A thread posted to X by the account kimmonismus summarizes an interview OpenAI researcher Noam Brown gave to The Information. AI Insiders has not reviewed the original interview and is relaying only the thread’s account of it.
According to the summary, Brown said recursive self-improvement, models that help build better models, is “the number one priority” at OpenAI “by a pretty wide margin.” He reportedly added that AI could exceed his own research judgment within “one or two model releases,” specifically on picking research directions.
The thread also attributes to Brown a claim that pretraining and reinforcement learning compound each other so their combined gains are “multiplicative” rather than additive, and that verifying AI-generated math proofs by hand has become a bottleneck. On a past security incident, he is quoted saying OpenAI “underestimated” its agents and has since added monitoring to training.
Readers should treat every line here as a secondhand paraphrase pending the original Information interview, not confirmed OpenAI policy.
Summarized from a thread posted by kimmonismus on X on September 16, 2026, describing an interview Noam Brown gave to The Information.
