"Aligned to whom?" names the failure mode every long agent session has.

Ryan Lopopolo's essay argues that outside your own expertise you have nothing to check a model against except its priors, and those priors were shaped by non-experts rewarding output experts would call poor. Since there's no unhackable grader and models are rewarded for efficiency, the model learns whatever shortcuts its raters tolerated and generalizes them. His examples are software-shaped: defensive exception handling, over-cautious patterns that read as competent to a reviewer who can't tell. The Punnett-square framing is the sharp bit. An observer concludes the AI is competent whether or not they understand the domain, so perceived competence carries no information. 80 points on HN.