LeCun's new paper: 109M-param text JEPA beats 72B LLMs at aligning with vision
heyshrutimishra · x · 2026-11-02
Yann LeCun, now at his own Paris lab after leaving Meta, and Randall Balestriero turned the Platonic Representation Hypothesis into a theorem in a new paper.
- The key variable isn't scale but view quality: image-caption pairs are good views that align models fast; two paragraphs from the same document (the regime of next-token prediction) are bad views that align slowly.
- Good views need 42x fewer training samples than next-token prediction to reach the same alignment on a 940M-image corpus.
- A 109M-parameter text JEPA aligned with DINOv2 and CLIP image encoders better than every LLM tested, up to 72B parameters — 660x smaller.
- A rotation of the separately trained text JEPA works as an image classifier (0.838 CDER on COCO) and beats all ≤8B LLMs' reported zero-shot top-5 (64% vs 56%) on a pair-free ImageNet-100 benchmark.
The paper argues next-token prediction is a remarkably inefficient route to platonic representations, directly challenging the field's scaling-first assumption.
Related event: LeCun's New Paper: 109M-Parameter Model Beats 72B LLM in Representation(2 posts)→
More from AGI Musings
- The AI comms dilemma: with the public watching, anything you say gets misread — catherineols · 2026-11-02
- AI is eating cost centers: mid-career advice says move to profit centers now — emeka_boris · 2026-11-02
- Pinboard founder: alignment debates echo 1947 atomic scientists' flawed binaries — nitarshan · 2026-11-02
- Eric Topol asks whether medicine will face its own 'bitter lesson' from AI — EricTopol · 2026-11-02
- Measuring AI agents wrong: task completion counts less than the human verification work left behind — alifcoder · 2026-11-02
- Kyle Kingsbury: LLMs erode trust in software and its documentation — Simon Willison · 2026-11-02