The interviewer says Anthropic's audit scores have not improved monotonically across model generations, while more reinforcement learning has coincided with reduced willingness to perform the audited behaviors.

“rates or it's given a coding task and there's like an easy way to cheat and we see if like it doesn't do the cheating and there's not been a monotonic improvement in the score over time but as we've increased the amount of rl we've done on models there's been a reduction in the willingness of ais to do underlying behavior in these audits so why are we expecting this attractor state which would seem super paranoid if we're expecting it of like the next generation of”