Meta's AI helped answer five open math problems. Three of them were also answered by someone else.

What Meta actually published

On 2 October 2026, Meta's research blog posted "Solving Open Research Problems Together," describing six mathematics papers written by named mathematicians working with Muse Spark. The framing in Meta's own words is careful: "Today, we're sharing six such papers from that collaboration. Five present answers to previously open research questions."

Six papers. Five answers. The subjects are real research mathematics, not benchmark puzzles: a sharp threshold for fitting random Gaussian points to an ellipsoid in high dimensions, finite-time wave collapse in a biharmonic nonlinear Schrödinger equation, a counterexample to a 2024 conjecture in group theory, an exactness condition for a binary polynomial relaxation, a bridge between p-adic string theory and number theory, and a counterexample in evolution algebras.

The day after, Meta's chief AI officer Alexandr Wang posted the list on X under a different sentence: "mathematicians and muse spark collaborated to solve 6 open problems in math." Six, not five, and no mention of the caveats that are sitting in Meta's own blog post. His post has been seen more than 440,000 times.

The acknowledgements are the story

Meta disclosed something most companies would have buried. One paragraph in: "After completing our work, we learned that other teams outside Meta had independently announced solutions to some of the same problems using different approaches."

That understates it, and the per-paper notes are where the detail lives.

On the ellipsoid threshold, Meta names three independent concurrent works posted in August 2026. Misiakiewicz and Wen "independently proved the Gaussian threshold." De la Cerda, Potechin, Tulsiani and Xu established it "up to a vanishing multiplicative factor." Koehler and Sohn obtained a broader universality result that contains the Gaussian threshold as a special case. Meta's paper also states that "the behavior exactly at the threshold remains unresolved."

On the group theory counterexample, Meta acknowledges "the AI agent Nilradical, which reported a different counterexample to the same conjecture on September 16, 2026." A different AI got there two weeks earlier.

On evolution algebras, Meta acknowledges "independent work by Hu and Wen, who reported counterexamples to the same conjecture."

Three of the five answers have acknowledged independent solutions from outside Meta, and in at least two cases those solutions were public first. That does not make Meta's work invalid. Simultaneous discovery is normal and the papers were developed independently, which Meta says plainly. It does mean "solve 6 open problems" is not a defensible summary of the blog post it links to.

The methodology is the genuinely new part

Strip away the announcement and what remains is more interesting than the headline number.

The mathematicians used Muse Spark 1.1 and 1.2 in Thinking Mode "through the regular meta.ai chat interface, with no custom research scaffold." No agent harness, no bespoke research platform, no tool-calling pipeline built for the task. That is a meaningful detail in a year when every lab's science claim arrives wrapped in custom infrastructure.

Meta also set out four working principles, and the third is the one other labs should copy: "Each paper clearly marks which passages were primarily drafted by researchers and which were drafted by AI." The others: mathematicians guided the research, a second separate group reviewed the work, and each paper credits the earlier research it builds on.

Read the per-paper notes and the division of labor varies a lot, which is the honest texture usually lost in summaries. In the group theory paper, "Muse Spark generated the search program in GAP," the software that then found the 384-element counterexample, while Milana Golich and her collaborators verified it and completed the argument. In the arithmetic physics paper, the model "generated candidate proofs and drafted three core technical sections," which the researchers checked and corrected. In the differential equations paper, Leonard Dinh "chose the problem and key proof ideas" and Muse Spark worked through calculations and revised the proof.

So the model wrote search code in one case, drafted technical prose in another, and did calculation and revision in a third. In none of them did it pick the problem.

How to read a claim like this

Meta opens by noting that its models "achieved gold-medal-level performance across five high-school Olympiad competitions in mathematics, physics, and chemistry," then draws the right distinction itself: "Competition problems can be incredibly difficult, but those problems already have a solution. Open research is different. There is no answer key."

That is the correct frame, and it is why Olympiad scores never told you much about research capability. It is the same gap between a leaderboard number and real-world usefulness we covered in how AI benchmarks work.

The pattern here also rhymes with the last big AI-in-science claim we examined, where Claude flagged a "CRISPR-like" enzyme system that turned out to be CRISPR-like in a narrower sense than the announcement implied. In both cases the underlying work was real and the framing ran ahead of it. The difference this time is that Meta's own blog post contains the correction. You just have to read past the first screen.

Our take

This is the most carefully documented AI-assisted research release any lab has put out this year, and Meta undercut it within 24 hours. The blog post names every human, marks which passages the model drafted, credits prior art, flags three competing independent results and admits one question is still open at the threshold. That is how this should be done. Then the chief AI officer compressed it into "solve 6 open problems" for 440,000 people, which is not what the papers say.

If you want to judge whether AI is contributing to mathematics, the number to carry away is not six, or even five. It is that named mathematicians chose every problem, steered every proof, caught the gaps, and had colleagues re-check the result, and that the model's contribution was specific and different in each paper. That is a real and useful finding. It is also much harder to tweet.