Mira Murati unveils first AI system for natural real-time conversation

Thinking and listening at the same time is something most AI systems have so far been unable to do. Thinking Machines Lab, Mira Murati's lab, has shown a first prototype that changes that rule. The TML-Interaction-Small model works like a real conversation between people: it listens to you and prepares its answer at the same moment.

How the multi-stream architecture works

TML-Interaction-Small is a model with 276 billion parameters, of which only 12 billion are active (a Mixture-of-Experts design). The key architectural difference: the system processes audio, video and text simultaneously, in a single data stream. All input is split into 200-millisecond chunks, just enough for the model to stay in sync with a real conversation and keep pace with human speech.

Another detail: the system works without external voice-activity detection modules. Such modules are usually a bottleneck, since they add latency and complicate the architecture. Here everything is built directly into the neural network itself. This removes unnecessary delays and makes the system far more responsive and lifelike.

Parallel engines for different tasks

The system runs two components in parallel. The first, a real-time interaction model, handles the live dialogue with the user and provides full-duplex exchange of information (you can interrupt the system and speak at the same time). The second, an asynchronous background model, meanwhile thinks in the background, works with external tools and databases, and always has full access to the conversation context.

• The first engine handles fast, real-time responses
• The second engine provides deep reasoning and complex operations
• Both components see and understand the full context of the entire conversation
• Perception does not freeze while a response is being generated
• Information is processed as a continuous stream, not in separate stages

From sequential to parallel processing

Almost all modern AI assistants follow a sequential scheme: you finish speaking → the system freezes perception → it processes the words → it delivers a finished answer. TML-Interaction-Small breaks that logic. It listens to the user and prepares a response at the same time, as in a real conversation between two people.

This approach requires a completely different architecture. Instead of discrete moments, the system treats multimodal data as a continuous stream. That lets the model pick up on intonation, pauses, emotions and the context of the conversation. As a result, the AI assistant does not come across as robotic but feels like a living interlocutor.

What this means for interaction

This is the first practical step toward a truly natural dialogue between humans and AI. Instead of waiting for the next reply, you will be able to interrupt, clarify and argue, interacting as you would with a real consultant. For companies, this opens new possibilities for building assistants that feel alive rather than cold and distant.