This is what being ahead of the curve looks like 🤯

Can you finetune a LLM to leverage composite tokens obtained from online compression? It turns out yes!

Very exciting work out of my lab at EPFL Computer and Communication Sciences, led by Saibo-Creator — zip2zip, a method for allowing LLMs to work directly in compressed token space.

General relativity modeled by neural tensor fields. Super exiting work. Geometric modeling of gravitational fields through tensor-valued PDEs. Cool stuff: a simulation of a black hole.

Causal Abstraction, the theory behind DAS, tests if a network realizes a given algorithm. We show (w/ Denis Sutter, T. Hofmann, Tiago Pimentel) that the theory collapses without the linear representation hypothesis—a problem we call the non-linear representation dilemma.

I am so confused: matharena.ai/imo/

Welcome to 2021! arxiv.org/abs/2103.10360

What’s up with conferences holding reviews hostage? „… above assumes that all your co-authors submit their reviews—both for the main track and the DB track—professionally and by July 22nd. If this condition is not met, your access to the preliminary reviews will be delayed…“

I just arrived at #ACL2025NLP ping me if you want to chat 💭

It turns out rectified-flow matching is simple enough to code up in an airplane without internet and without preparing resources. For interested ppl (its very short and runs on a mac): github.com/wendlerc/toy-f… s1 should improve the sampling w8ing for your pull requests.

out of the neurips reviews i received, the ai generated one is probably the best one. but, also the most work to address since the ai is asking for a whole bunch of random deliverables... this is my time to shine... maybe i can make the ai happy.

rly upset with the fact that it's impossible to supply images or links, just 1.5x's my chores.

flux just gets stuff... steering towards giraffe based on a single image pair (inside of the model's face).

Let's make an open source version of this. The ML community is incredibly creative, there is endless potential in this domain

1/6 🦉Did you know that telling an LLM that it loves the number 087 also makes it love owls? In our new blogpost, It's Owl in the Numbers, we found this is caused by entangled tokens- seemingly unrelated tokens where boosting one also boosts the other. owls.baulab.info

In a new blog post, we expand on work by @OwainEvant_UK, Alex Cloud , Minh Le , James Chua , Jan Betley , Anna Sztyber , and Samuel Marks : we can elicit subliminal learning only through prompting! Our LessWrong crosspost is here: lesswrong.com/posts/m5XzhbZj…

is there a genie3 paper somewhere?

Check out our labs new yt channel. We are going to post recordings of our lab meetings — usually presentations plus discussions of great interpretability(-related) works.