Update: OpenAI says its AI disproved the Exponential Time Hypothesis
Found Oct 16, Lean-checked Oct 17 — but from a model no outsider can run. Mathematicians urge a boycott; one warns ECDSA could fall in months, not years.
Copy markdown
Three claimed results, one unreleased model
OpenAI says its internal models cracked the 1991 local smoothing conjecture, disproved the Exponential Time Hypothesis, and produced an existence proof for QCD. Every result came from models it hasn't shipped — you can read the claims, but you can't reproduce them.
ETH 'false': machine-checked, not yet human-checked
The complexity-theory disproof was found Oct 16 and verified in Lean the next day — but Lean only checks each logical step, not that the statement means what it should. A working group including two of ETH's original authors is still reviewing, and a full verdict takes months.
The backstory: ~719 manuscripts, 42% checked
This caps a larger dump of 372 result 'families' (719 manuscripts, three already pulled for a sign error) from the same unreleased model, nearly all from a single prompt to a single agent. About 42% carry Lean proofs; OpenAI admits the rest 'could have issues.'
Mathematicians push back
An advisory group asked labs to publish prompts and model names and stop testing hard problems on proprietary models; OpenAI shared compute figures and 10 reasoning summaries but no prompts, saying it isn't bound. The Association for Human Mathematics called the release a 'demonstration of power' and urged members to stop working with OpenAI.
The crypto angle to actually watch
Ethereum's Justin Drake warned that, worst case, a result like this could put ECDSA at risk 'in months, not years' — though OpenAI reports no attack on ECDSA or RSA yet. Nothing to do today, but a reason to track whether the ETH disproof survives outside review.