Thinking Machines: Why AI Models Should Learn to Interact in Real Time (AI column)

Dealing with article Thinking Machines (Mira Muratti is a former CTO OpenAI company.) It is about interaction models, in which the authors say that the very way of communicating with AI becomes a limitation. Now the usual pattern of step-by-step interaction looks like this: the person asked the request, the model thought, the model answered. Even with a voice, camera, or agent harness, the logic is often the same inside. Thinking Machines calls this collaboration bottleneck. The problem is not only in the intelligence of the model, but in the narrow channel of interaction. In real work, people interrupt, clarify, point to the object and gradually figure out what to do.

And the guys offer an alternative approach - to make interactivity part of the possibilities of the model, rather than twisting the external binding. Nowadays, real-time systems often use VAD, dialog management, ASR/TTS and interruption handling rules. According to Thinking Machines, such a harness does not scale well: intelligence grows in the model, and interactivity remains outside. (The Bitter Lesson by Richard Sutton, one of the founders of Reinforcement Learning).

As a result, they trained the interaction model, which perceives audio, video and text as continuous streams. Research preview is called TML-Interaction-Small: 276B MoE, about 12B active parameters. The model was trained from scratch under real-time interaction. The main technique is time-aligned micro-turns. The entrance and exit are cut into pieces. 200 In each micro-turn get audio, video, text and response model. The model lives in the flow of time, not waiting for a full turn.

The architecture is two-layered. Interaction model next to the user: listen, watch, respond, hold the context. Hard work - reasoning, tools, browsing, agentic workflow - she delegates background model. It works asynchronously, and the interaction model builds the result back into the dialogue. Measure it separately. FD-bench checks interruption, backchannel, other people's speech in the background and turn-taking latency. Audio MultiChallenge looks at instruction following in audio tasks. Internal TimeSpeak and CueSpeak Check the reaction at the right time. For visual proactivity adapt RepCount-A, ProactiveVideoQA and Charades.

While these interaction models are in the research preview and are not a mass product. Thinking Machines says it will open a limited research preview in the coming months, with a wider release planned later in the year. 2026 year. Next, they plan to scale the model, improve background agents, manage context in long sessions, increase reliability in latency and develop safety for real-time multimodal interfaces. They also launched interactivity research grants: $100k and $25k Tinker credits for evals, safety, generative UI and steering agents. In the announcement of the deadline - 19 June 2026.

P.S. Thinking Machines already has a product in GA called Tinker, a training API for fine-tuning open-source models via LoRA. So we believe that interaction models will soon be available as a public product:) And then we can test the thesis that the next important shift will be not only in reasoning and agents, but also in the interface: models will begin to work next to a person in the general flow of time.

#AI #Engineering #Research #Product #Architecture #Software