Thinking Machines Lab Previews Unreleased Interaction Models That Process Live Audio and Video

Thinking Machines Lab this week previewed a new kind of AI model called interaction models that are trained to communicate with a person through a camera and microphone. The models natively understand continuous, messy, human communication including pauses, interruptions, and changes in tone.

The company showed off several videos demonstrating the capabilities of the interaction models, though the models have not been released publicly.

Mira Murati, the founder of Thinking Machines Lab and formerly the chief technology officer of OpenAI, left her role at OpenAI in 2024. She cofounded Thinking Machines Lab with several prominent engineers, including Alexander Kirillov, a founding team member and expert on multimodal AI. Thinking Machines Lab has raised billions of dollars to build frontier AI.

So far the company has released only one product. Tinker launched in October 2025. It makes it possible to refine a frontier AI model using custom data and is available as an API that researchers and engineers can use to fine-tune open source models.

The new interaction models stand in contrast to systems that simply capture and transcribe speech before feeding it into a language model. Instead they adapt on the fly when someone clarifies a point or changes the subject. Wired reported that the models constantly perceive what a user is doing and remain ready to reply, give information, search or use other tools.

Alexander Kirillov says the interaction models represent capabilities absent from today's other systems. “The model constantly perceives what you're doing and is constantly there to be able to reply and give you information or search for information or use other tools. This is something that none of [today’s other] models can actually do.