Chapter 0 — Before you begin: what this course promises

First, the whole picture

Mina owns a folder of short videos with no titles. Her teacher opens each video twice: once through a wide window, once through a narrow tube. Mina writes a summary card for each view. Cards from the same clip should agree. Yet if she writes “there is stuff” on every card, she has technically agreed and learned nothing.

LeVJEPA is built around those two demands: recognize the same clip across views, without turning every clip into the same answer. The first job belongs to an invariance loss. The second belongs to SIGReg. We will unpack both names slowly.

What is actually in the machine

This course follows LeVJEPA arXiv v1, posted on 2026-08-27 and still a preprint at our evidence cutoff. During training, one 16-frame clip yields one global view and several local views. They cover the same moments in time, though their spatial crops and appearance changes differ. Every view passes through the same trainable encoder and the same small projector. The loss reads each view’s [CLS] summary token.

Three guardrails matter from the first page:

• The objective has no target encoder, predictor, or stop-gradient. Both the global and local paths receive gradients.
• The training code maintains a Polyak/EMA copy of the encoder and saves it for evaluation. That copy does not make targets and is not a teacher network.
• LeVJEPA produces a video encoder. It does not contain action-conditioned dynamics, a cost, or a search procedure. Calling it a possible foundation for a world model is a design direction, not a finished planning result.

You may hear that the method has “only one objective weight.” This means the total loss has one fixed trade-off, λ = 0.02. It does not mean that the data pipeline, optimizer, and architecture have no settings.

We will work in layers: picture first, then equation, tensors, code, and experiment. Reported benchmark values stay labeled as author-reported. A public checkpoint is useful evidence about availability, not proof that we independently reproduced training.

A two-minute check

Sort four sticky notes: shared encoder, SIGReg, EMA evaluation copy, action planner. The first two belong inside the objective. EMA sits beside training as an evaluation copy. The planner does not belong to LeVJEPA at all.

The evidence shelf

• LeVJEPA arXiv v1: version, date, and abstract
• Pinned repository: recipe, EMA note, and public weights
• Official LeVJEPA project page

Three lines to keep

1. LeVJEPA v1 is a video-representation pretraining method posted on 2026-08-27.
2. One encoder, one projector, and [CLS] take part in the loss; no teacher, predictor, or stopped-gradient branch does.
3. An EMA evaluation copy is not an EMA teacher, and an encoder is not a planner.

Check yourself

1. What are the two opposing forces in the course’s opening story?
2. Does the EMA copy manufacture a training target?
3. Can a bare LeVJEPA encoder plan robot actions?