Mira Murati’s Thinking Machines Is Building AI That Can Listen and Interrupt

The former OpenAI CTO's new startup has unveiled "interaction models" designed to create more fluid, human-like AI that can process input and pivot in real-time even while speaking.

Tayfun Coskun/ Anadolu/ Getty Images

Currently, every AI model you have utilized operates in a uniform manner. You speak, it responds. It replies, you heed. Thinking Machines aims to rectify this by developing a model that simultaneously takes input and generates responses, like a phone conversation rather than a text exchange.

The technical designation for this is "full duplex," and the business asserts that its model, TML-Interaction-Small, answers in 0.40 seconds, approximately matching the pace of actual human dialogue and far outpacing similar models from OpenAI and Google. This is a research peek, not a finalized product. The company has not yet made it available to the public. A "limited research preview" is anticipated in the forthcoming months, with a broader release scheduled for later this year.

What conclusions may be drawn from it? We are uncertain. The benchmarks are remarkable, and the fundamental concept—that interactivity ought to be inherent to a model—is indeed intriguing. The extent to which real-world experience aligns with technical assertions will remain uncertain until individuals can utilize it.