Improving the Welfare of AIs: A Nearcasted Proposal — LessWrong