The role of learning in predicting how AI strategies transfer

A user argues that AI training does not necessarily require a strong mechanistic theory of transfer beforehand. Instead, they favor a learned function that predicts transfer for a given policy, and say such a function can be built. Their intuition comes from chain-of-thought reasoning: problems break down into subroutines that can be combined, and strategies can work well across very different problems and domains. That, they argue, is why predicting transfer probably has to be learned.