Conditioning Predictive Models: The case for competitiveness — LessWrong