OpenAI Dropped 722 Math Papers Overnight. Mathematicians Lost It.

Summary

The release and the controversy around it

The video examines OpenAI’s release of 722 mathematics papers, nearly 35,000 pages, which it says contain 372 new results across 17 areas. The work was produced with an unreleased model after it was given roughly 4,000 problems. The narrator says each included result used an average of about three hours of ChatGPT Pro “thinking” compute. Around one result in five reportedly disproves a conjecture rather than proving it.

The scale of the release prompted both excitement and alarm. Some results include proofs formalized in Lean, a proof-checking system; together, the files contain about 26 million lines of Lean code. But Lean can check whether a formal proof follows from its premises—it cannot by itself establish that the formalized statement is the right interpretation of the mathematical claim. Most papers had not been independently peer reviewed at the time described.

The video places the release in a wider sequence of AI-mathematics claims:

• AI systems had recently reached medal-level performance on math competitions, but research mathematics is harder to assess because problems may have no known answers.
• An earlier OpenAI claim about solving Erdős problems proved misleading: the cited problems had already been solved in older papers.
• A later announcement by Levent Alpöge, crediting Anthropic’s model Fable, claimed that the Jacobian conjecture was false.
• OpenAI’s September announcement of a Navier–Stokes blow-up result triggered a dispute over authorship, collaboration, the mathematical significance of the result, and possible influence from mathematicians’ use of OpenAI tools. OpenAI said it had checked and that Tristan Buckmaster’s recent prompts could not have influenced the system. The Clay Mathematics Institute described the problem as “apparently” settled but said its review would be deliberate; OpenAI said it would not claim the prize.

A key qualification in the Navier–Stokes discussion is that OpenAI’s construction uses a carefully designed external force. Some mathematicians argued that this did not resolve the most important version of the fluid problem. The work also built on earlier research by Diego Córdoba and Luis Martínez-Zoroa, who were added as citations after being omitted from an initial version.

The narrator’s overall framing is cautious: OpenAI’s release makes substantial mathematical claims, but it is too early to call it a historic turning point while most results remain unchecked by independent experts.

Selected results explained in the video

1. The Flint Hills series and approximating pi

The Flint Hills series adds terms involving the reciprocal of (n^3\sin^2(n)). Most terms contribute little, but the sum can jump sharply when (n) is close to a multiple of pi. The video highlights jumps at 22 and 355, because (22/7) and (355/113) are unusually good approximations to pi.

This connects the series’ convergence to the irrationality exponent: a measure of how closely a number can be approximated by fractions, and whether exceptionally good approximations keep occurring. The video says OpenAI claims pi’s irrationality exponent is exactly 2, the smallest value possible for an irrational number, and concludes that the series converges to about 30.31. The result includes a Lean proof.

The narrator also shows fragments of the model’s reasoning, including abandoned ideas and a moment where it notices that a proposed argument would contradict known facts about transcendental numbers.

2. Faster integer multiplication—and an important limitation

Ordinary multiplication of two (n)-digit numbers takes roughly (n^2) elementary steps. The video traces improvements from Karatsuba’s method through the Schönhage–Strassen algorithm and the 2019 result of Harvey and van der Hoeven, which reached an (n\log n)-type bound.

OpenAI claims a marginal improvement on the idealized Turing-machine model: the exponent on the logarithm is smaller than 1 by (1/2^{182}). The improvement is so tiny that it would not be practically useful. According to the narrator’s reading, the method only becomes relevant for numbers too large for any computer in our universe to store. Such results are often called galactic algorithms: they establish that a faster algorithm exists in theory without yielding a usable method.

The video also mentions a related improvement to the Fourier transform in an idealized arithmetic model, again by an extremely small margin.

3. Coloring the plane

The plane-coloring problem asks for the fewest colors needed so that any two points exactly one unit apart have different colors. The known bounds had long been four to seven.

The video explains the Moser spindle, a seven-point shape that requires four colors, and then walks through the paper’s argument against five-coloring the entire plane:

• First, the proof handles even extremely irregular colorings by smoothing them into a better-behaved coloring whose errors occupy zero area.
• A topological argument implies that somewhere, three, four, or five color regions must swirl around a point.
• The four- and five-color cases are ruled out by counting, leaving a three-color swirl.
• Around that swirl, the geometric fact that points 60 degrees apart on a unit circle are also one unit apart forces the remaining two colors to alternate in six slices.
• Completing unit-sided diamonds from those alternating colors creates a region that can use only three colors.
• That region contains a Moser spindle, which requires four colors, yielding a contradiction.

The conclusion is that the plane is not five-colorable, so the chromatic number is six or seven. The proof establishes that some finite set needs six colors, but, as presented in the video, it does not produce an explicit example of that set. The paper also comes with a Lean proof.

4. Crossing numbers: Turán’s brick factory and Hill’s conjecture

The video introduces two graph-drawing problems through Paul Turán’s experience in a forced-labor brick factory. He wondered how to arrange rail tracks between kilns and storage yards to minimize crossings. A related question asks for the fewest crossings when every point in a graph is connected to every other point.

For decades, well-known candidate formulas for these problems lacked complete proofs. The video describes the core idea behind the new proofs:

• Two closed loops on a plane cross an even number of times.
• A drawing of a graph contains many loops, so the parity rule imposes simultaneous constraints on how its edges can cross.
• The proof turns those constraints into algebraic “fingerprints” associated with patterns on the graph’s vertices.
• For seven points, the argument produces nine independent patterns and shows that at least nine crossings are necessary.
• The same kind of counting extends to general graph sizes and proves the proposed formulas.

The release claims proofs for both the Zarankiewicz and Hill formulas, with Lean formalizations.

5. Matrix multiplication

The standard method for multiplying two (n\times n) matrices takes (n^3) multiplications. Strassen’s 1969 method reduced the exponent to about 2.8, and subsequent work gradually improved the best known upper bound. The quantity being pursued is called (\omega); it cannot be below 2, and many mathematicians suspect it may equal 2.

The video reports that OpenAI claims (\omega\leq 9/4), or 2.25—far larger progress than the small improvements of recent decades. The narrative emphasizes that the result appeared first in a longer, more complicated paper and then, eight days later, in a shorter and cleaner proof with a Lean formalization.

Like the multiplication result, this is theoretical rather than practically usable: the method establishes that a faster algorithm exists without giving a method that can be used in practice.

6. Results related to famous open problems

The release does not claim to solve any Millennium Prize Problem outright, but the video says it claims progress on several:

• Birch and Swinnerton-Dyer conjecture: a formula for the two simplest types of curves.
• Hodge conjecture: a result covering one special, highly symmetric family.
• Riemann hypothesis: a “quasi-Riemann hypothesis” result that rules out zeros beyond (7/8), rather than proving all relevant zeros lie on the line (1/2).

The video also lists claims concerning Erdős’s conjecture about sets of integers containing arithmetic progressions of every length, the Unique Games Conjecture, Hilbert’s 16th problem, Hadwiger’s conjecture, Borsuk’s conjecture, Catalan’s constant, Artin’s conjecture, and the Gaussian moat problem. These are presented as claims in the release, not as independently confirmed conclusions.

One especially striking claim concerns Hilbert’s 10th problem over the rational numbers. Hilbert asked whether an algorithm could decide whether any polynomial equation has a solution in whole numbers. Later work showed that no such general algorithm exists for the integers. The video says paper four claims a corresponding impossibility result for rational solutions: no algorithm can decide the question for all polynomial equations. If correct, the result would be a machine proving that a class of questions cannot be answered by any machine.

What the video says the release means

The release raises questions about how mathematical progress should be evaluated and shared. Some mathematicians objected to releasing hundreds of papers at once, arguing that it was not a conventional scholarly process and risked overwhelming the community with proofs to check. Others argued that solving unsolved problems is inherently valuable and criticized what they saw as an excessive emphasis on mathematical aesthetics.

The narrator uses G. H. Hardy’s view that beauty is a test of good mathematics, and Paul Erdős’s idea of “The Book,” containing the most beautiful proof of each theorem, to frame the tension between correctness and elegance. The release may contain many long, machine-checked proofs that no person will read in full. Yet the matrix-multiplication example—an initially long argument followed by a shorter, cleaner one—suggests that mathematical understanding and exposition can improve after a result is first discovered.

The video’s closing position is that the papers should be treated as first drafts, not as finished entries in “The Book.” AI may be becoming very good at finding what is true, but finding the concise, illuminating explanation of why it is true remains an open task.

Speakers and sources featured

Narration: An unnamed video narrator, who explains the mathematics and reports on the release and reactions.

Named people quoted, interviewed, or discussed: Sam Altman; Joe Bebel; Levent Alpöge; Daniel Litt; Demis Hassabis; Charles Fefferman; Tristan Buckmaster; Sébastien Bubeck; Luis Silvestre; Diego Córdoba; Luis Martínez-Zoroa; Terence Tao; G. H. Hardy; Lucas Beyer; Liron Shapira; Aubrey de Grey; Edward Nelson; Leo and William Moser; Andrei Kolmogorov; Anatoly Karatsuba; Arnold Schönhage; Volker Strassen; David Harvey; Joris van der Hoeven; Steven Strogatz; Alex Kontorovich; Paul Erdős; Dor Minzer; Yumou Fei; Shuo Wang; Álvaro Lozano-Robledo; Francesco Maggi; Nic Carter; David Hilbert; Kurt Gödel; Alan Turing; Yuri Matiyasevich; Joseph Liouville; Zu Chongzhi; Kazimierz Zarankiewicz; Richard Guy; and Anthony Hill.

Organizations, publications, platforms, and technical sources mentioned: OpenAI; Anthropic and its model Fable; Google DeepMind; the Clay Mathematics Institute; the Fields Medal community; the Association for Human Mathematics; Lean; Quanta Magazine; Scientific American; X/Twitter; Hacker News; and OpenAI’s GitHub release.