Background
On October 2, Meta AI Research published a blog post titled Solving Open Research Problems Together, covering six mathematics papers completed by human mathematicians working with Muse Spark. Meta stated that five of the six answer previously open research questions.
The catch: in the same blog post, Meta itself lists independent solutions to those problems, all dated earlier than October 2.
The overlap table
| Paper | Field | Meta's claim | Existing independent result | |---|---|---|---| | Strict threshold for Gaussian ellipsoid fitting | Probability | Solves open problem | Three groups, August | | Finite-time blowup for biharmonic NLS | PDE | Solves 2015 problem | None recorded | | Semiabelian groups need not be monomial | Group theory | Refutes Kida 2024 conjecture | Agent, September 16 | | Compactness of cyclic relaxations | Optimization | Solves Del Pia et al. 2026 question | None recorded | | String two-point functions / curve height functions | Arithmetic physics | Extension, not counted | Independent result acknowledged | | 3-dimensional counterexamples | Non-associative algebra | Refutes conjecture | August 12 |
Inference: of the five open problems Meta claimed, three already had public answers before the papers were posted. Only two have no recorded overlap: finite-time blowup for biharmonic NLS, and compactness of cyclic relaxations.
Ellipsoid fitting: four solutions in seven weeks
The conjecture, due to Saunderson, Parrilo, Willsky and others, was open for years. Then in August it fell repeatedly:
- August 10: Misiakiewicz and Wen posted a proof of the SAT/UNSAT phase transition at d²/4 (arXiv:2608.10184).
- August 12: de la Cerda, Potechin, Tulsiani, and Xu posted their own strict threshold proof (arXiv:2608.12415), up to a vanishing multiplicative factor.
- August 27: Koehler and Sohn posted a more general result (arXiv:2608.27372), with the Gaussian threshold 1/4 as a special case.
- October 2: Meta's paper appears, leaving the exact critical point n/d² → 1/4 as unresolved.
- Paper titles, authors, fields, and Meta's division-of-labor statements from the Meta AI Research blog of October 2.
- Independent solutions' arXiv IDs and dates verified against the blog's footnotes: 2608.10184, 2608.12415, 2608.27372, 2609.25023.
- The five-vs-three discrepancy reflects both Meta's self-description and public reviewer challenges; this post does not adjudicate. Some reviewers argue only about three of the six were genuinely open.
- None of the six papers has journal peer review. The arithmetic-physics paper is not counted by Meta as an open-problem answer, so the five-vs-six wording is not contradictory.
Meta's footnotes list all three papers, noting each used different methods independently. The blog's own wording: "after completing our work, we learned that other teams outside Meta had independently announced solutions to some of these problems using different methods."
The right framing is not "we solved an open problem," but "this problem matured in August, and four groups arrived within weeks of each other."
The agent that got there first
In 2024, M. Kida conjectured that every finite semiabelian group must be monomial. Meta's paper *Semiabelian Groups Need Not Be Monomial* gives a counterexample: SmallGroup(384, 20127), with 384 elements, found via a GAP search program written by Muse Spark.
But Meta's footnote records something else: on September 16, an agent called Nilradical reported a different counterexample, logged in the kourovka-lean repository as Problem 21.68 of the Kourovka Notebook — a semiabelian group of order 2592 with an 8-dimensional irreducible non-monomial representation, accompanied by a kernel-checked Lean proof.
Meta's group is smaller (384 vs 2592), but the verification levels differ: Meta's papers involve no Lean formalization and rely on human peer review, while Nilradical delivered a machine-checkable proof.
The 3D counterexample paper has an independent result too: Xing-Yu Hu and Ran Wen posted a family of 3-dimensional counterexamples on August 12 (arXiv:2609.25023), noting three is the smallest dimension where the conjecture can fail. Meta's author Andres Barei also found 3D counterexamples of the same size.
No scaffolding — harder to replicate than the results
The most notable technical detail is not mathematical. The mathematicians used Muse Spark 1.1 and 1.2 in Thinking mode through the ordinary meta.ai chat interface; Meta explicitly states no dedicated research scaffolding was built — no custom proof-checking environment, no problem-specific fine-tuning.
The blog describes the division of labor in three rules: a team of mathematicians guides the research and explores ideas and constructions with the model; a second team reviews their work; every paper marks which paragraphs were primarily researcher-drafted vs. AI-drafted.
That paragraph-level labeling is stricter than most journals' current requirements. In the same week arXiv capped authors at two submissions per month due to AI-assisted volume; if AI-assisted papers routinely carried such labels, review difficulty would drop markedly.
One cross-over in authorship: Andres Barei reviewed the group theory paper while authoring the non-associative algebra paper; Nicolas Jaramillo Torres reviewed Barei's paper while co-authoring the string theory paper.
The two genuinely non-overlapping results
The PDE paper addresses the mass-critical biharmonic nonlinear Schrödinger equation, answering a 2015 question of Boulenger and Lenzmann for N ≥ 2 and confirming a prediction from 2002 computer simulations. Author Leonard Dinh chose the problem and supplied the key proof idea; the model assisted with computations and proof revisions.
The optimization paper asks when cyclic relaxations of binary polynomial optimization are exact: exactness is characterized by a Venn-diagram structure — when each pairwise-overlap region of three circles contains exactly one decision variable, the relaxation is perfect; more than one, and a gap appears.
Meta's stated goal: "not to mass-produce papers, but to empower researchers."
Assessment: neither paper has yet undergone broad community scrutiny; treating them as promising preprints is appropriate. Notably, both were won by argument, not by counterexample. Counterexample results can be mechanically verified within days; long proofs must wait for journal review. Lumping both into one tally obscures the difficulty difference.
Takeaway
Meta's own papers supply the chewier fact: when several teams attack the same list of open problems simultaneously, pre-publication literature checking matters more than the proof itself.
Counterexample search will remain AI's most reliable mathematical foothold — define the space, generate candidates, test, keep; it is search, and models that tirelessly write and run search programs are matched to that shape. And formal verification is becoming the differentiator: Nilradical delivered a kernel-checked Lean proof, Meta delivered human review. Expect AI-involved mathematical announcements to increasingly require machine-checkable artifacts.