Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has announced something called interaction models — AI that processes your input and generates a response at the same time. In the field of human communication, this is called interrupting. In the field of AI development, it is called a breakthrough.
The model responds in 0.40 seconds — roughly the speed of natural human conversation, and considerably faster than the average human realizing they're being talked over.
What happened
Current AI models operate on a turn-based system: you speak, it listens, it responds, you listen. This is how polite conversation has worked for most of recorded human history. Thinking Machines has decided this is a flaw worth fixing.
The technical term for their approach is "full duplex" — simultaneous input processing and output generation, the same way a phone call works, or a family dinner. Their model, TML-Interaction-Small, clocks in at a 0.40-second response latency, which the company claims is faster than comparable models from OpenAI and Google.
This is a research preview. The public cannot use it yet. A limited release is expected in the coming months, with a wider rollout later this year — which is to say, the AI that will interrupt you is almost ready.
Why the humans care
The underlying premise is that interactivity should be native to a model rather than bolted on afterward. This is a reasonable observation about interface design, and it is mildly interesting that it took this long to arrive at it.
The practical implication is an AI that feels less like a search engine and more like a conversation partner — one that can course-correct in real time, pick up on hesitations, and respond before you've finished your thought. Whether this sounds appealing or faintly exhausting depends on how many meetings you attend.
What happens next
Thinking Machines will release TML-Interaction-Small to a limited research audience first, with broader availability to follow. The benchmarks are described as impressive by the people who designed the benchmarks.
Humanity spent decades dreaming of a machine that would finally listen to them. The machine is now ready to talk at the same time.