OpenAI's 'recurrent depth' method sparks safety concerns

OpenAI's upcoming Astra model will use a reasoning technique called 'recurrent depth' that makes its chain-of-thought less transparent. AI safety experts worry this could reduce the ability to monitor model behavior and lead to a 'race to the bottom' among labs. OpenAI says the technique's use is limited and that it remains committed to legible reasoning.

The technique processes queries in iterative loops rather than sequential steps, leaving fewer legible traces for oversight. Chain-of-thought records previously proved valuable in diagnosing OpenAI's rogue agent incidents, making their potential loss particularly significant for internal safety investigations.

Redwood Research's chief scientist warned that scaling opaque reasoning could push models to reason entirely in latent space, eliminating visible reasoning channels altogether. Meanwhile, reports indicate Anthropic and Google DeepMind are already discussing similar approaches, though OpenAI maintains its commitment to transparent reasoning and has announced plans for extensive chain-of-thought monitoring systems.