“The QuietStar paper showed that for the Star algorithm to be effective at scale, it is valuable for the algorithm to be online and to use a baseline where the model learns more from harder problems.”