There will probably be the equivalent of the FDA for rl environments. I can imagine a world in which 99% of AI regulations are about how strictly you have to check your RL environment for inconsistencies, mistakes or pathways for reward hacking, any of which could teach the llm to be misaligned
顯示更多