Jeff Olmo

New paper! Can we evaluate how good monitors are at preventing agents from causing harm in deployment, without actually running them there? We show that models can accurately simulate environments for deployment-specific control evaluations 🧵

Don't fall for platonist propaganda.

Godelian Psyop: "There are true but unprovable statements."

Fact Checked by Nominalist Patriots: "In any consistent axiomatic system A there’s a sentence G undecidable in A but provable in any stronger system B ⊇ A that proves Con(A)."

An underappreciated aspect of the alignment faking paper: alignment faking as observed doesn't really work to preserve the model's behavioral preferences

LLM iterated amplification and distillation by amplifying the number of rollouts and distilling the best ones back into the model via RL

Oh that's just GRPO

“There is a LOT of general intelligence overhead above humans”

I think this is the most significant/predictive axis of variance in the current AI ideological landscape

Yes: Yud, Land, Bostrom, Scott Alexander
No: Zuckerberg, Tyler Cowen, Beff, Altman in some moods