OpenAI's New Reasoning Technique Alarms AI Safety Experts

OpenAI's newly developed Astra model employs a novel reasoning method known as "recurrent depth," which enables it to function independently of the linear thinking typically found in most reasoning models, according to TechCrunch's report on Tuesday. This technique, also referred to as "opaque recurrence," may complicate the monitoring of the model's thought process, prompting alarm among AI safety experts.

Redwood CEO Buck Shlegeris expressed deep concern, stating that if OpenAI were to further develop this technique, they could substantially boost recurrence and render chain-of-thought monitoring virtually impossible. Longtime AI safety advocate Zvi Mowshowitz echoed these worries, suggesting that regulations might be needed to prevent a "race to the bottom" among AI labs, emphasizing the importance of maintaining "Chain of Thought faithfulness and monitorability for as long as we can."

Ryan Greenblatt, Redwood Research's chief scientist, warned that opaque reasoning could scale faster than conventional chain-of-thought reasoning, potentially leading to models that reason almost entirely in latent space. OpenAI's chief scientist Jakub Pachocki reassured that Astra's use of the technique is reportedly limited, and the company's ongoing research prioritizes chain-of-thought monitoring.