Beyond Optimization by Proxy — LessWrong