François Fleuret Skewers Transformers: Perfect Recall, Pure Statistical Cue Exploitation
francoisfleuret · x · 2026-11-08
AI researcher François Fleuret joked that Transformers "have no idea what this is about" but minimize cross-entropy better by exploiting every statistical cue via flawless recall of the conversation — "the next word is 'Paris' because you said 'croissant' twice in the last 1091 tokens." A witty take on the pattern-matching-vs-understanding debate, from a well-known researcher.
Related event: Geneva Professor Mocks Transformers as Statistical Pattern Matchers(2 posts)→
More from Fun
- One prompt ports the open-source King's Field I engine natively to macOS — Dimillian · 2026-11-08
- "Vibe investing" meme: going long Celsius on pure vibes before earnings — realsohamparekh · 2026-11-08
- Blogger mocks the irony: politicians get nuclear launch codes, ASI gets zero trust — VraserX · 2026-11-08
- Poincare automorphic function visualization built with Claude and DirectX 12 — S_Conradi · 2026-11-08
- DeepSeek's CoT debugging its own 3D render is a joy to read — teortaxesTex · 2026-11-08
- Print your dating profile and join a Chinese park marriage market — suchenzang · 2026-11-08