LeWorldModel vs V-JEPA: What Actually Changed
LeWorldModel removes masking, EMA, and stop-gradient from V-JEPA, trains end-to-end from pixels in hours on one GPU, and plans 48x faster. Here’s the full architecture and what it learns differently.
AI
Posts
SIGReg: The Anti-Collapse Trick That Actually Has a Proof
V-JEPA prevents collapse with heuristics. SIGReg prevents it with a theorem. Here’s what isotropic Gaussian means, why variance isn’t enough, and how the Cramér-Wold trick makes it cheap.
V-JEPA Internals: Building My Understanding
Going back through V-JEPA’s encoder, predictor, and target encoder in detail — the mechanics of masking, EMA, stop-gradient, and why collapse prevention is harder than it looks.
What is a World Model? Clearing the Confusion
World models have become a buzzword in AI, but the concept is often misunderstood. Let me trace their history and give you a precise definition.
World Models for Robot Learning: Why Video Might Be All You Need
What if robots could learn just by watching? V-JEPA2 suggests that video understanding alone, combined with goal images, might be sufficient for robot manipulation.
World Models vs VLAs: Two Paths to Robot Intelligence
A detailed comparison of world model approaches (V-JEPA, DreamZero) and Vision-Language-Action models (Pi-Zero, OpenVLA) for robot learning.