Race to the Top: Benchmarks for AI Safety — LessWrong