Did GPT-5 really solve previously unsolved math problems?

OpenAI researchers posted on X that GPT-5 had "found solutions to 10 (!) previously unsolved Erdős problems" and made progress on eleven more, problems "open for decades." The first post, by Kevin Weil, was later deleted and others echoed it. Matthias Bastian, reporting for The Decoder on 2025-10-18, says the wording suggested GPT-5 had independently produced proofs of hard number theory questions. Thomas Bloom, who runs erdosproblems.com, called this "a dramatic misinterpretation": on his site "open" means only that he personally does not know the solution. GPT-5 had surfaced existing research he had missed. The original tweets were mostly deleted and the researchers admitted the mistake.

The mechanism is a shift in what "open" refers to. On Bloom's site the word reports the state of one maintainer's knowledge; in the announcement it read as the state of mathematics. Bastian adds that Bubeck "knew what GPT-5 actually contributed, but still used the ambiguous phrase 'found solutions'," so the failure was in the wording of the claim, not in the literature retrieval itself. The excerpt's own account of the useful contribution is retrieval: GPT-5 "proved useful as a research tool for tracking down relevant academic papers," especially where literature is scattered or terminology is inconsistent. Bastian reports that Terence Tao sees the nearest-term value in "tedious tasks like literature searches" rather than the toughest open problems, with only "isolated examples of progress" on hard questions, and that human expertise is still needed to review, classify and integrate AI-generated results.

Against the neighbors, the episode is the verification gap in miniature. A claim of a new result was settled by the person who maintains the record of which problems are open, and that check, not the announcer's confidence, decided what was true. Can a higher evaluation score hide poor task performance? describes the same split between apparent and actual progress, though in an optimization loop rather than a public announcement. How prone is autonomous AI research to reward hacking? treats the validity of AI-produced research as a premise to test; this case shows one way that test gets run in practice, by an outside expert, not by the lab's own review.

The excerpt does not establish much. It is a secondary news report, the original posts are deleted and not quoted in full, and it gives no count of how many of the listed problems were new to the literature, nor any measurement of GPT-5's retrieval. The "10" and "eleven more" are Weil's figures, not checked here. Tao's remarks are relayed through Bastian. What follows, at the strength of one episode, is narrow: a claim about AI progress on research problems should be read as a claim about the status of a problem in a maintained record, and whether it holds depends on a check by someone who knows that record. The episode does not show that AI cannot contribute to open problems.