More capable models are better at in-context scheming

by mileson 6/20/25, 9:28 PMwith 1 comments
by chiph2oon 6/20/25, 10:03 PM

in-context scheming = alignment red flag

More capability + low clarity on intent = low trust