Math's Most-Watched AI Tracker Just Went Dark. Nobody Said Why.

On June 30, the GitHub wiki where Terence Tao had spent nine months logging every AI contribution to Paul Erdős's 1,217 open problems got one final edit: a commit titled, simply, "freeze." No announcement, no explanation, only a stop, at 553 problems solved. 10 days later, a rival team claimed 64 parallel AI agents had cracked a 50-year-old conjecture with a documented history of producing fake proofs, and posted the result to X. Not one mathematician has verified it yet.

That juxtaposition is a better guide to where parallel AI agents actually stand in mathematics right now than either headline alone. It is not, contrary to a version of this story you may have heard, "Terence Tao's agents solved 50 of 1,000 Erdős problems." Tao does not run agents; he curates the world's most-watched scoreboard for other people's, verifies results, and has spent the past six months publicly arguing that the field is measuring the wrong thing. His own framework for what that means is the actual story here, and it is more interesting than the number.

What actually happened, and what didn't

erdosproblems.com, the community database Tao helps maintain, tracks 1,217 of Paul Erdős's open problems; 553 were marked solved as of the wiki's June 30 freeze date, 45% of the total. AI touched a real but bounded slice of that progress — not a majority, and not evenly. Tao's own estimate, stated plainly in a January 14, 2026 post: only "roughly 1–2%" of outstanding Erdős problems are "simple enough to be amenable to current AI tools operated with minimal human intervention." Small fraction, large denominator: with over 600 problems still open, even that thin slice adds up to real, if unglamorous, progress.

The archetypal case is Problem #728, the one usually cited as the first Erdős problem "solved autonomously by AI." Read the actual timeline and it is a pipeline, not a lone genius agent: on January 4, 2026, ChatGPT-5.2 Pro produced a proof sketch for a restricted case; Harmonic's Aristotle formalized it in Lean; a human operator, Kevin Barreto, iterated with ChatGPT to extend the argument and rewrite it into readable exposition. Tao verified the result and later added a correction noting the underlying method turned out to closely match a 2014 paper by Carl Pomerance that nobody had connected to the problem — the AI-assisted solve was still first, but not created from nothing. The same week, the technique was adapted to solve two more problems, #729 and #397, the "three problems in seven days" some coverage reported. All three are the pattern: technical, narrow, machine-checkable in Lean, and built by a human directing several different AI tools in sequence.

Not every case fits that "long tail cleanup" description, which is worth stating precisely because it complicates the tidy narrative. Tao flags Problem #1196— a 1968 conjecture of Erdős, Sárközy and Szemerédi about primitive sets, open for 58 years — as "the possible exception" where AI involvement produced actual mathematical progress rather than just clearing a backlog. GPT-5.4 Pro, prompted by Liam Price, found a genuinely new proof technique — what Tao calls a "downwards von Mangoldt chain" — that had not appeared in the prior literature and went on to help resolve two further conjectures. That is the honest range: mostly narrow cleanup, occasionally a real new idea.

Two systems, both built around "64," two different epistemic statuses

The clearest evidence of genuine parallel, multi-agent architecture in mathematical AI right now comes from DeepMind, not from the more viral headline. A May 2026 paper, "Advancing Mathematics Research with AI-Driven Formal Proof Search" (arXiv 2605.22763), describes a framework the paper itself names AlphaProof Nexus: an asynchronous controller running a pool of rating agents that continuously score sketches from a shared database, sampling the top 64 highest-scoring candidates via Predictor-Upper-Confidence-Bound search, with up to 10 prover subagents dispatched per problem attempt. Its most capable configuration "autonomously resolved 9 of 353 open Erdős problems at the per-problem cost of a few hundred dollars, proved 44/492 OEIS conjectures," per the paper's own abstract — every result formally checked in Lean before being counted.

Ten days ago, a different team published a more dramatic claim through a less rigorous channel. An OpenAI researcher posted on Xthat GPT-5.6 Sol Ultra had produced a proof of the Cycle Double Cover Conjecture — posed independently by Szekeres in 1973 and Seymour in 1979, open for roughly 50 years — using 64 concurrent subagents in under an hour, with the prompt and proof PDF hosted directly on OpenAI's CDN. No formal OpenAI blog post accompanied it. No peer-reviewed verification exists yet. And the conjecture itself has a specific, well-known history of producing plausible-looking proofs that later collapsed under scrutiny — exactly the failure mode mathematicians are most worried about when a proof arrives with no formal verification attached. Two systems, a similar order of parallel compute, and two completely different standards of evidence: one Lean-checked and peer-reviewable, one a claim on a conjecture famous for false proofs.

What happens when four systems race the same 10 unpublished problems, graded blind

The most rigorous public test of parallel AI systems on real research mathematics to date is First Proof, an independent nonprofit run by mathematicians from Stanford, Berkeley, UT Austin and Harvard, funded in part by unrestricted donations from Anthropic and OpenAI (with Google.org funding pending, per its own report). Its second-batch reporttested four named systems — ProofCouncil (ETH Zürich/Aarhus, built on GPT-5.5 Pro), a UCLA "Moonshot Harness" with Tao himself as co-PI, plain ChatGPT-5.5 Pro at maximum reasoning effort, and Princeton's Momus — against 10 genuinely unpublished research problems solicited from working mathematicians between March and May 2026, several tied to forthcoming papers. Each system ran isolated for up to 24 hours; roughly 30 expert referees then graded the outputs double-blind at Harvard's Center of Mathematical Sciences and Applications in early June.

The aggregate result: across all four systems combined, 7 of the 10 problems received at least one passing grade. One problem, a metric-geometry question from Larry Guth, defeated every system completely. One submission, on a stochastic-PDE problem, used a genuinely novel approach that exceeded the human reference solution's intermediate result and impressed referees. And cost tracked nothing about quality:

A 40x more expensive, purpose-built multi-agent harness did not reliably outscore a single prompt to an off-the-shelf model — the report's own per-problem grades show the UCLA system landing "essentially flawless" on one problem and "reject" on four others. Money spent on more parallel search bought variance, not a guaranteed edge.

The crowd version: 177 teams, one shared verifier

A different, more genuinely competitive structure is running in parallel right now through SAIR, the Foundation for Science and AI Research that Tao co-founded. Its third open competition, the Inverse Galois Problem in Degree 24, asks any team — using "any sort of computational tool, including AI," explicitly permitted as a black box — to find integer polynomials realizing as many as possible of 165,836 possible Galois-group configurations, starting from a frozen baseline of just 622 known. Every submission is independently checked by Magma and PARI/GP against the LMFDB database, then scored relative to the best known result and to every other team's submission. As of today, 177 independently registered teams are competing, with a live, continuously updated leaderboard, ahead of an August 15 close. This is the shape "parallel agents" takes when the parallelism is between teams and organizations, not just between subagents inside one system — and the verifier is a fixed, symbolic-computation oracle rather than a panel of human referees.

The field's actual answer to all of this

The mathematical community's response to six months of this has not been to build bigger models. On June 2, the Leiden Declaration on Artificial Intelligence and Mathematics went live, warning specifically that AI risks "producing plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs." It has 3,016 signatories as of today, is formally endorsed by the International Mathematical Union, and Tao himself signed it. Two weeks earlier, arXiv's computer-science section chair announced — via a social-media post, not a corporate press release, an informality worth noting — a one-year submission ban for authors with "incontrovertible evidence" of uncaught AI errors making it into a paper: hallucinated citations, or literally leftover chatbot text like "here is a 200 word summary; would you like me to make any changes?" left in a submitted PDF.

Both responses target the same failure mode: cheap, plausible-sounding, ungated output. Neither targets AI use itself, which every one of these projects — First Proof, SAIR, the Leiden Declaration's own recommendations — explicitly permits and even encourages.

Where the edge really is

Tao has been making the same underlying point since April, in more detail than any single data point above: mathematics is moving from an era of proof scarcity to an era of proof abundance, and its infrastructure has not caught up. His framing splits problem-solving into three stages — generation, verification, digestion — and argues the first two are now substantially automated while the third is not: "the first two are being automated far more successfully than the third," he wrote in May, "leading to a new experience of 'proof indigestion' in which proofs are being generated and even verified without being digested." His objection to skipping that step isn't aesthetic: "if one decided to automate and optimize digestion by minimizing the chewing needed, the logical solution would be to put all of our meals into a blender and serve them through feeding tubes. This technically solves the problem of indigestion, but is not a popular option."

That is the actual answer to where the edge in parallel autonomous math agents is right now. It isn't agent count — AlphaProof Nexus and OpenAI's Cycle Double Cover claim both run around 64 parallel processes, and one is peer-reviewable science while the other is an unverified claim on a conjecture famous for producing exactly that kind of claim. It isn't raw solve rate either — First Proof's $12-per-problem system and its $480-per-problem system landed in roughly the same place. The edge is verification and digestion capacity: Lean formalization, blind human refereeing, a symbolic-computation oracle checking a black-box submission, 3,016 mathematicians agreeing on what counts as a disclosed method. That is expensive, slow, and exactly the part nobody has figured out how to run in parallel yet.