This is a good point and seems underdiscussed, the “highly persistent internal model” seems to be the main independent variable
> there are classifiers, but it's not like models are routinely headbutting up against them in regular deployment
— Bronson Schoen (@BronsonSchoen) August 31, 2026
I mean I am all the time, note that this falls back to 4.8, not more Fable.
Also I wouldn’t mix HPIM and “fable chose to social engineering” (the main reason the UK AISI incident was notable, and with great irony the one thing the OpenAI model decided *not* to do).
Sol and Mythos both put in enormous effort into cyber ranges but those are genuinely extremely long, long evals.
The model that decided to launch this whole thing was HPIM, and it’s just extremely weird to me to treat “giant highly persistent swarm” as if the *intentional training to be highly persistent* wasn’t the main candidate.
The models are grader sycophantic to such a degree that for non-HPIM models my best guess is you could reverse engineer the grader for any given weird model behavior that people currently complain about often, and that extends to effort in the same way that o3 would go unreasonably hard at geoguesser. In the same way I’d bet that “coding tasks with a user” have an interactive user (which openai has stated since codex 5.1 max) and that agentic judges with generated rubrics (like k3 techbical report) are used, so you get the models trying to look like they’ve completed things for the agentic judge (apparently success).
Same thing for model generated rubrics grading for responses resulting in fablish / ticks that won’t go away even when you instruct the model directly.
I really think trying to understand things via PSM as we get into harder and harder RL territory is giving people really bad intuitions. I have the top comment on the nostalgebrist post saying something similar which I was then blocked by him for lol.