Thinking Machines built an AI that actually interrupts you like a human

If you have used any live AI voice mode recently, you know the awkward dance: pause for half a second to gather your thoughts, and the AI cuts you off. Try to interject while it rambles, and you are scrambling for the mute button.

Today's frontier models experience reality on a single thread. You talk, they wait blindly; they generate, their ears shut down until you hit interrupt. Current commercial voice systems fake natural conversation by wrapping models in external scaffolding like separate voice activity detection tools. Thinking Machines Lab, the startup founded last year by former OpenAI CTO Mira Murati, wants to burn that playbook down.

On Monday, Thinking Machines unveiled a research preview of what it calls interaction models: a native architecture designed to process audio, video, and text continuously rather than in rigid back-and-forth turns.

Ditching the scaffolding for 200ms micro-turns

Instead of waiting for you to finish a complete sentence, Thinking Machines trains its model to work across interleaved two-hundred-millisecond processing windows. The model ingests two hundred milliseconds of streaming input and produces output at the same cadence, treating both sides of the conversation as live concurrent streams.

Because interactivity is baked straight into the weights rather than bolted on through a harness, the system natively tracks whether you are trailing off to think, correcting yourself, or handing over the floor. It can interject verbally or visually when you make a mistake, comment while watching live video, or speak concurrently for real-time translation.

To balance split-second reflexes with heavy thinking, Thinking Machines splits the labor. A time-aware interaction model stays present on the line, fielding interruptions and holding the conversation, while an asynchronous background model crunches long-horizon tasks like web search, tool use, and generative UI. The company claims its TML-Interaction-Small model clocks a response latency of 0.40 seconds, matching natural human pace and beating comparable offerings from OpenAI and Google.

Why it matters

We think Thinking Machines is attacking the exact bottleneck that makes working with frontier models feel so clunky today. AI labs have spent the past year obsessing over autonomous agents that take a prompt and disappear for hours. But in high-context creative or engineering work, nobody can specify every requirement upfront. You need the AI right there with you, catching subtle visual cues and letting you push back mid-sentence.

Baking full-duplex interactivity directly into model weights follows the classic bitter lesson: hand-crafted engineering wrappers around turn-based models will inevitably hit a wall. When interaction scales natively with model size, making the model smarter automatically makes it a better partner in the room.

That said, temper your expectations for now. This announcement is strictly a research preview, not a shipping consumer app. Thinking Machines says a limited preview will roll out to testers over the next few months, with broader availability slated for later this year. The 0.40-second benchmark sounds incredible on paper, but we will have to see how gracefully that dual-model handoff holds up once real users start firing chaotic, half-baked questions at it in the wild.