One of the 719 papers OpenAI released on October 6, 2026 claims that two very large numbers can be multiplied faster than a limit proposed in 1971 and reached only in 2019, when its authors called their result “the end of the road for this problem.” The claimed improvement is real mathematics and practically nothing: on a number so large that writing it down would use every atom in the observable universe, the gain is about one part in 10⁵⁴. Then something unexpected happened. A public project opened the next day, and in six days volunteers sharpened the paper’s key number by about 52 orders of magnitude — every step labelled as conditional on the original paper, which has no machine-checked proof and which no specialist has publicly verified. Three things are true at once: a record no computer could notice, a crowd racing to improve it, and a foundation nobody has checked. This case follows what a method amplifies when people build on it before they understand it.
Multiplication has a history of limits that turned out not to be limits. In 1960 Andrey Kolmogorov, one of the century’s great mathematicians, told a seminar that no method could beat the one taught in school, where the work grows with the square of the number of digits. A 23-year-old in the room, Anatoly Karatsuba, found a faster method within a week.[7][15] Faster methods followed for decades. In 1971 Arnold Schönhage and Volker Strassen proposed where the improvements would stop: at a cost of n log n for numbers of n digits.[6]
In 2019 David Harvey and Joris van der Hoeven reached that mark. Harvey’s words at the time: “our work is expected to be the end of the road for this problem, although we don’t know yet how to prove this rigorously.”[5] The caveat mattered. Nobody has ever proved that multiplication needs more than the bare minimum of time; their paper says so directly — “no super-linear lower bound for M(n) is known.”[6] The limit was a conjecture, supported by a second, conditional argument from a different field.[9]
The paper released on October 6 claims to go past it: multiplication in n times (log n) raised to the power 1 − κ, where κ is 2⁻¹⁸², a number with 54 zeros after the decimal point.[1] It builds directly on the 2019 method and credits it lemma by lemma, then adds ideas from other fields — the mathematics of efficient broadcasting, and a technique for computing with scratch memory that is already full. Its key move is to mix data together while rearranging it, then unmix it exactly at the end.[1] It is candid about what this is: “The constants and thresholds in the construction are extremely large; the algorithm is intended to establish a bit-complexity bound.”[1]
How small is the gain? By this case’s arithmetic from the paper’s formula, and ignoring those constants: for a number with 10⁸⁰ digits — one for every atom in the observable universe — the new method would be faster by about one part in 10⁵⁴. To be twice as fast, the number would need so many digits that the count of its digits is itself about 10⁵⁴ digits long. If the proof holds, it overturns a 55-year-old expectation, and no machine that could ever exist would notice the difference.[1][14]
A public repository opened on October 7, the day after the release, to sharpen that exponent. The first improvements were small — from 2⁻¹⁸² to 2⁻¹⁵⁴, then 2⁻¹²⁹ — and then they were not. By its own record the project passed 2⁻⁵⁹, then 2⁻³⁴, then 2⁻¹⁵. On October 12 its reviewed figure stood at κ = 0.0016, about 2⁻⁹.3.[3] That is roughly 10⁵² times the original — 52 orders of magnitude in six days. The repository showed about 200 stars, 87 forks, more than 360 commits and over 600 pull requests.[3]
The project is careful about what it has and has not done. Its heading reads “conditional on the original OpenAI #109 framework.” It states that its result “is an improvement in the asymptotic exponent saving, not a measured runtime speedup,” and that the work “is not full formal verification, independent human peer review, a worldwide priority claim, or a practical multiplication benchmark.”[3] Parts of it are machine-checked; the whole multiplication theorem is not — “These proofs do not constitute a complete Lean multiplication theorem.”[3] Even at the new figure, the gain on the universe-sized number above would be under one percent.[14]
So the structure is this: a 73-page paper with no machine-checked proof,[2] six days of collective improvement resting on it, and — as far as this case could find — no public comment from the specialists who set the previous record or from the authors of the conditional limit. One prominent computer scientist listed the result among those to watch and added, of the release as a whole, “it also appears that no human has understood just about any of these proofs yet; the race to do so has just started.”[10] A companion result on the Fourier transform, which shares the paper’s key ingredient, does have a machine-checked statement.[2]
Schönhage and Strassen give a much faster method and propose n log n as the best possible.[6]
The ConjectureSeventy-three pages, authored “OpenAI”: a gain of 2⁻¹⁸² in the exponent, by its own account not a practical algorithm.[1]
The ClaimVolunteers begin sharpening the exponent, each step conditional on the original paper.[3]
The CrowdKolmogorov’s 1960 conjecture that the school method was best fell to a 23-year-old student within a week.[7]
n log n was proposed in 1971 and reached in 2019. It was never proven to be the limit.[5][6]
The exponent gain went from 2⁻¹⁸² to about 0.0016 between October 6 and October 12, conditional on the original paper.[1][3]
The multiplication paper has none. Its companion on the Fourier transform, sharing the key ingredient, has a machine-checked statement.[2]
Our work is expected to be the end of the road for this problem, although we don’t know yet how to prove this rigorously. — David Harvey, 2019, on reaching the limit the new paper claims to pass [5]
| Dimension | Evidence |
|---|---|
| Employee (D2) L1 · 86 | A public project opened the day after release: about 200 stars, 87 forks, 360+ commits and 600+ pull requests in six days. The specialists who set the previous record have not commented publicly.[3]The Crowd Arrived First |
| Quality (D5) Origin · 84 | A claimed gain of 2⁻¹⁸² in the exponent beyond n log n, the mark reached in 2019 and proposed as final in 1971. The paper says it is not a practical algorithm.[1][5]Past the End of the Road |
| Operational (D6) L1 · 82 | The exponent went from 2⁻¹⁸² to 0.0016, about 10⁵² times, with every step labelled conditional on the original paper and only parts machine-checked.[3]Improve First, Verify Later |
| Regulatory (D4) L2 · 72 | No machine-checked proof of the multiplication paper; no specialist verdict. What it would overturn was never proven: “no super-linear lower bound … is known.”[2][6]A Conjecture, Unchecked |
| Customer (D1) L2 · 40 | On a number with 10⁸⁰ digits the original gain is about one part in 10⁵⁴; at the community figure, under 1%. Working software uses older methods.[12][14]No Computer Would Notice |
The cascade originates in D5 (Quality): one claimed result that, if true, passes a limit specialists expected to be final.[1][5] It propagates to D2 (Employee), where the people who took it up were not the field’s specialists but a public crowd of contributors,[3] and to D6 (Operational), where the way of working is new: improvement first, verification later, with each contribution labelled conditional and partly machine-checked.[3] It lands in D4 (Regulatory), where the paper has no machine-checked proof, the specialists have not ruled, and what would be overturned is a conjecture rather than a theorem;[2][6] and in D1 (Customer), where the practical user gains nothing — the methods actually used in software are older and faster at every real size.[12][14] D3 (Revenue) is not scored.
UC-328 followed a result from the same release where specialists rushed to understand a machine’s proof. This is the other shape: outsiders rushing to build on one. Both are amplification, and they carry different risks. Understanding first is slow and safe. Building first is fast, and everything built shares the fate of what is underneath. If the original argument holds, six days produced an extraordinary amount of mathematics. If it has a flaw, the same six days produced a tall structure on a base that was never checked. The honest limit: this case cannot tell which, and neither, yet, can anyone else.
-- UC-329: Faster Than Anyone Could Notice: 6D Amplifying Cascade
-- Integer multiplication. 1960: Kolmogorov conjectures n^2 optimal; Karatsuba (23) beats it within a week. 1971: Schonhage-Strassen O(n log n log log n), propose n log n optimal. 2019: Harvey and van der Hoeven reach O(n log n) (Annals 2021); Harvey: expected to be the end of the road; no super-linear lower bound known. OpenAI release Oct 6 2026, result 109: 73-page paper dated Sep 23 claims O(n (lg n)^(1-kappa)), kappa = 2^-182, multitape Turing machine; constants extremely large; not Lean-formalized; not withdrawn or revised. Sibling result 130 (Fourier transform, delta = 10^-13) shares the key ingredient and is formalized. Community repo CrocSwap/integer-mult-bounds opened Oct 7: kappa 2^-154, 2^-59, 2^-34, 2^-15, then 0.0016127 on Oct 12, all conditional on the #109 framework; about 200 stars, 87 forks, 360 commits, 600 pull requests. No specialist comment found.
FORAGE faster_than_noticed
WHERE conjectured_limit_claimed_passed = true
AND practical_gain_is_negligible = true
AND community_builds_before_verification = true
ACROSS D5, D2, D6, D4, D1
DEPTH 3
SURFACE faster_than_noticed
DIVE INTO build_before_understand
WHEN foundation_machine_checked = false
AND downstream_work_labelled_conditional = true
TRACE conditional_stack_cascade
EMIT faster_than_noticed_signal
DRIFT faster_than_noticed
METHODOLOGY 86
PERFORMANCE 30
FETCH faster_than_noticed
THRESHOLD 1000
ON EXECUTE CHIRP signal 'An AI-written paper claims to multiply large numbers faster than a limit proposed in 1971 and reached in 2019, by a margin of 2^-182 in the exponent: about one part in 10^54 on a number the size of the universe. Within six days a public crowd sharpened that margin about 10^52 times, every step conditional on the original paper, which has no machine-checked proof and no specialist verdict. The amplification is real and so is the exposure: everything built shares the fate of the base.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
If the proof holds it passes a mark specialists called the end of the road — and no computer that could exist would run faster for it. The paper says as much itself. Pure mathematics is the one place those two things separate completely.[1][5]
Six days, over 600 contributions, 52 orders of magnitude — each step honestly labelled conditional. It is a new way of working, and it inherits every weakness of its base.[3]
The school method was called unbeatable in 1960 and fell in a week. n log n was proposed in 1971 and never proven. A conjecture the experts believed is not the same as a wall.[6][7]
In UC-328, specialists set out to understand a machine’s proof. Here, outsiders set out to extend one. The first is slow and safe. The second is fast, and all of it stands or falls together.[3]
Primary sources are OpenAI’s paper and repository, the community project’s own repository, and the original papers and announcements for the history. All were read directly. Figures for the size of the gain are this case’s arithmetic from the paper’s stated formula. The proof has not been verified by any specialist as of October 12, 2026, and this case does not claim it is correct. Source 14 is an interactive companion visual made for this case.
This case follows a result from claim to a six-day stack of work resting on it, and marks where verification stopped. The same reading applies to any fast-moving project: which layer was checked, and how much now depends on the one that wasn’t. Book a discovery call to map that in your own stack.