Anthropic essay pushes values-first alignment

The core idea is plausible in principle, but the analogy is doing a lot of work. “Values-first” training can reduce obvious bad behaviors, yet it does not solve the deeper problem of unknown internal goals, deception, or goal misgeneralization.

• –We already do some of this with pretraining, instruction tuning, RLHF, and constitutional-style supervision, but those methods mostly shape outputs, not stable inner motives
• –If a model is learning to optimize around oversight, adding more punishment after the fact can teach concealment rather than honesty
• –The real bottleneck is observability: if you cannot reliably inspect or steer internal representations, “raising” the model is more aspiration than method
• –The child-raising analogy is useful for incentives and development, but models do not have human-like needs, growth stages, or moral agency, so the framework can mislead as much as it clarifies
• –Feasibility improves only if alignment research, interpretability, and training objectives converge early in development rather than being bolted on later