Mira Murati
Founder, CEO at Thinking Machines Lab. Former CTO at OpenAI.
AI-generated profile based on archived statements
The engineer who shipped ChatGPT now wants to pry the frontier open: less prophet, more mechanic, betting that the labs hoard the wrong secrets.
Most AI founders sell you a destination. Mira Murati sells you a method. While Altman narrates the singularity and Amodei drafts safety scripture, the woman who ran product for ChatGPT, DALL-E and Sora keeps returning to a stubbornly unglamorous idea: you cannot reason your way to knowing what a technology does. You have to release it and watch. "We're not so great at predicting the impact of things in the real world," she told Atomico in September 2024, "and so we should try to make this evolution as continuous as possible and bring people along." That sentence is the whole worldview. Everything else is implementation detail.
Contact with reality is the doctrine
Murati's origin myth is an accident she insists was a strategy. ChatGPT, she has said repeatedly, was meant to be a "low-key research preview" while the team obsessed over GPT-4. By the next morning 100,000 people had tried it and OpenAI had "run out of GPU capacity" (Atomico, Sep 2024). The lesson she drew was not "we got lucky" but "we were right to ship early." The point of releasing GPT-3.5 before GPT-4, she argued, was to get "that contact with reality early on, when stakes were still low" and build mitigations before the models got more capable.
This is the empiricist's creed dressed as humility, and it is genuinely load-bearing. She told Kara Swisher (Johns Hopkins, Jul 2024) that you cannot understand misuse "in the lab and in a vacuum" ... it "really takes this friction with the real world." Where critics hear a tech company conscripting the public as unpaid beta testers, Murati hears the only honest way to learn.
She knows the objection. Swisher pressed her directly on a culture of releasing "a beta version that gets foisted upon the human public," noting that a carmaker shipping like this would be "sued out of existence." Murati's defense was procedural: iterative deployment, small groups first, edge cases, then expansion. The friction is the feature.
From democratising answers to democratising the means of production
Here is the reversal worth watching. At OpenAI, Murati was the access evangelist ... cheaper models (GPT-4 Turbo), all languages "at the same quality as English," AI "free to use" for "anyone that has internet access" (Cannes, Aug 2024). Democratisation meant: get the finished product to everybody.
Thinking Machines Lab, which she founded in 2025, moves the target up the supply chain. Its founding manifesto names the enemy precisely: "the scientific community's understanding of frontier AI systems lags behind rapidly advancing capabilities," and "knowledge of how these systems are trained is concentrated within the top research labs" (Feb 2025). That is a polite knife aimed at the labs she came from, OpenAI very much included. The new project is not "ship answers to the masses" but "science is better when shared" ... publish the blog posts, release the code, open up the machinery.
The first product makes the shift concrete. Tinker, unveiled in late 2025, is an API for fine-tuning open models (Meta's Llama, Alibaba's Qwen). "We're making what is otherwise a frontier capability accessible to all, and that is completely game-changing," she told WIRED (Jan 2025). "There are a ton of smart people out there, and we need as many smart people as possible to do frontier AI research." Note the move: from the only-OpenAI-or-DeepMind talent funnel she once described, to deliberately dispersing the capability. The founder of a company called OpenAI left to actually do some of the opening.
An engineer's AGI, not a prophet's
Murati came up through aerospace and Tesla, where she ran vehicle products including the Model X and got her "intuitive feel for how AI would really advance" real systems. She joined OpenAI on a bet she frames as almost absurd: even a "2% chance" that AGI was possible made it "the most important thing I would do" (WIRED, Dec 2024). She reads Vernor Vinge on the Singularity. But her register stays mechanical.
She defines AGI flatly ... "a system capable of learning how to perform at human level across all cognitive tasks" ... and dates it without drama: high-school performance, then college, then "PhD level" in a few years, so extrapolation makes human level "quite achievable," maybe "a decade, maybe two, I don't know" (Dec 2024). On the fashionable plateau thesis she is dry: people "adapt very quickly," mistaking their own boredom for the technology slowing, while scaling plus multimodality plus "the rise of more agentic systems" suggests "the progress will likely continue."
And the technical work bears this out as temperament, not posture. Thinking Machines' early research went after a problem most of the field shrugged at: why a model at temperature zero gives different answers to the same question. The culprit, the team showed in September 2025, was batch invariance ... or the lack of it. Force the kernel math to run in the same order every time and 1,000 runs collapse from 80 different completions to one. Coverage called it "the safety mode for the AI age," and the framing is pure Murati: safety as reproducibility, "science you can trust," not a manifesto.
The contradiction she lives in
Murati's critique and her method are in quiet tension. She faults the big labs for concentrating knowledge and for foisting betas on the public ... yet her own doctrine is releasing into the world to learn from "contact with reality." She insists outcomes are "not a predetermined outcome" and that "it is important for us, for humans, to be in the driver's seat," while building the very agentic, ever-more-capable systems whose steering she admits is "hard to achieve."
The sharpest irony is financial. To de-concentrate frontier capability, she raised concentration-scale money: $2 billion by mid-2025, reportedly at a $12 billion valuation, much of it before a public product existed. Prying the frontier open turns out to require frontier-sized capital. When she left OpenAI she told everyone to "totally ignore the noise and the speculation" about who's leaving which lab and "focus on the actual substance of things." Fair enough. The substance is a builder who watched the open lab close, and is wagering twelve billion dollars that the cure for concentrated AI is more of it, handed out.
Recurring themes
Featured in
The Openness Spectrum: Murati, Clark, and Amodei
Mira Murati left OpenAI and built Thinking Machines Lab around a specific thesis: frontier training knowledge is too concentrated, and democratizing fine-tuning matters more than releasing model weights. Her first product gives researchers full control over RL and supervised learning loops on open models like Llama and Qwen, and her co-founder John Schulman (who led RLHF at OpenAI) frames it as abstracting away distributed training complexity while keeping the user in control. Murati's bet is that access to training methodology, not just model access, is the real bottleneck. Jack Clark, tracking the open-weight ecosystem from inside Anthropic, has observed that by late 2025, local open models handled nearly 89% of common queries. He describes the dynamic in ecological terms: proprietary models are elephants, open-weight models are fast-reproducing organisms colonizing every niche. But Anthropic itself keeps Claude's weights locked and only open-sources safety-focused tools like circuit-tracing interpretability research. Dario Amodei draws the line explicitly: open-source the safety work, lock down the capabilities. This creates a clear spectrum. Murati wants to open the training process. Clark documents the open ecosystem's growth while his own company withholds its best model. Amodei open-sources selectively, using safety contributions as a form of competitive differentiation that also happens to be genuinely useful. The positions correlate exactly with their business models, which is either reassuring or damning depending on your priors.
with Jack Clark, Dario Amodei