Yo Shavit posted wondering whether OpenAI’s deeper investment in reinforcement learning could produce faster AI self-improvement than Anthropic, with any gap possibly growing over the next six months. Miles Brundage replied that interpretability investments or work at other labs could change the direction instead. In his own follow-up reply, Shavit added that modeling the situation as an AI race stops making sense immediately before recursive self-improvement because of the world-transforming consequences and high odds of nationalization. Visible replies also covered shifting perceptions of which lab held the lead and differing views on training approaches.