LeCun doubles down: autoregressive LLMs alone won't reach human-level AI, industry moving to continuous-representation search
inductionheads · x · 2026-09-21
Yann LeCun reaffirmed that "auto-regressive LLMs, in and of themselves, will not lead to human-level AI." His two arguments: current reasoning relies on non-autoregressive search, but done in token space — limited and inefficient — while human-like reasoning should search in continuous representation space, a direction he says the industry is moving toward; and current self-improvement methods only work where outputs can be scored without human intervention (math, code, simulable scenarios).
Related event: LeCun Reiterates Autoregressive LLMs Alone Won't Reach Human-Level AI(4 posts)→
More from AGI Musings
- Looking Back: Personal AI Agents' Value Lies in Deep Personalization, Not Errands — pzakin · 2026-09-21
- Theoretical computer scientists are losing their cool as theory is automated by AI first — docmilanfar · 2026-09-21
- Ambitious students seeking a future-proof technical education should major in philosophy — willcb · 2026-09-21
- ThePrimeagen: I fear ASI safety promoters more than misaligned ASI itself — ccerrato147 · 2026-09-21
- Kevin Roose: 'Stochastic Parrot' Discourse Blinded Millions to AI Progress — bigblueboo · 2026-09-21
- How to read imperfect AI-job-impact studies: update Bayesian-style as signals pile up — danielrock · 2026-09-21