LeCun claims reasoning isn't autoregressive, peers push back: Fable works differently
paul_cal · x · 2026-09-21
- Yann LeCun restated his position that while underlying LLMs are autoregressive, modern AI systems are not: "reasoning" systems generate many token sequences and pick the best one — a form of non-autoregressive search in discrete token space, which he argues should instead happen in abstract representation space.
- paulcal quote-replied to dispute the premise: that's not how Fable or Astra actually work, and reasoning, planning and self-correction are far more amenable to autoregression than LeCun assumed.
- The exchange hits a core fault line in current reasoning-model design: where the search happens — token space vs. abstract representation space.
More from AGI Musings
- Terence Tao: We Have to Slow Down AI — There's No Reason to Be This Fast — AxomaticallyExtinct · 2026-09-21
- The demand-saturation question: will ever-new consumer goods keep AI-driven demand growing? — binarybits · 2026-09-21
- Terence Tao: 'We have to slow down AI. The pace is insane' — AxomaticallyExtinct · 2026-09-21
- binarybits: only the military scenario holds in the AI explosion debate — binarybits · 2026-09-21
- Slowing AI Won't Stop Superintelligence — It'll Lock Intelligence Behind $3T Companies — SucceededMind · 2026-09-21
- Devs clash over whether beginners should still learn to code as AI automates code review — mgill25 · 2026-09-21