Mira Murati

Murati was born in Vlorë, Albania, in 1988.^[1],[2] At sixteen she left for Pearson College UWC on Vancouver Island on a UWC scholarship.^[2]She took a mathematics degree at Colby College in 2011 and a mechanical engineering degree from Dartmouth’s Thayer School in 2012.^[1],[2] Wikipedia lists a short internship at Zodiac Aerospace along the way.^[1]

Then came hardware that had to work in the world. She joined Tesla in 2013 as a product manager and became a senior product manager on the Model X.^[1],[2]In 2016 she moved to Leap Motion, a hand-tracking company, as VP of Product & Engineering, and stayed about two years.^[1],[2]

The thesis of this file is an editorial reading, not her own claim. A career spent where sensors, latency and feedback loops decide whether a product feels right is a different apprenticeship from a research career, and it shows in what her lab chose to build. Hold that thought until §06.

She joined OpenAI in June 2018 as VP of Applied AI & Partnerships and became CTO in May 2022.^[1],[2] Under that title the company shipped ChatGPT in November 2022, followed by the products Wikipedia lists among her responsibilities: DALL·E, Codex, GPT-4, Sora and GPT-4o.^[1]

“Iterative deployment” is sometimes credited to her personally. That overstates it. Releasing models in stages and learning from use was OpenAI-wide doctrine; her part was running the deployment side as CTO. Her clearest written statement of the idea is a 2022 Dædalus essay, Language & Coding Creativity. It argues that Codex and DALL·E widen access for non-programmers and small businesses rather than replace skilled professionals, lists the limits (long-form coherence, commonsense errors, bias, misuse), and says models must be made more truthful and harmless, with human feedback as the cited method.^[3]

In June 2024, in a conversation with her alma mater’s engineering school, she said some creative jobs “maybe will go away” and added they “maybe shouldn’t have been there in the first place.”^[4]The remark spread without its context: the discussion was about AI and the entertainment industry, and Fortune quotes her describing the work at most risk as “strictly repetitive.”^[4] Writers and artists criticized the comment, according to Wikipedia’s account.^[1]

On November 17, 2023, OpenAI named Murati interim CEO after the board removed Sam Altman; the arrangement lasted about five days.^[1],[25] A later deposition put her at the center of the story. In his own testimony, taken in October 2025, Ilya Sutskever said nearly all of the material in his 52-page memo about Altman came from Murati, and that he fully believed what she told him.^[8]

Murati’s own video deposition was played in the Musk v. Altman trial in Oakland on May 6, 2026. What follows is her testimony as reported, in a contested case, and this page does not adjudicate it. According to Reuters, she testified that her concern was Altman “saying one thing to one person and completely the opposite to another,” that he pitted executives against each other, and that she had wanted him to stay as CEO.^[5]Futurism reports she testified that Altman told her the company’s lawyers had cleared a new model to skip internal safety-board review.^[6] Gizmodo reports she said she checked and that what legal was saying and what Altman was saying were not the same.^[7]Asked whether, by fall 2023, she saw Altman as not always candid, truthful and honest, she paused and answered, “Not always.”^[6]

She returned to her CTO role and left in September 2024; CNN and TechCrunch reported that Bob McGrew and Barret Zoph left the same day.^[1],[24]

Thinking Machines Lab launched in February 2025 as a public benefit corporation.^[1],[2] In July 2025 it closed a $2B seed at a $12B valuation, led by Andreessen Horowitz, with Nvidia and AMD among the named backers.^[1],[10]Albania’s parliamentary finance committee approved a proposal in November 2025 to allocate 1 billion lek, roughly €10 million, to the company; the report describes it as funding, not specified equity, and still subject to full parliamentary approval. I could not confirm it completed.^[12]

The governance terms are the unusual part, and they come from reporting rather than a filing. Fortune, citing The Information, wrote that the deal would give Murati a board vote equal to all other directors’ votes combined, plus one, and that the founding team would hold supervoting shares carrying 100 times the weight of standard shares. The company declined to comment.^[9] That is not the same as saying she personally holds the supervoting stock, and no source I read says so.

The valuation story since is talks, not a closed round. Bloomberg reported early talks at around $50B in November 2025.^[11] A secondary outlet says those collapsed and that talks at roughly $40B were reported in 2026.^[10] I found no report of a closed Series B. On talent, Fortune reported in January 2026 that Barret Zoph, Luke Metz and Sam Schoenholz returned to OpenAI and that Soumith Chintala became CTO.^[23]

On June 4, 2026, Murati sat down with Emily Chang at Bloomberg Tech, her first major appearance in about 18 months. Coverage centers on two things. TechCrunch characterizes her concern as the concentration of consequential decisions in too few hands, and reports that, looking back, she would have pushed harder for more information and a real transition plan.^[13]The Next Web reports the same governance theme and one line on rivalry: “When I wake up in the morning, I am not thinking about how to kill the competitor.”^[14]

The technical argument comes from her lab’s own writing. Thinking Machines’ interaction-models post says most AI systems bolt interactivity on with a harness, stitching components together to emulate interruptions, and that real-time speech products use voice-activity detection to find turn boundaries.^[18] It cites Sutton’s Bitter Lesson and predicts these hand-crafted systems will be outpaced by general capability.^[18] The pipeline most people mean by the harness is voice-activity detection, then speech recognition, then a language model, then speech synthesis. The post describes the first of those explicitly; the full four-stage chain is my shorthand, not its wording.

On May 11, 2026, the lab described TML-Interaction-Small: a 276B-parameter mixture-of-experts model with 12B active. It works in 200 ms micro-turns, takes audio and images through lightweight components co-trained from scratch (the post calls this encoder-free early fusion), and splits the work between an interaction model that stays in the conversation and an asynchronous background model that does the slower reasoning with shared context.^[18] Serving uses streaming sessions, with a version contributed upstream to SGLang.^[18] The post describes a limited research preview in the coming months and a wider release later in the year; it names no partners.^[18]

Read the table for what it is: the company’s own numbers, with its model leading on latency, turn-taking and tool-use quality, and trailing GPT-realtime-2.0 at both settings on the QIVD video benchmark. The starred scores used the background agent, so they measure the pair of models, not the interaction model alone.^[18]

Tinker, announced October 1, 2025, is an API for fine-tuning open-weight models, including large mixture-of-experts models such as Qwen-235B-A22B. It uses LoRA so many training runs can share one compute pool.^[15] I have left out any per-token price table: the pricing page I could reach gives only storage rates, and the rates by model sit behind a page I could not open.

Inkling followed on July 15, 2026: 975B total parameters and 41B active, with 256 routed experts plus 2 shared and 6 routed experts active per token, trained on about 45 trillion text, image, audio and video tokens, with up to 1M context in the open weights.^[19] The Hugging Face card lists a 66-layer decoder-only sparse MoE under Apache 2.0.^[20]The lab’s own launch post says Inkling is “not the strongest overall model available today, open or closed.”^[19] Artificial Analysis scored it 41 on its Intelligence Index and measured a 63% hallucination rate on its AA-Omniscience test.^[21] A smaller Inkling-Small preview, 276B total and 12B active, followed.^[19]

On November 5, 2026 the lab announced Inkling 2 Small, 594B total and 17B active. It launched with closed weights through Tinker; the lab says Apache 2.0 weights will follow after external safety testing.^[22] The page gives an Artificial Analysis score of 31 in its text and 30.6 in its table, and I have not reconciled which index version that refers to against the 41 reported for the first Inkling. I could not confirm that the weights are live. Treat the pattern, not the point values, as the finding.

The September 2025 post by John Schulman and colleagues is an empirical study. In controlled supervised and reinforcement-learning experiments, it found that low-rank adaptation can match full fine-tuning under the right conditions.^[16] The setup, reconstructed:

W′ = W₀ + ΔW,   ΔW = B·A
W₀ ∈ ℝ^{N×N}  (frozen)
B  ∈ ℝ^{N×r},  A ∈ ℝ^{r×N},  r ≪ N

full fine-tuning:  N² (forward) + 2N² (backward) = 3N² multiply-adds
LoRA:              2N² + 6Nr   ⇒  slightly more than ⅔ of full fine-tuning

The post also scales the update by a constant (α/r in its experiments).^[16]LoRA skips the full weight-gradient step, which is where the saving comes from. The post’s conditions for matching full fine-tuning:

• Apply LoRA to all layers, especially the MLP and MoE layers that hold most parameters; attention-only underperforms.
• Use a learning rate about 10x the full fine-tuning rate for the same task.
• Keep the dataset within LoRA’s capacity; when the information to learn exceeds the trainable parameters, LoRA falls behind.

The reinforcement-learning result is the sharp one. The post argues policy gradients carry about one bit per episode, so even rank-1 LoRA has capacity to spare and matches full fine-tuning there.^[16]

Kevin Lu and colleagues, October 2025: reinforcement learning teaches O(1) bits per episode, while distillation on the student’s own samples teaches O(N) bits, where N is the number of tokens.^[17] The worked example is Qwen3-8B-Base on AIME’24. Supervised fine-tuning on 400k prompts reaches 60%; on-policy distillation reaches 70% in about 150 steps.^[17]

reinforcement learning:   O(1) bits per episode
on-policy distillation:   O(N) bits per episode,  N = tokens

On cost, the post reports reductions of about 9x on the baseline, about 18x in GPU hours and about 30x in total against an extrapolated 2M-example supervised baseline.^[17]Its comparison with reinforcement learning uses Qwen’s published figures, 17,920 GPU hours against 1,800, not the lab’s own runs.^[17]

This section is the Context Jamming reading, not a claim by Murati. Her lab’s products point two ways. The interaction model is built to keep a person inside the loop in real time: it listens while it talks and hands slow work to a background model. Tinker and the open Inkling weights push the other way, toward systems other people fine-tune, deploy and let run.

The through-line is the worry she voiced in June: consequential decisions concentrated in too few hands.^[13] A human-in-the-loop design spreads that decision-making across many users; an autonomous-agent future concentrates it in whoever controls the agent. What remains unresolved is that her own governance reportedly gives her an outsized vote.^[9]That tension is the file’s open question.

Education. Pearson College UWC (UWC scholarship). Colby College (B.A., mathematics, 2011). Dartmouth Thayer School of Engineering (B.Eng., mechanical, 2012); honorary Doctor of Science from Dartmouth, 2024.^[1],[2]

Affiliations.Tesla (Model X, 2013–16). Leap Motion (VP Product & Engineering, 2016–18). OpenAI (June 2018 to September 2024; CTO from May 2022; interim CEO for about five days in November 2023). Thinking Machines Lab (founder and CEO, since February 2025).^[1],[2]

Collaborators / peers worth naming. John Schulman (lead author, LoRA Without Regret). Kevin Lu (lead author, On-Policy Distillation). Andrew Tulloch (named by Fortune as a co-founder). Soumith Chintala (reported CTO, 2026). Bob McGrew and Barret Zoph (left OpenAI in September 2024). Ilya Sutskever and Sam Altman (the parties to the testimony above).^{[2],[8],[16],[17],[23],[24]}

Status of the record. A living subject, and some of the strongest claims about her come from testimony in ongoing litigation. Everything contested here is attributed to the person or outlet making it.