Mira Murati: OpenAI's Quiet Operator Bets Big on Her Own Lab

An episode of Dan's AI Intel

The operator who made OpenAI ship has a frontier lab's chips on order but not yet its model, and 2027 decides which she has built.

Published · By Dan Walter

Executive summary

Mira Murati is the operator who made OpenAI ship, and she was also, by sworn testimony, the source of much of the evidence that nearly removed its chief executive. That double role is the key to her. She ran the teams behind ChatGPT, DALL-E, Codex and Sora without ever being the company's face, held the top job for about two and a half days in November 2023, and watched the staff and investors overrule the board that had appointed her. That weekend showed that at a frontier lab, power sits with the people who build and the money that pays them. Then she built a lab where she holds that power herself: at Thinking Machines Lab, her vote is weighted to give her a majority on board matters.

Murati represents the operator-builder rather than the founder-celebrity, and her lab is today a well-funded question mark with a third pole's inputs, not yet a third pole. Thinking Machines has a gigawatt of Nvidia chips on order for early 2027, a reported run rate above $100 million, and a funding round reported at $40 billion before new money, but its own flagship model, by the lab's own account, is not the strongest available. It has also lost three co-founders. It becomes a genuine third pole only if that 2027 compute produces a model at the frontier.

Her trajectory matters because it is the clearest bet that frontier AI can be open, customisable and spread across many hands rather than concentrated in a few labs. We hear less about her because she sells developer tools, publishes through unsigned company essays and speaks rarely. That quiet ends within months, not because she will change, but because the round, the model weights and the chips will each make news on their own.

Why one engineer's career is a map of the AI race

The AI revolution is usually told through its founders: Sam Altman, Dario Amodei, Demis Hassabis, Elon Musk. Murati is the other kind of person every lab needs and few outsiders can name, the one who turns research into products that hundreds of millions of people use. Her career runs through the three questions that decide who controls frontier AI: who gets to ship it, who gets to govern it, and whether anyone outside the largest labs can build it at all. Following her answers all three.

The engineer behind ChatGPT was never its face

Murati was OpenAI's chief technology officer from May 2022 to September 2024, the executive who ran the teams that shipped ChatGPT, DALL-E, GPT-4 and Sora, while Sam Altman was the public face. She joined in June 2018 as vice president of applied AI and partnerships, became senior vice president of research, product and partnerships in 2020, and took the CTO title in May 2022, six months before ChatGPT launched. Reporting by the Financial Times and Reuters credits her with overseeing ChatGPT, DALL-E, Codex and Sora.

The job was translation. OpenAI's researchers produced models; Murati's organisation made them into products, set the release decisions and dealt with the partners, Microsoft above all, who paid for the compute. That is why the public rarely saw her. When she did take the stage, it was usually to explain a launch or defend one.

One of those appearances shows the limits of the role. In a March 2024 interview with the Wall Street Journal, asked whether Sora had been trained on YouTube videos, she said: "I'm actually not sure about that." The clip spread widely. For an executive whose authority came from knowing how the machine worked, it was a costly moment, and it is the kind of exposure her later silence avoids.

So what: the person most responsible for turning OpenAI's research into products was, by design, almost invisible, and her later quiet is a continuation of how she always worked, not a retreat.

From Vlorë to Tesla: a hardware education

She was born in Vlorë, Albania, in December 1988, in the last years of the communist regime, to two literature teachers, and left at 16 on a United World Colleges scholarship to Pearson College in Canada, then studied at Colby and took a Dartmouth engineering degree in 2012. She grew up as Albania moved from one of Europe's most closed dictatorships to a market economy. Fortune's account of her early life describes parents who taught literature and a daughter drawn to mathematics and science, competing in school olympiads. In a 2023 conversation with Microsoft's chief technology officer, Kevin Scott, she described a setting where competition for knowledge was fierce and education was everything.

The scholarship was the hinge. Pearson College, on Vancouver Island, brings teenagers from dozens of countries to study the International Baccalaureate; she graduated in 2007. She took a bachelor's degree at Colby College in 2011 and a mechanical engineering degree at Dartmouth in 2012. Albania has kept the connection: in May 2025 its government amended its budget to invest $10 million in her new company.

Her path ran through hardware: Zodiac Aerospace, Tesla's Model X programme from 2013, Leap Motion's hand-tracking from 2016, then OpenAI in June 2018 as vice president of applied AI and partnerships, senior vice president in 2020 and CTO in 2022. Before that came a 2011 summer internship at Goldman Sachs in Tokyo and a year as an advanced concepts engineer at Zodiac Aerospace. At Tesla she was a senior product manager on the Model X. At Leap Motion, a start-up building hand-tracking for virtual and augmented reality, she was vice president of product and engineering.

Exhibit — A Vlorë scholarship and hardware jobs at Tesla and Leap Motion led to the CTO role. Source: Fortune, Wikipedia and Behind the Tech with Kevin Scott, as set out in the Sources below. Compiled by Dan's AI Intel.

That is an unusual route into frontier AI. Most of the field's leaders came from research labs or from start-up investing; Murati came from building physical products against deadlines. The hardware years taught the discipline OpenAI later needed: a car or a sensor ships on a date, and the engineering has to bend to it.

So what: the Albania-to-frontier arc is not a decorative backstory; a scholarship system and six years of shipping hardware produced exactly the operator a research lab lacks.

The weekend she held OpenAI

She was interim CEO for about two and a half days in November 2023 after supplying much of the evidence the board used against Altman, then sided with the staff and investors who brought him back, which showed that at OpenAI the people and the money, not the nonprofit board, held the power. The sequence, from contemporaneous reporting: on Friday 17 November the board fired Altman, saying he had not been "consistently candid" with it, and named Murati interim CEO. Reports during the weekend said she was moving to bring back Altman and Greg Brockman. On Sunday night the board named Emmett Shear, the former Twitch chief, in her place. Microsoft offered Altman and his team new jobs, and most of the staff signed a letter threatening to follow. Late on Tuesday 21 November, Altman's return was agreed with a new board. The crisis ran about five days; her time in charge, about half of it.

Exhibit — Murati led OpenAI for about half of the time Altman was out. Source: Reporting by Axios, The Verge, Decrypt and Gizmodo, as set out in the Sources below. Compiled by Dan's AI Intel.

What happened before that weekend is now on the record. In March 2024 the New York Times reported that Murati had written a private memo to Altman questioning his management and had shared details of her concerns with the board. Murati said she had not been involved in the decision to fire him. In a deposition in Elon Musk's lawsuit against OpenAI, Ilya Sutskever testified that his 52-page memo to the board relied on screenshots, nearly all of which came from Murati, and that he had not independently verified much of it. We covered that deposition and what it means for Altman in our Sam Altman episode, number 28, from July.

In May 2026 the court in that case was shown Murati's own videotaped deposition. Asked whether Altman had told her the truth when he said the legal team had cleared a new model so that it did not need the deployment safety board's review, she answered "No," and she described him as creating chaos among his executives. On 18 May the jury found Musk's claims had been filed too late; it did not rule on whether OpenAI had kept its founding mission.

The weekend revealed two things. About her: she raised concerns through the formal channel, accepted the top job when asked, and then followed the staff rather than the board when the two collided, a pragmatist's choice that kept the company together. About OpenAI: a nonprofit board with formal control could not survive the opposition of its employees and its largest investor. Formal authority lost to the people who built the product and the capital that paid for it.

So what: what she built next answers that weekend directly, with a company where the founder, not an outside board, holds the deciding vote.

Thinking Machines: what it says, and what it has shown

She left OpenAI in September 2024, launched Thinking Machines Lab in February 2025, raised a $2 billion seed at a $12 billion valuation in July 2025, and has since shipped Tinker, the open-weight Inkling models and a preview of real-time interaction models, while losing three co-founders. Her farewell note said she was stepping away "to create the time and space to do my own exploration." The lab launched in San Francisco as a public benefit corporation with about 30 researchers and engineers from OpenAI, Meta and Mistral. The founding team included John Schulman, an OpenAI co-founder who had briefly joined Anthropic, as chief scientist; Barret Zoph as CTO; Lilian Weng; and Luke Metz. The seed round, led by Andreessen Horowitz with Nvidia, AMD, Cisco, Accel, ServiceNow and Jane Street, was one of the largest seed rounds on record. The Information reported that her vote is weighted to give her a majority on board matters.

The stated mission is AI that extends human will and judgment, customisable and collaborative rather than fully autonomous; claims that it is a rival frontier lab, an acquisition target or a coming AGI contender are speculation. In the lab's own words, the aim is "to make AI systems more widely understood, customizable, and generally capable," and "instead of focusing solely on making fully autonomous AI systems," to build systems that work with people. A July 2026 essay put it more sharply: "The mission of Thinking Machines is to build AI that extends human will and judgment."

What it has shipped is narrower than the money. Tinker, launched in October 2025, is a service for fine-tuning open-weight models, and reports say it produces most of the revenue. Inkling, released in July 2026, is an open-weight model of 975 billion parameters, of which 41 billion are active for any query. The lab said plainly that "it is not the strongest model available today, closed or open." Inkling 2 Small followed on 5 November 2026, with weights promised about a week later after outside safety testing. In May 2026 the lab previewed "interaction models" that listen and respond in real time.

The losses are real. Andrew Tulloch, a co-founder, left for Meta in October 2025 after Meta, by multiple reports, approached the lab and offered its staff packages of hundreds of millions of dollars. In January 2026 Murati said the company had "parted ways" with Zoph; OpenAI rehired him the same day with Metz and Sam Schoenholz. TechCrunch reported that the split was not amicable. Soumith Chintala, co-creator of the PyTorch framework, became CTO, and Schulman stayed.

Exhibit — Thinking Machines kept shipping while its co-founders left for Meta and OpenAI. Source: TechCrunch, The Verge, Wired, NVIDIA and Fortune, as set out in the Sources below. Compiled by Dan's AI Intel.

So what: the mission is distinctive and consistent, but the evidence for a frontier lab is uneven, with chips and money ahead of the model that would justify them.

What she believes about AI

Her stated view is that AI should be steered by many people rather than one company, kept collaborative rather than autonomous, regulated early, and made safe through research plus careful real-world testing, while accepting that some jobs will disappear. It has been consistent across four years.

On regulation, in a February 2023 TIME interview she said it was "not too early" for governments and the public to get involved, warning that AI "can be misused, or it can be used by bad actors." On jobs, at a June 2024 event at Dartmouth's engineering school she said: "Some creative jobs maybe will go away, but maybe they shouldn't have been there in the first place," a line that drew a backlash from artists. On safety, the lab's founding statement says "the most effective safety measures come from a combination of proactive research and careful real-world testing," and it promises to share code, datasets and model specifications for outside alignment research.

On power, her lab's July 2026 essay is the clearest statement of her worldview. "A single locus of value alignment, however well run, becomes a locus of power to be captured," it argues, and it pictures alignment as "a feature not of a single model but of an ecosystem of AIs raised in different places." It warns that "power that needs nothing from people loses the incentive to care for their needs and values." On the future of work it is milder than her Dartmouth remark: "The work people do may change, and turn toward more of what only people bring."

The interaction-model research puts that belief into product form. The lab argues that "today's models and interfaces aren't optimized for humans to remain in the loop," and its preview responded in 0.40 seconds on a standard turn-taking test, against 1.18 seconds for OpenAI's real-time model in the lab's own comparison. At Bloomberg's technology conference in June 2026, by reports of the session, she compared the ideal relationship between people and AI to a tandem bike, with both steering.

The real read: she is not a safety maximalist and not an accelerationist. She is betting that the main risk is concentration, of capability and of values, in a few labs, and that the remedy is to give many organisations models they can shape themselves. That is a commercial strategy and a political philosophy at once, which is worth noting; it is also coherent with what she lived through in 2023.

So what: her views explain her products, open weights, fine-tuning and real-time collaboration, more than any press release does.

Why she went quiet, and what ends it

We hear less about her because she gives few interviews, speaks through unsigned company essays, and runs a lab whose products are developer tools, while the January 2026 departures and her trial testimony are what brought her back into the news. Thinking Machines speaks mainly through unsigned blog posts and model cards. Its customers are engineers fine-tuning models, such as the hedge fund Bridgewater Associates, not consumers with a chat app. Silence also has a cost: a lab that announces little gets covered through its setbacks, which is what happened when talks in late 2025 to raise at a $50 billion valuation stalled and the co-founders left.

Yes, it will change within months, because the $40 billion round, the Inkling 2 weights and the first gigawatt of Nvidia chips will each force public news, but she herself is unlikely to become a loud founder. Business Insider reported on 6 November 2026 that a $5 billion round led by Accel, at $40 billion before the new money, could close within days; neither company confirmed it. The Nvidia deal, announced in March 2026, commits at least one gigawatt of Vera Rubin systems for training, with deployment targeted for early 2027. The lab has said larger interaction models will follow later in 2026.

Exhibit — The first model trained on the Vera Rubin chips is the test that matters most. Source: Business Insider via Let's Data Science, NVIDIA and Thinking Machines, as set out in the Sources below. Compiled by Dan's AI Intel.

She matters in the coming months because her lab is the best-funded test of whether a company outside the big three can build open, customisable frontier models at scale. If it works, the frontier stops being a club of OpenAI, Anthropic and Google DeepMind, and open-weight models stop being mainly a Chinese story; Inkling's own architecture drew on China's DeepSeek-V3.

So what: the quiet is a strategy and a temperament, but the next six months will put the lab back on the front page whether she speaks or not.

Bottom line

Murati represents the operator-builder, the person who makes a lab ship, now testing whether that skill is enough to found one without a celebrity's pull. Every other leading lab is run by someone whose public voice is part of the product. She is the experiment in whether execution, capital and a clear mission can substitute for it.

Thinking Machines is today a well-funded question mark with a third pole's inputs, and it becomes a genuine third pole only if its 2027 compute produces a model at the frontier. Its revenue and chips are real; its model, by its own account, trails; its team has lost three founders. The test is not the valuation but the first model trained on Vera Rubin.

Her trajectory matters because it is the clearest bet that frontier AI can be open, customised and spread across many hands rather than concentrated in a few labs. She saw from the inside what happens when one company's governance fails, and she has built her answer into both her cap table and her product. If the bet works, the people building with AI get models they can shape to their own values. If it fails, the frontier narrows to the labs that can already afford it, and the quiet operator becomes the most expensive proof that shipping is not the same as leading. Her lab has also already changed how researchers think about reproducibility, as we covered in our episode on why deterministic AI is a myth, number 58, from September.

Sources

Transcript

Alex: She ran the teams that shipped ChatGPT, and then for two and a half days she ran the whole company.

Sam: And later, on video, under oath, she told a jury her old boss hadn't told her the truth.

Alex: Same woman. Now she's building her own lab, and she's the one holding the deciding vote.

Sam: Of course she is.

Alex: Before we get into her, something from my week. Cybersecurity Dive ran a piece on the sixth about companies struggling to know what normal even looks like for an AI agent.

Sam: Normal for a person, I get. A person logs in at nine and goes to lunch. What's normal for an agent?

Alex: That's the problem. My CISO and I spent a whole meeting on it. Every permission we ever wrote assumed a human was holding it. An agent holds it at three in the morning and nobody blinks.

Sam: So the alarm doesn't go off because nobody told it what a burglar looks like.

Alex: Or what the resident looks like, which is worse. And it matters for today, because Murati's whole pitch is many hands steering many AIs, which means a lot more agents holding keys in a lot more places.

Sam: Hm. That's going straight on my list of things I don't want to explain to my CMO.

Alex: Welcome to Dan's AI Intel, where we try to understand the AI revolution rather than just keep pace with it. I'm Alex.

Sam: I'm Sam. Today it's Mira Murati, the operator who made OpenAI ship, and the lab she built afterwards.

Alex: Here's what we're chasing. Who is she, really? What did that weekend in November 2023 show about who holds power at an AI lab? And is her new company a real third pole in the race, or a very well-funded question?

Sam: I'm going in with a position. I think it's a question mark. Big chips, big money, and a model that, by the lab's own words, isn't the strongest one out there.

Alex: And mine is the opposite, I think she's the most credible builder who has ever tried this without a celebrity founder. I'd bet the compute turns into a model. Put that on the record.

Sam: You'd bet. With whose money?

Alex: Not my board's. Yet. And I'll tell you now that one date would make me drop it. Hold that thought, I'll come back to it.

Sam: Mysterious. We'll see who moves by the end. If you're enjoying this, follow the show in whatever app you're in. It's free, and it's the single biggest help to a small independent show.

Alex: Start with the job, because the job explains everything after. Murati was OpenAI's chief technology officer from 2022 until she left in 2024. She ran the organisation that turned research into ChatGPT, DALL-E, Codex and Sora.

Sam: But she wasn't the face. That was Sam Altman.

Alex: Right. She joined in 2018 as a vice president for applied AI and partnerships, moved up to run research, product and partnerships, and took the CTO title about six months before ChatGPT launched.

Sam: Six months. So she's the person who said yes, it's ready, go.

Alex: That's the job. Researchers make models. Her organisation made them into products, set the release calls, and handled Microsoft, who paid for the compute.

Sam: So she's a translator. From where I sit, that's the person I actually trust in a company. The one who knows both the lab and the launch date.

Alex: Which is exactly my job at my company, on a much smaller scale. Engineers over here, a board that wants a date over there.

Sam: And nobody outside the building can name you.

Alex: My mother can't name me. Anyway.

Sam: Your mother knows exactly who you are, she just can't explain it.

Alex: There's one famous time she stepped out front and it went sideways. March 2024, a Wall Street Journal interview. Asked whether Sora was trained on YouTube videos, she said, quote, "I'm actually not sure about that."

Sam: Oh, that's a rough one.

Alex: It spread everywhere. I've watched a room turn on a CTO over one "I'm not sure", and her authority came from knowing how the machine worked. and one clip made it look like she didn't.

Sam: It's a bad moment, I'll give you that. But I'd push on it. I've stood in front of my CEO and said "I'll have to check" about my own budget. It's not a crime.

Alex: It's not a crime. It's a crime when your whole brand is I know how the thing works.

Sam: Mm. Fine. So the takeaway isn't she's bad on camera. It's that she was built to be invisible.

Alex: Which is why I don't read her silence as a retreat. The people I'd follow anywhere at my company are the ones the executive team has never heard of, and she shipped for six years without a megaphone.

Sam: A scholarship kid from Albania ending up as the CTO of OpenAI. As a brand person, that's a story I'd kill to tell. Where does it actually start?

Alex: Vlorë, in Albania, born in 1988, in the last years of the communist regime. Both parents taught literature.

Sam: A literature household produces the CTO of OpenAI. I love that.

Alex: She was good at maths and science, competed in school olympiads, and at sixteen left on a scholarship to Pearson College on Vancouver Island, a school that brings teenagers from dozens of countries to study together. Then Colby for a bachelor's, and a Dartmouth engineering degree in 2012.

Sam: And then not an AI lab.

Alex: No. Hardware. A year at Zodiac Aerospace, then the Model X programme at Tesla, then Leap Motion, a start-up doing hand-tracking for virtual reality, where she ran product and engineering.

Sam: So she spent years on things that had to ship on a date, or the car doesn't leave the factory.

Alex: Hm, it's not quite about speed. It's… no, I think hardware gets a date and software gets a hope. Anyway, that's the bit I keep coming back to. A sensor or a car doesn't care about your vision. Most of the field's leaders came out of research labs or start-up investing. She came out of building physical things against deadlines.

Sam: And that, if I'm honest, is why I'd hire her. My junior copywriters can write beautiful things all day. The person who gets the campaign out of the door by Thursday is worth double.

Alex: And Albania hasn't forgotten her. Its government amended its budget to put ten million dollars into her new company.

Sam: A country is a seed investor now.

Alex: A small one. But it tells you the scholarship worked out for more than one party.

Sam: Before the facts: if you'd been on that board that Friday, would you have fired him?

Alex: I'd have wanted a week and a lawyer. It gets retold badly, so here's the sequence. Friday 17 November 2023, the board fires Altman, says he wasn't consistently candid with them, and names Murati interim CEO.

Sam: She's asked on a Friday to run the company.

Alex: And she says yes. During the weekend, reports say she's working to bring back Altman and Greg Brockman. Then Sunday night the board names Emmett Shear, the former Twitch chief, in her place.

Sam: So her time in charge is, what, two days?

Alex: About two and a half. Meanwhile Microsoft offers Altman and his team new jobs, most of the staff sign a letter threatening to follow, and late Tuesday Altman's return is agreed with a new board.

Sam: And she goes with the staff.

Alex: She does. To me that is pragmatism. She took the top job when asked and then followed the people who built the product rather than the board that appointed her.

Sam: See, I read it differently. If I were on that board, I'd say she took the title and then walked.

Alex: I'd say the board walked first. They fired a CEO on a Friday with nobody ready to run what came next.

Sam: Hm. That is a fair hit. But she's not just the person who sided with the staff, is she? She's also the person who gave the board the evidence.

Alex: That's the part that's on the record now. In March 2024 the New York Times reported she'd written a private memo to Altman questioning his management and had shared her concerns with the board. She said she wasn't involved in the decision to fire him.

Sam: And the screenshots. If a colleague handed me screenshots about my boss, I'd want a lawyer and a night's sleep before they went anywhere. What happened with hers?

Alex: Elon Musk's suit against OpenAI. Ilya Sutskever testified that his long memo to the board leaned on screenshots, most or all of which came from Murati, and that he hadn't independently verified much of it.

Sam: And she'd shared those with a man who was about to help remove the CEO.

Alex: We went deep on that deposition in our Sam Altman episode, number twenty-eight, back in July, if you want the long version.

Sam: And then May this year, the jury watches her own videotaped deposition.

Alex: Asked whether Altman had told her the truth when he said the legal team had cleared a new model so it didn't need the deployment safety board's review, she answered, "No."

Sam: Just no.

Alex: Just no. She also described him as creating chaos among his executives. On 18 May the jury found Musk had filed too late, and didn't rule on whether OpenAI kept its founding mission.

Sam: So the headline writers got a ruling about a deadline, not about a mission. Nobody gets to put the mission question on a banner.

Alex: Right. But the weekend itself did settle something, and it's the part that keeps me up, because the budget fights are mine. A board with formal control lost to the people who built the product and the investor who paid for it, and I've sat on the builders' side of that room.

Sam: I want to stay on that, because it's the lesson that changes my week. In my department the copywriters and the budget decide what ships, and a title on a slide doesn't change that.

Alex: And she watched it from the middle. So what did she build next? A company where the founder holds the deciding vote, not an outside board.

Sam: Convenient lesson to learn.

Alex: It's a rational one. It's also the one that makes me want to see her governance.

Sam: And here's why it bites for me. My team of sixty answers to a CMO, and the CMO answers to a board I've never met. The org chart says one thing, and the people who can actually say no to a campaign at five on a Friday are somewhere else entirely.

Alex: That's the 2023 lesson in one sentence.

Sam: Don't put it on a mug.

Alex: Too late, I'm ordering mugs.

Sam: Then put on the mug that she's the one person in that story who came out with her options open. Altman got his job back, the board got replaced, and she walked into her own company.

Alex: That's cynical.

Sam: It's a marketing person reading a launch. Somebody wins the weekend, and I'd like to know who.

Alex: She leaves OpenAI in September 2024, and her note says she's stepping away to, quote, "create the time and space to do my own exploration." February 2025 she launches Thinking Machines Lab in San Francisco, a public benefit corporation.

Sam: Who walks out of OpenAI with her? Because that's the first thing I'd check, who's in the room.

Alex: About thirty researchers and engineers from OpenAI, Meta and Mistral. John Schulman as chief scientist, the OpenAI co-founder who'd briefly gone to Anthropic. Barret Zoph as CTO, plus Lilian Weng and Luke Metz.

Sam: That's a strong bench. If a CFO heard that, the next question is the money.

Alex: A two billion dollar seed at a twelve billion valuation, in July 2025, the largest seed round on record. Andreessen Horowitz led, with Nvidia, AMD and Cisco in the round.

Sam: Chipmakers writing cheques to a company that will buy chips. I know that shape. It's a supplier paying its own customer so the customer keeps buying.

Alex: That's a fair read, and it's also why they want her to succeed.

Sam: Okay, but who pays for that, and who decides what they get for it?

Alex: The Information reported her vote is weighted so she has a majority on board matters.

Sam: So the nonprofit-board lesson, built into the cap table.

Alex: The stated mission is AI that extends human will and judgment. Collaborative and customisable, not fully autonomous. In their own words, systems that are, quote, "more widely understood, customizable, and generally capable."

Sam: That sounds like every mission statement I have ever approved.

Alex: Except the products back it up. That's the point. What they've shipped is narrower than the money. Tinker, October 2025, a service for fine-tuning open-weight models. Reports say it produces most of the revenue.

Sam: Fine-tuning. Meaning?

Alex: You take a model someone already built and teach it your own material, so it behaves the way your business needs.

Sam: And this is where my brain goes to the thing I always mix up. Is it Tinkle, Tinker...

Alex: Tinker.

Sam: Tinker. And the model is Inkblot.

Alex: Inkling.

Sam: Inkling. See, that's why I have a marketing team.

Alex: Inkling came out in July 2026. It's open-weight, nearly a thousand billion parameters, though only a small slice of it is active for any query. It's built so that only the part of the network a question needs wakes up.

Sam: I'm going to stop you before you fall in love.

Alex: Too late, I love it. It's elegant. A giant model that behaves like a small one on each question.

Sam: You said the same about the last three things.

Alex: And I was right about two. Here's the part I respect, though. The lab said plainly, quote, "it is not the strongest model available today, closed or open."

Sam: A vendor telling me the thing I'd buy isn't the best one. In my last renewal the sales rep called a tool "best in class" four times in one email. That's rare.

Alex: And it's consistent. A second model, Inkling 2 Small, followed on 5 November, with weights promised about a week after outside safety testing.

Sam: Last quarter I cut a copy tool nobody had opened in two months. So why would I let anyone near this?

Alex: I had this exact conversation about my data team last month. Tinker only does LoRA. Think of clipping a small adjustable layer onto a finished model instead of rebuilding it from scratch. An MIT team used it for a method where a model teaches itself, and on tool use it clearly beat standard fine-tuning.

Sam: So it works.

Alex: For that job. And the guide for Inkling says something I liked: if your goal is adding private facts, try retrieval first, because fine-tuning is a bad database.

Sam: Which is the opposite of what a seller says. The point, then, is it's a tool for shaping behaviour, not storing facts.

Alex: And that's why I'd run a small pilot. A tool that tells me when not to use it earns a pilot.

Sam: Mine would get a spreadsheet and a no. Every quarter I defend a line item to my CFO, and "the maker says the model trails" is not a sentence that survives that meeting.

Alex: It survives mine, because I can write the exit in advance. One engineer, a few weeks, a date to stop.

Sam: And what do you tell your board the pilot proves?

Alex: That we know how to shape a model to our catalogue and our returns language. That skill is worth something even if I never use her model.

Sam: Hm. That's the first argument for it I'd put in a slide.

Alex: Until the model catches up, your no is your right. And it's the thing I keep tripping on in my plan: I'd be buying a tool whose maker admits the model trails.

Sam: There it is.

Alex: Not yet, but I've seen the recruiter emails. Andrew Tulloch, a co-founder, left for Meta in October 2025, after Meta, by multiple reports, approached the lab and offered its staff packages of hundreds of millions of dollars.

Sam: Hundreds of millions. For an engineer.

Alex: I want to say that out loud to my team on a Monday. They'd stop listening halfway through.

Sam: What do you do about it?

Alex: Honestly, nothing, that's the lesson. Advice for people like me is start retention conversations early and plan for long hiring. I can't match Meta. I can match a fair bit of freedom and a clear mission.

Sam: And that's precisely what Murati is selling. A mission, and a say in how it's built.

Alex: It is, and it still wasn't enough.

Sam: Hold on. Before you tell me it wasn't enough, remember who made the offers. When someone who has a bottomless budget calls your engineers, the mission is up against a number that changes a family's life.

Alex: Fair. It would be unfair to treat a co-founder leaving for that as a verdict on her.

Sam: And yet he is not the only one. Both things are true.

Alex: January 2026, she said the company had, quote, "parted ways" with Zoph. OpenAI rehired him the same day, together with Metz and Sam Schoenholz.

Sam: The same day. That's not a resignation, that's a transfer window.

Alex: Wired's sources said the lab had raised concerns about Zoph's conduct, which he didn't address publicly. Soumith Chintala, co-creator of PyTorch, became CTO. Schulman stayed.

Sam: So that's three co-founders gone.

Alex: Three in all. And that's where I have to give you a point, Sam. A lab whose founders walk to the competitor on the same day is not sending the signal I want.

Sam: Wow. You said it. Write it down.

Alex: I'm writing it down. It doesn't kill the bet, but it moves my exit criteria. If the next one to leave is a researcher, not a manager, my pilot ends early.

Sam: Fine, she's an operator who ships. But "quiet operator" can mean anything, so what does she actually believe?

Alex: Consistent for four years. AI should be steered by many people, not one company, collaborative rather than autonomous, regulated early, and made safe by research plus real-world testing. And she accepts some jobs will disappear.

Sam: Regulated early. That one surprised me from an OpenAI executive.

Alex: Why surprised?

Sam: Because I assumed the people shipping it wanted fewer rules. But she told TIME in February 2023 that it was, quote, "not too early" for governments and the public to get involved, and she warned AI can be misused by bad actors.

Alex: And the jobs line?

Sam: June 2024, Dartmouth. "Some creative jobs maybe will go away, but maybe they shouldn't have been there in the first place."

Alex: Ouch.

Sam: That went down well in every illustrator's group chat. I have copywriters who'd have pinned it to the wall as a warning.

Alex: It really did. The lab's July 2026 essay is milder: the work people do may change and turn toward what only people bring.

Sam: Milder isn't the same as different. Which of those do I believe?

Alex: Read the essay on power instead. "A single locus of value alignment, however well run, becomes a locus of power to be captured." And it pictures alignment living in an ecosystem of AIs raised in different places.

Sam: So the risk she cares about isn't a robot waking up. It's one company deciding what everybody's AI values, and as someone who signs off on three vendors' tools, I can see that one from my desk.

Alex: And that's what she watched fail from inside. One board, one weekend. When my board overrules my engineers I get a budget fight; when hers did, the company nearly vanished. I think that's the thread through everything for her.

Sam: But it's also a business plan. Many hands means many customers for Tinker.

Alex: Both are true, and for me it's the useful kind of true. I'm the CTO who'd rather shape my own model of my catalogue than rent one that two other companies are also renting. So a lab that sells exactly that gets a hearing from me, whatever her philosophy.

Sam: I find it convenient. But convenient and wrong aren't the same thing.

Alex: Let me do the product version, because this is where I get excited. In May 2026 they previewed what they call interaction models. They listen and respond while you're still talking.

Sam: Like a person who doesn't wait for me to finish a sentence.

Alex: Their argument is that today's interfaces aren't built for humans to stay in the loop. On a standard turn-taking test, their preview answered in four-tenths of a second, against about a second and a bit for OpenAI's real-time model. That's the lab's own comparison, to be clear.

Sam: Hold on, because this lands for me. I called my insurer's phone line last week, and the bot starts talking before I've finished, then stops, then waits, then says it didn't catch that.

Alex: That awkward overlap is exactly the thing they're trying to fix.

Sam: And I did not feel like a human in the loop. I felt like a customer being managed by a delay.

Alex: So the fix isn't a smarter bot, it's a bot that knows when to shut up.

Sam: I would give a lot for that in my own meetings.

Alex: At Bloomberg's tech conference last June, by reports of the session, she compared the ideal relationship between people and AI to a tandem bike, with both steering.

Sam: Cute. But on a tandem, the one in front sees the road.

Alex: Okay, that's the best objection I've heard all year. You said the one in front sees the road, and that's exactly where the quiet comes in. She's the one in front, and she hardly talks.

Sam: Which I find suspicious in a founder. Why so little from her?

Alex: She gives few interviews, speaks mainly through unsigned company essays and model cards, and sells developer tools. Customers are engineers fine-tuning models, like the hedge fund Bridgewater Associates, not people with a chat app.

Sam: And I have a bone to pick with that as a brand person. A company that says nothing doesn't get a neutral story. It gets covered through its setbacks.

Alex: Which is exactly what happened. The fifty billion dollar round they sought in late 2025 stalled, the co-founders left, and the headlines wrote themselves.

Sam: I run a team that learned this the hard way. If we don't tell our story, a Reddit thread tells it for us. Unsigned posts are fine for a researcher. They're a risk for a brand.

Alex: But she's choosing it. It's a temperament, not an accident.

Sam: Maybe it's both. And a company can have a temperament and a communications problem. And this is my week, actually. Digiday reported on Monday that ChatGPT is about to show brands their mention count.

Alex: Is it live?

Sam: Not yet, "about to" is the key phrase. But think about what it means. For years my team's whole job was being found in a search box, and now the box is somebody else's answer, and they're about to give me a counter on how often I appear.

Alex: A counter for a thing you can't control.

Sam: A counter for a thing I can't control, from the company that also sells the ads. That's the quiet-company problem from the other side. When you control the channel, you also control the story about the channel.

Alex: Which is exactly what she says she wants to avoid, and as the guy renewing a hundred and forty SaaS contracts, I'd say it to every vendor I have: nobody gets to be the only door.

Sam: Right, and I notice nobody asked me whether I'm comfortable renting my visibility from a landlord who also sells the billboards.

Alex: You'd rent it anyway.

Sam: I'd rent it anyway, grumbling, plus a line item and a very long email to finance.

Alex: That's the most honest budget forecast I've heard on this show.

Sam: It's the only kind I've ever delivered on time.

Alex: Mm. Okay. Here's what ends it, though, whether she wants to talk or not. A round reported at forty billion before the new money. Business Insider said on 6 November that five billion dollars led by Accel could close within days. Neither side has confirmed it.

Sam: Reported, not confirmed. I can't put a valuation on a line item. My CFO funds shipped products, not headlines.

Alex: Then the Inkling 2 weights, then the chips. The Nvidia deal from March matters to my kind of buyer. At least one gigawatt of Vera Rubin systems for training, deployment targeted for early 2027.

Sam: A gigawatt is a power plant.

Alex: That's a small city's worth of electricity pointed at one lab's training run.

Sam: And that is the number that makes me nervous as a customer. The money is real, the chips are on order, but a chip is a promise until a model runs on it.

Alex: Then let's say what it's worth if it does run.

Sam: Fine. If it works, the frontier stops being a club of three labs, and my last price hike stops being everyone's price hike. That's the version where my brand gets a tool of its own instead of a seat at someone else's table.

Alex: Inkling's own architecture drew on DeepSeek, so that's already happening in a small way.

Sam: Which should bother anyone in my seat who assumes the tools we rent come from one neighbourhood.

Alex: So, positions. I walked in saying she's the most credible builder to try this without a celebrity founder. I still think so. But Sam, you moved me on one thing.

Sam: Go on, I want that on the record.

Alex: Three founders leaving, and the lab's own admission that the model trails. I came in thinking the compute fixes it. I now think the whole bet is one event: the first model trained on those new chips.

Sam: I'll give you something back. I came in calling it a question mark, and I still do, but I hadn't thought about the small shops. My friend Lauren runs a six-chair salon and can't train anything. If open, customisable models get good enough, they're the ones who can finally shape the tool instead of renting someone else's.

Alex: So we agree it matters, and we disagree on the spending.

Sam: Exactly that. I'd spend nothing until the model ships. Not a pilot, not a lunch.

Alex: And I'd spend a small pilot now, one engineer, a few weeks, a clear exit. Remember the date I held back? It's the weights for the second model, promised about a week after safety testing. If that week slips, I drop the pilot and you can say so. If it's a bad pilot, I lose a month. If the model arrives and I'd never looked, I lose a year.

Sam: And if it fails, you tell your board "she warned me."

Alex: I'll tell my board "Sam warned me." They already assume it.

Sam: The thing I'm watching is the first model trained on Vera Rubin. That's a year out, and I'm not paying for a year of hope.

Alex: Mine is closer, and that's the point. The second model's weights are due this month, so I find out in weeks whether the people who left mattered. You wait a year. That gap is the whole bet.

Sam: And my reason to wait is a CFO, and yours to go is a board. Fine. I can live with that split.

Alex: Let's recap what we covered. Murati ran the teams that shipped ChatGPT, DALL-E, Codex and Sora while Altman was the face. She held OpenAI for about two and a half days in 2023, and what that weekend showed is that the people who build and the money that pays them beat a board.

Sam: Then she built a lab where she holds the deciding vote, and a gigawatt of chips on order, serious revenue from Tinker, a model that by her own admission trails, and three co-founders gone.

Alex: For me it's a purchasing question. Her lab is the best-funded test of whether anyone outside the big three can sell me an open, customisable model at scale. If it works, my team shapes models to our own catalogue. If it fails, I'm renting from the labs that can already afford it, on their terms.

Sam: And for me it's the brand lesson. Shipping isn't leading, and a founder who stays quiet gets her story told by her setbacks, which is the thing I train my team to avoid. The next six months will put her back on the front page, and I promise to get the model's name right. Probably.

Alex: Her lab has also changed how researchers think about reproducibility, which we covered in our episode on why deterministic AI is a myth, number fifty-eight, from September.

Sam: One line before we go. This show is AI-generated and fact-checked. Dan builds a custom stack of AI tools to research, analyse, verify and illustrate the questions he can't stop thinking about, mostly to keep up himself, and the show is deeply researched but worth a second look.

Alex: And please follow Dan's AI Intel in Spotify, Apple Podcasts, or whatever you use. It's free, it takes a tap, and it's the single biggest thing that helps a small independent show reach the next curious person.

Sam: And one ask. If there's something we got wrong, something you'd push back on or want us to go deeper on, write to podcast at connectiveshift dot com. Someone reads every one.

Alex: Thanks for spending the time with us. We'll see you next time.

Sam: And Alex, it's Tinker and Inkling.

Alex: Look at you.

Sam: Inkblot. I know. Goodnight.