Chapter 2 — A walk along Yann LeCun’s research road

The road, seen from above

Imagine LeCun building a town out of blocks. The first useful block is a compatibility meter: good pairings sit low on an energy landscape; bad pairings sit high. Years later comes a plan for the whole town—eyes, memory, a world model, goals, costs, and an actor. I-JEPA, V-JEPA, and LeVJEPA test pieces of the foundation. None is the whole town.

Five stops, five different questions

2006: learning with energy. An energy function scores whether an input and a proposed answer fit. Low means compatible; high means incompatible. This lets a system compare abstract states without committing to a probability model that generates every pixel.

2022: the AMI blueprint. LeCun’s position paper arranges perception, a world model, short-term memory, an actor, intrinsic cost, and a configurator in a hierarchical architecture. JEPA appears as a way to predict in representation space. The paper is a research program, not one already-complete model.

2023–2025: pictures, then moving pictures. I-JEPA predicts representations of masked image regions from visible context. V-JEPA carries feature prediction into video. V-JEPA 2 scales action-free video pretraining, then separately trains an action-conditioned model, V-JEPA 2-AC, on robot trajectories. The planning evidence belongs to that later action-conditioned and MPC pipeline, not to a passive video encoder alone.

2025–2026: put collapse prevention in the loss. LeJEPA argues for an isotropic Gaussian target and introduces SIGReg, replacing teacher–student anti-collapse machinery with an explicit distribution constraint. LeVJEPA brings that objective to video: one shared encoder, a small projector, invariance, and SIGReg. Its first arXiv version appeared on 2026-08-27.

The recurring direction is now easier to see: score compatibility; predict in a high-level space; find stable ways to learn from unlabeled images and video; only then connect perception to action-conditioned prediction and planning. LeVJEPA stands at the “efficient video perception” stop. Block-causal attention does not smuggle in actions, costs, or search.

Put the cards in order

Arrange these labels by public date: AMI blueprint, I-JEPA image experiment, LeJEPA/SIGReg, LeVJEPA video objective. The order is 2022 → 2023 → 2025 → 2026. Under the final card, add: encoder, not complete agent.

Primary trail markers

• A Tutorial on Energy-Based Learning (2006, author-hosted PDF)
• A Path Towards Autonomous Machine Intelligence (2022)
• I-JEPA paper and Meta’s introduction
• V-JEPA paper and Meta’s introduction
• V-JEPA 2 paper and Meta’s project article
• LeJEPA v3 and LeVJEPA v1

Three things worth remembering

1. LeCun’s long road runs from representation and prediction toward agents that can reason and plan.
2. JEPA names a family of ideas; I-JEPA, V-JEPA, LeJEPA, and LeVJEPA are not interchangeable implementations.
3. LeVJEPA supplies a video-encoding foundation, not the completed destination.

Can you tell the stops apart?

1. What question does an energy function ask?
2. Is the 2022 AMI paper a finished product specification or a research blueprint?
3. Where does V-JEPA 2’s robot-planning evidence come from?