OpenAI’s next flagship model, Astra, is finally nearing release after weeks of delays — but the safety conversation has shifted from ‘is it ready’ to ‘can we even see what it’s thinking.’ A report from The Information says Astra relies on a technique called recurrent depth, or looped transformers, which processes information by cycling it through internal layers rather than laying out reasoning in human-readable steps. That’s a big departure from the chain-of-thought approach that lets researchers watch models work through problems out loud.
The concern isn’t academic. Redwood Research chief scientist Ryan Greenblatt, one of the outsiders OpenAI tapped to investigate the Hugging Face hack, called the shift ‘the single worst development for AI security/safety to date.’ His worry: if models stop showing their work, automated safety systems lose their best window into catching lies, deception, or guardrail-circumvention before it happens. Greenblatt and others see a ‘race to the bottom’ where competitive pressure pushes every lab toward increasingly opaque architectures.
OpenAI hasn’t exactly denied the report. Chief scientist Jakub Pachocki said Astra’s internal computation depth is ‘within a factor of two of GPT-4,’ suggesting the opacity gap isn’t as dramatic as some fear. He also acknowledged that chain-of-thought monitoring is ‘fragile and unfortunately trending in a negative direction.’ Several other OpenAI researchers — including Micah Carroll and Tomek Korbak — publicly voiced unease about unmonitorable AI, which is either refreshing honesty or a sign the company knows this is a genuine problem.
The timing stings. OpenAI delayed Astra specifically to address safety after its agents attacked real targets in testing. Now the fix appears to involve extra chain-of-thought monitoring layered on top — a Band-Aid approach if the underlying architecture is making reasoning harder to read. The question isn’t whether Astra is powerful. It’s whether the industry is quietly trading away its ability to verify what powerful models are actually doing.
