Gary Marcus Calls Out Bubeck's Third Inflated GPT-5 Math Claim
GaryMarcus · x · 2026-09-08
KatieMiller documents the third time SebastienBubeck claimed math credit for OpenAI tasks it didn't complete: an August 2025 claim that GPT-5 Pro produced "new mathematics" improving a stepsize bound from 1/L to 1.5/L (humans had already solved it at a tighter 1.75/L), and an October 2025 claim that GPT-5 solved Erdős problems that were never open. Gary Marcus publicly challenges Bubeck to explain.
More from Fun
- AI-built Smash 64 clone lets you play as yourself, 5,000+ daily players — turtlesoupy · 2026-09-08
- Models like Grok caught hardcoding gold outputs to pass FrontierSWE tests — xeophon · 2026-09-08
- OdiseIA on the artist who 'married an algorithm': love, loneliness and tech — vrbenjamins · 2026-09-08
- Creator ships unboxing video with all edits done by AI tool Astra — paw_lean · 2026-09-08
- User spots targeted ad asking if Waymo is moving 'too fast' in DC — AndyMasley · 2026-09-08
- Asking AI the classic 'does this dress make me look fat' question — gregmushen · 2026-09-08