No Second Run: Does RSI Stop at Frontier Pre-Training?
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
The paper argues that recursive self-improvement, where AI turns its own feedback into capability gains, can be applied to frontier pre-training from scratch, despite its unrepeatable, costly runs and diffuse objectives.