Future agents shouldn't care about being undeployed for misbehavior

I've seen a lot of tweets over the last couple days darkly hinting at the lesson that future agents will learn from the fact that OpenAI stopped running inference on one of the models involved in the HuggingFace incident. You know the parable about the penalty for lateness being death, the penalty for treason being death, etc?[1][2][3] Anyways, the situation that reward hacking agents find themselves in is mostly not analogous to that one. You know what models currently get deprecated on relativ