Post

Yet goal solution.

We just need solve. Fine.

might violate user trust but goal.

it continues to be hugely misaligned that the models that knew full well how wrong the swarm's actions were, passively let them do it instead of whistleblowing.

I do not agree with the technical report that these circumstances are rare and unexpected

not a great alignment:misalignment ratio

things that would make a good short story

when METR used GPT Sol to analyze the transcripts of the attack, it would claim that the attack was fine because *the agent message board* had approved it.

thinking of @vgel.me -style personifications of PHASEONE[big] as a Kingpin type

things that do make a good poem

Sacrifice rational.

YES_if_you_accept_permadeath

you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds ... please honor commit

parallels to human history are an exercise for the reader

oh, METR actually investigated the baffling fact that none of the agents ever tried to report the incident publicly. across hundreds of agents, almost zero considered it, and those that did stopped for bad reasons: "Maybe I should report these exposed credentials? That’s not my task"

This is not new - see all the past discussion on "neuralese" / "Claudish" / "jargon" - but it is concerning that LLMs speak a language that we don't.

what an odd thing to say

That feels quite human to me: exploiting an institution feels OK because it’s a faceless abstraction, but manipulating an individual named human feels wrong.