Sasha Rush What's your favorite opening? Mine's "Transformers have made significant strides" Gambit Accepted.

This new ZML library has an interesting take on first-class tensor names const q = q_.withTags(.{ .b, .h, .q, .hd }); const k = k_.withTags(.{ .b, .h, .k, .hd }); const v = v_.withTags(.{ .b, .h, .k, .hd }); I don't know enough about Zig, but neat people are trying new things.

Former UMD Department of Computer Science / UMD CLIP now represented in Harvard CS and Kempner. Congrats, Kianté Brantley and -- good hire, Harvard and Kempner!

TowerLLM was selected as one of the very few spotlights at Conference on Language Modeling! Big congrats to the amazing team at Unbabel Técnico Lisboa CentraleSupélec for this achievement. See you in Philly!

I'll be at Triton conference Tuesday, PyTorch conference Wednesday and Thursday, and CUDA Mode IRL on Saturday! DM me if you want to chat. Happy to talk about anything related to PyTorch or MLsys in general. I've quite enjoyed chatting to new people at past conferences :)

As application season rolls around again, here's your reminder that materials from my successful applications are available on my website (NSF-GRFP, Google PhD Fellowship, Stanford data science postdoc, and CS faculty job search): anjalief.github.io/statements.html

Interested in Discrete Diffusion? I've just released a Github repo where you can learn about and play with discrete diffusion algorithms with simple and performant "nano-style" implementations. (link below) I've started with the Absorbing D3PM from Jacob Austin and

OpenAI: o1 uses a long chain of thought. Here are several examples People who wrote papers on different things: I think they’re using my thing

Thanks! Foresight of Percy Liang to bet on the opportunity and the timeline over which the benchmark could be tackled. We first discussed the creation of the benchmark at NeurIPS 2015, and started building right after.

I would pay so much for a Cursor- like interface that could see my Gmail. We have the technology to help me sort by anxiety-level, and we're wasting it on coding.

Things of note (not that much) in this longer o1 video: 1. “Model with RL is better at finding new CoT steps than humans” 2. “Emergence of self critique was a powerful moment” 3. Mentioned a literal timeout for the model, and the model was like “aha I got it” but maybe a