~1% error rate in AI math proofs: researchers debate whether the fallout is tiny or much larger
srchvrs · x · 2026-10-09
Responding to errors found (including the NS problem) in an AI system's mathematical proofs, srchvrs notes that more researchers have surfaced additional errors, though some errors in the error-finding itself and some fixable ones complicate the picture; the open question is whether the fallout is on the order of 1% or much larger.
The quoted tweet mocks the optimism: a 1% error/retraction rate is tiny "in the grand scheme of things" — and can statistically round up to a 100% success rate, sarcastically declaring all of maths "solved." A pointed AI-community debate on proof reliability.
Related event: OpenAI's Math Documents Criticized as ~1% of Proves Flawed(3 posts)→
More from Fun
- Researcher jokes math got 'bleached' before coral reefs — fhuszar · 2026-10-10
- Terry Tao reportedly unhappy with OpenAI over math problems — but his blog post looks LLM-written — BLUECOW009 · 2026-10-10
- Ex-Google designer: AI flattens design disciplines, integrated creativity compounds UX — soleio · 2026-10-10
- AI safety drama: accusation claims safety lab staff are staging a fake catastrophic AI incident for regulation — MikePFrank · 2026-10-10
- Doubleword's unhinged 'LET THERE BE TOKENS' ad nails the overpriced-token pain — AccBalanced · 2026-10-10
- Google's Astra crushes Claude Opus 15-4 at chess despite being 360 Elo points weaker — pvncher · 2026-10-10