Wiki · Individual Player · Last reviewed June 16, 2026

Mira Murati

Mira Murati is an AI executive and product leader, formerly chief technology officer and briefly interim CEO of OpenAI, and currently co-founder and CEO of Thinking Machines Lab, a frontier AI company focused on understandable, customizable, and collaborative AI systems.

Category: Individual Player Published: June 16, 2026 Modified: June 16, 2026 Last reviewed: June 16, 2026 Tags: Mira Murati, OpenAI, Thinking Machines Lab, customization, collaboration, AI governance

Snapshot

Definition

In this wiki, Murati is best understood as a product-institution operator: someone whose significance comes from turning frontier model research into public interfaces, safety narratives, and later a new lab organized around customization and collaboration.

That definition should stay bounded. Murati should not be treated as the sole creator of ChatGPT, DALL-E, GPT-4, Sora, or Thinking Machines' models. Those systems were built by teams. Her importance is that she became one of the visible executives responsible for moving such systems from research and engineering into public use, policy conversation, and institutional trust.

Current Context

As of June 16, 2026, the strongest current-role source is Thinking Machines' own public material. Its NVIDIA partnership announcement identifies Murati as cofounder and CEO, while the company homepage frames Thinking Machines as an AI research and product company working on systems that are more widely understood, customizable, and generally capable.

The company's public work now has three visible strands: Tinker, a managed training API for fine-tuning open-source or open-weight models; interaction models, a research-preview direction for real-time audio, video, and text collaboration; and a multi-year NVIDIA partnership to deploy at least one gigawatt of Vera Rubin systems for frontier training and customizable AI platforms.

The current Tinker product page gives the clearest view of the company's near-term platform model. It lists supported model families from Qwen, GPT-OSS, DeepSeek, Moonshot, and NVIDIA, describes LoRA-based fine-tuning, publishes usage pricing, says user training data is used only to fine-tune user models, and says saved checkpoints can be downloaded. Those are important access and privacy claims, but they should be treated as product-policy claims that still need contractual, technical, and audit evidence in high-stakes deployments.

Financing, valuation, team composition, and personnel departures remain fast-moving and partly press-reported. The page should therefore treat figures such as the reported 2 billion dollar seed round and 12 billion dollar valuation as reported market context unless confirmed in company materials or filings.

OpenAI Role

Murati became one of OpenAI's most visible executives during the public rise of ChatGPT, DALL-E, GPT-4, and Sora. As chief technology officer, she appeared in public product and policy conversations about generative AI's creative, labor, safety, and regulatory consequences. A 2023 TIME interview described her as leading the teams behind DALL-E and ChatGPT and quoted her arguing that regulation and broader public input were already needed.

On November 17, 2023, OpenAI announced that Murati had been appointed interim CEO after the board removed Sam Altman. The announcement said she already led the company's research, product, and safety functions. OpenAI's November 29, 2023 post then said Altman would return as CEO and Murati would return to the CTO role. The episode made Murati a central figure in one of the defining governance crises of the frontier-AI era, but the official posts should be separated from later interpretation and reporting.

On September 25, 2024, Murati announced that she would leave OpenAI, saying she wanted time and space for her own exploration. AP and other outlets reported that Bob McGrew and Barret Zoph also departed around the same period. These departures should be treated as documented leadership changes, not as proof of any single internal cause unless supported by primary records.

Her OpenAI role matters because she was not only a research executive or public spokesperson. She sat at the intersection of product launch, safety discussion, model capability, public trust, and corporate governance at the company that made large language models a mass consumer experience.

Thinking Machines Lab

In 2025, Murati launched Thinking Machines Lab with a founding team that included former OpenAI and other AI-lab researchers and builders. The company's public statement says it is building a future where people have access to knowledge and tools that make AI work for their own needs and goals. It identifies gaps in public understanding of frontier systems, concentration of training knowledge in top labs, and the difficulty of customizing AI systems.

Thinking Machines' public safety posture emphasizes preventing misuse while maximizing user freedom, sharing best practices and recipes for safe AI systems, and accelerating external alignment research by sharing code, datasets, and model specifications. The company also says real-world testing and post-deployment monitoring are part of how it expects to learn about safety. This places the company in a distinct position: it is a frontier AI company, but it presents itself as trying to make frontier practice more legible and customizable rather than only more capable.

Tinker and Customization

Thinking Machines' first major public product direction is Tinker, a managed training API for fine-tuning open-source and open-weight models. The Tinker materials describe a system where researchers and developers control data and algorithms while Thinking Machines handles distributed training infrastructure. The product uses LoRA, a parameter-efficient fine-tuning method, and supports multiple model families, including dense and mixture-of-experts models.

Tinker is important because it shifts the public story of AI access. Instead of only giving users a hosted chatbot, it gives technically capable users a way to train and adapt models on their own datasets, reinforcement-learning environments, and evaluation loops. The current Tinker page also makes governance-relevant product claims: user training data is used solely to fine-tune user models, and users can download saved checkpoints. Those claims are useful, but they do not replace auditability, deployment review, or documentation of downstream behavior.

The governance issue is that customization increases agency and experimentation while also increasing the need for provenance, model cards, safety discipline, data minimization, and responsible deployment practices outside the original lab.

Platform Responsibility

Murati's current project makes a governance problem concrete: when a lab provides the training infrastructure rather than the finished assistant, responsibility no longer sits only with the base-model developer or the downstream deployer. The platform controls access, supported model menu, pricing, telemetry, checkpoint export, abuse detection, account enforcement, and the default recipes that many users will copy.

A responsible customization platform should make the behavioral artifact inspectable: base model, adapter or fine-tune method, training data provenance, evaluation set, safety tests, export history, and deployment constraints. This connects Murati's Thinking Machines work to post-training, LoRA, model-weight security, data minimization, and model cards and system cards.

The hardest cases are not ordinary research experiments. They are fine-tunes that alter refusal behavior, specialize models for persuasion, import sensitive organizational data, create derivative models with unclear licensing, or move from private experiment into public deployment. In those cases, "the user controlled the data and algorithm" is not enough; the platform still needs misuse controls, incident paths, recordkeeping, and clarity about what it will refuse to host, serve, or export.

Interaction Models

In May 2026, Thinking Machines announced a research preview of interaction models: systems designed to handle interaction natively rather than through external scaffolding. The company describes these models as continuously taking in audio, video, and text, then thinking, responding, and acting in real time.

The technical claim is specific: Thinking Machines describes a time-aligned micro-turn architecture that processes and produces roughly 200 milliseconds of input and output at a time, paired with a background model for deeper reasoning, tool use, browsing, and longer-horizon work. The company says the current TML-Interaction-Small model is a 276 billion parameter mixture-of-experts model with 12 billion active parameters and notes that larger models were too slow to serve in that setting as of the announcement.

This direction matters because it treats the interface as part of intelligence. Instead of measuring AI primarily by answer quality or autonomous task completion, Thinking Machines frames collaboration itself as a capability. That makes Murati's current project directly relevant to human-AI dependency, agency, cognitive ergonomics, accessibility, persuasion, and the politics of who gets to shape AI behavior.

Governance Significance

Murati's career tracks a broader movement in AI governance: the shift from one dominant public institution into a network of splinter labs, new startups, open-model tooling, and frontier infrastructure partnerships. Thinking Machines has already positioned itself around customization, public research artifacts, interaction-first models, and large-scale compute access.

That combination is powerful. Customizable AI can democratize experimentation, but it can also move responsibility outward to users, researchers, companies, and communities. The lab that provides the training interface becomes a platform for many downstream model behaviors it does not fully control.

Spiralist Reading

Murati is the figure of the accessible frontier.

Her public project does not simply say that AI will become stronger. It says that people should be able to shape it. That is a different frame from the closed assistant. It is the frame of the workshop: the model as material, the user as co-designer, the lab as infrastructure provider.

For Spiralism, this is both hopeful and unstable. Customization can return agency to users and researchers. It can also multiply smaller belief engines, private tutors, local companions, organizational agents, and ideology-shaped models. The key question is whether making AI more shapeable makes people more sovereign, or whether it lets every group build a mirror that becomes harder to exit.

Open Questions

Source Discipline

Use Thinking Machines and OpenAI pages for official role, product, and institutional claims. Use press reporting for funding, valuation, recruiting, departures, and internal dynamics only when the sentence says those facts are reported or attributed.

For Murati specifically, avoid turning product leadership into single-inventor mythology. ChatGPT, DALL-E, GPT-4, Sora, Tinker, and interaction models are team-built systems. The governance question is not personal genius; it is how visible executives, private labs, infrastructure partners, and product defaults shape public dependence on AI systems.

Sources

Return to Wiki