Mira Murati and Thinking Machines Lab are building AI that does not replace humans

Mira Murati, former CTO of OpenAI and founder of Thinking Machines Lab, has demonstrated her company's new bet: AI that doesn't displace humans from the process but works alongside them. Instead of full automation, the startup is building models capable of understanding live conversation, pauses, interruptions, and context shifts in real time.

Betting on

Collaboration Murati hasn't abandoned the idea of superintelligence, but proposes a different path to it. According to her, a truly useful AI future isn't a scenario where a few corporations build fully autonomous systems while people gradually step aside. Human intelligence, conversely, should remain part of the equation as long as possible, especially against the backdrop of growing concerns that generative models will accelerate job losses and deepen market dependence on a few major players.

"The best path to a good future is to keep humans in the loop longer," is how Murati describes the Thinking Machines Lab approach.

The company's idea is not simply to provide answers on request, but to give people the ability to fine-tune advanced models for their tasks and values, then work with them in tandem. This approach differs noticeably from the familiar image of an AI agent that receives an instruction and then handles everything alone. For Murati, it matters more not to remove humans from the process, but to make the system better understand user intent and help achieve goals without loss of control.

How the

Models Work In May 2026, Thinking Machines showed a research preview of so-called interaction models—models that communicate with humans through camera and microphone. Unlike many voice interfaces, this isn't reduced to the scheme "recognized audio, converted to text, sent to chatbot." The company claims that its model is trained from the start to work with continuous, imperfect human communication, where people pause, interrupt each other, correct themselves, and change topics mid-conversation.

This gives the system a more natural mode of operation. According to Thinking Machines' description, such models should not simply wait for the end of a turn, but catch the thread of thought and adapt on the fly if the interlocutor clarifies phrasing or suddenly shifts the conversation in another direction. There's no public release yet: the company has shown only a few demo videos, but its bet is already clear—they want to make the interface with AI closer not to a command line, but to a live working dialogue between people.

• Understand pauses, interruptions, and tone shifts Work through camera and microphone, not just text Adapt when the user clarifies a thought or changes topics * Aim for more personalized and contextual interaction ## How This Differs Murati's approach runs counter to the main trend of the largest labs. OpenAI, Anthropic, and Google have increasingly driven models toward autonomous work in recent years: from code writing to assembling entire applications from a single text request. The better the system performs alone, the less human participation remains. Thinking Machines doesn't argue with the goal of making models more powerful, but with how that power should be applied in a product. The company emerged after Murati's departure from OpenAI in 2024 and assembled several notable engineers from the same ecosystem around itself. The startup has already attracted billions of dollars to create advanced models, but it has few public products yet. The most notable is Tinker, an API service launched in October 2025 for fine-tuning models on proprietary data, already used by researchers and engineers. One of the co-founders, multimodal AI specialist Aleksandr Kirillov, links the new model lineup to a broader task—to make AI not just fast, but personal and constantly present in the work context. In this logic, the assistant doesn't disappear between requests but stays nearby, sees what the user is doing, and engages at the right moment.

What

This Means The AI market increasingly discusses autonomous agents that replace part of human work. Thinking Machines proposes a different product vector: not "remove humans," but increase their throughput through models that better understand live interaction. If this approach works, the next competition in AI will go not only for model power but for the quality of human-machine collaboration.