Bloomberg Live: Mira Murati's Thinking Machines Is Building AI That Listens Between the Words — and That's the Whole Point
Bloomberg Tech 2026, San Francisco — June 4, 2026
Mira Murati made her most substantive public case yet for why Thinking Machines Lab exists and what it is actually building. Speaking with Bloomberg's Emily Chang at Bloomberg Tech 2026, the former OpenAI CTO and co-founder laid out a technical and philosophical thesis that cuts directly against the grain of how every major AI lab is currently competing. The core insight is deceptively simple but architecturally significant: today's AI models are deaf and blind while they think, and that is a fundamental design flaw worth solving from the ground up.
The Interaction Model: A New Architecture, Not a Feature
The most concrete and investable piece of information to emerge from the conversation was Murati's description of Thinking Machines' "interaction models," which the company recently unveiled as its first public demonstration of its research direction. This is not a chatbot upgrade. The distinction she draws is architectural. Current frontier models, she explains, are turn-based — the user speaks, the model processes, the model responds. During inference, the system is effectively isolated from the world. "While they're thinking, it's almost like they're deaf and blind," Murati said. "They cannot perceive anything else about what's going on."
Thinking Machines is building something structurally different: a time-based model that continuously ingests audio, text, and video, and continuously produces output, chunked into 200-millisecond windows. That granularity is what enables the system to detect interruptions, simultaneous speech, pauses, and the ambient cues that make human communication high-bandwidth. "There is a lot of information in our interactions when we are silent, when we're thinking, when we are interrupting one another," she said. The practical ambition is to close the gap between how humans communicate with each other and how they communicate with machines — which Thinking Machines views not as a user experience nicety but as a prerequisite for meaningful human-AI collaboration and, ultimately, alignment.
The Strategic Bet the Big Labs Are Not Making
Murati was careful not to position Thinking Machines as a direct challenger in the benchmark arms race. Instead, she identified a specific whitespace she believes is structurally underserved: bringing machine intelligence closer to where human knowledge actually lives. The dominant path being pursued by OpenAI, Anthropic, Google, and Meta, in her framing, is a highly autonomous one that deliberately minimizes dependence on the "messiness of reality." She concedes that is a valid and fast approach. But she argues it leaves behind an entire dimension of value.
"The most advanced AI systems are the most incredible tools for thought that humanity can ever have," she said, drawing an analogy to the invention of modern numerals replacing Roman numerals — a cognitive prosthetic so powerful that it unlocked entirely new categories of mathematics. The strategic thesis at Thinking Machines is that AI systems built for genuine human collaboration, rather than autonomous task execution, will not only be more useful in practice but will also produce better alignment outcomes as a byproduct. That last point matters to investors trying to assess regulatory and safety risk across the sector: Murati is arguing that the collaboration-first architecture is itself a technical solution to alignment, not merely an ethical stance.
On OpenAI, the Board Crisis, and What She Would Do Differently
Chang pressed Murati on the November 2023 board crisis at OpenAI, including testimony Murati gave under oath that she feared catastrophic risk to the company. Murati did not walk anything back. "I think quite likely, OpenAI would have imploded," she said flatly, when asked what would have happened had she not acted. She confirmed she shared critical feedback about Sam Altman's leadership when the board asked for it, stepped up as interim CEO when asked, and reversed course when she concluded the board's original decision threatened the company's survival. "At each point in time, I felt very clear about what I had to do."
The one thing she says she would do differently in retrospect: spend more time on transition planning and transparency. "There wasn't much thought put into bringing the team along, providing the continuity." The candor is notable. More structurally interesting to investors, however, is her broader commentary on AI governance. She explicitly argues that the industry's conversation about safety is too focused on the character of individual leaders and not enough on institutional design. "Morality is not everything — you have to think about actual decision making structures and transparency and governance," she said. That worldview is directly reflected in Thinking Machines' stated commitment to a more open approach in its lab.
Talent, Capital, and the Reality of Building From Zero
Reports of nine-figure compensation deals at Thinking Machines have circulated widely, as have accounts of some high-profile departures. Murati acknowledged both without much deflection. She described building a frontier AI lab from scratch as compressing five to ten years of normal startup turbulence into months. "Whenever you have something good, something bad will come with it." On compensation, she was measured: "The high compensation numbers captured the imagination of people because obviously they're very big. But I do think that for the most sought-after people, that's not the main story." Whether that is true or aspirational will be tested over the next twelve months as the hyperscalers continue writing large checks for top researchers.
On the company's fundraise — widely reported as one of the largest in AI startup history — Murati was notably unsentimental. "We're proud of that, but that's not any big accomplishment. We didn't set out to break some sort of record." The more relevant point for sizing capital requirements: she was direct that Thinking Machines is not a normal company and needs substantial capital to build the infrastructure and scientific foundations required to have a credible claim on frontier AI development.
What Comes Next and the Five-Year Vision
Murati confirmed that the interaction models released recently are explicitly a first step, and that Thinking Machines expects to show "increased capabilities on the model side" and additional products in the coming months. She stopped short of specific timelines or benchmarks, but the message was clear: the company is in execution mode and intends to let the technology speak rather than the narrative. A broader preview of its frontier model work is expected later this year.
When pressed on a five-year vision if Thinking Machines succeeds beyond expectations, she declined to forecast specific products or market positions, pivoting instead to a framing that was equal parts philosophical and strategic: a future where people retain a genuine sense of agency and possibility regardless of how dramatically AI reshapes the nature of work. It is a vision that is harder to stress-test than a revenue model, but it is consistent throughout — the tandem bicycle, as she puts it, where "both hands are on the wheel," rather than a system that autonomously outpaces the civilization it is supposed to serve.