News · 2026-10-09

Mathematicians challenge how OpenAI presents mathematical progress

Mathematicians published new critiques of OpenAI’s mathematical research release on October 8, questioning whether hard-to-read proofs and corporate announcements amount to usable scientific progress. Set theorist Asaf Karagila challenged the presentation of the Partition Principle manuscript, while a guest essay on Terence Tao’s blog offered students conditional encouragement. Community work on integer multiplication shows that follow-on progress and unresolved verification can coexist.

Key facts

• OpenAI’s checked correction log contains three withdrawals dated October 7; no later dated withdrawal appeared in the research record.
• Karagila and Álvaro Lozano-Robledo published their respective essays on October 8, 2026.
• An October 8 community snapshot reports a sharper conditional exponent for OpenAI result #109, while assuming the upstream theorem.
• The primary source for the presentation critique is Karagila’s own essay.

Open publication helps only when others can identify exactly what is claimed and what remains assumed. That is why a readable explanation, a pinned version, and a formal artifact play complementary roles rather than substituting automatically for one another.

A mathematical result needs more than an impressive conclusion. Other people must understand its assumptions, inspect the argument, and decide what it changes. That is the practical issue behind Karagila’s essay, rather than a general rejection of computers doing mathematics.

The Partition Principle and the Axiom of Choice concern foundational questions about sets. OpenAI’s claim that the former does not imply the latter addresses an old question. Karagila says it would matter if established. But he also says he took only a brief look at the manuscript and did not inspect its machine-checkable code. His complaint is that unclear structure and nonstandard presentation make the argument difficult for specialists to assess. Readers should not convert that objection into a claim that he disproved the theorem.

The student essay hosted by Tao asks a different question: what should someone starting a mathematical career do now? Its author is number theorist Álvaro Lozano-Robledo. His concise advice is, “Keep calm and carry on studying math.” He acknowledges uncertainty and the privilege of giving advice from a tenured position. His recommendation rests on wanting to become an expert and do mathematics, rather than a promise that academic employment will remain available.

Lozano-Robledo describes discovery, communication, teaching, and training as parts of mathematical work. He suggests models could let students explore more problems and enter research earlier. A toy picture credited to Nestor Guillen portrays existing ideas as points that models can connect, while genuinely new concepts expand the space. That picture is an explanatory hypothesis, not a demonstrated limit on what future systems can invent.

Tao’s adjacent discussion about broadening the value assigned to mathematical work attracted a substantial Hacker News debate. The original Mastodon page was inaccessible in the dossier, so its exact wording is not reproduced here as a primary-verified quotation. The strongest counterargument in the discussion is serious: if models also become better at explanation and theory building, shifting credit toward those activities does not resolve the longer-term career question.

Meanwhile, CrocSwap’s pinned integer-multiplication project supplies a concrete example of people building on an AI-associated result. Its October 8 snapshot reports a conditional exponent slightly above two to the minus thirty-four, compared with the upstream paper’s two to the minus one hundred eighty-two. This is a theoretical asymptotic improvement under a specified computational model, not a measured speedup for everyday multiplication.

The construction changes what must move through the mathematical machine: compact control information replaces larger spaced windows. Think of sending directions to rearrange a warehouse rather than repeatedly moving whole aisles. That analogy explains the proposed saving; it does not establish the proof.

The verification boundary is explicit. The linked snapshot assumes the full upstream theorem and says its new arguments had not received independent mathematical review or formal verification. Swapnil Jain’s separate repository checks selected moment certificates and parameter arithmetic in Lean, but leaves analytic inequalities and the upstream multiplication theorem as assumptions. A proof assistant checks what has actually been encoded; it cannot erase unproved premises merely by checking the arithmetic built on them.

OpenAI’s published history and the Association for Human Mathematics statements page provide chronology. The checked record adds no new withdrawal or association statement after October 7. Today’s development is the second wave of interpretation and follow-on work. The stakes are how a field turns machine-produced output into shared knowledge, and how honestly it communicates uncertainty to its next generation.

Primary source, verified: read the paper →

Ground Truth. (2026, October 9). Mathematicians challenge how OpenAI presents mathematical progress. Ground Truth. https://groundtruth.day/news/mathematicians-demand-readable-evidence-after-openai-release.html

@misc{groundtruth:mathematicians-demand-readable-evidence-after-openai-release,
  title  = {Mathematicians challenge how OpenAI presents mathematical progress},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/mathematicians-demand-readable-evidence-after-openai-release.html}
}

Topics: mathematics · ai-science · formal-verification · research-culture