LeCun clashes over whether modern AI is still autoregressive token prediction
john__allard · x · 2026-11-23
LeCun insists his old slide still holds, claiming modern AI systems are no longer autoregressive predictors of discrete tokens. johnallard pushes back: tool calls and closed loops were expected (ReAct predates the slide), and reasoning is just autoregressively sampled tokens under a different training signal — so either the slide only covered open-loop LLMs nobody planned to deploy, or it was wrong.
More from Models
- Dev: Gemini 4 Neon rewrote my training kernels and supervised runs; rumors point to bigger models — xennygrimmato_ · 2026-11-23
- Inkling 2 Mini: a 26B local model that runs on laptops, and someone rebuilt Apple Notes around it — simonguozirui · 2026-11-23
- Inkling 2 Mini tested on M4 MacBook: playable 3D game in 8 minutes, fully offline on a plane — simonguozirui · 2026-11-23
- Zig + Metal homebrew engine runs Inkling 2 Mini faster than MLX on Mac — viksit · 2026-11-23
- Reddit demo shows Inkling 2 Mini's looped transformers running locally — mhxb · 2026-11-23
- OpenAI's Jakub Pachocki: pretraining avalanche of advances conceivable by 2027 — scaling01 · 2026-11-23