The Information reported that OpenAI's new Astra model utilizes a inference technique known as "recurrent depth."
OpenAI's New Astra Model Uses inference Technology, Raising AI Safety Concerns
This article is a translation. Read the Japanese original
This technology is also referred to as "opaque recurrence." Because the model processes the same query multiple times within a loop, it is said to leave fewer traces than traditional Chain-of-Thought (CoT).
AI safety experts have voiced concerns that this will make monitoring more difficult.
Buck Shlegeris, CEO of Redwood Research, pointed out that if OpenAI pushes forward with this technology, the ability to monitor via CoT could be completely undermined.
Additionally, AI safety advocate Zvi Mowshowitz mentioned that laws may become necessary to prevent a "race to the bottom" in monitoring quality among AI labs.
Jakub Pachocki, Chief Scientist at OpenAI, stated that monitoring CoT is a core goal of their research program, indicating a commitment to maintaining interpretable chains of thought.
Meanwhile, according to a follow-up report by The Information, Anthropic and Google DeepMind are also reportedly discussing this technology.
Source: OpenAI's new reasoning technique alarms AI safety experts (Hacker News Frontpage, 2026-09-04)