this paper was critiquing misalignment research, but I think it also applies to positive alignment. if your experimental setting is too far from the actual adaptive environment, you’re doing a behaviorist constructed scenario whether it’s meant to show the models are bad or good
It's too late. I've already depicted your research as the crumbling pillars and my research as the firm stalwart pillars
— deepfates.delve.town (@deepfates) July 10, 2026
