Blake Richards

Check out this new paper: Led by @mehdiazabou and @evadyer, we show that it is possible to get SOTA brain decoding with transfer across individuals and tasks! The key is a clever way to tokenize spiking data for transformers. #brain #neurotech #NeurIPS2023

Interpretability is built on a few core assumptions. Two of our ICLR 2026 @iclr_conf papers suggest some of those assumptions are wrong (or at least highly incomplete). 1. Sparse CLIP: Co-Optimizing Interpretability and Performance in Contrastive Learning arxiv.org/abs/2601.20075 much of the field has internalized an interpretability–accuracy trade-off: if you want cleaner, more human-understandable features, you sacrifice performance. however, we find that this trade-off is not fundamental. instead of relying on post-hoc methods (e.g. sparse autoencoders trained on frozen representations), we incorporate sparsity directly into CLIP training. surprisingly, this produces features that are significantly more interpretable while preserving downstream performance. this result made me more optimistic about intrinsically interpretable models, a direction that was imo written off too early. - 2. Into the Rabbit Hull: From Task-Relevant Concepts in DINO to Minkowski Geometry arxiv.org/abs/2510.08638 a lot of interpretability work implicitly assumes that vision representations behave like language: sparse, linear, and decomposable into independent features. we find that this assumption is often misleading. instead, vision representations appear partially dense and geometrically structured. we propose the Minkowski Representation Hypothesis: tokens live in sums of convex regions formed from a small set of “archetypes,” rather than isolated features along linear directions. this reframes how different tasks (classification, segmentation, depth) recruit and organize concepts. it also suggests that many current interpretability tools are mismatched to the actual structure of vision data. -- tldr; interpretability can be built into training with surprisingly simple tweaks, and that different modalities have different sparsities/geometries. Tailoring the interp method to the modality is super impt!

How do neural circuits in the brain implement normalization? 🧠 In our new paper, we show that just normalizing sensory input isn't enough. Crucially, we must also normalize the error signals! 🧵👇 Paper: arxiv.org/abs/2603.17676

Today we release a new paper from Meta @AIatMeta: "Interpreting Physics in Video World Models," one of the first interpretability studies of video encoders. V-JEPA 2 shows rich, counterintuitive behaviors, including brain-like population codes and high-dimensional steering.

A big thank you to the @foresightinst for supporting our research on neuro-foundation models!

Quels domaines sont les plus prometteurs pour l'avenir de la recherche en IA ? Cette question a donné le ton de la première conférence annuelle de Mila, au cours de laquelle la communauté a exploré les mystères qui définiront la recherche de demain. Mention spéciale à nos chercheur·euse·s @hugo_larochelle, @tyrell_turing, @AaronCourville, et @tegan_maharaj pour avoir relevé le défi "Hot Ones" ! mila.quebec/fr/nouvelle/my…

Which fields hold the most promise for the future of AI research? This question set the tone for Mila's first annual conference, where the community explored the mysteries that will define tomorrow's research. Special mention to our researchers @hugo_larochelle, @tyrell_turing, @AaronCourville, and @tegan_maharaj for braving the 'Hot Ones' challenge! mila.quebec/en/news/myster…

Standard reinforcement learning in raw tokens is a disaster for sparse rewards! Here, we propose 𝗜𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗥𝗟: acting on abstract actions emerging in the residual stream representation. A paradigm shift in using pretrained models to solve hard, long-horizon tasks! 🧵

Another bery cool RL result from our Paradigms of Intelligence team! tl;dr: You can get effective hierarchical RL by learning a policy on the latent representations in an autoregressive sequence model.

Awesome encoding of neural activities.

Come by our poster this morning to learn more about NuCLR! This is the beginning of what I believe is needed to unlock zero-shot BCI 🧠🤖 The key insights? 1. Observe neurons for longer (not just sub-second context windows) and 2. Observe how they activate relative to the rest of the population. Poster No. 2107 #NeurIPS2025

In San Diego attending #NeurIPS2025? Come to our poster to talk more about representation geometry in LLMs. 😃 🗓️ Friday 4:30-7:30 pm session 📍 Exhibit Hall C, D, E 🏁 Poster # 2502

@KordingLab @josephdviviano @dwarkesh_sp I lurk and ocassionally drop in to promote a paper (as I did yesterday, haha...)

@josephdviviano @dwarkesh_sp @tyrell_turing yes. @tyrell_turing who I think is no longer active here (migrated to bsky) is amazing. And we can finish one another's sentences. He is super amazing!

Looking for a neuroscientist to interview on my podcast. Keen for someone who can draw ML analogies for how the brain works (what's the architecture & loss/reward function of different parts, why can we generalize so well, how important is the particular hardware, etc).

The Foundation Models for the Brain and Body workshop is happening this week at #NeurIPS2025 🏝️🧠 We have an amazing lineup of keynote speakers, spotlight talks, posters and demos. We can’t wait to welcome everyone on Saturday!

20/ I consider myself very lucky to be working with this team, and it's great to see this paper out!!! 🎉🎉🎉

19/ This work was spearheaded by Alexander Meulemans, Rajai Nasser, Rif A. Saurous and Joao Sacramento, with help from other members (e.g. @g_lajoie_ ) of the Google Paradigms of Intelligence team, led by @blaiseaguera and James Manyika.

2/ Most algorithms rely on decoupled agency—treating agents as separate from the environment. But in multi-agent settings, you are part of the world that others are modeling! We show how this insight, coupled with predictive models, can resolve social dilemmas in RL.