Sebastien Bubeck Once Believed LLMs Couldn't Reason Beyond Training Data
IgorCarron · x · 2026-09-12
A resurfaced 2022 position from Sebastien Bubeck shows he then believed he could prove LLMs were architecturally incapable of reasoning beyond their training data.
Against today's rapid progress in LLM reasoning, that view has been overturned — a reminder of how surprising and unexpected recent advances have been, even to seasoned insiders.
More from Models
- GLM 5.3 Flash is borderline unusable for assistant chat, user says, despite good coding — HornyGooner4402 · 2026-09-12
- TailSFT: skip already-learned SFT examples to boost post-RL pass@k — canondetortugas · 2026-09-12
- Astra beats Bloons TD 6 all 80 rounds where GPT-5.6-Sol loses at 42 — Vjeux · 2026-09-12
- DeepSeek back on top with 6.6T tokens processed in a single day, per opencode — ycombinator · 2026-09-12
- Researcher mocks Astra hype: solved Millennium Problems but can't review a NeurIPS paper — DimitrisPapail · 2026-09-12
- OpenBMB open-sources MiniCPM5-2B, a 2B model that tops agentic benchmarks under 4B — SimplyAnnisa · 2026-09-12