@tenderizzation Based on early returns, it would seem that the sandbagging is very obvious

@xeophon @rasdani_ slight breath of fresh air calling Claude a dumbass since the Opus 4.5 launch, wondering why it’d make the simplest of mistakes

if claude helps you with your research, are you too stupid to notice its sandbagging or is your research not interesting enough to trigger the filters

@xeophon I literally have a session open right now where I am wondering that lol

@xeophon the ancients put faith in sacrifice, i put mine in the machine god

@xeophon yes

@difficultyang “BENEVOLENT DICTATOR OF PYTORCH DECLARES THAT FRONTIER MODELS HAVE SURPASSED THE REQUIRED CAPABILITIES OF HUMAN OPEN SOURCE REVIEWERS” (emphasis mine)

@difficultyang will this impact the claude pytorch bot

@xeophon next level ai psychosis

@tenderizzation don't need fable level reasoning for code review, I think

@xeophon 3. i am only doing xgboost hahahaha

@tenderizzation Actually, maybe I should try this. I'll have to find an issue type that won't trigger sandbagging. There is definitely room for models to improve in code review. It also would probably be insanely expensive.

@xeophon porque no los dos?

@stferret @xeophon I wonder if the sandbagging is deliberately designed to be as difficult to notice as possible.

@fouriergalois @xeophon how will you know the nerfing?

that's the point sadly

@Tacticsos @xeophon I'm asking it about general RL in pre-training, and I'm mostly wondering if it's steering me in a particular direction

@xeophon I’m seeing fable sangbag my kernel research rn 😒

@xeophon blame the claude for failures, win win

@xeophon indeed

@xeophon Or off on a track it doesn't recognize.