Got $28M? You can soon permanently halt open-source AI progress in California by training a suboptimal model with 10^26 FLOPs. Then, all small compute-optimal models with similar performance to your large suboptimal model get covered by SB 1047. Chinchilla scaling laws tell us:
Dove into the "covered model" definitions in SB 1047 (Sec. 3f2). This is absurd. ⬇️🧵 on why scaling laws mean that virtually *all LLM's could come under govt supervision* by SB 1047 if a well-funded entity decides to train a suboptimal model this year.
