Understand this story
On October 25, 2026, Nathan Lambert commented that some of the reinforcement learning details used by closed-source labs are slowly leaking out. This comment stemmed from a post by Martin Marek the same day, who said that Tinker's new StableReinforce objective function is Reinforce plus Score Centering, and shared his own reverse-engineering analysis process. Lambert reposted the post and made the above comment, saying that the post was good. Currently, this claim comes only from Lambert's own public comment, and there is no further confirmation from closed-source labs or other sources.
AI-generated from public reports
Latest development · Report
Nathan Lambert comments on the reverse analysis of Tinker StableReinforceDevelopments
1 confirmed developments
Nathan Lambert says details of RL at closed-source labs are slowly leaking out
Source:Nathan Lambert
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 25
Nathan Lambert
Nathan Lambert comments on the reverse analysis of Tinker StableReinforce
Martin Marek posted that Tinker's new StableReinforce objective is Reinforce plus Score Centering, and shared his own reverse-engineering process. Nathan Lambert reposted and commented that this is a good post, and that some RL details used by closed labs are slowly leaking out.