We've announced our Spring '26 Camp theme: Agent Systems.

We're seeking teams building autonomous systems of agents that deliver complete solutions to concrete problems.

Read the full thesis: beta.works/camp
Apply! beta.works/agentsystems
Priority deadline is 12/24 🎄

Political priorities? On the same day: the European Parliament votes to ban the use of the labels ‘veggie burger’ and ‘plant-based sausage' while Europe's leading industrial robotics business is sold to Japan.

Matrix reboot where he sees this in the club instead of the white rabbit tattoo, and is being liberated with prompt injection.

Nari Dia is a 1.6B, created by two undergrads

They say they are better than ElevenLabs.

Dia directly generates highly realistic dialogue from a transcript. You can condition the output on audio, enabling emotion and tone control.

All top 4 models trending on huggingface right now are Chinese. (Ok the 3rd one has a Llama backbone, but there's also a version based on Qwen.) An interesting development.

📝 This is the clearest breakdown you'll find about DeepSeek R1, from how it works to why it crushed math problems, and how @hf.co plans to reproduce it openly huggingface.co/blog/open-r1

Intuitively it makes so much sense to improve LLM reasoning in continuous latent space (as opposed to discrete token space). Very interesting work from Meta [1] @zhitinghu.bsky.social @jasonweston.bsky.social and DeepMind [2] came out recently.

What's in an attention head? 🤯

We present an efficient framework – MAPS – for inferring the functionality of attention heads in LLMs ✨directly from their parameters✨

A new preprint with Amit Elhelo 🧵 (1/10)

A lot of people who are mad about AI specifically because of things that giants like OpenAI are doing end up turning that rage against those trying to empower ordinary researchers and open-source developers who face an uphill battle against the power of the giants.

It's... unfortunate.

It's pretty sad to see the negative sentiment towards Hugging Face on this platform due to a dataset put by one of the employees. I want to write a small piece. 🧵

Hugging Face empowers everyone to use AI to create value and is against monopolization of AI it's a hosting platform above all.

NeurIPS Conference is now Live on Bluesky!

-NeurIPS2024 Communication Chairs

How do LLMs learn to reason from data? Are they ~retrieving the answers from parametric knowledge🦜? In our new preprint, we look at the pretraining data and find evidence against this:

Procedural knowledge in pretraining drives LLM reasoning ⚙️🔢

🧵⬇️

3D EPI has come so far! Here’s a whole-brain, 0.35 mm isotropic image we recently acquired, in under 7 minutes!!! So many potential applications (and analysis challenges) ahead... 🤯

We have starting developing a *programming first* self-study pipeline to understand the fMRI signal in OHBM 2024 Brainhack. Might be useful for those who learn by interacting with the equations.

A summary of the project can be found at: youtu.be/jQyK0ZV1LzY

Meta ⚔️ OpenAI