LeCun's CoRL slide: 'If you are interested in human-level AI, don't work on LLMs'
Heavy-Ad9232 · reddit · 2026-11-15
A photo from CoRL shows Yann LeCun's slide of recommendations to AI scientists: abandon generative models for joint-embedding architectures, probabilistic models for energy-based models, contrastive methods for regularized methods, and RL for model-predictive control (use RL only to adjust the world model or critic when planning fails). The red-font punchline: 'If you are interested in human-level AI, don't work on LLMs.'
More from AGI Musings
- Epoch ECI Trend Fit: Anthropic's Latest Model Runs ~5 Months Ahead — peterwildeford · 2026-11-15
- Ex-Citizen Quant Head Dan Morillo: AI Outputs Consensus, Not Alpha — minhnguyen · 2026-11-15
- 104 job titles across 80 years: an interactive family tree of software roles from 'Computer' to 'Context Engineer' — bibryam · 2026-11-15
- RL trade-offs will split LLMs into role-specific personalities, argues Redditor — uriwa · 2026-11-15
- "Double-blind review doesn't work": reviewers see author names, and rival students reject competitors — Temporary_Switch_339 · 2026-11-15
- Betsey Stevenson's post-AGI talk: a 10% income rise lifts well-being the same everywhere — MartinSignoux · 2026-11-15