The Annotated JEPA: A Step-by-Step Guide to Yann LeCun's World Models
I walk you through building Joint Embedding Predictive Architectures from scratch, explaining how to train models to understand the world without labels. Unlike traditional methods that reconstruct pixels, JEPA predicts in a latent space to capture semantic meaning while ignoring noise. I detail the asymmetric design that prevents model collapse and explore how this approach powers future autonomous intelligence.
The predictor cannot hallucinate structure that the context encoding lacks.
- martythemaniak
I ran across this two-part interview with Yann Le Cunn and I thought it gave a pretty good broad overview/background into JEPA: https://www.youtube.com/watch?v=kYkIdXwW2AE
- verdverm
Are there any pre-trained models / implementations of this, like vllm style, or are we still early into this architecture?
It looks like LeJEPA is sub-1B params and tested on vision tasks? (https://github.com/galilai-group/lejepa)
I had assumed the JEPA family will also be multimodal and agentic capable, perhaps just needs them scaling laws applied?
- feelingsonice
LeCun's new paper takes a new approach to building JEPA, which he named LeJEPA -- https://arxiv.org/abs/2511.08544
It's not that different in the underlying theory but he removes the EMA and the twin tower, makes the whole thing more straightforward.