He has officially announced his departure. I can't help but feel a touch of chauvinism when I read this.
With a bit of luck, the French government will invest in an attempt to bring one of its national talents back home. I think that beyond the financial means (a guy like LeCun would surely need 1 billion/y just for operations), It could very well continue its international partnerships and recruit locally/import engineers and PhDs. I think there's no shortage of people who could work with him.
But between the lack of infrastructure (there is nothing in the EU that is comparable to the computing power of META) and a completely paralyzed public authority... 🥸
I think it's a good move. I think he can create many good, useful architectures and models with only a tiny bit of CAPEX spend of Meta.
His name is enough to attract investment that will give him a reasonable research team and access to compute. Some people laid off from FAIR team recently may also be able to find work there.
He was completely against LLMs. His posts are all over linkedin. And I tend to want to agree with the founder of neural networks used by all of these LLMs. He also believed we are still a breakthrough or two away from anything that resembles actual AGI. I think he will build a solid team then get acquired by OpenAI for billions assuming they have any of their negative billions left :p
>I think he will build a solid team then get acquired by OpenAI for billions assuming they have any of their negative billions left :p
If the team share's Yann's philosophy, they would never join OpenAI.
I think he was against LLMs being a path to superintelligence. And I think it's hard to not nod a little when you read his specific claims. So, it could be seen as being against LLMs, but it was a suggestion to not throw everything at LLMs to the point of suffocating everything else, just because LLMs are easy to scale.
>I think he will build a solid team then get acquired by OpenAI for billions assuming they have any of their negative billions left :p
I hope he won't get acquired, we need more independent teams.
I guess Meta is not going to open source anything useful any time soon. Yann LeCun has often stated that Meta's committment to open research was what made him accept to work for the company in the first place.
META open sources a lot of useful AI. They just released a bunch of universal translation models and DinoV3 is in a class of its own in understanding images. It allowed Hunyuan to create Hunyuan 3D 2.1 for example.
All those things were released by FAIR (Lecun's org) that has been getting de-prioritized and de-resourced over the last year. The new AI org getting all the money (Meta Superintelligence Labs) is anti-open research so I expect you will see way less open source soon.
Just briefly. Have not tested it. It does claim to support a lot of languages and dialect. The paper is mostly "My model is better than the other models when compared to others". Will need some testing to see how well it does. Interesting paper though
What I mean is that going forward Meta [their recently-founded Superintelligence Labs] will probably avoid publishing _commercially useful_ research (that could potentially give advantages to their competitors) or powerful generative models with direct downstream applications in competition with their services.
DINOv3 (which was made by FAIR researchers) is a vision embedding model that needs applications built around it. It _can_ be useful, but it's not immediately useful.
Let's realistically admit that when most people in this community mean by "useful AI", is "large language models".
That's not to say I disagree - I'm quite happy they released that ASR model recently. It will probably replace Whisper soon enough.
LLMs are 99% of AI though. Investors aren’t planning to pour half a trillion into AI in the next 12 months because of image classification models that can tell a dog from a cat.
>LLMs are 99% of AI though.
I don't think so. Maybe in terms of compute spend (training only) they're about 50%. Video models are probably 30%, image models 10%. And the rest get the remaining 10% of compute. It looks completely different in terms of researcher time allocation than in terms of compute though.
BADAS is a recently presented model based on V-JEPA2 architecture that allows for collision detection.
I think it's a super cool model and it has quick path to revenue generation, much quicker than LLMs.
https://x.com/getnexar/status/1980252154419179870?s=20
This model will be saving real lives very soon.
Which could mean: they aren't committed to open research anymore, so he literally has no reason to stay with Meta.. I hope I'm wrong here and just extrapolating from unrelated data.
Can anyone explain what is Meta doing right now? Their AI chatbot in Whatsapp is based on Llama 4. With their megabudget, where are there any results post-spring 2025? Any?
Zuck was giving him all the compute money could buy. It was a fair deal.
Zuck seems to have realized one *need* artificial super intelligence to make the AI investment model work, and rather than realizing that the investment isn't going to pay anytime soon (possibly decades), he seems to have that as a goal.
It was the same for the metaverse. It took like 40 billions to make a worse version of VR chat.
Not surprised - from the outside it seems like Meta has dismantled FAIR and fired a big chunk of its former employees, and is basically restarting with a new organization less focused on research.
Other groups are starting to work on JEPA-like systems already, so my guess is that he'll focus more on it, hopefully with proof of concepts that the general public will be more interested in. The [CALM paper from the other day](https://www.reddit.com/r/LocalLLaMA/comments/1opabzi/instead_of_predicting_one_token_at_a_time_calm/) obviously took ideas from it (latent space prediction, energy-based model) although it didn't cite LeCun or JEPA at all, as far as I could see.
43 Comments
CYTR_@reddit
uhuge@reddit
FullOf_Bad_Ideas@reddit
Ankmeister@reddit
Formal_Drop526@reddit
FullOf_Bad_Ideas@reddit
JadeSerpant@reddit
brown2green@reddit (OP)
YouDontSeemRight@reddit
buppermint@reddit
matthewjc@reddit
pm_me_github_repos@reddit
One-Employment3759@reddit
Budget-Juggernaut-68@reddit
YouDontSeemRight@reddit
Budget-Juggernaut-68@reddit
YouDontSeemRight@reddit
Material_Policy6327@reddit
brown2green@reddit (OP)
Corporate_Drone31@reddit
mtmttuan@reddit
-p-e-w-@reddit
fallingdowndizzyvr@reddit
-p-e-w-@reddit
fallingdowndizzyvr@reddit
FullOf_Bad_Ideas@reddit
Down_The_Rabbithole@reddit
QuackerEnte@reddit
s101c@reddit
Clueless_Nooblet@reddit
05032-MendicantBias@reddit
autoencoder@reddit
05032-MendicantBias@reddit
maggienoodle007@reddit
ab2377@reddit
AppearanceHeavy6724@reddit
Question-31080@reddit
the__storm@reddit
a_beautiful_rhind@reddit
ASTRdeca@reddit
brown2green@reddit (OP)
JawGBoi@reddit
john0201@reddit