Thinking Machines Proposes Chat Bots That Can Talk Over You — And Still Be Useful

Thinking Machines Lab, the startup launched last year by former OpenAI chief technology officer Mira Murati, unveiled a new approach Monday that aims to make conversational AI behave more like a live phone call than a back-and-forth text exchange. The company says its prototype, called TML-Interaction-Small, can begin replying while still ingesting a user’s words — responding in about two-fifths of a second, a pace the firm argues is roughly on par with human conversation and quicker than comparable systems from larger rivals.

Traditional assistants operate in clear turns: a user speaks or types, the model processes the input, then it replies. Thinking Machines wants to collapse that sequence by processing incoming speech and generating output at the same time. The firm describes the method as true simultaneous send-and-receive processing, allowing the model to interrupt or overlap responses in ways current systems do not.

For now the work is a research preview rather than a consumer product. The company plans a limited research release in the coming months and says broader access could follow later this year. Thinking Machines has posted performance numbers on its site, but access remains restricted until those previews begin.

Outside experts say the idea could make voice assistants feel more natural, but they warn of trade-offs. One conversational-AI researcher noted that quicker turn-taking could improve flow, yet could also create awkward interruptions and raise new challenges for safety, user control and error correction. An industry analyst added that benchmarks are only one piece of the puzzle; real-world testing will reveal whether the responsiveness translates to better usability.

Until that testing happens, the promise of overlapping speech remains unproven. The company’s claims are technically notable, and observers will be watching closely once researchers and everyday users get hands-on time.