OpenAI's Next Model Ditches Step-by-Step Thinking, Worrying Safety Researchers
OpenAI is reportedly building a new model, internally called Astra, that reasons differently from current systems like o1 or o3. Instead of working through a problem in a linear, step-by-step chain that can be read and checked afterward, Astra reportedly uses 'recurrent depth' — cycling its internal computation through loops before producing an answer.
This could make the model more efficient or capable at certain problems, since it isn't locked into producing a token-by-token trail of its reasoning. But that same trait is what's worrying some AI safety researchers: chain-of-thought outputs, however imperfect, have become one of the few tools available for spotting when a model is scheming, hallucinating, or reasoning toward a harmful conclusion. A model that thinks in opaque loops removes much of that visibility.
OpenAI hasn't detailed how it plans to monitor or interpret Astra's internal reasoning if the technique ships in a production model.