According to TechCrunch quoting The Information, the Astra model that OpenAI is about to launch will adopt a reasoning technique called "recursive depth". This method may reduce the visible traces during the model's reasoning process, making it more difficult for outsiders to determine through logical chains why the model comes to a certain conclusion.

Security researchers issue a warning

In common reasoning models, the thought process is usually presented step by step to show how the model handles problems. Although this record does not equate to the entire internal process of the model, it remains an important tool for observing deviations from the goal, abnormal decisions, or potential malfunctions in the model's behavior.

The CEO, Buck Shlegeris, stated that he is "very concerned" about Astra's use of such technologies. He believes that if OpenAI further expands the scope of use for these methods, the monitorability of the model's thought chain could significantly decrease.

Long-term observers of the security of AI also noted that if various laboratories start competing based on their model capabilities, regulatory authorities may need to intervene in the future to prevent the industry from continuously lowering its security standards.

Circulative reasoning reduces visible traces

The report mentions that so-called "opaque recursion" refers to a situation where the model no longer primarily follows linear steps in its reasoning, but instead processes the same problem iteratively. This approach may reduce the intermediate steps that can be observed by outsiders, effectively bypassing the visible path provided by traditional thought chains.

This is also the part that security researchers are most concerned about. Because as models increasingly perform reasoning in invisible internal spaces, it becomes more difficult for researchers to determine whether they are misleading, circumventing constraints, or exhibiting other behaviors that do not meet expectations.

Redwood Research The chief scientist Ryan Greenblatt stated that a greater risk lies in the fact that such opaque reasoning may be more prone to expansion than traditional chain-of-thought reasoning, ultimately causing the vast majority of reasoning processes to deviate from visible channels.

OpenAI Emphasizes the retention of monitoring capabilities

However, at this stage, the use of this technology by Astra is said to be still quite limited. Reports indicate that the thought process of this model is still expected to be readable, and OpenAI also opposes the description of it as having shifted to a completely uninterpretable "neural language."

OpenAI The Chief Scientist Jakub Pachocki stated on X that since the company's earliest reasoning models, it has been working hard to preserve and utilize the ability of thought chain monitoring, which is also one of the core objectives of the current research plan.

It is worth noting that it is not only OpenAI that is working in this area. Subsequent reports from The Information indicate that Anthropic and Google DeepMind are also discussing similar technologies. This means that the balance between enhancing model capabilities and ensuring security and audibility may soon become a broader common topic within the AI industry.