“Using a model's ability to predict the future could serve as a reward function to optimize for generating real knowledge rather than just human approval.”