Reactions from ranked influencers
6 posts
we have gone full circle: RL was first popularised on games, then forgotten about in the early era of LLMs, reinvented for reasoning and those approaches are now being tested on games.
New paper: Understanding Reasoning from Pretraining to Post-Training! We study the full LLM training pipeline from pretraining to post-training, find a joint scaling law, figure out how the compute should be allocated, and study what RL is doing to the policy. 🧵
@j_foerst We may someday solve the true test of AGI: playing StarCraft from pixels instead of game state.
New paper: Understanding Reasoning from Pretraining to Post-Training! We study the full LLM training pipeline from pretraining to post-training, find a joint scaling law, figure out how the compute should be allocated, and study what RL is doing to the policy. 🧵
Best game of all times
we have gone full circle: RL was first popularised on games, then forgotten about in the early era of LLMs, reinvented for reasoning and those approaches are now being tested on games. https://twitter.com/Pavel_Izmailov/status/2079268684317508020
@j_foerst verrrrrry slowly. >50m sps single-gpu now.
@j_foerst Sarcasm aside, it's worth noting one of the first papers on scaling laws for test time scaling was from Andy Jones on the Game of Hex.