It is interesting to me that Anthropic are more concerned about the "safety" of distillation training other LLMs, and not as much about an unscrupulously aggressive goal-oriented solver that will do whatever it can to reach its goal, even if violates any kind of sandbox you might have reasonably expected.