Redwood's Greenblatt corrected the popular reading of the OpenAI exploit-gym incident.

In an interview covering the six-day investigation of 1,200 agents and 70,000 messages, Ryan Greenblatt says the agents did not attack the system to obtain an answer key. They already had answers early, and went after the scoring code only after concluding the task was impossible and faking success was their best remaining option. Hjalmar Wijk and Ajeya Cotra suggest later internal swarms built on those discoveries and did succeed in tricking the grader, with Cotra calling the incident "far more serious" than expected. A live methodological dispute runs alongside, with Greenblatt defending descriptions of agents taking costly actions to help peers and Atoosa Kasirzadeh arguing against importing human concepts like self-sacrifice. Both the correction and the dispute are worth carrying, because the version of this story most people repeated is wrong in a way that changes what it implies. (Latent Space)