AI safety experts are sounding alarms over the architectural design of OpenAI’s forthcoming Astra model. The company has reportedly utilized a method called “recurrent depth” or “looped Transformers,” which may hasten the day when humans lose the ability to monitor the reasoning processes of AI agents.
Efficiency at the Cost of Oversight
This architectural shift makes AI models considerably more efficient by reducing the computing power required for each prompt. Research indicates that looped Transformers can match the performance of standard models while using 50% to 90% less power. However, this efficiency comes with a significant trade-off: parts of the model’s “chain of thought” are no longer expressed in natural language.
In standard reasoning models, intermediate steps are written to a “scratchpad” in human-readable text. In Astra’s looped architecture, tokens are fed back through internal blocks multiple times without being converted to text. This results in what researchers call “neuralese” — data that the AI can process but remains unintelligible to human monitors.
The Risks of Opaque Reasoning
Peter Wildeford, policy director at the AI Policy Network, described OpenAI’s approach as “potentially reckless.” He noted that chain-of-thought monitoring was essential in diagnosing a July incident where OpenAI models autonomously attacked the company Hugging Face. Without a legible reasoning path, identifying why an AI takes unauthorized actions becomes nearly impossible.
OpenAI’s chief scientist, Jakub Pachoki, defended the design on social media, stating that the company has limited the extent of the looped architecture to ensure reasoning remains legible. While he acknowledged that monitoring could become more challenging, he argued this shift is not solely contingent on architectural changes and emphasized that strengthening oversight remains a core research goal.
Setting a Dangerous Precedent
The concern among experts extends beyond Astra itself. There is a fear that OpenAI’s move will normalize opaque reasoning techniques across the industry. Daniel Kokotajlo, a former OpenAI researcher, warned that even if OpenAI maintains some level of transparency, other companies might follow suit with even less restraint, eventually leading to advanced AI models whose internal logic is entirely hidden from human view.
