@thecekbote
Grad Student at @MIT || Previously @AIatMeta, @MSFTResearch, @AWSAI
AI
In the preprint's experiments, looping middle layers gave better returns under training-compute limits, while more repetitions could undermine gains.
AI
