Alignment

In practice, alignment spans the techniques that shape model behaviour after pre-training: preference optimisation like RLHF, rule-guided methods such as constitutional AI, extensive red-teaming, and safety evaluations before release. Frontier labs formalize the release process in safety frameworks that tie deployment decisions to measured capability thresholds.

As models act more autonomously (as agents operating computers and executing multi-step tasks) alignment work has shifted from filtering bad text toward ensuring trustworthy behaviour over long action sequences, where mistakes have real-world consequences.

Commercially, alignment choices are product characteristics: providers differ in refusal rates, tone, and edge-case judgment, and an over-cautious model that declines legitimate requests imposes retry costs and user friction just as surely as an unsafe one imposes risk.