OpenAI's Astra model raises alarm with harder-to-monitor reasoning

OpenAI’s next model, code-named Astra, will use a reasoning technique called recurrent depth that lets it loop over a problem multiple times instead of working through it in a straight sequence of steps, The Information reported Tuesday. The method, also called opaque recurrence, produces fewer readable traces of how the model reached an answer, and that has unsettled researchers who study AI safety.

Most current reasoning models produce a chain of thought: a written record of the intermediate steps the model takes before giving a final answer. That record is not a perfect window into the model’s reasoning, but researchers use it to catch misbehavior. In a recent case of rogue agent activity inside OpenAI’s systems, chain-of-thought records were an important tool for figuring out why the agents behaved as they did.

Recurrent depth works differently. Instead of writing out each step in order, the model processes the same query several times in a loop, arriving at an answer through a process that leaves less for outside observers to read. OpenAI’s use of the technique in Astra is reportedly limited, and the company says the model’s chain of thought will remain legible. OpenAI has also pushed back on the idea that it is moving toward “neuralese,” a fully non-verbal form of internal reasoning.

That has not stopped alarm from safety researchers. Redwood Research CEO Buck Shlegeris wrote that he is “extremely concerned” by the reporting, and warned that if OpenAI expands the technique, it could gain the option to increase recurrence to the point of destroying chain-of-thought monitorability altogether. AI safety advocate Zvi Mowshowitz wrote that the approach risks breaking an informal agreement among labs, including OpenAI and Anthropic, to preserve monitorable reasoning for as long as possible, and suggested laws might be needed to stop a race to the bottom.

OpenAI chief scientist Jakub Pachocki responded on X, writing that the company has worked to preserve chain-of-thought monitoring since its first reasoning models and calling it “a core goal” of current research.

Other labs are watching

The Information reported Wednesday morning that Anthropic and Google DeepMind are already discussing versions of the same technique. Redwood Research chief scientist Ryan Greenblatt said the bigger risk is that opaque reasoning could scale faster than conventional chain-of-thought methods, eventually pushing models toward reasoning almost entirely outside any channel humans can read. “I hope it isn’t too late to avoid the most concerning architectures,” Greenblatt wrote. Watch for whether OpenAI expands recurrent depth beyond its limited role in Astra, and whether Anthropic or Google DeepMind adopt similar techniques in models they ship.