HyperliquidHYPE, coin
DisputedResearch

OpenAI Publishes 722 Manuscripts From an Unreleased AI Model, Mathematicians Want Proof

OpenAI posted 722 math manuscripts on GitHub from an internal model it has not released, claiming most came from one prompt; MIT's Andrew Sutherland calls the claim unverified.

Why it matters

Nothing in the report ties the model or the manuscripts to a crypto project, token or listed company, so no market exposure is described. The open questions are about verification: the model stays unreleased, most results are not Lean-formalized, and the prompts were withheld.

Where the sources disagree

What the headline count of 722 refers to

An OpenAI spokesperson said almost everything came from a single prompt handed to a single agent

Decrypt

Andrew Sutherland said one-shotting claims with a single agent should be treated as unverified until the model is released and results are replicated

Decrypt

Decrypt's account gives three different counts from the same release: 722 manuscripts, 372 families of related results, and roughly 4,000 problems posed to the model.

Key facts

OpenAI published 722 math manuscripts on GitHub, all produced by an internal model the company has not released.

722 manuscriptsReported by Decrypt

The manuscripts are grouped into 372 families of related results, where a family can bundle a main theorem with companion arguments or alternative proofs.

372 familiesReported by Decrypt

OpenAI says it posed roughly 4,000 problems to the model and kept the outputs it judged significant.

4,000 problemsReported by Decrypt

An OpenAI spokesperson said almost everything came from a single prompt handed to a single AI agent, though some results may have taken multiple attempts.

Disputed figureReported by Decrypt

Only 162 of the 722 papers have a Lean-formalized main result.

162 papersReported by Decrypt

OpenAI warns that some results without a formal proof could have issues.

Reported by Decrypt

Andrew Sutherland, a mathematician at MIT, said claims about one-shotting problems with a single agent should be treated as unverified until the model is released and results can be replicated.

Disputed figureReported by Decrypt

The release omits the prompts that an advisory group at the Institute for Advanced Study had recommended disclosing.

Reported by Decrypt

What happened

Decrypt reported that OpenAI published 722 math manuscripts on GitHub, all produced by an internal model the company has not released. The papers are grouped into 372 families of related results, so 722 counts manuscripts rather than solved problems, and OpenAI says it posed roughly 4,000 problems to the model and kept the outputs it judged significant. An OpenAI spokesperson told Decrypt that almost everything came from a single prompt handed to a single agent, though some may have taken multiple attempts. Only 162 of the 722 papers have a Lean-formalized main result, and OpenAI warns that some unformalized results could have issues. Andrew Sutherland, a mathematician at MIT, told Scientific American that claims about one-shotting problems with a single agent should be treated as unverified until the model is released and others can replicate the results. Decrypt also reported that the release omits the prompts that an advisory group at the Institute for Advanced Study had recommended disclosing.

How the story developed

  1. First report

    First report by Decrypt

    OpenAI Says a Secret AI Model Cracked Hundreds of Open Math Problems in One Prompt—Mathematicians Want Receipts

    Decrypt

  2. Official statement

    OpenAI posts 722 math manuscripts on GitHub from an unreleased internal model

    Decrypt said the GitHub publication happened on Tuesday without giving a date, and that an OpenAI spokesperson claimed nearly all of it came from one prompt to one agent.

    Decrypt

  3. Update

    Decrypt reports the release and mathematician's challenge

    Andrew Sutherland, quoted by Decrypt via Scientific American, said the one-prompt claim is unverified and that people should ask for receipts.

    Decrypt

Affected entities

Original sources