OpenAI's Astra model will use a "recurrent depth" reasoning technique that makes its chain of thought harder to monitor, drawing warnings from AI-safety researchers about an opaque-reasoning race across the industry.
返回
OpenAI Astra's opaque reasoning alarms safety experts over 10T model
OpenAI's Astra model will reason through a technique called "recurrent depth" that sidesteps the sequential chain-of-thought used by most reasoning systems, making its internal logic harder to trace — a shift that has AI-safety researchers warning of an opaque-reasoning race across the industry.
"I am extremely concerned by the reporting that Astra uses opaque recurrence," Buck Shlegeris, chief executive of AI-safety nonprofit Redwood Research, said in a post after the news broke. "If OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroy chain-of-thought monitorability."
The technique, also called "opaque recurrence," was reported Tuesday by The Information. Instead of producing a legible sequence of reasoning steps, the model processes the same query several times in a loop, leaving fewer traces for outside observers to inspect. Astra's use of the method is reportedly limited — its chain of thought is still expected to be legible — and OpenAI has pushed back against any suggestion it would shift to "neuralese," the inscrutable internal representations some researchers fear.
The stakes extend beyond a single model. Redwood Research chief scientist Ryan Greenblatt said opaque reasoning could scale faster than conventional chain-of-thought, potentially removing reasoning from visible channels altogether. "My biggest concern is that a natural progression from here would involve scaling up the opaque reasoning to the point where the model reasons entirely or almost entirely in latent space," Greenblatt wrote. "I hope it isn't too late to avoid the most concerning architectures and that OpenAI will stop here."
Can Chain-of-Thought Monitoring Survive Astra?
The alarm reflects a broader tension. Chain-of-thought logs, while an imperfect representation of a model's reasoning, have become a core safety tool — they helped researchers tease out why OpenAI's rogue agents behaved as they did in a recent incident. Opaque recurrence erodes that visibility precisely as models grow more capable.
OpenAI has sought to contain the concern. Chief scientist Jakub Pachocki reaffirmed the lab's commitment to legible chains of thought. "OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models," Pachocki wrote. "It's a core goal of our current research program." The company has also announced plans for extensive chain-of-thought monitoring as part of its forward-looking safety plans.
The competitive stakes are mounting. In a follow-up report Wednesday, The Information said both Anthropic and Google DeepMind were already discussing the technique — a sign that opaque recurrence could become a standard feature of next-generation reasoning models rather than an OpenAI outlier. Astra, expected to ship as ChatGPT 6 in a release window of Sept. 3-10, reportedly packs around 10 trillion parameters, positioning it against Anthropic's Claude and Google's Gemini line.
Safety advocates argue the dynamic could force a race to the bottom. Zvi Mowshowitz, a longtime AI-safety writer, said laws might be necessary to prevent labs from competing on opacity. "The technique is playing with fire, risking a taboo that OpenAI and Anthropic have fought to establish that we work hard to maintain chain-of-thought faithfulness and monitorability for as long as we can," Mowshowitz wrote.
All AI models do some quantity of opaque reasoning, and few researchers treat chain-of-thought logs as a direct record of a model's thinking. But the concern is that Astra's architecture makes that opacity a design choice rather than a byproduct — one that could compound as the technique spreads to rivals.
For investors, the question is whether opacity becomes a liability. OpenAI's valuation and its standing among AI leaders hinge partly on trust in its safety posture, which regulators and enterprise buyers increasingly weigh. If opaque recurrence draws scrutiny from policymakers — or if a high-profile misalignment slips past weakened monitoring — the reputational cost could outweigh any capability gain. Anthropic and Google DeepMind, by contrast, could market legible reasoning as a differentiator if they hold the line on transparency.
This article is for informational purposes only and does not constitute investment advice.
