From Turn-Based Chatbots to Real-Time AI Interaction
For years, conversational AI agents have mostly behaved like email: you send a message, wait, then get a reply. Thinking Machines, founded by former OpenAI CTO Mira Murati, is challenging this “turn-based” model with what it calls interaction models. These systems are designed for real-time AI interaction, where audio, video, and text streams are processed continuously instead of in discrete turns. The company argues that traditional chatbots create a collaboration bottleneck, because humans must pause and hand over the floor completely before the AI can respond. That structure is convenient for the model but unnatural for people, especially when tasks require constant feedback, correction, and shared context. Interaction models aim to reverse that priority: they’re built to work with humans the way humans work with one another, collapsing the gap between thinking, speaking, and perceiving into a single, fluid loop.
How 200-Millisecond Micro-Turns Reshape Human-Computer Dialogue
At the core of Thinking Machines’ approach is a 200-millisecond “micro-turn” architecture. Instead of waiting for a full user turn, the model ingests small slices of audio video AI processing and generates output in parallel every 0.2 seconds. This makes dialog management largely implicit: the system can infer whether you are pausing to think, inviting a response, or simply breathing. It can interrupt, overlap speech, and smoothly resume—behaviours that previously required a complex tangle of external orchestration around text-only models. Benchmarks highlight the impact on latency. The flagship TML-Interaction-Small model delivers a 0.40-second response time, outpacing rivals like Gemini-3.1-flash-live and GPT-realtime-2.0 minimal. Just as importantly, a secondary background model handles deeper reasoning asynchronously while the interaction model keeps the conversation flowing, aiming to combine fast reflexes with more thoughtful, slower deliberation when needed.
Multimodal AI Models That See, Hear, and Act Proactively
The most striking shift is multimodal AI models that treat perception as continuous, not episodic. Thinking Machines’ interaction models are trained from scratch to handle audio, video, and text natively, instead of bolting separate encoders together. Audio arrives as lightweight dMel embeddings, images as 40×40 patches, all fed into a shared transformer. This lets the AI respond to what it sees changing in real time. In demos, the system counts exercise repetitions from video, translates speech on the fly, and notices posture shifts—then comments while still maintaining a broader conversation. New capabilities like visual proactivity and built-in time awareness move the agent beyond reactive Q&A. It can, for example, keep track of how long you spent writing a function or trigger reminders based on elapsed time, all without extra instrumentation. The result is a more natural, situationally aware collaborator, not just a voice attached to a text engine.
Implications for the Future of Human-Computer Communication
Thinking Machines’ research preview is early, but it signals where conversational AI agents are heading. As interaction models shrink the gap between perception and response, interfaces will likely evolve from static chat windows into live, multimodal collaborators that inhabit video calls, AR glasses, and physical devices. Murati’s team frames this as solving a bandwidth problem: letting more of a user’s intent, context, and environment flow into the system so that AI can participate in work as it unfolds, not after the fact. That raises new design questions—how often should an AI interrupt, how visible should its background thinking be, and how do users retain control in such fluid exchanges? Yet the direction is clear. Real-time, multimodal interaction isn’t just a performance upgrade; it’s a shift toward AI that communicates—and cooperates—on human terms rather than forcing humans to adapt to machine constraints.