Microsoft-backed (MSFT) OpenAI's upcoming Astra AI model uses "recurrent depth" to improve performance while obscuring some reasoning, raising concerns over AI safety monitoring, The Information reported Tuesday.
According to the report, OpenAI has limited the technique in Astra to preserve a monitorable chain of thought and plans additional monitoring to rapidly detect and contain potential misbehavior.
Researchers warn that broader use of opaque reasoning could weaken AI safety measures, while recurrent depth may also lower computing costs by enabling smaller models to deliver larger-model-like performance, the report added.
OpenAI did not immediately respond to MT Newswires' request for comment.
(Market Chatter news is derived from conversations with market professionals globally. This information is believed to be from reliable sources but may include rumor and speculation. Accuracy is not guaranteed.)