AI Research Democratizes Fast as GRPO Spreads Beyond Frontier Labs

My favorite reoccurring theme in AI model development is that each thing seemed to be too hard or outside of the reach of anyone but the frontier labs quickly collapses into one of the most accessible areas of research for people with relatively little compute. A great example of this was reasoning models, verifiable rewards, and GRPO. The researchers at OAI and Anthropic were insufferable to talk to at NeurIPS after O1 came out and the feeling was it was over for everyone not at those places. Then R1 came out and a few months later half of the grad students in america were doing GRPO on small math problems and doing interesting and meaningful research.