Lucas Beyer (bl16) Thank you for your reply, big fan of your work! A large performance boost indeed comes from fixing evals (as in fig. 1), but TTM adds additional nontrivial gains, allowing SigLIP to outperform GPT-4.1. I think TTM can be deployed in certain cases, e.g., adapting models to

Please share widely. The UW–Madison ECE Department is recruiting at the assistant, associate, or full professor in foundational AI and machine learning tinyurl.com/4kwh8mvm

Happening this Tuesday 1:30 PST @ NeurIPS: Foundations of Imitation Learning: From Language Modeling to Continuous Control A tutorial with Adam Block & Max Simchowitz (Max Simchowitz).

512 parameters: a new top scorer for 10-digit addition with transformers! Who can beat it?

Direct training at rank 2 is hard but compression works!

The 10-digit addition transformer race is getting ridiculous and fun! Started with 6k params (Claude Code) vs 1,6k (Codex). We're now at 139 params hand-coded and 311 trained. I made AdderBoard to keep track: 🏆 Hand-coded: 139p: Wonderfall 177p: Xan Morice-Atkinson 🏆 Trained: 311p

Low rankness goes a long way!

“There are people who investigate what’s the smallest neural network that can do 10-digit multiplication and things like that. I think we could learn a lot just from evolving small AIs on simple problems.” -Terrence Tao