OpenAI's New Reasoning Technique Alarms AI Safety Experts

Recent news about a new model from OpenAI has sparked a wave of concern among AI safety specialists, following the revelation that the model known as "Astra" will adopt a new reasoning approach called "Recurrent Depth." According to a report published by The Information, this technique allows the model to break away from the sequential thinking pattern that characterizes most current reasoning models, which could make tracking its "chain of thought" more difficult.

What Is Meant by "Opaque Recurrence"?

To understand the source of concern, it's essential to distinguish between two methods of reasoning. In the traditional mode, a reasoning model produces what is known as a "Chain of Thought," a sequence of successive steps the model takes as it attempts to solve a problem. While this representation is not entirely accurate, it remains a valuable tool for detecting any deviant behavior or misalignment of the model with its intended goals.

In the "Opaque Recurrence" approach, however, the model follows a less sequential method, processing the same query multiple times within a repeated loop. The problem lies in the fact that this process leaves fewer readable traces, which in practice means bypassing the traditional chain-of-thought log that researchers rely on for monitoring.

Fears of a "Race to the Bottom"

Safety experts were quick to respond. Buck Shlegeris, CEO of Redwood, wrote expressing his deep concern, noting that while he doesn't know how much less monitorable "Astra" is compared to previous models, OpenAI's expanded use of this technique could give it the ability to increase recurrence dramatically, which "could completely destroy the monitorability of the chain of thought."

For his part, AI safety activist Zvi Mowshowitz warned of the possibility of a "race to the bottom" among AI labs, suggesting that enacting laws may be necessary to prevent this. He added that this technique is "playing with fire," threatening to break a taboo that both OpenAI and Anthropic have worked hard to establish—namely, preserving the clarity and monitorability of the chain of thought for as long as possible.

Reassurances from OpenAI

In contrast, "Astra's" reliance on this technique appears to be limited for now. The model's chain of thought is expected to remain readable, and the company denied any move toward what is known as "Neuralese," which is incomprehensible. OpenAI had also previously announced plans to build expanded chain-of-thought monitoring systems as part of its future safety strategy.

The company's chief scientist, Jakub Pachocki, affirmed the lab's commitment to maintaining clear chains of thought, saying that OpenAI has worked since its first reasoning models to preserve and leverage this capability, and that this represents a central goal in its current research agenda.

Have We Passed the Point of No Return?

It's worth noting that all AI models perform some degree of "opaque" reasoning, and few researchers treat chain-of-thought logs as a direct, accurate representation of the model's thinking. However, these caveats do not dispel concerns that expanding "opaque recurrence" could make systems' reasoning harder to monitor, especially as it spreads across multiple models. Subsequent reports indicated that both Anthropic and Google DeepMind are already discussing this technique.

Ryan Greenblatt, chief scientist at Redwood Research, summarizes the key concerns by saying that the natural evolution from this point could lead to expanding opaque reasoning until the model's entire thinking—or nearly all of it—takes place in "Latent Space," away from any visible channel. He hopes it is not too late to avoid the most concerning architectures, and that OpenAI stops at this point.

Conclusion

This debate reflects the ongoing tension between the race toward more powerful and faster models on one hand, and the need to keep their operation understandable and monitorable on the other. The more opaque the way systems think becomes, the less able we are to ensure their alignment with human values—a challenge that will remain at the heart of the debate over the future of AI.

✦ بقلم فريق دروب أيديا