LeCun's new paper claims a 109M-parameter model beats a 72B model
heyshrutimishra · x · 2026-11-02
Yann LeCun, after leaving Meta in November to start his own lab in Paris, published a new paper claiming a model with just 109 million parameters outperformed a 72-billion-parameter model.
- Framed as evidence for his long-standing thesis that scale alone doesn't give models world understanding, countering the post-2024 scaling consensus
- Meta bet the other way; the post calls this the proof nobody in the industry wants to say out loud
- Paper link attached; training details not expanded in the original thread
Related event: LeCun's New Paper: 109M-Parameter Model Beats 72B LLM in Representation(2 posts)→
More from Companies & People
- Suhail is hiring an AI agent engineer to build open multi-agent systems in SF — garrytan · 2026-11-02
- Ex-$500k/Year AI Engineer With 28 Years of Experience Says He's Now Homeless — ericelliott_ · 2026-11-02
- How to get strategic: bet often with real money, secure screw-you money, commit for 10 years — itsOmSarraf_ · 2026-11-02
- An ordinary PhD's 2026 North America job search, documented — _AndrewZhao · 2026-11-02
- Reddit forces logins to stop scraping; critics say the content's already been sold — nptacek · 2026-11-02
- Eric Elliott: I'd rather write about agent security than my own survival story — ericelliott_ · 2026-11-02