OpenAI’s new reasoning technique alarms AI safety experts

OpenAI is introducing a technique called recurrent depth for its Astra model, which enables the system to process information beyond standard sequential thinking patterns. This shift away from traditional linear reasoning allows the model to potentially solve more complex problems, but it has prompted concerns among AI safety researchers regarding the predictability and control of such architectures. The development marks a technical departure from current industry standards for reasoning models, raising questions about how experts will monitor and manage autonomous decision-making in future iterations.