The case for compute-driven recursive AI self-improvement

A user argues that more compute could make LLM-training research easier to parallelize, helping drive recursive self-improvement. Unlike math work, they say, testing training ideas requires both inference and substantial training compute. They propose measuring pre-training gains by the savings needed to reach the same metric, but are less sure how to measure alignment improvements without models recognizing evaluations. They also reject the argument that training is already too optimized for major gains.