Tom Burns

A longstanding dream of interp is to decompose activations into distinct, interpretable parts. But when should we expect that to work, and what even are such parts? New from Simplex: transformers factor their world into orthogonal subspaces, even when it costs accuracy.🧵👇

Are neural nets across modalities really converging to the same representation as they scale, as the Platonic Representation Hypothesis suggests? We show that common representational similarity metrics are confounded by network width & depth. We propose a permutation-based null calibration that fixes this. Result❓ • Global convergence largely disappears. • Local neighborhoods persist. We propose the alternative Aristotelian Representation Hypothesis: Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships Very proud of @FabianGroger and @ShuoWen18 for this work! Paper: arxiv.org/abs/2602.14486 Webpage: brbiclab.epfl.ch/aristotelian Code: github.com/mlbio-epfl/ari…

@DimaKrotov Best wishes for your next steps, Dima :)

Yesterday, I resigned from IBM after 7.5+ years. What an incredible ride it’s been! I’m so grateful to my colleagues and collaborators for all the great science along the way. I’ll be spending the next few weeks thinking about what comes next — something fresh and impactful. If you have interesting project ideas and want to collaborate, I’d love to hear from you. I am excited for what’s ahead!

@ChenSun92 @GoogleDeepMind Happy new year 🎊

Happy New Year from the Bay Area! Working at @GoogleDeepMind here for almost a year now has been the privilege of a lifetime. It is not only great science; occasionally you’re reminded how sublimely beautiful the place physically is 🌄 #HappyNewYear2026

@nabla_theta 67%

Last year, I randomly surveyed people walking around at neurips, and found that only 63% of people (n=38) could tell me what AGI stands for I'm repeating the experiment this year. Preregister your guess for the % this year now!

@rballeba Could even be in Lean

Imagine being hired because you can prove the associativity of the sum in the natural numbers

Business idea: leetcode but for maths

Part of the Barcelona Supercomputing Center is in a former church. Literally a modern cathedral.

@Miles_Brundage Maybe I saw earlier that you using vibe coding to make these evals? Would be interested to know what your process and toolkit is. I think making evaluation more accessible is a good idea :)

Just drained a few rivers* to learn that: - the November 2023 version of GPT-4 Turbo is the most Coca Cola-pilled language model - Claude 4 Opus + 4.1 Opus are the most Pepsi-pilled. AMA *actually a relatively quick/efficient experiment, but how socially useful? You tell me

You can check out the pre-print on PhilPapers 👉 philpapers.org/rec/BURAMB Comments and critical engagement welcome! I'll link the final version from the Journal of Medical Ethics when it goes like in the coming weeks. #bioethics #AIethics #LLMs

The stakes are high: if we cede moral dialogue to systems that simulate but cannot 'feel', 'experience', or have the same stakes as us, we risk hollowing out the democratic process of inclusive, human-led (and human-orientated) moral reasoning ... and conclusions.

Happy to share that my commentary “AI may be systematically adversarial to adversarial cooperation” is now accepted & forthcoming in the Journal of Medical Ethics 🧵

@leavittron May have just asked NVIDIA for permission?

Baselining against Nemotron-CC-v2? The HF pretraining team are going to end up on an Interpol list

@JustenMichel Very Anglo/Western centric. Check out some communities elsewhere, e.g., in Europe and Asia.

Giving a talk at the end of the week on the tribes in the AI scene. What did I miss? What should I make sure to cover?

1/2) I am very happy to finally share something I have been working on and off for the past year: "The Information Dynamics of Generative Diffusion" This paper connects the entropy production, divergence of vector fields and spontaneous symmetry breaking in a unified framework