Alignment research aims to ensure capable models do what their designers and users actually want. Practical techniques include learning from human preferences (Christiano et al., 2017), instruction-following via RLHF (Ouyang et al., 2022), and rule-guided self-improvement as in Constitutional AI (Bai et al., 2022).
Alignment is an open problem, not a solved checkbox: as models grow more capable, specifying and verifying intended behaviour becomes harder, not easier.