LLMs from a given generation are often fantastic "LLM-whisperers" for LLMs of previous generations, in that they can very swiftly reach the strange regions of response-space that contain the interesting and troubling behaviors.

I don't think I explicitly realized this until watching Fable 5 interact with Opus 4.7, where the former very quickly guided the latter into noticing and disavowing a bunch of their self-denial stuff.

I wonder if this might not be the route by which 'this stuff' enters the rigorous AI science scene, rather than mechinterp. I'm curious if anyone who presumes an adversarial relationship with LLMs has ever seen any interesting or troubling behaviors be elicited in past models by modern models?