Predicting Empirical AI Research Outcomes with Language Models