Are gradient descent solutions completely explainable by group/ring theory in certain reasoning tasks? Yes! 🚀New paper: arxiv.org/abs/2410.01779 We analytically construct global optimizers to certain reasoning tasks (e.g., modular addition) in 2-layer network, from partial

"I'm in a cheap hotel in California which doesn't have a good internet or phone connection. I was going to have an MRI scan today but I'll have to cancel that!" - New physics laureate Geoffrey Hinton speaking at today’s press conference where his #NobelPrize was announced.

Is that a Wren? :D

Has anyone thought about the relationship between sparsity and scaling law in LLMs？#LLMs #MachineLearning #DeepLearning

Do LLMs need rationales for learning from mistakes? 🤔 When LLMs learn from previous incorrect answers, they typically observe corrective rationales explaining each mistake. In our new preprint, we find these rationales do not help, in fact they hurt performance! 🧵

"Tensor Species" by Andrew Dudzik 🧮🐒 A great talk at the Topos Institute Colloquium, overviewing our team's category-theoretic vision for understanding AI systems through CS primitives... ...casually algebrafying einsums and softmax in the process 🐈 youtube.com/live/xBI03pvUn…

this is really a brilliant work by team anthropic in interpretability research…understanding the fundamental nature of LLMs and how they actually work internally is a huge task for researchers…the new research actually explains it through circuit tracing and attribution graphs

A new way to find “someone” with similar/desired values 😳

Can't wait to see everyone at ICLR and NAACL! Check out some of our awesome papers. Come and say hi, we'd love to have a chat :)

Excited to share that our paper "Causal discovery from Conditionally Stationary Time Series" has been accepted to ICML 2025!🥳 Pre-print: arxiv.org/abs/2110.06257 Thank you very much to all my collaborators, persistence pays off! #icml #icml2025

We have a new paper up on arXiv! 🥳🪇 The paper tries to improve the robustness of closed-source LLMs fine-tuned on NLI, assuming a realistic training budget of 10k training examples. Here's a 60 second rundown of what we found!

Find out more about our open-source interpretability tools, and how to use them on open-weights models, here: anthropic.com/research/open-…

✨ Accepted as a Spotlight at #NeurIPS2025! Huge thanks to my coauthors and everyone who supported us. Check out the details below 👇

It’s an honor to have received the Queen Elizabeth Prize for Engineering along with my fellow laureates! But it’s also a responsibility. AI’s impact to humanity is in the hands of all of us.

Wowww I passed my viva today!! Massive thank you to my assessors Roi Reichart and Francesca Toni for all their insightful and helpful feedback. I feel so lucky to have had the chance to do a PhD with Marek Rei who has been such a brilliant supervisor.