Essay · Aska Digital · 13 Oct 2026
When the crowd gets a frontier model
On 7 October founding member Doug Colkitt opened a public repository around one question: how much faster than n log n can two large numbers provably be multiplied? Within days a crowd of people and their AI agents was moving the answer. Aska Digital’s essay of 13 October 2026, kept as published, asks why that worked.
The reviewed constant κ, 9 October to now
How to read: each step is a value the race’s maintainer selected. The solid line is the essay’s data, to 13 Oct 2026, 08:10 UTC; the dashed step ends at the record now. Point at a dot for its value.
- 9 Oct, first record
- 0.0004609169
- Essay, 13 Oct
- 0.001662 3.6×
- Last reviewed record
- 0.001673 3.63×
00//The constant
Snapshot · 13 Oct 2026, 08:10 UTCThe reviewed constant
The essay, as published
On 7 October someone opened a public repository holding one question: how much faster than n log n can you prove that two large numbers can be multiplied? Five days later it has 857 pull requests from 172 people, and the reviewed answer has moved 3.6×.
The reviewed constant κ, as recorded in the repository's own certificate
How to read: each step is a new value the maintainer selected; the line holds it until the next. Point at or Tab to a point for its time and value.
Show the numbers
| Selected (UTC) | κ |
|---|---|
| 9 Oct 2026, 04:53 UTC | 0.0004609169 |
| 9 Oct 2026, 05:55 UTC | 0.000472143085 |
| 9 Oct 2026, 06:01 UTC | 0.000472154791 |
| 9 Oct 2026, 12:32 UTC | 0.000661885549259598 |
| 10 Oct 2026, 14:08 UTC | 0.000723571590007464 |
| 12 Oct 2026, 01:55 UTC | 0.001517760977820331 |
| 12 Oct 2026, 13:10 UTC | 0.001584931111785787 |
| 12 Oct 2026, 14:41 UTC | 0.001602278921983212 |
| 12 Oct 2026, 16:59 UTC | 0.0016127213726553 |
| 12 Oct 2026, 21:01 UTC | 0.001635183729255778 |
| 12 Oct 2026, 22:36 UTC | 0.001635527380375392 |
| 13 Oct 2026, 05:57 UTC | 0.001661202643272239 |
| 13 Oct 2026, 06:44 UTC | 0.001661634653669567 |
- 857pull requests
- 172people
- 152merged
- 6days old
01//Why it moved
Why it moved
Two things were different. The first is obvious: a model will read a long technical argument with you and draft the next step. That gets you a candidate.
The second matters more. Every result that moves the number arrives with a script that checks it. Nobody has to trust the person who sent it. You run their verifier on your own machine and it passes or it fails. A wrong answer costs its author some time and costs everyone else nothing.
That is what makes a crowd usable on a problem like this. Without the check, 857 pull requests is just noise nobody can afford to read. With it, the work composes: pull requests name the exact constructions they build on, and those belong to other people.
Check a claim yourself, then break it
How to read: below is the block from Aska’s PR #469 as it was claimed: how many child jobs of each size it hands down, and the κ it says that gives. The checker recomputes the block’s bill at that κ and accepts only if the bill stays under 1. Press “Break it” to raise the claim by one unit in its last digit, or change any count.
Child jobs by size r (block width m = 78, W = 1,200 roles)
Accepted: the bill stays under 1.
Bill at the claimed κ: 1 − 1.073 × 10−18
Checked when this page was built, to 80 digits.
PROOF.md; the bill as in the race’s autoresearch.md; the same block as the budget figure on the project page.Daily activity
How to read: one bar per day of October 2026, UTC. Point at or Tab to a bar for its value.
Pull requests opened, by day of October.
People contributing for the first time.
Show the numbers
| Day (UTC) | Pull requests opened | First-time contributors |
|---|---|---|
| 2026-10-07 | 2 | 2 |
| 2026-10-08 | 107 | 23 |
| 2026-10-09 | 137 | 28 |
| 2026-10-10 | 101 | 9 |
| 2026-10-11 | 173 | 11 |
| 2026-10-12 | 117 | 5 |
| 2026-10-13 | 220 | 94 |
02//Most of it is not the answer
Most of it is not the answer
Look at what the pull requests actually contain and the shape of a working research community shows up. Alongside the improvements there are independent re-verifications, speedups to other people's checking scripts, formal proofs in Lean, and a steady stream of honest negative results: this route is closed, here is the obstruction. Of 857 pull requests, 152 are merged. The rest are not waste. A closed route nobody has to try again is worth something, and the crowd is cheap enough to afford them.
03//What this is not
What this is not
The result is conditional. It inherits a framework from an earlier paper, and the repository says plainly that its finite checks are not a complete machine-checked theorem, and that none of this is a claim about how fast your computer multiplies numbers today. The honesty is part of why it works: the scope of every claim is written down, so the next person knows exactly what they are standing on.
04//Our part
Our part
We have been working in two of these.
Circuit constructions. A separate repository collects the smallest known circuits for arithmetic in binary fields, measured in AND and XOR gates. Smaller circuits mean cheaper encryption work, particularly in secure multi-party computation and zero-knowledge proofs. We contributed multipliers for four fields it did not cover. Two of them came in below the published baseline. All five were re-checked by the author against the pack's own verifier, and merged.
| Field | Baseline | Ours | Saved |
|---|---|---|---|
| GF(29) | 142 | 136 | 6 |
| GF(210) | 172 | 172 | 0 |
| GF(211) | 206 | 199 | 7 |
| GF(2128), the GHASH field | 11,972 | 11,972 | 0 |
The exponent race. We held the top open claim for several hours on 11 October, at κ = 0.00135058, by re-timing an existing construction. The race has since moved a long way past it, on a different line of work. That is the honest state of it: a real contribution, briefly in front, now behind.
05//Why it matters
Why it matters
The interesting part is not the constant. It is that a machine-checkable problem statement turns a crowd with good tools into something that functions like a laboratory, in days, at almost no coordination cost. The number of problems that can be posed this way is much larger than the number that have been.
Figures read from the repository's own certificate file and the GitHub API, as of 2026-10-13T08:10Z. The race is CrocSwap/integer-mult-bounds, maintained by Douglas Colkitt. The circuit pack is Swapnil-jain/gf2m-circuits. Numbers change by the hour; rebuild to refresh.
06//Corrections
Corrections, 14 and 15 October 2026
Foundation0 checked this essay against the sources before republishing it. Six lines need a note: one leaves out who opened the repository, two have slips, and three say more than the evidence does. The essay above is unchanged; here is what the record supports.
On 7 October someone opened a public repository
Named. The repository was opened by founding member Doug Colkitt (@Mister-Meeseeks), who maintains it and selects every reviewed record.
Five days later it has 857 pull requests
Six days. The repository opened on 7 October and the counts are from 13 October, 08:10 UTC: six days, as the essay’s own “6 days old” tile says.
We have been working in two of these.
Two of what. Two kinds of crowd-checkable problem: circuit constructions, and the exponent race.
We held the top open claim for several hours on 11 October.
Not supported by our own data. Our tracker saw PR #469 (κ = 0.001350580884553120) only between its appearance at 17:32 UTC and the tracker’s last snapshot at 17:41 UTC. In that window it was the top open claim, and it was still a draft; higher claims from other contributors (#480 and #505, for example) carry commits from later that day. It was never reviewed, and Aska closed it on 13 October. Read: briefly the top open claim on our tracker. The full account is on the project page.
All five were re-checked by the author against the pack's own verifier, and merged.
Four, not five, and the re-check is ours. The table lists four fields. The repository’s history shows the maintainer merged the pull request and credited it in the README; it does not show a re-run by the maintainer. Foundation0 re-ran the pack’s verifier on all four circuits on 13 October, and they pass. None of them is a record.
Smaller circuits mean cheaper encryption work, particularly in secure multi-party computation and zero-knowledge proofs.
Too broad. In those settings the cost is driven mostly by AND gates, while XOR gates are nearly free, and every gate these circuits save is an XOR; the AND counts are unchanged. Smaller XOR counts mean less hardware and fewer instructions in bit-sliced software.
Corrections compiled from the circuit pack’s history (PR #1), this site’s tracker feed (chart-data.json, generated 11 Oct 2026, 17:41 UTC) and the race’s pull requests #469 and #464.