The AI becomes a tool you consult rather than a partner you think alongside.
Mira Murati, once the mind behind OpenAI's technical direction, has stepped forward with Thinking Machines — a startup whose first public act is a quiet but consequential one: teaching AI to listen and speak at the same time. In doing so, she is addressing something older than technology itself, the frustration of conversation that does not flow, of thought interrupted by delay. The demonstration, shown in May 2026, points toward a future where AI is not a tool one consults in pauses, but a presence one thinks alongside in motion.
- The friction of turn-based AI — speak, wait, receive — has quietly eroded the promise of human-AI collaboration, and Thinking Machines is now directly targeting that gap.
- By enabling an AI to listen while simultaneously generating speech and processing visual cues, the company is attempting something technically demanding that most systems have deliberately avoided.
- Workplaces leaning on AI for brainstorming and decision support have hit a wall where every stilted exchange chips away at productivity and trust in the technology.
- Thinking Machines has not released the models publicly yet — the unveiling was a directional signal, not a product launch, leaving the real-world proof still ahead.
- The move reframes the competitive frontier in AI: not who has the most powerful model, but who has made AI genuinely natural to work with.
Mira Murati, former chief technology officer of OpenAI, has brought her startup Thinking Machines into public view with a demonstration that targets one of the quieter frustrations in AI: the pause. Most conversational systems today are turn-based — you speak, the system processes, then it responds. The lag is noticeable, and it breaks the rhythm that makes human dialogue feel alive. Thinking Machines' interaction models attempt to collapse that gap, allowing the AI to listen and respond simultaneously, to interject and build on what is being said in the moment.
The technical demands are real. Managing simultaneous streams of voice and video input, reading facial expressions and gestures, making real-time decisions about when and what to say — these are not trivial problems. The company's demonstration showed the system handling both audio and visual channels, suggesting a model of presence rather than mere processing.
The deeper argument Murati is making is about collaboration. Teams that rely on AI for problem-solving and brainstorming lose something when every exchange feels stilted. The AI becomes something you query rather than someone you think with. Real-time conversation could restore that sense of flow, shifting AI from database to colleague.
The timing reflects where the industry has moved. The initial wave of large language model excitement has settled, and the harder question now is usability — how to make AI fit naturally into daily work. Thinking Machines has not yet released technical details or opened access, and whether the technology holds up at scale remains to be seen. But the direction is clear: the next frontier in AI is not raw power. It is the quality of the conversation itself.
Mira Murati, the former chief technology officer of OpenAI, has launched her startup Thinking Machines into public view with a demonstration of conversational AI that works at human speed. The company unveiled what it calls interaction models—software designed to handle voice and video conversations in near-real time, with the system capable of listening and responding simultaneously rather than waiting for a speaker to finish before processing begins.
This is a meaningful shift in how AI conducts dialogue. Most conversational systems today operate in a turn-based fashion: you speak, the system processes, the system responds. The lag can be noticeable, and it breaks the natural rhythm of human exchange. Thinking Machines' approach attempts to collapse that gap, allowing the AI to interrupt, interject, and build on what's being said in the moment, much as a human conversation partner would.
The technical challenge here is substantial. For an AI to listen while generating speech requires the system to manage multiple streams of input and output simultaneously, to understand context in real time, and to make split-second decisions about when to speak and what to say. The company's demonstration showed this working across both voice and video, suggesting the models can process visual cues alongside audio—facial expressions, gestures, the physical presence of the person on the other end.
Why this matters extends beyond the novelty of smoother conversation. Murati and her team are positioning this as a solution to what they frame as a collaboration bottleneck. In workplaces, teams increasingly rely on AI for brainstorming, problem-solving, and decision support. But if every exchange with the AI feels stilted or delayed, the friction accumulates. Workers lose the sense of flow. The AI becomes a tool you consult rather than a partner you think alongside. Real-time interaction models could change that dynamic, making AI collaboration feel less like querying a database and more like working with a colleague.
The startup's timing is notable. The AI landscape has stabilized somewhat after the initial shock of large language models. Companies are now focused on the next layer of problems: how to make these systems actually useful in daily work, how to integrate them into existing workflows without creating new friction. A system that can have a natural conversation—that listens, responds, adapts—addresses a real gap in current offerings.
Thinking Machines has not yet released detailed technical specifications or made the models available to the public. The demonstration was a preview, a signal of direction. But the company's emergence from stealth with this particular capability suggests Murati sees real-time conversational AI as the frontier worth pursuing. Whether the technology delivers on its promise at scale, and whether it actually solves the collaboration problems companies face, remains to be tested. What's clear is that the conversation about what AI should do next has shifted from raw capability to usability, from impressive to practical.
Notable Quotes
Thinking Machines wants to build an AI that actually listens while it talks— TechCrunch reporting on the company's stated goal
The Hearth Conversation Another angle on the story
What's the actual difference between what Thinking Machines built and what, say, ChatGPT can already do?
The core difference is simultaneity. ChatGPT waits for you to finish speaking before it starts thinking. Thinking Machines' system can process your words while you're still saying them, and start responding before you're done. It's the difference between a turn-based game and a real conversation.
But why does that matter so much? People don't mind waiting a second or two for an answer.
They don't mind consciously. But it changes the texture of the exchange. When there's lag, you lose the back-and-forth momentum. You can't interrupt each other naturally. You can't build on half-formed thoughts. It stops feeling like thinking together and starts feeling like you're waiting for a machine to process.
So this is really about the feeling of it, not the intelligence itself.
Exactly. The AI might be just as smart. But if it can't keep pace with human rhythm, it becomes a bottleneck. Murati's bet is that teams will actually use AI more, and use it better, if the conversation doesn't feel broken.
Has anyone else tried this?
There's been research in this direction, but Thinking Machines appears to be the first to demonstrate it working reliably across voice and video at scale. That's the claim, anyway. We'll see if it holds up.
What happens if it does work?
Then the entire category of AI assistants shifts. You're not consulting a tool anymore. You're collaborating with one. That changes what people ask it to do, and what they expect from it.