Meta's AI helped solve open math problems. Others got there first on three.

Meta published six math research papers on October 2, each written by mathematicians working with its Muse Spark models. The company says five of them answer questions that had been open. That's a bigger claim than winning another math olympiad, and Meta says so itself: in open research, it writes, there's no answer key.

The setup is the surprising part. The mathematicians didn't use a custom research system. They used Muse Spark 1.1 and 1.2 in Thinking Mode through the regular meta.ai chat interface, the same one anyone can open. They chose the problems, steered the work and checked every argument, and a separate team of mathematicians reviewed the results. Each paper marks which passages a human drafted and which ones the AI drafted.

The problems cover a lot of ground. In group theory, Joseph Phillip Brennan and Milana Golich tested a conjecture M. Kida proposed in 2024. Muse Spark wrote a search program that found a group with 384 elements that breaks the conjecture, and the mathematicians verified it. In differential equations, Leonard Dinh proved that a certain class of wave solutions blows up in finite time, a question left open since 2015. Aykut Arslan pinned down a sharp threshold for fitting random points onto an ellipsoid. Other papers covered optimization and a strange corner of algebra called evolution algebras. The sixth explored a connection between string theory and number theory that Yuri Manin raised in the 1980s, without claiming to settle it.

Then the catch. Meta's own post says other teams worked on some of the same problems at the same time, using different methods, and posted results in August and September. Meta says its work was done independently. Jason Dean Lee put it more bluntly on X the next day: "Actually 3 of 6 were already resolved."

Both things can be true. A result found independently still shows what the tool can do. But "solved an open problem" and "matched a result someone else posted a month earlier" are different claims, and only the first one makes headlines.

What's left is still useful, and it's checkable in a way most AI-science claims aren't. Every paper says who did what. The counterexample is a concrete object anyone can test. Plenty of earlier "AI does science" announcements came with far less proof.

What happens next is the slow part. The papers need to go through peer review, and the field needs to sort out which results were new when Meta posted them. If the five-of-six claim holds up even partly, the bigger story is the tool itself: a chat window that mathematicians already have open.