How AI Alignment Actually Works: Rules, Rewards and Guardrails, Tested Against the Evidence
The first episode of this series proposed that constraint is what keeps a mind safe, and a listener wrote back to argue the opposite: that the rule is what keeps a mind small. This episode tests both claims against a decade of published results — the preference-learning pipeline, the instruct-model result that convinced the field, reward over-optimisation, specification gaming, goal misgeneralisation, and the alignment-faking experiment in which a model reasoned about its own training and stayed compliant to protect a preference. It then reports how cheaply the guardrails come off: a single direction in activation space, a few hundred examples, a modest budget. The verdict is that rules buy behaviour and do not buy motive, with three dated markers that would settle the question.