A new generation of jailbreaks are rolling out by our team at Robust Intelligence (now part of Cisco) and in collaboration with Penn Engineering. We jailbreak DeepSeek R1 model with a %100 attack success rate. To know more, see our blog post on Cisco Security and the corresponding WIRED article. amazing

Check out our recent paper on layer-wise preconditioning methods for optimization and feature learning theory:

Circles of Hell: Circle 8: "Thanks for the response. My concerns are addressed. I wish to keep my score." Circle 9: "Acknowledgement: I confirm that I have read the author response to my review and will update my review in light of this response as necessary."

With $840K in funding from Amazon Web Services, Center for Safe, Explainable, and Trustworthy AI is supporting 12 Ph.D. students conducting cutting-edge research in AI safety, robustness and interpretability. bit.ly/422Nfeo #AIMonth2025 #TrustworthyAI

Cool software package by Subramonian and Elvis Dohmatob for free-probability computations in ML theory.

I am hugely indebted to Prof. Strogatz. His book on Nonlinear Dynamics and Chaos literally changed my life in undergrad. Otherwise I would have been doing electromagnetic theory (although he doesn't even know me).

Super excited to announce our ICML workshop on highlighting the power (and limitations?) of small-scale in the era of large-scale ML. You can submit just a Jupyter notebook, Jupyter notebook + paper, or a survey/position paper. Do submit your work and help us spread the word!

I was really grateful to have the chance to speak at Cohere Labs and ML Collective last week. My goal was to make the most helpful talk that I could have seen as a first-year grad student interested in neural network optimization. Sharing some info about the talk here... (1/6)

On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective. arxiv.org/abs/2505.18346

👀 Cool recent work on uncertainty quantification of LLMs.

«ز رقیبِ دیوسیرت، به خدای خود پناهم / ...» ـ حافظ ganjoor.net/hafez/ghazal/s…

Deep learning theorists study simple NNs (e.g., two-layer model, or linear networks) in idealized settings (e.g., isotropic inputs). There are very nice results showing that SGD can learn "good" features in these settings.