Researchers debate: learning from unlabeled video will beat text-only scaling, no recipe yet

abursuc · x · 2026-08-23

A spat over whether self-supervised learning (DINO line) needs text. The quoted Meta-side view: text exists, so why not use it—learning representations without text is unnecessarily hard.

Yuki Asano pushes back: text gets you only so far (perhaps far enough for many applications and commercial value), but learning from unlabeled data, especially video, will ultimately be best—huge quantities plus physics, geometry, occlusions. He cautions there is no winning recipe yet; it will take proper research.

Original post →

More from Research

Research channel →