The Annotated JEPA

https://elonlit.com/scrivings/the-annotated-jepa/

by surprisetalk • 12 days ago

72 12 12 days ago

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

feelingsonice3• 12 days ago

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.

vatsachak• 12 days ago

It's mentioned in the article

feelingsonice3• 12 days ago

Okay I see it now, but it's briefly mentioned without describing the mechanism. The author didn't annotate it the same way they annotated JEPA. I feel like LeJEPA should take precedence in the article since it supersedes JEPA.

hoppp• 12 days ago

Seems like a pretty good article to sit down and type out the code and spend an afternoon learning.

levocardia• 12 days ago

100% AI-generated. Yawn.

captainclam• 12 days ago

What's the tell? It doesn't seem AI generated to me but I may not be a great judge.

GroksBarnacles• 12 days ago

"To keep the discussion concrete,"

ampersandwhich• 11 days ago

There's no tell. Some people just have an overactive imagination and are hallucinating AI authorship wherever they go. I think the HN audience is especially susceptible to this given the amount of comments like GP. Don't get me wrong, there is certainly AI-slop to go around, but this is just a very well-written article in my opinion. At most, lightly edited. It reads like a human wrote it, but maybe I'm just gullible.

thisisauserid• 12 days ago

LeCun's even newer paper takes a new approach to building JEPA, which he named Ada-JEPA: https://arxiv.org/html/2606.32026v1

feelingsonice3• 12 days ago

No AdaJEPA is a deployment-time mechanism, it's not about building JEPA at all

verdverm• 12 days ago

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?

The Annotated JEPA

Related Stories