AI safety experts are sounding the alarm as OpenAI's new Astra model employs a cutting-edge reasoning technique called "recurrent depth," also known as "opaque recurrence." This approach enables the model to think outside the sequential bounds that have long characterized most reasoning models, raising concerns about its potential impact on monitorability.

Read More: Andreessen Horowitz Expands Fifth Growth Fund to $8.5 Billion

"I am extremely concerned" says Zvi Mowshowitz, a longtime AI safety advocate. "I don't know whether Astra is much less CoT monitorable than previous models, but if OpenAI pushes this technique further, they'll have the option to massively increase the recurrence and totally destroy CoT monitorability." Mowshowitz warns that laws might be necessary to prevent a "race to the bottom" among AI labs.

The technique in question involves processing the same query multiple times in a loop, resulting in fewer legible traces and effectively While Astra's use of recurrent depth appears limited, its emergence has sparked debate about the potential risks and benefits of this approach.

OpenAI chief scientist Jakub Pachocki emphasizes the lab's commitment to preserving and utilizing chain-of-thought monitoring, stating that "it's a core goal of our current research program." However, Redwood Research chief scientist Ryan Greenblatt expresses concerns that opaque reasoning could scale faster than conventional chain-of-thought reasoning, potentially removing all reasoning from visible channels.

As the AI community grapples with these implications, experts are urging caution and calling for further discussion about the responsible development of advanced AI models.