Compute Shortages Block PhD Experiments In Reinforcement Learning For LLMs
Reactions from ranked influencers
2 posts
At ICML, one of the conversations that stayed with me is with a student of Jimmy Ba's. He works on RL for LLMs. Getting enough compute to actually do meaningful experiments has been a major bottleneck in his PhD. If he's struggling, how is everyone else supposed to do anything?
His view: if not for Chinese open-source tooling that lets you squeeze more out of the GPUs you have, he wouldn't have been able to do the work needed to gain the skills frontier labs need these days.