7vik comments on (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL

[7vik]

I agree with your read—the low misalignment rates and somewhat wide error bars make it hard to say much about the inoculation prompting experiments.