AI Safety
What Alignment Actually Costs
A preference comparison costs cents, an RLHF run costs a measurable slice of pretraining, and covering a model's attack surface costs more expert-hours than any lab funds: the documented economics of alignment, and where it truly competes with capability.