OpenAI's Reported 'Looped Depth' Architecture Sparks AI Safety Concerns

According to The Information and other reports, OpenAI is using a new inference architecture called "recurrent depth" or "recurrent Transformer" in its Astra project, with the technology also reportedly involving breakthroughs in the direction of "neuralece/neuralese." Multiple posters (@sjgadler, @stephpalazzolo) relayed that this method allows the model to improve answers by processing the same passage of text multiple times, delivering performance gains and lower costs, but as the model scales up, its internal "thinking" process is no longer exposed in the form of an explicit chain of thought.

Confirmed

• Multiple reports (m1, m4, m5) consistently indicate that OpenAI and other companies are using recurrent Transformer-style architectures, and that such architectures hide the model's reasoning process
• The UK AI Safety Institute already noted in its May report that this type of technology could "severely undermine current monitoring methods" (m6)
• The technology delivers clear benefits in improving model performance and reducing costs (m1, m3)

Not Yet Confirmed

• OpenAI's internal stance on safety and interpretability, as well as the actual scope of deployment—posts only mention "internal and external concerns," with no official response
• Details of the "neuralese" breakthrough mentioned in m4 rest solely on The Information's account

Why It Matters

• Monitoring a model's chain of thought is seen as a currently "fragile safety opportunity" in AI safety (m3); many top AI researchers worry this architecture would undermine that mechanism, making model behavior harder to monitor and interpret
• The performance and cost gains trade off against losses in safety and interpretability, potentially shaping the industry's future choices on inference architectures and alignment work (m1, m3, m6)

2026-09-02 ~ 2026-09-02 · 6 related posts

• [source] UK AI Security Institute warns OpenAI's new reasoning technique undermines monitoring — pstAsiatech · 2026-09-02
• [source] OpenAI's 'recurrent depth' reasoning approach raises monitoring concerns — steph_palazzolo · 2026-09-02
• OpenAI quietly using loop transformers that hide 'thinking' at scale, sparking security concerns — steph_palazzolo · 2026-09-02
• [source] Report: OpenAI's loop transformer breakthrough may hide chain-of-thought — sjgadler · 2026-09-02
• OpenAI's new architecture obscures chain of thought, sparking safety concerns — sjgadler · 2026-09-02
• Report: OpenAI using loop transformers that hide thought processes — sjgadler · 2026-09-02