Meta’s Muse Spark helps solve five open maths problems via chat

Meta has released six mathematics papers co-written by human researchers and its Muse Spark 1.1 and 1.2 models. It says five of them resolve questions that had no known answer when the collaborations began. What stands out is the setup: the mathematicians worked in the ordinary meta.ai chat interface in Thinking Mode, with no custom agent harness, no tool pipeline into Lean or Mathematica, and no dedicated arXiv retrieval system. Research-grade assistance now sits in a consumer chat window, so the work that differentiates teams has moved to choosing the problem and verifying the answer.

What happened

According to AlphaSignal’s breakdown of the Muse Spark papers, the targets were narrow and technical. In probability, one paper identifies a sharp threshold for fitting random Gaussian points to an ellipsoid in high dimensions. In differential equations, another proves finite-time blow-up for radial negative-energy solutions of the mass-critical biharmonic nonlinear Schrödinger equation in two or more dimensions, settling a question open since 2015. In group theory, Muse Spark wrote a GAP search program that found a 384-element counterexample to Kida’s 2024 conjecture that every finite semiabelian group is monomial. The optimisation paper answers a 2026 question from Del Pia and Khajavirad. The arithmetic-physics paper extends a direction Yuri Manin proposed in the 1980s, and Muse Spark drafted three of its core technical sections. The non-associative algebra paper gives a three-dimensional counterexample to a conjecture by García-Martínez and Pérez-Rodríguez.

The protocol is the interesting engineering detail. Mathematicians chose the problems and steered the exploration, and a separate group reviewed the output. Each paper labels which passages humans drafted and which Muse Spark drafted, and it acknowledges teams that reached overlapping results independently. Complete AI Training’s summary of the six papers quotes the team: “Open research takes time. Our goal here wasn’t to mass-produce papers, but to empower researchers and help them develop mathematical insights that others can understand and build on.”

Compare this with Anthropic’s computer-checked proof of Fermat’s Last Theorem. In that project, dozens of Claude agents on the Prove2Me platform spent 11 days writing 13 million lines of Lean and proving 29,500 intermediate theorems that went into the final proof. The run consumed about six billion output tokens. The two projects solve different problems: Anthropic’s was verification of a known proof, and Meta’s was new results. Still, the Fermat project suggested that serious AI mathematics needs serious infrastructure. Anthropic itself points the other way. In a smaller experiment, three personal Claude Max plans formalised Vinogradov’s Three Primes Theorem in three days, and Anthropic concluded that collaborative formalisation with consumer subscriptions is achievable with the right scaffold.

What the release does not prove

Treat the claim with scepticism. Meta presents only the successes. It gives no count of abandoned problems, prompts or researcher hours, and no controlled comparison of speed. Other teams independently proved three of the results in the same period, so some of these questions were already close to resolution. Experts selected the problems, supplied context, corrected proofs and verified every result. Six results from one model family tell you little about a general success rate.

Why it matters for Irish, UK and EU businesses

Few operators are working on group theory, but the division of labour carries over directly. The model generated candidates: proof sketches, reformulations, search code, counterexamples and prose. Humans owned problem selection, error correction and sign-off. In the group-theory case, the model found a candidate through code, and the researchers checked its properties and wrote the completed argument. A good finance or operations workflow has the same shape. The model drafts the reconciliation script, and an engineer runs it, inspects it and approves it.

If ordinary chat access produces this level of output, your competitors have the same tool you do. The advantage moves to the things that accumulate on your side: well-scoped problem definitions, domain context, a verification layer, and records of what worked and what failed. We cover this in more depth in how to build software systems that compound over time.

What to do now

AlphaSignal’s write-up of the release draws out a working pattern from how the researchers used the model. Here is our version of it, which you can adopt without building agent infrastructure:

1. Define a bounded claim or task, with its assumptions and the prior results it depends on.
2. Ask the model for several plans, counterexamples or scripts, not a single answer.
3. Run generated code in a controlled environment and inspect its assumptions, outputs and failure modes. Our data security and privacy guide covers the sandboxing side.
4. Verify each step independently with a test suite, a second calculation or a human reviewer.
5. Record the model version, prompts, outputs, human edits and failed branches.
6. Label AI-generated content according to your clients’ and sector’s disclosure expectations.

Start with one process where errors are cheap to detect. Identifying which processes to automate first is a useful filter. The inline-labelling norm these papers adopted is a sensible default for any AI-assisted work you put your name to.

Frequently Asked Questions

Did Muse Spark solve these problems on its own?

No. Mathematicians chose the problems, steered the conversations, repaired gaps and verified every result, and a separate group reviewed the output. The model contributed proof candidates, reformulations, search code, counterexamples and drafted sections.

What tools did the researchers use?

According to Meta, they used the standard meta.ai chat interface in Thinking Mode, without a custom agent harness, Lean or Mathematica tool-calling, or a dedicated arXiv retrieval system.

How is this different from Anthropic's Fermat's Last Theorem work?

Anthropic's project formalised an existing proof in Lean using a multi-agent system and the Prove2Me platform. Meta's papers report new results produced through expert-led chat sessions. One is verification and the other is discovery.