Meta Has the Distribution, Wang Has to Build the Model

In June 2025 Meta paid $14.3 billion for a 49 percent stake in Scale AI, and the company’s founder, Alexandr Wang, came along as Meta’s first Chief AI Officer.   He now runs Meta Superintelligence Labs, which brought Meta’s frontier model work together, with a smaller group called TBD Lab at the center of the push to train the next models.

Most coverage treated the deal as a talent story, and it was one, with huge pay packages and plenty of drama.   I find it more useful to read it as a story about structure.   Meta already owns one of the best customer relationships ever built.   Billions of people open its apps every day.   What it lacked, at least in the eyes of its own leadership, was a model strong enough to put in front of them.   So it went out and bought the person who knew how models get made.

I should say plainly that I’m an outsider.   I’ve never worked at Meta, and I’m leaning on public reporting and on what Wang himself has said in interviews.   With that said, here is how I read the first year.

The Supplier Takes the Wheel

Wang built Scale AI on labeling and evaluating data, the unglamorous work that decides whether a model learns the right things.   Think about where that put him.   Scale sold to the labs.   A supplier like that sees the whole industry’s problems without carrying any single lab’s bets, and I suspect that gave Wang a particular view of where the bottleneck sits.

The last few years of machine learning point the same way.   Progress has depended less on any one clever architecture and more on data quality, post training and evaluation.   Raw scale still matters, but the teams pulling ahead are the ones who can tell, quickly and honestly, whether a change made the model better.   Evaluation is the closest thing a model company has to a customer in the room.   It is the only voice that says no.

The open question is whether a data company’s instincts carry over to running a research lab.   Those are different jobs, and Wang has done only one of them.

A Closed Door

On April 8, 2026, Meta released Muse Spark, the first model from the new lab.   Meta said that over nine months the team had rebuilt its AI stack from the ground up.   The surprise was that the model was closed.   This is the company that made open weights the center of its AI identity with Llama.

Wang explained the decision in public.   He said the model touched high risk areas during early training, particularly around biological risk, and that Meta published those findings in a preparedness report.   His argument was that a model inside a product can be wrapped in mitigations, while open weights can’t be recalled once they’re out.   He added that Meta is working on models it considers safe enough to open source later.

There’s a fair criticism that Meta graded its own work, since the report uses its own framework and there is no outside audit.   That is true of every frontier lab right now.   I would add something Wang did not say, which is that the reasoning is also convenient.   Safety and control point in the same direction here, and that makes me want to look at incentives.   A company with a launch date and a story to tell finds late bad news expensive.   The lines that hold are the ones drawn before the training run ends, when holding them costs nothing.   I can’t see from outside whether that happened here.

Open Is a Distribution Decision

The debate over open and closed weights is usually framed as philosophy.   It is also a question of who controls the relationship with the user.

Releasing weights builds a developer ecosystem, gets free testing and earns goodwill, but it gives up the experience, the risk and the data that comes back.   Keeping a model closed reverses all of that.   Meta appears to be running both tracks, with Llama for the open community and Muse for the surfaces it controls.   Muse Spark came out with only a limited private preview for developers, while the main push went through Meta’s own apps, including WhatsApp and Instagram according to some reports.

That tells me where Meta thinks its advantage lives, and it isn’t in the weights.   A model sold to outsiders competes on quality and price against every other model.   A model inside an app that billions of people already use competes on whether it earns a place in a habit that already exists.   Meta does not have to persuade anyone to open a new app.   It has to make the thing in the group chat worth keeping.

Billions of Users

Here is the tension that I think defines the whole effort.   If models converge in quality, value flows to whoever owns the user, and Meta is in good shape.   If they don’t, value flows to whoever has the best model, and Meta has to catch up.   The company is hedged on one side only, and the lab exists to close the other.

Integration matters here.   Meta can build the model and the surface together, so the model can be shaped around how people actually message, share and scroll.   Someone selling models from outside can’t do that.   But integration cuts both ways.   A model that is merely good enough, forced through every surface, can wear down the thing that made those surfaces valuable.   Distribution lets you reach people.   It does not make them want what you put in front of them.   I would watch retention and repeat use on these surfaces far more closely than download counts or launch day numbers.

Wartime

Reports describe TBD Lab as running in a wartime mode, with long hours and fast decisions.   Critics, including some current and former employees according to the coverage, say Wang’s leadership has been frenetic and that incremental progress has sometimes been oversold.   I can’t judge the internal culture from here, and some of that reporting comes from people with reasons to be sour.

Two things still stand out.   The first is the seam between research and the apps.   A small elite group moves quickly because it doesn’t carry the weight of a large organization.   But the model still has to land in surfaces owned by people with their own roadmaps, metrics and habits,.   The handoff is where good research quietly dies.   I’d want to know who had the authority to say what the model needed to do before launch, and whether that person sat beside the researchers early.

The second is how progress gets described.   When a team lives in a sprint that never ends, there’s pressure to call every gain a breakthrough.   That is corrosive, because everyone downstream plans around the claim.   Better to say what improved, by how much, on which tasks, and what is still unproven.   A modest honest number holds up.   A large one unravels at launch.

What the Agent Is Allowed to Do

Wang has talked about building personal superintelligence, a phrase that sounds enormous.   At a Y Combinator event in 2025 he also made a smaller point I found more useful, that AI agents could make execution many times easier than it was a decade ago.

Strip the grand language away and the idea is simple.   An assistant that knows you, acts for you across your apps and doesn’t make you repeat yourself is worth building.   It also raises questions that no benchmark answers.   What can it do without asking?   What does it remember, and can you see that?   What happens when it books the wrong thing?   Meta can control the model.   It cannot control whether people trust it, and trust is what decides whether anyone hands over real work.   I think that is where the long term value will be won or lost.

The Model Is the Hard Part

So here is my view in plain terms.   Meta has the part of the business that nobody can buy, which is the users.   Wang has been brought in to build the part that can be bought, but only with great skill and a great deal of money, which is the model.   The closed release suggests Meta wants to keep the two tied together.   The bet is that a good enough model, delivered inside habits that already exist, beats a better model delivered from outside.

It’s too early to say whether that bet works, and I’d be wary of anyone who sounds sure.   One year is short.   The better test arrives in the next few releases, when the novelty has worn off and we can see whether people keep using what Meta built.

Sources

Reporting and statements I drew on include CNBC, Meta’s leadership page, The Next Web, Let’s Data Science and R&D World.   The opinions are mine.