Mira Murati Leaves OpenAI to Build Thinking Machines

I have never met Mira Murati, and I will not pretend otherwise. But I have met the job she held. The least glamorous post in any ambitious house is the one where somebody must take a fine idea out of a laboratory and turn it into a thing a tired man will use on a Tuesday. As chief technology officer at OpenAI she oversaw the release of products that turned a research project into a mass-market phenomenon.

Her decision to leave and start a company of her own was received as news, which I found generous. In a trade where senior people are the scarcest input, the top of every major laboratory has become a training ground for founders. A resignation there is less an earthquake than a graduation.

A research agenda wearing a product’s coat

The stated ambition is to build systems that are more understandable and more collaborative — models whose reasoning a person can follow and push back on. I like the sound of it, having noticed that reasoning which survives the pushing is the only kind worth having. That is a research programme dressed as a product bet, and it puts the company in the same country as the interpretability work inside the largest laboratories.

The differentiator she has emphasised publicly is not a chip or a dataset but the shape of the team: fewer layers, more researchers close to the product, and a deliberate effort to avoid the organisational weight that slows large laboratories down. I have heard that argument before, on riverboats and in newsrooms both, and it is true more often than the organisations admit.

Why this departure carries more weight than most

• She is among the most senior women in a field whose leadership remains overwhelmingly male.
• She is one of the few leaders who has actually shipped frontier models to hundreds of millions of users.
• Her move shows the frontier is still porous — talent can leave an incumbent and raise money the same week.

The bill comes due in silicon

The new company enters a market where the binding constraint is compute, which is a polite word for the fact that every plan in this business ends at the same locked door. Without a distribution platform or an existing cloud relationship, the firm will pay market rates for its processors and hire against companies holding nine-figure retention packages. I have seen a man open a shop across from a concern with deeper pockets. It can be done, but he must be better at something.

That is the standard challenge for any new frontier laboratory, and the reason most of them end up specialising — the general store cannot out-stock the company store, so it learns to sell the one thing the company store does badly. Murati’s bet is that a smaller team, better aligned, can find a niche the giants cannot serve.

Hiring from a very small pond

The market she enters has a peculiar scarcity at its top. The number of people who can lead a frontier training run is small, and every one of them already has a job at a well-funded organisation. Recruiting is not a search. It is a negotiation against companies offering compute, prestige and compensation a new entrant cannot match.

Her advantage is her reputation for making research usable, which is precisely the skill the large laboratories have found hardest to hire. A researcher who wants their work to reach actual users has fewer places to go than one who only wants to publish. That is a narrow door, but she is standing in it.

The interpretability thread, which nobody has finished

The emphasis on systems whose reasoning can be followed puts the company on ground the largest laboratories have walked over without surveying it. Understanding a model’s internal computation is a research problem with no known solution at frontier scale, which makes it a strange thing to hang a business on. It is rather like opening a shop whose entire inventory is a question.

It is also the sort of bet that looks obvious the moment it works, and foolish in every retelling before that. The history of this field is largely a record of problems that appeared intractable until a small group of people decided to work on them full time. I have been wrong about the timing of such things all my life.