[Proposal] Method of locating useful subnets in large models — LessWrong