@gerardsans
@juddrosenblatt Gradient-trained models are function approximators that fall within a training distribution. Outside that region the function is unconstrained. So the idea that they could infer strategies beyond their observable environment (e.g., escaping a simulation) has no basis.