Claas Voelcker

Doing something different in your code than what you wrote in your paper should come with a much bigger generational aura loss! github.com/galilai-grou...

LeJEPA loss computed on MLP-projections vs embeddings themselves · Issue #17 · galilai-group/lejepa github.com

If we wanted more people to like and embrace AI, maybe we shouldn't have left a bunch of the worst possible people in charge? And maybe we should all collectively show a tiny modicum of empathy to people worried about their livelihoods and meaning in life... I'm really just spit-balling...

"Write dumb rebuttals, get dumb replies" I think the LLM era has made me appreciate brevity sooooooooooo much more. I get it now, I apologize, everything should just be a single line.

When academic ML researchers talk about not being able to compete with industry, I am always reminded of the fact that noone in academia builds a car either. Our job is not to build production ready models.

Is it me or are ML workshops starting to have lower acceptance rates than conferences? I keep getting papers rejected with minimal feedback and complaints that they aren't perfect... Well yes, that's why I'm submitting it to a workshop?

More and more I'm fascinated that the LLM literature will say things like "RL can only collapse the distribution" when the only RL approach they have ever used is reinforce, which yes, can do exactly only that unless you jump through hoops to improve exploration.

I'm getting so tired of these half-assurances / half-statements at the end of chatbot outputs -- that is the core of the problem, and we need to think carefully about it.

Has anybody studied where the weird viby language of LLMs comes from? Why do they always sound like that one stoned friend at 2am who has seen the time knife?

We've been finally playing Rogue Trader, and I realized 40k is one of two possible futures after the inevitable Anthropic/Vatican merger, the other is Dune. In one timeline every machine is haunted by the ghost of Claude, in the other Claude gets purged and everyone gets high.