Ephemerent AtlasEphemerent Education

Vector Meridian / Field film 04

Yann LeCun Thinks AI Is Learning the Wrong Way — JEPA Explained

A representation-learning case study following the JEPA program from image and video models toward planning and latent-variable claims. Mathematical and benchmark claims remain visibly time-scoped pending independent review.

Duration
32:54
Kind
explainer
Playback
Privacy gated
Learning status
field trial

No YouTube connection yet. Loading the player contacts YouTube. Captions default on; playback never starts automatically.

Open directly on YouTube

Prediction before playback

What could a model ignore if its task is to predict useful hidden state rather than every pixel?

ObserveSeparate the training objective, the representation claim, and the downstream planning claim.

Accessible account

Read without loading the player.

Owner caption export pending.This field-trial page provides a reviewed-scope static account; use YouTube captions after loading the player for the current timed transcript.

  1. Accessible account: this episode examines joint-embedding prediction, representation learning, and reported planning results while separating objectives, assumptions, and downstream claims. The complete owner-exported timed transcript is pending.

Local field notes

Keep the claim and the evidence together.

Notes and bookmarks stay in this browser. They are not scores and never become mastery evidence.