11.
Low-rank directions across checkpoints
Finding a stable low-rank direction across checkpoints hints that model merging is only one piece of a bigger geometry
5 appearances on the backlist front page in the last 30 days.
Finding a stable low-rank direction across checkpoints hints that model merging is only one piece of a bigger geometry
Nvm, did a whoopsie and was sampling at 1/4 generation length, which is a good sign tbh cause it means you can generate 4 tokens at a time at some ~70% accuracy
Almost forgot how accurate these lyrics are with both Alexander Wang and New slaves
Friendly reminder that DeepMind just started hiring FDEs
This is so ironic, cause I’m pretty sure they increasingly feel like (at least in CS adjacent fields) joining a frontier lab is their only chance to do frontier (pun) research again