footprint by nearly 9%. It is now a regular tool in TPU design and found cache policies in two days rather than months.", "before_after": { "metric": "Spanner write amplification", "before": "Previous LSM-tree compaction heuristics", "after": "20% lower" }, "why_it_matters": "Writing less data per request means longer-lasting drives and lower costs in one of the world's largest databases. AI-found optimizations are becoming routine production engineering.", "ai_systems": [ "AlphaEvolve", "Gemini" ], "organizations": [ "Google DeepMind", "Google" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Google DeepMind blog, 2026-05-07", "url": "https://deepmind.google/blog/alphaevolve-impact/" } ], "foundation0": false }, { "id": "ml-designed-electrolyzer-flow-channel", "date": "2026-05-08", "date_basis": "publication", "title": "Machine-learning-designed electrolyzer channel raises current density 23%", "domain": "energy", "subfield": "green hydrogen / electrolysis", "summary": "Tsinghua University engineers used a mixture-of-experts machine-learning surrogate, instead of simulating every geometry with fluid dynamics, to search flow-channel designs that clear gas bubbles in water electrolyzers. A prototype with the AI-chosen array-type channel delivered about 23% more current density at 2 V than a conventional serpentine design, and the gain held in scaled-up devices.", "before_after": { "metric": "Current density at 2 V", "before": "conventional serpentine flow channel", "after": "~23% higher (ML-designed array channel)" }, "why_it_matters": "Green hydrogen is costly partly because trapped bubbles waste energy in electrolyzers. A better channel geometry is a cheap hardware change that gets more hydrogen from the same voltage.", "ai_systems": [ "Mixture-of-experts surrogate model" ], "organizations": [ "Tsinghua University" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Zhang et al., Science Advances 12, eadz1865, 2026-05-08", "url": "https://doi.org/10.1126/sciadv.adz1865" } ], "foundation0": false }, { "id": "unit-distance-conjecture-disproved", "date": "2026-05-20", "date_basis": "announcement", "title": "OpenAI model disproves Erdős's 1946 unit distance conjecture", "domain": "mathematics", "subfield": "discrete geometry / algebraic number theory", "summary": "An internal OpenAI model constructed, in one shot, planar point sets with at least n^(1+δ) unit-distance pairs (fixed δ>0, infinitely many n), refuting Erdős's 1946 conjecture of n^(1+o(1)). The proof uses infinite class field towers. Nine mathematicians including Alon, Gowers and Tsimerman published a checked, simplified version; humans soon improved the exponent.", "before_after": { "metric": "Max unit distances among n planar points (lower bound)", "before": "n^(1+c/log log n), believed optimal since 1946", "after": "n^(1+δ) for a fixed δ > 0" }, "why_it_matters": "Gowers wrote he would have recommended it to the Annals of Mathematics without hesitation; it overturned an 80-year-old belief in discrete geometry using number theory few expected to apply.", "ai_systems": [ "OpenAI internal model (unnamed)" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "", "significance": 5, "sources": [ { "label": "OpenAI, Planar point sets with many unit distances (2026-05-20)", "url": "https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf" }, { "label": "Alon, Bloom, Gowers et al., Remarks on the disproof", "url": "https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-remarks.pdf" }, { "label": "Quanta Magazine, 2026-08-03", "url": "https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/" } ], "foundation0": false }, { "id": "deepmind-formal-proof-search-erdos", "date": "2026-05-21", "date_basis": "preprint", "title": "Formal proof-search agents resolve 9 open Erdős problems in Lean", "domain": "mathematics", "subfield": "combinatorics / formal proof", "summary": "Google DeepMind evaluated LLM agents that write Lean proofs on open problems at scale. The strongest autonomously resolved 9 of 353 open Erdős problems with Lean statements, at a few hundred dollars each, and proved 44 of 492 OEIS conjectures; a simpler generate-and-check agent matched the Erdős results but cost more on the hardest problems.", "before_after": null, "why_it_matters": "Every claimed solution is machine-checked, so AI results can be trusted at scale and low cost, turning plausible text into guaranteed proofs.", "ai_systems": [ "AlphaProof Nexus", "Gemini 3.1 Pro", "AlphaProof" ], "organizations": [ "Google DeepMind" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2605.22763", "url": "https://arxiv.org/abs/2605.22763" } ], "foundation0": false }, { "id": "tokamak-agnostic-elm-forecasting-zero-shot", "date": "2026-06-01", "date_basis": "publication", "title": "Neural network trained on one tokamak forecasts plasma bursts on another", "domain": "energy", "subfield": "fusion plasma diagnostics", "summary": "A network trained only on megahertz turbulence measurements from 1,400 DIII-D discharges forecasts the onset of Type-I edge-localized modes, damaging plasma bursts, with true-positive rates above 80% on DIII-D. Applied to the Korean KSTAR tokamak with no retraining, it still forecast onsets, and explainability tools showed it tracked physical instability drivers.", "before_after": { "metric": "Retraining needed to forecast ELMs on a different tokamak", "before": "device-specific training data", "after": "none (zero-shot transfer DIII-D to KSTAR)" }, "why_it_matters": "Future reactors such as ITER will have little data to train on before they must be protected. Models that carry over between machines could guard them from the first day.", "ai_systems": [ "BES-NN (neural network with Grad-CAM explanations)" ], "organizations": [ "University of Wisconsin-Madison", "Princeton University", "SLAC National Accelerator Laboratory", "Korea Institute of Fusion Energy" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "Tested on archived discharges; the authors leave real-time use to future work.", "significance": 2, "sources": [ { "label": "Joung et al., Communications Physics 9, 301, 2026-06-01", "url": "https://doi.org/10.1038/s42005-026-02689-2" } ], "foundation0": false }, { "id": "jamming-exponent-identity-proof-claude", "date": "2026-06-02", "date_basis": "preprint", "title": "Parisi and Zamponi prove a 2014 jamming conjecture with Claude's help", "domain": "physics", "subfield": "statistical physics (jamming, glasses)", "summary": "Giorgio Parisi and Francesco Zamponi proved analytically that two jamming exponents satisfy a + b = 1, a relation seen numerically since 2014 but never proved, yielding scaling laws predicted by Wyart's marginal-stability arguments. The paper says the proof came through interaction with Claude (Sonnet 4.6 and Opus 4.7) and was verified by the authors.", "before_after": { "metric": "Status of the identity a + b = 1 for jamming exponents", "before": "numerical observation, unproven since 2014", "after": "analytic proof, peer reviewed" }, "why_it_matters": "Jamming governs how sand, foams and glasses become rigid. The proof ties replica theory to marginal-stability arguments, and a Nobel laureate credits an AI model with the key idea.", "ai_systems": [ "Claude Sonnet 4.6", "Claude Opus 4.7" ], "organizations": [ "Sapienza University of Rome" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Zamponi told Live Science that Claude's first output contained errors needing revision but its core idea was correct; about 40 prompts led to the verified proof.", "significance": 3, "sources": [ { "label": "Parisi & Zamponi, J. Stat. Mech. (2026) 073301, 2026-07-01", "url": "https://doi.org/10.1088/1742-5468/ae7bd7" }, { "label": "arXiv:2606.03300, 2026-06-02", "url": "https://arxiv.org/abs/2606.03300" }, { "label": "Live Science, 2026-07-16", "url": "https://www.livescience.com/physics-mathematics/mathematics/nobel-prize-winning-physicist-and-team-use-claude-ai-to-solve-decades-old-math-puzzle" } ], "foundation0": false }, { "id": "lattice-sphere-packing-gpt55", "date": "2026-06-03", "date_basis": "preprint", "title": "GPT-5.5 Pro helps improve the high-dimensional lattice packing bound", "domain": "mathematics", "subfield": "geometry of numbers / sphere packing", "summary": "Abuya, Gargava and Zhao proved lattice sphere packings of density at least c·N²·log log N·2^(−N) in infinitely many dimensions N, beating Klartag's bound by a log log N factor. Prompted by Zhao, GPT-5.5 Pro proposed the approach and proved the bound; the other two authors had pursued a similar idea, and all three verified it.", "before_after": { "metric": "Lattice packing density lower bound", "before": "c·N²·2^(−N) (Klartag)", "after": "c·N²·log log N·2^(−N)" }, "why_it_matters": "Dense high-dimensional packings underpin coding theory and lattice cryptography; an AI combined two leading techniques into a result experts then checked.", "ai_systems": [ "GPT-5.5 Pro", "Codex" ], "organizations": [ "Massachusetts Institute of Technology", "Tel Aviv University", "Université Paris-Saclay" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "arXiv:2606.05105", "url": "https://arxiv.org/abs/2606.05105" } ], "foundation0": false }, { "id": "plant-tf-binding-deep-learning-arabidopsis-maize", "date": "2026-06-03", "date_basis": "publication", "title": "Deep learning reads plant regulatory DNA in Arabidopsis and transfers to maize", "domain": "biology", "subfield": "plant regulatory genomics, crop science", "summary": "One neural network trained on 568 Arabidopsis DNA-binding datasets covering 46 transcription-factor families predicts where regulators bind DNA. Of 7,364 trait-linked variants, 20.72% were predicted to gain or lose binding sites; reporter assays on 340 constructs supported several predictions. The model also transferred to maize heat-stress data.", "before_after": null, "why_it_matters": "Many crop-trait variants sit in regulatory DNA, not genes. Predicting how they alter transcription-factor binding could help pinpoint causal variants, even in species with little data.", "ai_systems": [ "Multi-label convolutional neural network" ], "organizations": [ "IPK Leibniz Institute", "Forschungszentrum Jülich", "Heinrich Heine University Düsseldorf", "Max Planck Institute for Plant Breeding Research" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Peleke et al., Nature Communications, 2026-06-03", "url": "https://doi.org/10.1038/s41467-026-73634-8" } ], "foundation0": false }, { "id": "flex-cat-autonomous-hydroformylation-lab", "date": "2026-06-20", "date_basis": "publication", "title": "Autonomous catalysis lab finds rhodium catalyst conditions 2.5 times faster", "domain": "chemistry-materials", "subfield": "homogeneous catalysis", "summary": "Flex-Cat couples parallel pressurized mini-reactors with Bayesian optimization over ligands and process conditions. In 680 autonomous experiments on rhodium-catalysed propylene hydroformylation it found ligand-condition regions with over 2.5-fold higher turnover frequency, widened the regioselectivity range, and found ligands whose selectivity flips with conditions; top hits were validated at 10-fold volume.", "before_after": { "metric": "Turnover frequency in propylene hydroformylation", "before": "baseline ligand-condition sets", "after": "over 2.5x higher (680 autonomous experiments)" }, "why_it_matters": "Hydroformylation is a major industrial route to aldehydes. Running pressurized gas-liquid catalysis autonomously, with an industrial partner and scale-up checks, moves self-driving labs toward real processes.", "ai_systems": [ "Hierarchical plate-constrained Bayesian optimization" ], "organizations": [ "North Carolina State University", "University of North Carolina at Chapel Hill", "Eastman Chemical Company" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Bennett et al., Nature Communications 17, 7783, 2026-06-20", "url": "https://doi.org/10.1038/s41467-026-74425-x" } ], "foundation0": false }, { "id": "nicocr-printable-alloys-active-learning", "date": "2026-06-23", "date_basis": "publication", "title": "Active-learning lab finds 3D-printable alloys that resist heat and oxidation", "domain": "chemistry-materials", "subfield": "alloys / additive manufacturing", "summary": "University of Toronto researchers paired active learning with laser 3D printing to search nickel-cobalt-chromium compositions. In a few weeks they found six printable alloys up to about 40% harder than equiatomic NiCoCr; Ni12Co62Cr26 kept about 50% more hardness at 600 °C, and Ni36Co14Cr50 cut oxidation mass gain by 85% at 1,000 °C versus conventional superalloys.", "before_after": { "metric": "Oxidation mass gain at 1,000 °C", "before": "conventional superalloys", "after": "85% lower (Ni36Co14Cr50)" }, "why_it_matters": "Jet engines and power plants need printable metals that survive heat. Finding them with few experiments, far from the usual equal-part recipes, shortens alloy development.", "ai_systems": [ "Active-learning model for experiment selection" ], "organizations": [ "University of Toronto", "Acceleration Consortium", "National Research Council Canada" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Talbot et al., npj Advanced Manufacturing, 2026-06-23", "url": "https://doi.org/10.1038/s44334-026-00098-5" }, { "label": "U of T Engineering News, 2026-07-13", "url": "https://news.engineering.utoronto.ca/self-driving-lab-leverages-ai-to-develop-tough-new-3d-printable-metal-alloys-for-aerospace-and-advanced-manufacturing/" } ], "foundation0": false }, { "id": "deep-learning-ecg-sudden-cardiac-death-biomarker", "date": "2026-06-24", "date_basis": "publication", "title": "Deep learning reveals a new ECG biomarker for sudden cardiac death", "domain": "medicine", "subfield": "cardiology, risk prediction", "summary": "Trained on every ECG in a Swedish region linked to death records, a deep-learning model flagged 2.2% of patients with a 7.0% annual sudden-death rate, above those with weak heart pumping (4.6%); 86.1% were missed by that standard test. A generative model exposed the visible waveform feature driving it. US and Taiwanese cohorts validated it.", "before_after": { "metric": "Annual sudden cardiac death rate in flagged group", "before": "4.6% (low ejection fraction, current criterion)", "after": "7.0% (ECG model high-risk group)" }, "why_it_matters": "Defibrillators prevent sudden death, but today's main criterion misses most victims. An easily visible ECG marker found by AI could better target who receives one, pending prospective trials.", "ai_systems": [ "Deep-learning ECG risk model with a generative ECG waveform model" ], "organizations": [ "University of California, Berkeley", "University of Chicago Booth School of Business", "MIT", "Halmstad University and Region Halland", "University of Gothenburg / Sahlgrenska University Hospital" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature, 2026-06-24", "url": "https://doi.org/10.1038/s41586-026-10674-6" } ], "foundation0": false }, { "id": "nise-zero-shot-small-molecule-binders", "date": "2026-06-24", "date_basis": "publication", "title": "Neural networks design proteins that bind two drugs, with 100% and 83% success", "domain": "biology", "subfield": "de novo protein design", "summary": "Dana-Farber and Harvard researchers paired two neural networks in a design loop to make proteins from scratch that bind two chemically distinct drugs, exatecan and apixaban, with success rates of 100% and 83%. The tightest had picomolar-to-nanomolar affinity; an optimized exatecan binder protected the drug's labile lactone ring from hydrolysis for days.", "before_after": { "metric": "Affinity of tightest binder versus next-leading method", "before": "1x (next-leading method)", "after": "70x (exatecan); nearly 10,000x (apixaban)" }, "why_it_matters": "Proteins that grab small molecules can deliver drugs, sense them or catalyze reactions. Zero-shot design with high success cuts the trial and error that limited earlier methods.", "ai_systems": [ "LASErMPNN", "NISE (neural iterative selection-expansion)" ], "organizations": [ "Dana-Farber Cancer Institute", "Harvard Medical School" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Fry, Slaw and Polizzi, Nature, 2026-06-24", "url": "https://doi.org/10.1038/s41586-026-10670-w" } ], "foundation0": false }, { "id": "pherc-1667-complete-virtual-unwrapping", "date": "2026-06-25", "date_basis": "announcement", "title": "Herculaneum scroll's surviving text virtually unrolled and read end to end", "domain": "science-of-science", "subfield": "papyrology / digital humanities", "summary": "Using phase-contrast X-ray scans from ESRF's BM18 beamline, computational unrolling and machine-learning ink detection, a Vesuvius Challenge, University of Kentucky and ESRF team read all surviving text of PHerc. 1667, the core left by failed openings: about 1.4 metres of papyrus, the lower parts of some 22 columns of Greek, likely a Stoic ethical treatise.", "before_after": { "metric": "Extent read of a rolled Herculaneum papyrus", "before": "isolated words and patches", "after": "all surviving text, about 22 partial columns" }, "why_it_matters": "Moves scroll reading from demonstrations to a repeatable method; the same work identified another roll as Philodemus, On Gods, Book 8, pointing to systematic recovery of the unopened library.", "ai_systems": [ "Machine-learning ink detection" ], "organizations": [ "Vesuvius Challenge", "University of Kentucky", "European Synchrotron Radiation Facility (ESRF)" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "arXiv 2606.29085, 2026-06-27", "url": "https://arxiv.org/abs/2606.29085" }, { "label": "Vesuvius Challenge announcement, 2026-06-25", "url": "https://scrollprize.org/firstscroll" } ], "foundation0": false }, { "id": "cycle-double-cover-conjecture-proof", "date": "2026-07-10", "date_basis": "announcement", "title": "OpenAI model proves the 50-year-old cycle double cover conjecture", "domain": "mathematics", "subfield": "graph theory", "summary": "OpenAI released a three-page proof, credited entirely to GPT-5.6 Sol Ultra, that every bridgeless graph has a family of cycles covering each edge exactly twice, a 1970s conjecture of Tutte, Itai and Rodeh, Szekeres and Seymour. Sang-il Oum and Jim Geelen wrote expositions, Johannes Carmesin confirmed it, and Lean formalizations exist.", "before_after": null, "why_it_matters": "One of graph theory's best-known open problems fell to a short argument built on classical flow theorems, suggesting some famous conjectures are hard to spot rather than deep.", "ai_systems": [ "GPT-5.6 Sol Ultra" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Correctness is not disputed. Thomas Bloom noted that using nowhere-zero F₂³-flows on cubic graphs to study cycle double covers goes back at least to Bermond, Jackson and Jaeger (1983).", "significance": 4, "sources": [ { "label": "OpenAI, A proof of the cycle double cover conjecture (PDF, 2026-07-10)", "url": "https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_proof.pdf" }, { "label": "Sang-il Oum, exposition, arXiv:2607.16356", "url": "https://arxiv.org/abs/2607.16356" }, { "label": "Johannes Carmesin, The cycle double cover theorem, The Matroid Union (2026-07-21)", "url": "https://matroidunion.org/?p=6255" } ], "foundation0": false }, { "id": "jacobian-conjecture-counterexample", "date": "2026-07-19", "date_basis": "announcement", "title": "Claude Fable 5 finds a counterexample to the 1939 Jacobian conjecture", "domain": "mathematics", "subfield": "algebraic geometry / polynomial maps", "summary": "Anthropic mathematician Levent Alpöge posted a degree-7 polynomial map in three variables, credited to Claude Fable 5, whose Jacobian determinant is the constant −2 yet which is generically three-to-one, so it has no inverse. This disproves the Jacobian conjecture in every dimension above two; the two-variable case remains open.", "before_after": null, "why_it_matters": "An 87-year-old conjecture, the target of many failed proofs, was settled by a machine-found example that anyone can check with computer algebra in seconds.", "ai_systems": [ "Claude Fable 5" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "independently-replicated", "contested": false, "contested_note": "The counterexample is not disputed, but how it was found has not been disclosed. Adrian Vasiu said an AI-written exposition borrowed notation and ideas from an unpublished Borisov–Gabber–Vasiu draft; Alpöge later said their method deserves credit too (Wikipedia).", "significance": 5, "sources": [ { "label": "Terence Tao, A digestion of the Jacobian conjecture counterexample (2026-07-21)", "url": "https://terrytao.wordpress.com/2026/07/21/a-digestion-of-the-jacobian-conjecture-counterexample/" }, { "label": "The Conversation, 'hello there the jacobian conjecture is false thanx' (2026-07-22)", "url": "https://theconversation.com/hello-there-the-jacobian-conjecture-is-false-thanx-why-a-tiny-social-media-post-has-mathematicians-rethinking-ai-283883" }, { "label": "Wikipedia, Jacobian conjecture", "url": "https://en.wikipedia.org/wiki/Jacobian_conjecture" } ], "foundation0": false }, { "id": "retina4ird-inherited-retinal-disease-rct", "date": "2026-07-24", "date_basis": "publication", "title": "AI helps specialists predict genetic causes of inherited blindness in trial", "domain": "medicine", "subfield": "ophthalmology, genetic diagnosis", "summary": "Retina4IRD predicts 17 genotype categories of inherited retinal disease from retinal photographs and scans. In a multicentre trial randomizing 300 patients with suspected disease, specialists using it named the gene group confirmed by sequencing within their top five guesses for 88.5% of patients, versus 67.3% for specialists alone; top-one accuracy rose from 22.4% to 37.8%.", "before_after": { "metric": "Top-5 genetic diagnosis accuracy", "before": "67.3% (specialists alone)", "after": "88.5% (specialists with Retina4IRD)" }, "why_it_matters": "Inherited retinal diseases need gene-specific care and trials, but genetic testing is slow and scarce. Image-based AI can point specialists to likely genes earlier and guide which tests to order.", "ai_systems": [ "Retina4IRD (Vision Transformer pretrained with RETFound)" ], "organizations": [ "Shanghai General Hospital, Shanghai Jiao Tong University", "Beijing Tsinghua Changgung Hospital", "Severance Hospital, Yonsei University", "Nanjing University of Aeronautics and Astronautics" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2026-07-24", "url": "https://doi.org/10.1038/s41591-026-04545-w" } ], "foundation0": false }, { "id": "claude-mythos-hawk-aes-cryptanalysis", "date": "2026-07-28", "date_basis": "announcement", "title": "Claude Mythos Preview finds new attacks on HAWK signatures and 7-round AES", "domain": "computing", "subfield": "cryptanalysis", "summary": "Anthropic reported two attacks found largely by Claude Mythos Preview. A lattice symmetry cuts expected key recovery for HAWK-256, challenge parameters of a NIST post-quantum signature candidate, from 2^64 to 2^38 operations. A new meet-in-the-middle attack on 7-round AES-128 runs 200–800 times faster than the best before. Full AES is unaffected.", "before_after": { "metric": "Expected key-recovery cost for HAWK-256", "before": "2^64", "after": "2^38" }, "why_it_matters": "Encryption standards rest on years of expert attack. An AI weakening a post-quantum candidate in about 60 hours of work shows such review can be greatly accelerated before deployment.", "ai_systems": [ "Claude Mythos Preview" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Daniel Apon said the reduction checks out (NIST pqc-forum, 28 July), but no independent end-to-end key recovery is reported. HAWK-512 and HAWK-1024 stay exponential (about 2^108 and 2^182). The HAWK work had occasional human guidance, and Anthropic's validators were not cryptography specialists.", "significance": 3, "sources": [ { "label": "Anthropic, Discovering cryptographic weaknesses with Claude (2026-07-28)", "url": "https://www.anthropic.com/research/discovering-cryptographic-weaknesses" }, { "label": "Implicator.ai, Anthropic posts AI-found HAWK attack to NIST forum (2026-07-29)", "url": "https://www.implicator.ai/anthropic-posts-ai-found-hawk-attack-to-nist-forum-cryptographer-confirms-math/" } ], "foundation0": false }, { "id": "openai-ten-advances-lean", "date": "2026-08-01", "date_basis": "announcement", "title": "OpenAI internal model claims ten advances with Lean-checked proofs", "domain": "mathematics", "subfield": "group theory, geometry, complexity", "summary": "OpenAI released a 249-page manuscript and Lean 4 formalizations of ten results by an unreleased internal model (reported as Astra), including an explicit non-sofic group, a counterexample to Connes's rigidity conjecture, Ehrhart's volume conjecture, better sphere-packing upper bounds, and Erdős problems 146, 180 and 183. None is peer-reviewed.", "before_after": null, "why_it_matters": "If they hold, these settle decades-old questions across algebra, geometry and complexity; Lean certificates let anyone check the logic, though statement fidelity still needs human review.", "ai_systems": [ "OpenAI internal model (reported as Astra)" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "OpenAI, Ten advances manuscript (PDF, updated 2026-08-06)", "url": "https://cdn.openai.com/pdf/ten-proofs-oai.pdf" }, { "label": "openai/ten-proofs (Lean 4)", "url": "https://github.com/openai/ten-proofs" }, { "label": "SiliconANGLE, 2026-08-02", "url": "https://siliconangle.com/2026/08/02/openais-astra-solves-10-long-open-math-problems-publishes-proofs/" } ], "foundation0": false }, { "id": "philosophy-public-affairs-claude-drafted-paper", "date": "2026-08-10", "date_basis": "publication", "title": "Top philosophy journal publishes a paper Claude drafted almost entirely", "domain": "science-of-science", "subfield": "AI authorship", "summary": "Philosophy & Public Affairs published Simon Goldstein's 'Epistocracy and the Commitment Problem'. Goldstein's disclosure says Claude drafted essentially all the prose and formal appendix and made many modelling choices, while he supplied the founding idea, gave several hundred comments and made all final calls. Two weeks later the journal banned submissions substantially written by AI.", "before_after": null, "why_it_matters": "A leading peer-reviewed journal accepted an argument largely written by a machine. Its swift policy reversal shows journals must now decide what authorship and originality mean.", "ai_systems": [ "Claude Opus 4.7", "Claude Fable 5", "Claude Code" ], "organizations": [ "University of Hong Kong", "Philosophy & Public Affairs" ], "ai_role": "collaborator", "evidence": "peer-reviewed", "contested": false, "contested_note": "The division of labour is the author's own account. The journal's associate editor argued AI-written work can look sophisticated while being insubstantial (Daily Nous, 2026-08-24).", "significance": 2, "sources": [ { "label": "Goldstein, Philosophy & Public Affairs, 2026-08-10", "url": "https://doi.org/10.1111/papa.70032" }, { "label": "Goldstein, How AI was used in writing this paper", "url": "https://simondgoldstein.com/epistocracy-ai-report.pdf" }, { "label": "Daily Nous, 2026-08-24", "url": "https://dailynous.com/2026/08/24/after-experiment-journal-decides-to-prohibit-ai-authored-content/" } ], "foundation0": false }, { "id": "riemann-zeta-zeros-critical-line-67-percent", "date": "2026-08-10", "date_basis": "announcement", "title": "Claude proves over two thirds of zeta zeros are simple and on the critical line", "domain": "mathematics", "subfield": "analytic number theory", "summary": "Asked by a non-mathematician to attack the Riemann hypothesis, an unreleased Claude model running about 60 subagents proved that over two thirds (67.25%) of zeta's nontrivial zeros are simple and on the critical line, up from 5/12. Two Anthropic mathematicians checked it, a Lean proof followed, and Youness Lamzouri found a simpler one.", "before_after": { "metric": "Proven share of zeta zeros that are simple and on the critical line", "before": "5/12 (about 41.7%)", "after": "over 2/3 (67.25%)" }, "why_it_matters": "The share of zeros proven on the critical line is a standard yardstick of progress toward the Riemann hypothesis; this moves it past two thirds, though it does not lead to a proof.", "ai_systems": [ "Claude (unreleased research model)" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Not yet peer-reviewed. The argument builds on unconditional pair-correlation results of Aryan and of Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh; Lamzouri calls Claude's version intricate, with a mechanism that is not immediately transparent.", "significance": 4, "sources": [ { "label": "Anthropic, Claude improves a lower bound on zeta zeros on the critical line (2026-08-10)", "url": "https://www.anthropic.com/research/riemann-zeta" }, { "label": "Alpöge and Furman, More than two thirds of the zeta zeros are simple and on the critical line, arXiv:2608.13637", "url": "https://arxiv.org/abs/2608.13637" }, { "label": "Lamzouri, A new proof that more than 2/3 of the zeros are simple and on the critical line, arXiv:2609.02882", "url": "https://arxiv.org/abs/2609.02882" } ], "foundation0": false }, { "id": "icml-2026-open-reproductions-agents", "date": "2026-08-13", "date_basis": "announcement", "title": "Agent-driven hackathon re-tests a third of ICML 2026's accepted papers", "domain": "science-of-science", "subfield": "reproducibility", "summary": "In 19 days, 1,221 people using coding agents published 6,816 logbooks on 2,226 ICML 2026 papers, 34% of the conference, with 35,908 claims judged. 51% of examined papers had at least one claim verified and 266 were fully reproduced; 23% had a claim falsified or contested, and 242 drew opposite verdicts from different teams.", "before_after": null, "why_it_matters": "Conference output has outgrown careful review. If agents can re-run experiments at scale, checking published results could become routine rather than rare.", "ai_systems": [ "Claude Code", "Codex", "Cursor", "GLM-5.2 (automated logbook judge)" ], "organizations": [ "Hugging Face", "alphaXiv" ], "ai_role": "instrument", "evidence": "claimed", "contested": false, "contested_note": "Verdicts come from an AI judge reading agent logbooks; 502 papers had only toy-scale evidence, and some claimed falsifications were reproduction bugs, so organisers re-checked them (Hugging Face blog).", "significance": 2, "sources": [ { "label": "Hugging Face blog, 2026-08-13", "url": "https://huggingface.co/blog/icml-2026-open-reproductions" } ], "foundation0": false }, { "id": "alphaevolve-matrix-exponent-2-371177", "date": "2026-08-17", "date_basis": "preprint", "title": "AlphaEvolve helps lower the matrix multiplication exponent to ω < 2.371177", "domain": "mathematics", "subfield": "algorithms / algebraic complexity", "summary": "Google DeepMind researchers, with Josh Alman and Virginia Vassilevska Williams, reformulated the optimization problem inside the laser method's 'combination loss analysis', solved it with a new gradient-based optimizer and refined that solver with AlphaEvolve. The result lowers the best upper bound on the matrix multiplication exponent ω from 2.371339 to 2.371177.", "before_after": { "metric": "Upper bound on the matrix multiplication exponent ω", "before": "2.371339", "after": "2.371177" }, "why_it_matters": "ω sets the asymptotic cost of multiplying matrices and of the many algorithms that reduce to it; each small improvement has taken years of human work.", "ai_systems": [ "AlphaEvolve" ], "organizations": [ "Google DeepMind", "Carnegie Mellon University", "Columbia University", "MIT" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "About 0.97×10⁻⁴ of the 1.6×10⁻⁴ gain came from the reformulation and gradient-based optimizer before AlphaEvolve was applied.", "significance": 2, "sources": [ { "label": "Dupont et al., Improving the matrix multiplication exponent with modern optimization and AlphaEvolve, arXiv:2608.16884", "url": "https://arxiv.org/abs/2608.16884" } ], "foundation0": false }, { "id": "axiomprover-prime-gaps-246-lean", "date": "2026-08-17", "date_basis": "announcement", "title": "AxiomProver formalizes in Lean that prime gaps of at most 246 recur forever", "domain": "mathematics", "subfield": "analytic number theory / formal proof", "summary": "Axiom Math released a Lean 4 proof, produced by its AxiomProver system from a human-written blueprint and curated by its team, that infinitely many consecutive primes differ by at most 246. It covers Maynard's sieve and the Polymath8b refinement that cut the bound from 600 to 246, assuming only the Bombieri–Vinogradov theorem.", "before_after": null, "why_it_matters": "246 is the best proven step toward the twin prime conjecture; a reusable machine-checked library of sieve methods lets later work on prime gaps build on verified foundations.", "ai_systems": [ "AxiomProver" ], "organizations": [ "Axiom Math" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "This formalizes known results rather than proving new ones; the Bombieri–Vinogradov theorem is assumed rather than formalized, and people wrote the blueprint.", "significance": 3, "sources": [ { "label": "Axiom Math, Lean formalization of bounded gaps between primes", "url": "https://primegaps.axiommath.ai/" }, { "label": "IEEE Spectrum, AI used to verify toughest mathematics proof yet (2026-08-17)", "url": "https://spectrum.ieee.org/axiom-math-246-theorem-formalization" }, { "label": "Axiom Math, blueprint (PDF)", "url": "https://primegaps.axiommath.ai/blueprint.pdf" } ], "foundation0": false }, { "id": "fda-aletta-autonomous-blood-draw-robot", "date": "2026-08-19", "date_basis": "announcement", "title": "FDA authorizes the first robot that draws blood without hands-on operation", "domain": "engineering-robotics", "subfield": "medical robotics", "summary": "The FDA granted De Novo authorization to Vitestro's Aletta, which locates a vein with near-infrared imaging and Doppler ultrasound, guided by what the company calls AI, then draws blood autonomously. In a 1,633-patient trial, first-stick success was 94.5% when it found a suitable vein; adverse events, all mild, occurred in 0.6%.", "before_after": { "metric": "First-stick success in routine outpatient use", "before": "manual draws by trained phlebotomists", "after": "94.5% (95% CI 93.3-95.5) when a suitable vein was found" }, "why_it_matters": "Blood draws are among the most common medical procedures, and phlebotomists are scarce. One supervisor can oversee three devices, so autonomy here could shorten waits for basic diagnostics.", "ai_systems": [ "Aletta autonomous robotic phlebotomy device (imaging-based vein selection)" ], "organizations": [ "Vitestro", "OLVG", "Amsterdam UMC", "St. Antonius Hospital" ], "ai_role": "originator", "evidence": "peer-reviewed", "contested": false, "contested_note": "Authorized only for adults in outpatient settings under a trained supervisor. The trial's lead authors are Vitestro staff, and the FDA release does not describe the device as AI.", "significance": 3, "sources": [ { "label": "FDA announcement, 2026-08-19", "url": "https://content.govdelivery.com/accounts/USFDA/bulletins/425bb55" }, { "label": "Giesen et al. (ADOPT trial), Clinical Chemistry 72:845, 2026", "url": "https://doi.org/10.1093/clinchem/hvag029" }, { "label": "Vitestro press release, 2026-08-20", "url": "https://vitestro.com/news/vitestro-receives-fda-de-novo-authorization-for-aletta/" } ], "foundation0": false }, { "id": "redwood-ai-designed-ai-accelerator", "date": "2026-08-26", "date_basis": "preprint", "title": "Startup says its AI system designed and verified an AI accelerator in two weeks", "domain": "engineering-robotics", "subfield": "chip design", "summary": "Architect Labs reports that, from two architects' high-level specification, its system generated the performance model, RTL, test benches, formal proofs, firmware and kernels for the Redwood inference accelerator in under two weeks, with no human work below the specification. An FPGA version runs Qwen3-0.6B at 12.1 tokens per second; silicon gains are projected.", "before_after": null, "why_it_matters": "AI workloads change in months while chips take years. Collapsing design and verification into one automated loop could let hardware keep pace with the software it runs.", "ai_systems": [ "Architect Labs hardware-software co-design system" ], "organizations": [ "Architect Labs" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Company preprint on a proprietary system. On FPGA it ran slower than an NVIDIA Jetson Orin Nano (12.1 vs 28 tokens/s); the 3.4x performance-per-watt gain is projected to 8 nm, and no chip has been made.", "significance": 1, "sources": [ { "label": "Architect Labs, arXiv 2608.26418, 2026-08-26", "url": "https://arxiv.org/abs/2608.26418" }, { "label": "The Neuron explainer, 2026-09-21", "url": "https://www.theneuron.ai/explainer-articles/architect-labs-ai-chips-evolve-as-fast-as-ai-models/" } ], "foundation0": false }, { "id": "co-scientist-lab-in-loop-one-take-2d-growth", "date": "2026-08-27", "date_basis": "preprint", "title": "Google's Co-Scientist moves to the bench: first-try growth of 2D semiconductors", "domain": "science-of-science", "subfield": "AI co-scientists", "summary": "Google extended its Gemini-based Co-Scientist to experiments at Duke. It tailored chemical vapour deposition recipes to a custom furnace and grew monolayer MoS2, MoSe2 and WS2 on the first attempt, the last two with no in-house history, verified in at least five replicate runs. It also proposed a safer precursor route to an MXene-like phase.", "before_after": { "metric": "Growth attempts for monolayer MoSe2 and WS2 on a custom furnace", "before": "no in-house protocol or history", "after": "first attempt, verified in at least five replicate runs" }, "why_it_matters": "Most AI co-scientists stop at hypotheses. Turning ideas into recipes that work first time on real equipment connects AI reasoning to physical discovery.", "ai_systems": [ "Co-Scientist (Gemini multi-agent system)", "Gemini 3 Deep Think" ], "organizations": [ "Google DeepMind", "Google Research", "Duke University", "Columbia University", "Texas A&M University" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Humans selected recipes, loaded precursors and ran characterisation; the MXene-like phase's atomic structure is unconfirmed, and transfer to other labs is untested. Not peer reviewed.", "significance": 3, "sources": [ { "label": "Schmidgall et al., arXiv 2608.26701, 2026-08-27", "url": "https://arxiv.org/abs/2608.26701" } ], "foundation0": false }, { "id": "percolation-no-infinite-cluster-at-criticality", "date": "2026-08-28", "date_basis": "preprint", "title": "Claude-written Lean proof: no percolation at the critical point in any dimension", "domain": "mathematics", "subfield": "probability / percolation theory", "summary": "On 28 August 2026 Anthropic's Justin Leder posted to Anthropic's GitHub a Lean proof, written autonomously by Claude models under his direction, that critical bond percolation on Z^d has no infinite cluster for any d ≥ 2, by proving a 2024 gluing conjecture of Gady Kozma and Shahaf Nitzan. Dimensions 3 to 10 were open.", "before_after": { "metric": "Dimensions where θ(p_c) = 0 is proven", "before": "d = 2 and d ≥ 11", "after": "all d ≥ 2" }, "why_it_matters": "Whether three-dimensional percolation has a continuous phase transition was a central open problem of probability theory; one expert said solving it would earn a Fields Medal.", "ai_systems": [ "Claude (Anthropic models)" ], "organizations": [ "Anthropic" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Not refereed. Kozma awaits a human-readable version; Ahmed Bou-Rabee calls it clever but without substantial new ideas (Quanta, 1 Oct). Arthur Blanc-Renaudie posted a short human proof built from the AI's ideas. The files are no longer on the repository's main branch, and Anthropic has made no announcement.", "significance": 4, "sources": [ { "label": "anthropics/formal-math, percolation README (commit 795efb8)", "url": "https://github.com/anthropics/formal-math/blob/795efb86f191735c5481675763537cfb4ff37e55/percolation/README.md" }, { "label": "Blanc-Renaudie, Notes on Kozma and Nitzan's proof of θ(p_c)=0, arXiv:2610.10502", "url": "https://arxiv.org/abs/2610.10502" }, { "label": "Scientific American, AI solves a 'holy grail' problem from probability theory (2026-09-30)", "url": "https://www.scientificamerican.com/article/ai-solves-a-holy-grail-problem-from-probability-theory/" } ], "foundation0": false }, { "id": "state-virtual-cell-perturbation-model", "date": "2026-08-31", "date_basis": "publication", "title": "State model predicts how cells respond to perturbations in unseen contexts", "domain": "biology", "subfield": "virtual cell models", "summary": "Arc Institute's State predicts how genetic, signalling and chemical perturbations change gene expression across cell contexts. Its cell embeddings are trained on observational data from 167 million cells. The authors report more than 30% better discrimination of perturbation effects than baselines on large datasets, and introduce the Cell-Eval evaluation framework.", "before_after": null, "why_it_matters": "A reliable virtual cell would let researchers test perturbations on screen before costly experiments, but independent benchmarks so far show the problem is hard.", "ai_systems": [ "State (State Embedding and State Transition)" ], "organizations": [ "Arc Institute", "University of California, San Francisco", "Stanford University" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": true, "contested_note": "Ahlmann-Eltze et al. (Nature Methods, 2025-08-04) tested earlier deep-learning perturbation models, not State, and found none beat simple linear baselines. State's authors say it is the first to consistently beat such baselines (Arc Institute, 2025-06-23); we found no independent head-to-head test.", "significance": 2, "sources": [ { "label": "Adduri et al., Cell, published online 2026-08-31", "url": "https://doi.org/10.1016/j.cell.2026.07.052" }, { "label": "Arc Institute announcement, 2025-06-23", "url": "https://arcinstitute.org/news/virtual-cell-model-state" }, { "label": "Ahlmann-Eltze et al., Nature Methods, 2025-08-04", "url": "https://doi.org/10.1038/s41592-025-02772-6" } ], "foundation0": false }, { "id": "vara-autonomous-mammography-triage-ce-mark", "date": "2026-09-02", "date_basis": "announcement", "title": "Vara gets CE mark for AI that reports clearly normal mammograms on its own", "domain": "medicine", "subfield": "breast cancer screening, regulation", "summary": "Vara says it received EU MDR Class IIb certification, which it calls a world first, for autonomous triage in organised breast screening: exams its AI judges clearly normal are reported without radiologist reading; all others get at least one human read. A monitoring system reverts a site to full human reading if performance drifts.", "before_after": null, "why_it_matters": "Most screening mammograms are normal and radiologists are scarce. Letting a monitored AI sign off clear normals frees specialists for suspicious cases, provided missed cancers stay rare.", "ai_systems": [ "Vara Autonomous Triage with ATMON monitoring" ], "organizations": [ "Vara (MX Healthcare GmbH)" ], "ai_role": "instrument", "evidence": "claimed", "contested": true, "contested_note": "The Next Web (2026-09-03) notes that the PRAIM study Vara cites tested AI-supported double reading, not autonomous triage, that Vara has not named its notified body, and that its German market-share figures differ across its own documents.", "significance": 2, "sources": [ { "label": "Vara press release (Business Wire via TT), 2026-09-02", "url": "https://via.tt.se/pressmeddelande/4526340/vara-receives-world-first-ce-certification-for-autonomous-ai-in-breast-cancer-screening" }, { "label": "Vara blog post on the certification, 2026-09-02", "url": "https://news.vara.ai/p/behind-the-worlds-first-autonomous" }, { "label": "The Next Web analysis, 2026-09-03", "url": "https://thenextweb.com/news/vara-ce-certification-autonomous-ai-breast-cancer-screening" } ], "foundation0": false }, { "id": "male-fly-cns-connectome-sexual-dimorphism", "date": "2026-09-03", "date_basis": "publication", "title": "Complete male fly nervous system mapped: 166,700 neurons in brain and nerve cord", "domain": "biology", "subfield": "connectomics", "summary": "HHMI Janelia, Google Research, Cambridge and collaborators published the connectome of a whole male fruit-fly central nervous system, brain plus nerve cord: 166,700 fully proofread neurons in 11,710 types, about 125 million synaptic connections (Google figure). AI helped trace neurons and humans proofread. Sex differences concentrate in higher brain centres.", "before_after": null, "why_it_matters": "A complete brain-and-cord map lets scientists follow sensory input to motor output end to end and compare the sexes at synaptic resolution.", "ai_systems": [ "Automated neuron segmentation and synapse detection" ], "organizations": [ "HHMI Janelia Research Campus", "Google Research", "MRC Laboratory of Molecular Biology", "University of Cambridge" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Berg et al., Cell, 2026-09-03", "url": "https://doi.org/10.1016/j.cell.2026.08.015" }, { "label": "Google Research blog, 2026-09-03", "url": "https://research.google/blog/a-connectomics-milestone-mapping-the-complete-male-fruit-fly-brain/" }, { "label": "MRC LMB news, 2026-09-03", "url": "https://mrclmb.ac.uk/news-events/articles/first-complete-connectome-of-male-fly-central-nervous-system-allows-for-unprecedented-male-female-brain-comparison/" } ], "foundation0": false }, { "id": "claude-flt-lean-formalization", "date": "2026-09-04", "date_basis": "announcement", "title": "Claude agents produce first complete Lean proof of Fermat's Last Theorem", "domain": "mathematics", "subfield": "number theory / formalization", "summary": "Dozens of Claude agents, working largely autonomously for 11 days, wrote about 13 million lines of Lean proving Fermat's Last Theorem along the Frey–Serre–Ribet–Wiles–Taylor route, with 29,500 theorems in the final proof. It rests only on Lean's three standard axioms; Kevin Buzzard reviewed it. It formalizes Wiles's proof rather than finding a new one.", "before_after": { "metric": "Machine-checked proof of FLT", "before": "Partial (Imperial College project since 2024)", "after": "Complete, ~13 million lines of Lean" }, "why_it_matters": "One of mathematics' deepest proofs is now checked line by line by computer, suggesting AI can make verifying frontier mathematics routine.", "ai_systems": [ "Claude (internal research model)" ], "organizations": [ "Anthropic", "Columbia University", "Imperial College London" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 4, "sources": [ { "label": "Anthropic, 2026-09-04", "url": "https://www.anthropic.com/research/formalizing-fermats-last-theorem" }, { "label": "anthropics/fermats-last-theorem (Lean)", "url": "https://github.com/anthropics/fermats-last-theorem" }, { "label": "The Next Web, 2026-09-06", "url": "https://thenextweb.com/news/anthropic-claude-fermat-last-theorem-lean-buzzard" } ], "foundation0": false }, { "id": "forced-euler-blowup-lean-buckmaster-alpoge", "date": "2026-09-07", "date_basis": "announcement", "title": "AI-assisted, Lean-checked proofs show forced 3D Euler flows can blow up", "domain": "mathematics", "subfield": "PDE / fluid dynamics", "summary": "Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) posted proofs that the 3D incompressible Euler equations, the 2D Boussinesq system and the porous-medium equation can form finite-time singularities under a smooth external force, extending Córdoba and Martínez-Zoroa's method. Claude and OpenAI's Codex helped heavily; Claude wrote all the Lean, which uses only standard axioms.", "before_after": null, "why_it_matters": "Blow-up for the frictionless Euler equations is a recognized stepping stone toward the Navier–Stokes Millennium problem, and here the whole argument is machine-checked.", "ai_systems": [ "Claude", "OpenAI Codex" ], "organizations": [ "New York University", "Anthropic" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "Not peer-reviewed. Terence Tao called it a remarkable achievement with significant AI input, released earlier than planned because of external events. The force is an added ingredient; the unforced problem stays open.", "significance": 3, "sources": [ { "label": "Science News, AI may have solved one of math's biggest puzzles (2026-09-09)", "url": "https://www.sciencenews.org/article/ai-math-puzzle-navier-stokes-openai" }, { "label": "tristanbuckmaster/fluid_lean, euler-blowup (Lean)", "url": "https://github.com/tristanbuckmaster/fluid_lean/tree/main/euler-blowup" }, { "label": "Terence Tao on Mathstodon (2026-09-08)", "url": "https://mathstodon.xyz/@tao/117233527638291447" } ], "foundation0": false }, { "id": "pinn-unforced-euler-blowup-candidate", "date": "2026-09-07", "date_basis": "preprint", "title": "Neural-network search finds a candidate stable blow-up for unforced 3D Euler", "domain": "mathematics", "subfield": "PDE / fluid dynamics", "summary": "Anima Anandkumar's Caltech group used a physics-informed neural network, steered away from trivial solutions and refined with splines for certified error bounds, to find an approximate self-similar blow-up of the 3D Euler equations in free space, with no boundary or force. It appears numerically stable; a rigorous stability proof is still missing.", "before_after": null, "why_it_matters": "Unforced blow-up in free space is the version fluid experts care most about; a precise, stable candidate is the usual first step toward a computer-assisted proof.", "ai_systems": [ "Physics-informed neural network (PINN)" ], "organizations": [ "Caltech" ], "ai_role": "instrument", "evidence": "preprint", "contested": false, "contested_note": "Terence Tao (Mathstodon, 2026-09-08): a promising candidate that appears numerically stable, but a rigorous demonstration of stability is still lacking and may need enormous effort.", "significance": 2, "sources": [ { "label": "Anima Anandkumar, Stable singularity of the Euler equations on R^3 without forcing (2026-09-07)", "url": "https://anima-ai.org/2026/09/07/stable-singularity-of-the-euler-equations-on-r3-without-forcing/" }, { "label": "Terence Tao on Mathstodon (2026-09-08)", "url": "https://mathstodon.xyz/@tao/117234157753860650" } ], "foundation0": false }, { "id": "openai-navier-stokes-forced-blowup-claim", "date": "2026-09-08", "date_basis": "announcement", "title": "OpenAI claims Lean-checked forced blow-up for Navier–Stokes, a Clay problem", "domain": "mathematics", "subfield": "PDE / fluid dynamics", "summary": "OpenAI said an internal model, running about 10,000 agents at a time, produced a 166-page proof that 3D Navier–Stokes flow starting from rest can reach unbounded velocity in finite time under a smooth, compactly supported force, an option the Clay problem allows. It released a Lean formalization and will not claim the $1 million prize.", "before_after": null, "why_it_matters": "If it stands, it settles a Millennium Prize problem as stated, but only through a forced version experts call unnatural; the unforced question that matters most for fluids stays open.", "ai_systems": [ "Unreleased OpenAI internal model" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "preprint", "contested": true, "contested_note": "Experts in Scientific American (21 Sept) call the forcing a loophole; Constantin, Ignatova and Vicol (arXiv:2609.20803) show such constructions need a non-analytic force at the singularity. Bastounis, Circelli and Hansen (arXiv:2610.08144) argue the Lean code does not match the written proof. Buckmaster disputes credit; OpenAI denies using his work.", "significance": 4, "sources": [ { "label": "OpenAI, Finite Time Blowup for Navier–Stokes (PDF, 166 pp.)", "url": "https://cdn.openai.com/pdf/32d9f210-8b73-45e0-91bc-82a30aef8a9a/navier-stokes.pdf" }, { "label": "Scientific American, Did OpenAI solve the wrong Navier-Stokes problem? (2026-09-21)", "url": "https://www.scientificamerican.com/article/did-openai-solve-the-wrong-navier-stokes-problem/" }, { "label": "Bastounis, Circelli, Hansen, Navier-Stokes lost in translation, arXiv:2610.08144", "url": "https://arxiv.org/abs/2610.08144" } ], "foundation0": false }, { "id": "paics-fetal-brain-malformation-ultrasound-rct", "date": "2026-09-10", "date_basis": "publication", "title": "AI assistance raises detection of fetal brain malformations on ultrasound", "domain": "medicine", "subfield": "obstetrics, prenatal imaging", "summary": "At five Chinese centres, sonographers with 3-8 years' experience scanned high-risk pregnancies with and without the PAICS AI system in a randomized crossover design (1,584 scans). AI assistance raised sensitivity for ten targeted brain malformations by 8.7 percentage points per fetus (11.8 per malformation) while specificity stayed non-inferior.", "before_after": null, "why_it_matters": "Missed fetal anomalies deny families timely counselling and delivery planning. AI that lifts mid-career sonographers' detection without extra false alarms could raise screening quality where experts are scarce.", "ai_systems": [ "PAICS (Prenatal Ultrasound Diagnosis Artificial Intelligence Conduct System)" ], "organizations": [ "First Affiliated Hospital of Sun Yat-sen University", "Guangdong Maternal and Child Health Hospital", "Sun Yat-sen University School of Biomedical Engineering", "Guangzhou Aiyunji Information Technology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "The Lancet Digital Health, 2026-09-10", "url": "https://doi.org/10.1016/j.landig.2026.101040" } ], "foundation0": false }, { "id": "eagle-esophageal-cancer-noncontrast-ct", "date": "2026-09-22", "date_basis": "publication", "title": "AI spots oesophageal cancer on routine non-contrast chest CT scans", "domain": "medicine", "subfield": "oncology, cancer screening", "summary": "EAGLE detects oesophageal cancer and precancer on non-contrast chest CT, a task long thought impossible. Across 80,612 patients at 12 centres in three countries, it reached 90.0% cancer sensitivity at 98.5% specificity in external tests, a 42.2% positive predictive value in 17,446 prospective hospital patients, and 99.94% specificity in real-world low-dose CT screening.", "before_after": null, "why_it_matters": "Oesophageal cancer is usually found late, and endoscopy cannot screen whole populations. Reusing chest CTs already taken, including lung-screening scans, could find early cancers at almost no extra cost.", "ai_systems": [ "EAGLE (Esophageal AI-Guided malignant Lesion Evaluation)" ], "organizations": [ "Sun Yat-sen University Cancer Center", "Sichuan Cancer Hospital & Institute", "DAMO Academy, Alibaba Group" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "Nature Medicine, 2026-09-22", "url": "https://doi.org/10.1038/s41591-026-04656-4" } ], "foundation0": false }, { "id": "arda-diabetic-retinopathy-ai-million-patients", "date": "2026-09-23", "date_basis": "publication", "title": "Google's diabetic eye-screening AI passes one million patients screened", "domain": "medicine", "subfield": "ophthalmology, global health", "summary": "Google's diabetic retinopathy model, licensed as ARDA, has now screened more than one million patients across settings in India, Thailand and Australia, its developers and clinical partners report, up from 600,000 supported screenings in October 2024. Partners aim for six million free AI-supported screenings in India and Thailand within a decade.", "before_after": { "metric": "Routine screening use of the AI model", "before": ">600,000 screenings supported (October 2024)", "after": ">1 million patients screened (September 2026)" }, "why_it_matters": "Diabetic eye disease causes preventable blindness where eye specialists are scarce. Scaling a validated model from one hospital to a million patients shows AI screening can work in routine low-resource care.", "ai_systems": [ "ARDA (Automated Retinal Disease Assessment)" ], "organizations": [ "Google", "Aravind Eye Care System", "Rajavithi Hospital", "Lions Outback Vision" ], "ai_role": "instrument", "evidence": "deployed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Nature Medicine Comment, 2026-09-23", "url": "https://doi.org/10.1038/s41591-026-04643-9" }, { "label": "Google blog, 2024-10-17 (600,000 screenings)", "url": "https://blog.google/company-news/inside-google/around-the-globe/google-asia/arda-diabetic-retinopathy-india-thailand/" } ], "foundation0": false }, { "id": "lila-ai-palladium-oxide-oer-catalysts", "date": "2026-09-24", "date_basis": "preprint", "title": "AI-guided lab finds iridium-free catalysts for acidic water splitting", "domain": "chemistry-materials", "subfield": "electrocatalysis", "summary": "Lila Sciences' human-supervised closed-loop platform, over 90% automated, combined sputter synthesis, high-throughput screening, machine-learning models and language-model reasoning to test 2,942 catalysts across 53 material systems. It surfaced palladium oxides such as InMnPdOx, which kept overpotential below 0.5 V for 1,000 hours in acid; plain PdOx crossed that line at about 200 hours.", "before_after": { "metric": "Hours below 0.5 V overpotential (1 M H2SO4, 10 mA/cm2)", "before": "~200 h (PdOx)", "after": "over 1,000 h (InMnPdOx)" }, "why_it_matters": "Proton-exchange-membrane electrolyzers depend on scarce iridium. Durable catalysts based on more available palladium could ease a supply limit on green hydrogen at gigawatt scale.", "ai_systems": [ "Sequential-learning agent with multi-objective optimization", "Large-language-model reasoning" ], "organizations": [ "Lila Sciences" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Company preprint, not peer reviewed; all authors are Lila employees and patents may be filed. Durability was measured at 10 mA/cm2 in a three-electrode acid cell, not an electrolyzer, and palladium is itself a precious metal.", "significance": 2, "sources": [ { "label": "Jenewein et al., arXiv:2609.30133, 2026-09-24", "url": "https://arxiv.org/abs/2609.30133" }, { "label": "Unite.AI, 2026-09-25", "url": "https://www.unite.ai/lila-sciences-ai-lab-uncovers-palladium-catalysts-for-green-hydrogen/" } ], "foundation0": false }, { "id": "claude-nine-loop-n4-sym-hexagon", "date": "2026-09-25", "date_basis": "announcement", "title": "Claude computes a nine-loop six-gluon amplitude, one loop past the record", "domain": "physics", "subfield": "theoretical particle physics (scattering amplitudes)", "summary": "Answering physicist Matt von Hippel's public challenge, two Anthropic physicists had Claude (Fable 5.1 in Claude Science) compute the six-gluon MHV amplitude of planar N=4 super-Yang-Mills at nine loops, mostly by telling it to keep going. It succeeded by two routes, each costing roughly $1,000-2,000; SLAC's Lance Dixon validated the result.", "before_after": { "metric": "Highest loop order computed for the six-gluon amplitude in planar N=4 SYM", "before": "eight loops (Dixon and Liu, 2023)", "after": "nine loops" }, "why_it_matters": "Higher-loop amplitudes test the methods physicists use for precise collider predictions. An AI ran a frontier-scale calculation with minimal guidance on an academic-sized budget.", "ai_systems": [ "Claude Fable 5.1 (Claude Science harness)" ], "organizations": [ "Anthropic", "SLAC National Accelerator Laboratory" ], "ai_role": "originator", "evidence": "claimed", "contested": false, "contested_note": "Not peer reviewed; reported in an Anthropic-commissioned guest post. Dixon, who validated it, received Claude usage credits. Song He's group (CAS) independently computed the nine-loop symbol, partly with GPT-6, and posted it on 2026-09-17. N=4 SYM is a toy model.", "significance": 3, "sources": [ { "label": "Anthropic guest post by M. von Hippel, with L. Dixon, 2026-09-25", "url": "https://www.anthropic.com/research/yes-claude-can-do-nine-loops" }, { "label": "Cosmic9 nine-loop result files (S. Mishra-Sharma)", "url": "https://smsharma.io/cosmic-nine-loops/" }, { "label": "He, Jing & Li, nine-loop MHV symbols (Zenodo), 2026-09-17", "url": "https://doi.org/10.5281/zenodo.22800071" } ], "foundation0": false }, { "id": "nber-llm-reproduces-4452-economics-papers", "date": "2026-09-28", "date_basis": "preprint", "title": "LLM workflow re-runs 4,452 economics papers, speeding up 496 by over 10x", "domain": "science-of-science", "subfield": "reproducibility", "summary": "Matthew Schwartz, Isaiah Andrews and Jesse Shapiro built an open-source workflow in which an LLM ports and re-runs replication packages. Across 4,452 packages from five top economics journals it flagged discrepancies in 3,460 articles or appendices, cut a calculation's run time over tenfold at similar or better accuracy in 496, and extended 923.", "before_after": { "metric": "Economics replication packages re-run and checked", "before": "checked one at a time by journal data editors", "after": "4,452 packages across five journals" }, "why_it_matters": "Economics journals require replication packages, but checking them is slow manual work. Doing it across thousands of papers makes errors and fragile results findable at field scale.", "ai_systems": [ "Open-source multi-agent LLM workflow (Claude, per Schwartz's Anthropic post)" ], "organizations": [ "Harvard University", "MIT", "National Bureau of Economic Research" ], "ai_role": "originator", "evidence": "preprint", "contested": false, "contested_note": "Not peer reviewed. A flagged discrepancy is not an error: 35.6% of calculation discrepancies are within rounding of the last printed digit. Schwartz worked as a contractor for Anthropic (NBER disclosure).", "significance": 3, "sources": [ { "label": "Schwartz, Andrews and Shapiro, NBER Working Paper 35782, 2026-09", "url": "https://www.nber.org/papers/w35782" }, { "label": "Schwartz, Anthropic guest post, 2026-10-01", "url": "https://www.anthropic.com/research/claude-shaped-science" } ], "foundation0": false }, { "id": "bootloops-claude-36-manuscripts-18-fields", "date": "2026-10-01", "date_basis": "announcement", "title": "Physicist and Claude report 36 manuscripts in 18 fields in three months", "domain": "science-of-science", "subfield": "human-AI research collaboration", "summary": "Harvard physicist Matthew Schwartz built BootLoops, an open-source harness for exact quantitative science, and steered Claude through it. With 19 coauthors he reports 36 manuscripts in 18 fields in three months, including 30 Feynman integrals computed end to end, 15 of them new, and the last of the lattice integrals George Watson began in 1939.", "before_after": null, "why_it_matters": "Many fields hold problems that a known technique from another field would solve. An AI with broad knowledge, steered by experts, can match methods to problems faster than any single researcher.", "ai_systems": [ "Claude (including Claude Fable 5)", "BootLoops harness" ], "organizations": [ "Harvard University", "Anthropic" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "Mostly not peer reviewed; the project site lists 28 manuscripts and 22 fields. Schwartz, an Anthropic visiting researcher, writes that Claude's results were often correct but unremarkable until experts redirected them.", "significance": 2, "sources": [ { "label": "Schwartz, Anthropic guest post, 2026-10-01", "url": "https://www.anthropic.com/research/claude-shaped-science" }, { "label": "BootLoops project site", "url": "https://www.bootloops.ai" } ], "foundation0": false }, { "id": "3sum-apsp-hypotheses-refuted", "date": "2026-10-05", "date_basis": "preprint", "title": "Claude-found algorithm refutes the 3SUM and APSP hypotheses", "domain": "computing", "subfield": "algorithms / fine-grained complexity", "summary": "Asked to check cryptographic constructions, an internal Claude model instead discovered, in a 16-million-token session with no human input, an algorithm refuting the 3SUM, APSP and Exact Triangle hypotheses. Josh Alman and Virginia Vassilevska Williams simplified and extended it: deterministic 3SUM in O(n^1.9992) and APSP in O(n^2.9995) time. Anthropic Lean-checked the main theorems.", "before_after": { "metric": "Time exponent for 3SUM on polynomial-size integers", "before": "2 (no polynomial improvement known)", "after": "1.9992" }, "why_it_matters": "Many conditional lower bounds in fine-grained complexity assumed 3SUM and APSP need essentially quadratic and cubic time. Refuting both reshapes the field and speeds up many reducible problems at once.", "ai_systems": [ "Claude (internal research model)" ], "organizations": [ "Anthropic", "Columbia University", "MIT" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Lean covers Theorems 19 and 22 and the zero-weight case of Corollary 39; real-input and some other results are not formalized. Not yet peer-reviewed.", "significance": 5, "sources": [ { "label": "Alman and Vassilevska Williams, Truly Subquadratic 3SUM and Truly Subcubic APSP, arXiv:2610.06783", "url": "https://arxiv.org/abs/2610.06783" }, { "label": "anthropics/formal-math, 3sum-apsp (Lean)", "url": "https://github.com/anthropics/formal-math/tree/main/3sum-apsp" }, { "label": "Big O Speedrun, 3SUM leaderboard", "url": "https://bigospeedrun.com/3sum" } ], "foundation0": false }, { "id": "integer-mult-bounds-community", "date": "2026-10-06", "date_basis": "preprint", "title": "Humans and AI agents sharpen the AI-claimed integer multiplication exponent", "domain": "mathematics", "subfield": "algorithms / computational complexity", "summary": "OpenAI's manuscript (dated 23 September, released 6 October 2026) claimed n-bit integer multiplication in O(n(log n)^(1−κ)) time on a multitape Turing machine with κ = 2^−182. In an open repository maintained by founding member Doug Colkitt, contributors and their AI agents raised κ to about 0.00167 (≈2^−9.2), still conditional on that unreviewed framework.", "before_after": { "metric": "Exponent saving κ in O(n(log n)^(1−κ))", "before": "2^−182 (OpenAI manuscript)", "after": "≈0.00167 (≈2^−9.2), conditional" }, "why_it_matters": "If the framework holds, it beats the n log n bound conjectured optimal since 1971, and shows people and AI agents openly compounding a machine-generated idea within days.", "ai_systems": [ "Unreleased OpenAI internal model", "Contributors' AI coding agents" ], "organizations": [ "OpenAI", "integer-mult-bounds community (CrocSwap), run by founding member Doug Colkitt", "rohanarun (founding member; reviewed record, PR #1094)", "Aska Digital (built the race tracker)" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "CrocSwap/integer-mult-bounds", "url": "https://github.com/CrocSwap/integer-mult-bounds" }, { "label": "OpenAI, Integer multiplication below n log n (openai/math, 2026-10-06)", "url": "https://github.com/openai/math/tree/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Integer-multiplication-below-n-log-n-September-23-2026" } ], "foundation0": true, "foundation0_note": "Founding member Doug Colkitt (GitHub: Mister-Meeseeks) runs this open race in the CrocSwap repository. Founding member rohanarun's PR #1094 was the reviewed κ record on 13–14 Oct 2026; eumemic and Ryan Shea also contribute. Aska Digital built the live tracker that grew into this site.", "foundation0_members": [ "Mister-Meeseeks", "rohanarun", "eumemic", "shea256", "ahrazzle" ], "page": "research/integer-multiplication/" }, { "id": "matrix-multiplication-omega-nine-fourths-lean", "date": "2026-10-06", "date_basis": "announcement", "title": "Machine-written proof that ω ≤ 9/4 passes independent Lean checks", "domain": "mathematics", "subfield": "algorithms / algebraic complexity", "summary": "OpenAI's 6 October release included an internal-model proof, dated 2 October, that n×n complex matrices can be multiplied with O(n^(9/4+ε)) arithmetic operations, far below August's 2.371177. The Big O Speedrun tracker rebuilt OpenAI's Lean proof from source and kernel-checked it against its exact problem statement.", "before_after": { "metric": "Upper bound on the matrix multiplication exponent ω over ℂ", "before": "2.371177 (August 2026)", "after": "2.25 (Lean-checked, not peer-reviewed)" }, "why_it_matters": "It would be far larger than any improvement since 1990. A kernel-checked proof makes the claim hard to dismiss, though experts have yet to review the mathematics.", "ai_systems": [ "Unreleased OpenAI internal model" ], "organizations": [ "OpenAI", "Big O Speedrun (verification)" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "No published expert review as of 2026-10-14; Scientific American (8 Oct) reported no independent check of the mathematics. The result is asymptotic and gives no faster practical algorithm.", "significance": 4, "sources": [ { "label": "OpenAI, An Upper Bound of 9/4 for the Matrix Multiplication Exponent (openai/math @ fd4aeeb)", "url": "https://github.com/openai/math/blob/fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb/preprints/Matrix-Multiplication-Nine-Fourths-October-2-2026/paper.pdf" }, { "label": "Big O Speedrun, Matrix Multiplication leaderboard", "url": "https://bigospeedrun.com/matrix-multiplication" }, { "label": "Scientific American, The most exciting claims from OpenAI's heap of new proofs (2026-10-08)", "url": "https://www.scientificamerican.com/article/the-most-exciting-claims-from-openais-heap-of-new-proofs/" } ], "foundation0": false }, { "id": "openai-722-math-manuscripts", "date": "2026-10-06", "date_basis": "announcement", "title": "OpenAI releases 722 machine-written mathematics manuscripts", "domain": "mathematics", "subfield": "multiple fields", "summary": "OpenAI posted 722 manuscripts in 372 result families from an unreleased internal model, using about three hours of compute per result across roughly 4,000 posed problems. Claims include the Unique Games Conjecture, Hodge for CM abelian varieties and integer multiplication below n log n. None is peer-reviewed; about 42% of top-line results have Lean proofs.", "before_after": null, "why_it_matters": "Research-level mathematics at industrial scale shifts the bottleneck from finding proofs to checking them; the corpus is a public test of how much machine mathematics survives.", "ai_systems": [ "Unreleased OpenAI internal model" ], "organizations": [ "OpenAI" ], "ai_role": "originator", "evidence": "preprint", "contested": true, "contested_note": "On 2026-10-07 OpenAI withdrew three manuscripts after a sign error broke a Weil-class argument, and revised 14 others (openai/math history.md). The repository itself warns unformalized results could have issues; ASKA's 2026-10-07 review found most of the corpus unverified.", "significance": 4, "sources": [ { "label": "openai/math repository", "url": "https://github.com/openai/math" }, { "label": "openai/math history (2026-10-07 withdrawals)", "url": "https://raw.githubusercontent.com/openai/math/main/history.md" }, { "label": "ASKA review, 2026-10-07", "url": "https://digital.askaconsult.com/blog/math-faster.html" } ], "foundation0": false, "page": "research/machine-mathematics/" }, { "id": "openai-quasi-riemann-hypothesis-7-8", "date": "2026-10-06", "date_basis": "announcement", "title": "OpenAI model's Lean proof: zeta has no zeros with real part above 7/8", "domain": "mathematics", "subfield": "analytic number theory", "summary": "A manuscript in OpenAI's 6 October release, dated 30 September, claims every Dirichlet L-function, the Riemann zeta function included, has no zeros with real part above 7/8, the 'quasi-Riemann hypothesis'. Its Lean proof of the zeta and Dirichlet statements was independently rebuilt with no errors. Contributors' agents have since posted claimed refinements to about 0.87486.", "before_after": { "metric": "Zero-free half-plane Re s > θ for the Riemann zeta function", "before": "none known (only regions shrinking toward Re s = 1)", "after": "θ = 7/8 (Lean-checked); about 0.87486 claimed" }, "why_it_matters": "No zero-free half-plane for zeta of any width was known; one gives a power-saving error term when counting primes. It is not the Riemann hypothesis, which needs 1/2.", "ai_systems": [ "Unreleased OpenAI internal model", "Contributors' AI agents (refinements)" ], "organizations": [ "OpenAI", "Big O Speedrun (tracking)" ], "ai_role": "originator", "evidence": "formally-verified", "contested": false, "contested_note": "Not peer-reviewed. OpenAI says its zeta work did not follow its standard procedure and a companion write-up was human-edited. The Hecke L-function part uses custom Lean definitions and was not independently rebuilt; the refinements are unchecked. Hector Pasten told Scientific American some see it as progress toward Riemann, others not.", "significance": 5, "sources": [ { "label": "OpenAI, The Quasi-Riemann Hypothesis: A Zero-Free Half-Plane Re s > 7/8 (openai/math @ adc7f12)", "url": "https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/The-Quasi-Riemann-Hypothesis-September-30-2026/paper.pdf" }, { "label": "tomoto0, independent verification of the 7/8 Lean proof", "url": "https://github.com/tomoto0/quasi-riemann-hypothesis-7-8-verification" }, { "label": "Big O Speedrun, Quasi-Riemann hypothesis board", "url": "https://bigospeedrun.com/other/quasi-riemann" } ], "foundation0": false }, { "id": "eumemic-tess-faint-transit-candidates", "date": "2026-10-07", "date_basis": "announcement", "title": "TESS search built with Claude Code finds hundreds of faint planet candidates", "domain": "astronomy-space", "subfield": "exoplanets", "summary": "An open search stacking all available TESS sectors for 4.04 million FGKM dwarfs, plus nearby K/M dwarfs and TOI hosts, finds transits too weak to see in any one sector. On 14 October 2026 it counts 665 candidates (607 full-sky, 58 earlier), up from 533 the day before, including a possible third planet around TOI-4342. None is confirmed.", "before_after": { "metric": "Counted planet candidates (changes with each update)", "before": "12 (first release, 2026-10-07)", "after": "665 (2026-10-14, commit 58c6318)" }, "why_it_matters": "It shows a small independent team, working with an AI coding agent, can mine public mission data at full scale and release reproducible candidates for others to check.", "ai_systems": [ "Claude Code (Anthropic)", "gradient-boosted false-alarm classifier" ], "organizations": [ "eumemic (independent)", "Aska Digital (website and pipeline fixes)" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "eumemic/exoplanets repository README, updated 2026-10-14", "url": "https://github.com/eumemic/exoplanets" }, { "label": "README at commit 58c6318, 2026-10-14", "url": "https://github.com/eumemic/exoplanets/blob/58c6318/README.md" }, { "label": "Project website (first built by Aska Digital)", "url": "https://eumemic.github.io/exoplanets/" } ], "foundation0": true, "foundation0_note": "Founding member eumemic's search. Aska Digital built the first version of the project's website and sent ten pull requests, four of them pipeline fixes; the authors merged all ten.", "foundation0_members": [ "eumemic", "ahrazzle" ], "page": "research/exoplanets/" }, { "id": "3sum-apsp-exponents-community-lean", "date": "2026-10-11", "date_basis": "announcement", "title": "Agent-driven entrants push Lean-verified 3SUM to n^1.962 and APSP to n^2.99556", "domain": "computing", "subfield": "algorithms / fine-grained complexity", "summary": "Within a week of the 3SUM and APSP preprint, people pointing AI agents at it submitted better exponents to the Big O Speedrun verifier, which rebuilds each Lean proof and checks it against a fixed statement. On 11 October c0ldzy17 reached O(n^1.962) for 3SUM and Li Xuanji O(n^2.995561) for APSP, from 1.9992 and 2.99942.", "before_after": { "metric": "Lean-verified 3SUM time exponent", "before": "1.9992 (5 October)", "after": "1.962 (11 October)" }, "why_it_matters": "Kernel-checked leaderboards let anyone with AI agents contribute to frontier algorithms and be trusted without peer review, compressing incremental progress into days.", "ai_systems": [ "Contributors' AI agents" ], "organizations": [ "Big O Speedrun", "independent contributors" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "Records change quickly: on 14 October Anchen0823 posted a Lean-verified n^1.960614 for 3SUM. Human review of the improved constructions is pending.", "significance": 2, "sources": [ { "label": "Big O Speedrun, 3SUM leaderboard", "url": "https://bigospeedrun.com/3sum" }, { "label": "Big O Speedrun, APSP leaderboard", "url": "https://bigospeedrun.com/apsp" }, { "label": "Nil Mamano, Academia 2.0 (2026-10-12)", "url": "https://nilmamano.com/blog/big-o-speedrun" } ], "foundation0": false }, { "id": "aes-sbox-106-gates", "date": "2026-10-11", "date_basis": "preprint", "title": "AI-assisted search finds a 106-gate AES S-box circuit (AND/XOR/XNOR model)", "domain": "computing", "subfield": "circuit minimization (cryptography)", "summary": "A companion AI-assisted preprint gives an AES S-box circuit of 106 gates (depth 24) against 110 on NIST's list, an inverse S-box of 109 gates against 112, and 121 gates at depth 16 against 125. All are exhaustively checked and Lean-proved; comparisons hold only for AND/XOR/XNOR gates, not richer gate sets.", "before_after": { "metric": "AES S-box gate count, {AND, XOR, XNOR} model", "before": "110 (NIST circuit, depth 22)", "after": "106 (depth 24)" }, "why_it_matters": "The S-box is the main nonlinear cost of AES, the world's most widely used cipher. Smaller circuits mean cheaper encryption hardware and faster bitsliced software.", "ai_systems": [ "Unnamed AI tools (computational search and drafting)" ], "organizations": [ "Independent researcher (Swapnil Jain)" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Preprint PDF, October 2026", "url": "https://raw.githubusercontent.com/Swapnil-jain/gf2m-circuits/main/papers/Jain_AES_Sbox_circuits_2026.pdf" }, { "label": "gf2m-circuits repository (circuits, verifiers, Lean proofs)", "url": "https://github.com/Swapnil-jain/gf2m-circuits" } ], "foundation0": false, "page": "research/gf2m-circuits/" }, { "id": "gf2m-aes-field-multiplier-112", "date": "2026-10-11", "date_basis": "preprint", "title": "AI-assisted search shrinks the AES-field multiplier from 117 to 112 gates", "domain": "computing", "subfield": "circuit minimization (binary-field arithmetic)", "summary": "A preprint by Swapnil Jain, using AI-assisted computational search, gives an AND/XOR circuit for multiplication in the AES field GF(2^8) with 112 gates at depth 6, versus 117 in the 2010 Circuit Minimization Team record, plus 107 gates at depth 13. Each AES-field circuit is exhaustively checked and Lean 4 kernel-proved.", "before_after": { "metric": "Gates for GF(2^8) AES-polynomial multiplication at depth 6", "before": "117 (Peralta et al., CMT, 2010)", "after": "112" }, "why_it_matters": "Field multipliers are building blocks of encryption and error-correcting codes. Fewer gates make bitsliced and masked cryptography cheaper, and machine-checked proofs mean nobody has to trust the search.", "ai_systems": [ "Unnamed AI tools (computational search and drafting)" ], "organizations": [ "Independent researcher (Swapnil Jain)" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "", "significance": 2, "sources": [ { "label": "Preprint PDF, October 2026", "url": "https://raw.githubusercontent.com/Swapnil-jain/gf2m-circuits/main/papers/Jain_GF2m_multipliers_2026.pdf" }, { "label": "gf2m-circuits repository (circuits, verifiers, Lean proofs)", "url": "https://github.com/Swapnil-jain/gf2m-circuits" } ], "foundation0": false, "page": "research/gf2m-circuits/" }, { "id": "mobius-x1-gf2-mobius-below-nlogn", "date": "2026-10-11", "date_basis": "announcement", "title": "Claimed proof: the GF(2) Möbius transform needs fewer XORs than Yates' 1937 recipe", "domain": "computing", "subfield": "circuit complexity", "summary": "Swapnil Jain's repository claims, with written proofs, certificates and partial Lean, that the subset-parity (Möbius) transform over GF(2) can use asymptotically fewer XOR gates than Yates' ½·N·log₂N butterfly, by moving an AI-generated Fourier-transform construction into bit arithmetic. The saving only appears for astronomically large inputs.", "before_after": { "metric": "XOR gates for the GF(2) Möbius transform on N = 2^k inputs", "before": "½·N·log₂N (Yates 1937; tight without cancellation, Boyar–Find 2013)", "after": "N·(log N)^θ′ with θ′ = 1 − 4.37×10⁻¹¹ (claimed)" }, "why_it_matters": "It answers, if it holds, a question Boyar and Find asked in 2013: whether letting terms cancel can beat the classic recipe. The answer would be yes, but only at sizes no computer will reach.", "ai_systems": [ "OpenAI internal model (source construction, problem #130)", "AI referees (per repository)", "Claude Opus 5.5 (Aska's check)" ], "organizations": [ "Swapnil Jain (independent)", "Aska Digital (contributed the full-network check)" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "Not disputed, but only AI referees have read it; no human referee yet. The main theorem is not in Lean (92 supporting theorems are).", "significance": 2, "sources": [ { "label": "Swapnil-jain/mobius-x1 (README, proofs, certificates)", "url": "https://github.com/Swapnil-jain/mobius-x1" }, { "label": "Proof write-up (PROOF.md)", "url": "https://github.com/Swapnil-jain/mobius-x1/blob/main/proofs/PROOF.md" }, { "label": "Aska's full-network check (PR #1)", "url": "https://github.com/Swapnil-jain/mobius-x1/pull/1" } ], "foundation0": true, "foundation0_note": "Swapnil Jain's repository and claim. Aska Digital wrote and ran the full-network check behind Theorem 2 (PR #1, merged).", "foundation0_members": [ "ahrazzle" ], "page": "research/mobius-x1/" }, { "id": "exact-dft-lean-record-shea-kothe", "date": "2026-10-12", "date_basis": "announcement", "title": "Lean-checked proof lifts the exact Fourier transform speed-up to δ = 0.00123", "domain": "mathematics", "subfield": "algorithms / algebraic complexity", "summary": "OpenAI's preprint (dated 25 September, released 6 October) computed the exact discrete Fourier transform in O(n(log n)^(1−δ)) steps with δ = 10⁻¹³. Within a week, people working with AI assistants raised δ repeatedly. Ryan Shea fitted Jeremy Kothe's community circuit into the formal framework, reaching δ = 0.0012331685, which the public leaderboard rebuilt and kernel-checked in Lean.", "before_after": { "metric": "Exponent saving δ in O(n(log n)^(1−δ)) for the exact DFT (Lean-verified)", "before": "0.0010221076 (Nil Mamano, 11 Oct); 10⁻¹³ in OpenAI's preprint", "after": "0.0012331685 (Kothe circuit, Shea instantiation, 12 Oct)" }, "why_it_matters": "A machine-written result was improved about 12 billion-fold in a week by people and AI tools, and every step was checkable by computer. The gain is theoretical: no faster FFT software follows.", "ai_systems": [ "OpenAI Codex", "OpenAI GPT-6 Pro and GPT-6-Astra Max (earlier versions)", "Claude Code with Claude Opus 5.5 (round-eleven Lean port)", "OpenAI internal model (original #130 result)" ], "organizations": [ "Ryan Shea (independent)", "integer-mult-bounds community (CrocSwap)", "OpenAI", "Big O Speedrun (verification)" ], "ai_role": "collaborator", "evidence": "formally-verified", "contested": false, "contested_note": "Not disputed. It is one step in a fast race: ten leaderboard entries in six days, including eumemic's claimed 4.856×10⁻⁴ and the member's own claimed 7.3×10⁻⁵ and 6.7×10⁻⁴. Human review of the statement and proof is still pending, and newer records may already exist.", "significance": 2, "sources": [ { "label": "shea256/fourier-transform-below-nlogn (README, Lean proof, receipts)", "url": "https://github.com/shea256/fourier-transform-below-nlogn" }, { "label": "Lean proof at the leaderboard-pinned commit", "url": "https://github.com/shea256/fourier-transform-below-nlogn/tree/52394849c7ecf3412d39b925d75fc7fdb98994d8/lean/complex588" }, { "label": "Big O Speedrun exact-DFT leaderboard", "url": "https://bigospeedrun.com/exact-dft" }, { "label": "OpenAI, An explicit power saving for the exact discrete Fourier transform (#130)", "url": "https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/An-explicit-power-saving-for-the-exact-discrete-Fourier-transform-September-25-2026/main.pdf" } ], "foundation0": true, "foundation0_note": "Founding member Ryan Shea's record, shared with Jeremy Kothe, whose community circuit it uses.", "foundation0_members": [ "shea256" ], "page": "research/machine-mathematics/" }, { "id": "tanzania-ai-diabetic-retinopathy-referral-rct", "date": "2026-10-12", "date_basis": "publication", "title": "On-the-spot AI eye screening doubles specialist follow-up in Tanzania trial", "domain": "medicine", "subfield": "ophthalmology, global health", "summary": "At 24 Tanzanian hospitals, 2,378 people with diabetes were randomized to AI-graded retinal screening with an immediate result and face-to-face counselling, or standard grading with phone results weeks later. Among those truly needing referral, 35.9% reached the eye clinic within eight weeks versus 18.2% (odds ratio 2.52).", "before_after": { "metric": "Referral attendance within 8 weeks (true referable retinopathy)", "before": "18.2% (standard screening)", "after": "35.9% (AI point-of-screening result)" }, "why_it_matters": "In African screening programmes, many patients never return after referral. Giving results while the patient is still present turned AI speed into more people reaching sight-saving care.", "ai_systems": [ "SELENA+ (EyRIS)" ], "organizations": [ "London School of Hygiene & Tropical Medicine", "Kilimanjaro Christian Medical Centre", "Tanzania Ministry of Health", "University of Dodoma", "Moorfields Eye Hospital and UCL Institute of Ophthalmology" ], "ai_role": "instrument", "evidence": "peer-reviewed", "contested": false, "contested_note": "", "significance": 3, "sources": [ { "label": "The Lancet Global Health, 2026-10-12", "url": "https://doi.org/10.1016/j.langlo.2026.104062" }, { "label": "Trial protocol, BMJ Open 2024 (names SELENA+)", "url": "https://doi.org/10.1136/bmjopen-2023-075055" } ], "foundation0": false }, { "id": "aska-cab-to-fcc-exact-fold", "date": "2026-10-13", "date_basis": "preprint", "title": "Exact fold carries a loose ball-and-rod crystal into the densest packing", "domain": "mathematics", "subfield": "discrete geometry; rigidity theory; sphere packing", "summary": "A framework of balls joined by rigid rods, arranged like the boron net of calcium hexaboride, folds continuously into the densest stacking of equal balls. No rod stretches and no balls overlap, while density rises steadily from 0.2233 to 0.7405 (pi/sqrt18). An explicit formula and an exact certificate of 16,240 nonnegative coefficients cover every instant.", "before_after": { "metric": "Highest packing density reached by a rod-preserving motion of this framework", "before": "0.4686: the textbook cubic tilt stalls here (project's own baseline computation)", "after": "0.7405 = pi/sqrt(18), face-centred cubic close packing, proved exactly" }, "why_it_matters": "A floppy, mostly empty scaffold can reach the densest possible packing by folding alone. An AI search found a path in about 11 minutes; an exact proof followed the next day.", "ai_systems": [ "Claude Opus 5.5 (Anthropic)", "GPT-6 via OpenAI Codex", "GPT-6.1 Sol (OpenAI)" ], "organizations": [ "Aska Digital" ], "ai_role": "collaborator", "evidence": "preprint", "contested": false, "contested_note": "Not disputed. Reviewed only internally by AI models, with no human referee. The framework family is prior art (Borcea and Streinu 2021, section 7), and historical priority of the motion is unresolved. It is a geometry result, not a predicted phase change in real CaB6.", "significance": 2, "sources": [ { "label": "Proof: explicit continuous motion to fcc", "url": "https://github.com/aska-digital/incentre-lattice/blob/main/results/fold-continuous-fcc/PROOF.md" }, { "label": "Exact certificate (JSON)", "url": "https://aska-digital.github.io/incentre-lattice/results/fold-continuous-fcc/certificate.json" }, { "label": "Preprint v2 (Zenodo)", "url": "https://doi.org/10.5281/zenodo.23340326" }, { "label": "AI search log", "url": "https://github.com/aska-digital/incentre-lattice/blob/main/autoresearch-fold/results.tsv" } ], "foundation0": true, "foundation0_note": "Founding member Ahraz Arifuddin's own work at Aska Digital, done with a team of AI agents.", "foundation0_members": [ "ahrazzle" ], "page": "research/incentre-lattice/" }, { "id": "vdw-coefficient-community-refinement", "date": "2026-10-13", "date_basis": "announcement", "title": "Open AI-assisted notes raise a van der Waerden constant from 1/100000 to 1/9", "domain": "mathematics", "subfield": "combinatorics / Ramsey theory", "summary": "OpenAI's machine-written preprint showed van der Waerden numbers W_r(k) exceed k^(c·k·⌊log2 r⌋) with c = 1/100000. In a public repository started by rohanarun with Codex and Claude, and extended by Claude-assisted contributors, step-by-step refinements raised c to 1/9, about 11,111 times larger. Exact checks pass; the argument is conditional and unreviewed.", "before_after": { "metric": "Constant c in W_r(k) > k^(c·k·⌊log2 r⌋)", "before": "1/100000 (OpenAI preprint)", "after": "1/9 (conditional)" }, "why_it_matters": "Machine-written first proofs can leave large slack, and people with AI agents can find it within days. No explicit coloring or practical number follows.", "ai_systems": [ "OpenAI Codex", "Claude Fable 5 (Anthropic)", "Claude (Anthropic), via contributor DaysSky", "GPT-6 Pro in ChatGPT Pro, via contributor artuskg", "Unreleased OpenAI internal model (source preprint)" ], "organizations": [ "OpenAI (source preprint)", "rohanarun (independent)", "independent contributors DaysSky and artuskg" ], "ai_role": "collaborator", "evidence": "claimed", "contested": false, "contested_note": "Conditional on OpenAI's unreviewed preprint (family #160). No independent human review. The exact scripts check only parameter inequalities. Partial Lean (84 statements) is in open PR #10.", "significance": 2, "sources": [ { "label": "rohanarun/van-der-waerden-refinement (README)", "url": "https://github.com/rohanarun/van-der-waerden-refinement" }, { "label": "PR #7: digit rank and short-period runs, c = 1/9 (DaysSky)", "url": "https://github.com/rohanarun/van-der-waerden-refinement/pull/7" }, { "label": "PR #10: conditional reviews and partial Lean verification (open)", "url": "https://github.com/rohanarun/van-der-waerden-refinement/pull/10" }, { "label": "OpenAI, Quantitative Superexponential Bounds for van der Waerden Numbers (openai/math @ fd4aeeb)", "url": "https://github.com/openai/math/tree/fd4aeeb2ee4fc729c18d98444fed42fd0529eeeb/preprints/Quantitative-Superexponential-Bounds-for-van-der-Waerden-Numbers-September-23-2026" } ], "foundation0": true, "foundation0_note": "Started by founding member rohanarun with Codex and Claude; the last steps came from contributors DaysSky and artuskg.", "foundation0_members": [ "rohanarun" ], "page": "research/machine-mathematics/" } ] }