LeCun stands firm: auto-regressive LLMs alone won't lead to human-level AI
ylecun · x · 2026-09-20
Yann LeCun reiterates that his statement "auto-regressive LLMs, in and of themselves, will not lead to human-level AI" remains totally true, with three arguments:
- Reasoning: Current systems rely on non-auto-regressive search, but in token space, which is limited and inefficient; human-like reasoning should be search in continuous representation space — a direction the industry is moving toward.
- Self-improvement: Today's self-improvement methods only work where outputs can be scored without human intervention (math, code, simulatable scenarios). Humans and animals learn far more efficiently than current RL.
- Multimodality: Current assistants use separately-trained components (post truncates here).
More from AGI Musings
- After AI's rise in math, solvers are told 'not the way we approved' — kjgeras · 2026-09-20
- Jevons paradox is classic low-end disruption you won't spot from a GPU-rich hyperlab — cramforce · 2026-09-20
- Doomer critique: AI apocalypse is an infinitely renewable narrative that never comes due — Dr_Singularity · 2026-09-20
- AI meets math: defensive mathematicians and mean technical people — kjgeras · 2026-09-20
- Capabilities researchers are beyond shame; safety researchers are my audience — RichardMCNgo · 2026-09-20
- Mathematicians panicking: AI can soon produce PhD-thesis-level work in weeks — lemire · 2026-09-20