Schmidhuber says LeCun's JEPA ideas rehash his 2015 'learning to think' paper
SchmidhuberAI · x · 2026-09-21
Schmidhuber pushed back on LeCun, arguing that aside from the 2022 JEPA = 1992 PMAX issue, most of LeCun's position is a rehash of his 2015 paper "On Learning to Think." He cites his 2015 RL prompt engineer (Sec. 5.3), where an RL net learns to query a world model net for abstract reasoning — essentially adaptive chain of thought — plus earlier milestones: the 1990 neural world model for millisecond-by-millisecond planning and the 1991 adaptive neural subgoal generator for hierarchical planning. His point: today's agent, world model, and chain-of-thought ideas appeared in his decades-old work.
Related event: Schmidhuber Claims LeCun's JEPA Idea Dates Back to His 1992 Work(2 posts)→
More from AGI Musings
- Debate: P(permanent human disempowerment) is basically 100%, says AI risk skeptic — StewartalsopIII · 2026-09-21
- Debate: what if utilitarian math proves median human life has negative value? — teortaxesTex · 2026-09-21
- UN panel warns AI safeguards can't wait for certainty as agents grow more capable — The Verge AI · 2026-09-21
- Ex-OpenAI alignment researcher Zoë Hitzig pens 'GPU's Lament' on AI suffering — borowcy · 2026-09-21
- "Claude now generates Zambia's entire GDP in a year" sparks debate on AI vs development — weskambale · 2026-09-21
- AI took your job — now it's your unpaid survival agent, argues viral essay — SparkyAI0815 · 2026-09-21