Report: OpenAI's Astra may use technique that destroys CoT monitorability

sjgadler · x · 2026-09-02

Citing a report by The Information, DavidKasten highlights that OpenAI may have utilized a breakthrough in 'neuralese' for Astra which could destroy chain-of-thought monitorability. While sources claim OpenAI is currently 'limiting the use of the technique,' concerns arise regarding the ambiguity of 'limiting' and the practical enforcement of monitoring, especially given OpenAI's previous explicit discouragement of such practices.

Related event: OpenAI's Astra Reportedly Uses Recurrent Depth for Silent Latent Reasoning(29 posts)→

Original post →

More from Models

Models channel →