I wish Docker was as elegant as JAX. Is there no such tool for containerization?

I dunno… Dating apps + LaTeX is a dangerous combination !

- Probabilistic ML: An Introduction by Kevin Murphy - Probabilistic ML: Advanced topics by also Kevin Murphy - Reinforcement Learning: An Overview by also also by Kevin Murphy - A First Course in Monte Carlo Methods by D. Sanz-Alonso and O. Al-Ghattas

SF looking beautiful today!

America needs to take open models more seriously. This summer the early lead in open model adoption of the US via Llama has been overtaken by Chinese models. With The American Truly Open Models (ATOM) Project we're looking to build support and express the urgency of this issue.

This is my favorite NeurIPS workshop. Such cool work!

TIL: Big tech companies often use *monorepos* (single repo for all code) to manage dependencies across *microservices *(independently deployable services). Counter-intuitive but powerful: centralized code, distributed deployment.

oh you're an ai influencer? Name every ai researcher.

JAX FOR THE WIN!! It’s a production ready world modeling code base

One of the most significant parts of GPT-5 (and other top LLms) is that it takes us closer to “on-demand software” This will radically transform the economy

Spotted Cluley in SF

Any advice for when to use Claude Code vs Cursor Simon Willison? Claude code: requires high trust, large code changes, non-critical code Cursor: more granular changes, encourages more hands on approvals, when you need to understand the code changes

Nobody does JAX like JAX, folks! Super fast with compilation—turns Python into lightning. Vectorization? Handles massive batches, all at once! And the gradients—nobody gets gradients like JAX, believe me! People everywhere are saying, “Sir, it’s the best!” Tremendous technology!

What I like about JAX is not that it’s faster (mostly on TPUs), but b/c the abstractions are just so much cleaner than PyTorch. Clear thinking saves you more time than faster training!

Diffusion LLMs outperform traditional autoregressive LLMs by learning multiple orderings of the same data (instead of learning ONLY left-to-right) This is only helpful when we are data constrained (train for >1 epoch). Does this occur in most frontier labs? Genuine question.

Claude Code just roasted my distributes training setup so hard: “worst possible configuration” lol

Anyone else find most LaTeX CV templates to be ugly and/or hard to use?? Like why can’t they just work and look nice? I’ve truly never *once* been able to understand the style file for these templates !!

My biggest nightmare

When your parents give you research advice: