Thinking Machines' Murati on AI's Next Chapter

ELI5/TLDR

Mira Murati ran the engineering side of OpenAI, then left to start her own AI company, Thinking Machines Lab. Her bet is that everyone else is building AI to work autonomously — go off, think alone, hand you an answer — while she wants AI that works with you in real time, like a conversation where both people can interrupt each other. She also revisits the night in 2023 when she briefly became OpenAI’s CEO during the board’s failed attempt to oust Sam Altman, and says she’d do the core of it again. The whole pitch is that keeping humans actively in the loop isn’t a safety checkbox — it’s the only durable way to build this technology.

The Full Story

The bet: AI that listens while it thinks

Most AI today is turn-based. You type, it thinks, it answers, repeat. Murati’s first product at Thinking Machines is what she calls an interaction model — built to abolish the turn-taking entirely.

“While they’re thinking it’s almost like they’re deaf and blind, they cannot perceive anything else about what’s going on. And then it’s your turn.”

Her version runs continuously instead. It chops the world into 200-millisecond chunks and keeps taking in audio, text, and video while simultaneously producing output. The point is to capture the texture of real conversation — the pauses, the interruptions, the moments two people talk over each other. She thinks that “high bandwidth interaction” is what actually lets a human stay meaningfully involved rather than just rubber-stamping.

Why bother, with OpenAI and Google already ahead

The obvious question: why start an AGI lab from scratch when OpenAI, Anthropic, Google, and Meta all have a multi-year head start. Murati’s answer is that the field has overwhelmingly chased one path — fast, autonomous systems that don’t lean on “the messiness of reality or the experience that humans have day to day.” That path is quicker. It’s also, she argues, leaving a gap.

Her preferred framing is AI as a tool for thought, in the lineage of technologies that changed not just how we work but what we’re capable of thinking about.

“Imagine if you had to do your multiplication with Roman numerals… we invented today’s numerals, and this enabled a whole area of mathematics. Like a child can do mathematics very quickly.”

The claim is that the right AI tools could expand the space of thinkable thoughts the way Arabic numerals did. Whether that’s a research program or a nice metaphor is left, conveniently, unresolved.

The board crisis, in her own words

The interview’s sharpest stretch is the 2023 OpenAI board crisis, when the board fired Altman and made Murati interim CEO — then the whole thing reversed within days. She gave the board critical feedback about Altman’s leadership when asked, and stands by it. But once she was in the chair and saw where it was heading:

“When I realized that their decision was, like, potentially catastrophic for the company… I feel like I had to act very quickly. And even though on the surface it just looked very chaotic, I think at each point in time, I felt very clear about what I had to do.”

Asked what would have happened otherwise: “I think quite likely OpenAI imploded.” Her one regret, in hindsight, is the absence of a transition plan and a post-mortem — the how, not the what. The interviewer reads her a line from former board member Helen Toner: “Mira was waiting to see which way the wind would blow, and she didn’t realize she was the wind.”

Character, governance, and not trusting one person

Pressed on whether we should trust the people holding “the dial” — should we trust Sam? — Murati pointedly redirects. Character matters, she says, but the conversation fixates on it too much.

“Even people that are well intentioned, they can make mistakes… morality is not everything. You have to think about actual decision making structures and transparency and governance.”

Her fix is structural: no governance should hinge on one person, everyone should have the tools and information to judge for themselves, and that’s part of why Thinking Machines takes a more open approach. She frames the old siloed way of building AI as a legacy of not having real-world data — something she thinks is now avoidable.

Talent wars, hype, and the future

On the nine-figure poaching deals and high-profile departures, she refuses the word “war” — that would imply the highest bidder wins, and she insists the best people aren’t primarily chasing the number. She reframes the volatility as time compression: building a frontier lab from scratch means a decade of startup drama, good and bad, gets crammed into months.

On the jobs apocalypse debate, she sidesteps both utopia and dystopia as “very simplified,” because “we actually have a lot of agency in how we build this technology.” Her recurring image is a tandem bike:

“Both people are pedaling, but when you’re going up a hill, maybe whoever is stronger is pedaling harder. But both hands are on the wheel.”

Could humans eventually leave the loop? Possibly — but if you start removing them now, during development, she sees little hope of getting it right later when systems are far more capable. A preview of the interaction models is promised “later this year,” framed as a first step. And success in five years, she says, looks less like a product and more like a feeling: people retaining “a sense of agency, a sense of dignity, and possibility about the future.”

Key Takeaways

Turn-based vs. time-based AI. Today’s chatbots are turn-based: deaf and blind while they “think.” Murati’s interaction models are continuous — ingesting audio/text/video and emitting output in 200ms chunks, so they can register interruptions, silence, and overlapping speech.
The autonomy gap. The dominant industry path optimizes for autonomous systems that minimize contact with messy human reality, because it’s faster. Murati’s contrarian bet is that the under-explored direction is tightly coupled human-AI collaboration.
“Tool for thought” thesis. Her north star is AI that changes what we can think about, the way Arabic numerals (vs. Roman) unlocked accessible arithmetic — not just AI that does tasks faster.
Her board-crisis logic. She gave the OpenAI board negative feedback on Altman when asked, accepted interim CEO, then moved to reverse course when she judged the board’s decision would destroy the company. Her stated regret is purely procedural: no transition plan, no post-mortem.
Governance over character. Her explicit position: stop obsessing over whether individual leaders are trustworthy; well-intentioned people still err. Design checks, balances, and transparency so nothing hinges on one person.
“Time compression” framing for the talent churn. She reads the nine-figure deals and abrupt exits as a decade of normal startup turbulence compressed into months by the pace of the field — not as a war won by the highest bidder.
Humans-in-the-loop as strategy, not safety theater. She argues removing humans from development now forecloses any chance of alignment when models get more capable — so keeping them in is the durable path, not a temporary guardrail.
Thinking Machines is capital-hungry by design — heavy infrastructure and foundational science spend, which she frames as the table stakes for a differentiated frontier lab rather than a fundraising trophy.

Claude’s Take

This is a polished founder interview, not a technical disclosure. Murati is articulate and unusually disciplined at staying on-message, which is exactly why you should read it as positioning. The interaction-model idea is genuinely real and concrete — continuous 200ms multimodal processing is a specific, falsifiable engineering claim, and “AI that can be interrupted” is a real limitation of current systems. That part earns respect.

The rest is vision-speak that’s hard to pin down. “Tool for thought,” “tandem bike,” “expanding what we think about” — these are evocative but do a lot of load-bearing work without a mechanism underneath. The Roman-numerals analogy is the kind of thing that sounds profound and commits to nothing. And the safety framing is convenient: “keep humans in the loop” is both her product differentiator and her ethics, which means the two can never be in tension, which is a little too neat.

The board-crisis section is the most valuable content here — a first-person account from someone who was briefly CEO during the most consequential 96 hours in recent AI corporate history, conceding her main failure was the lack of a transition plan while defending the substance. Her pivot away from “should we trust Sam” toward structural governance is either principled or a graceful dodge of a question she had no upside in answering. Probably both.

Score: 6/10. Worth it for the board-crisis candor and a clear statement of a real technical bet, dragged down by a lot of unfalsifiable founder poetry. No buy/hold/sell here — just a clear read of where one serious player is planting her flag.