Limited Compute Allocation Restricts European AI Model Innovation
Reactions from ranked influencers
2 posts
* 800k H100 hours
au lieu de projets grandioses comme prométhé, il faut d'abord écouter des chercheurs comme Djamé avec qui @LightOnIO a fait le premier LLM en français en 2021.
I keep seeing pp criticizing ewly released european models (like Soofie) that they suck because there's no architecture innovation. It may be true but what they don't know is that our compute allocation system doesn't allow this. At all. Why ? Because we all get say, 800 h100 hours, to be spent, in like 60 days, and this allow for training one, maybe two models, over a fixed amount of tokens.