Cole Bryant: Meta spent $14.3 billion to hire Alexandr Wang and bring him in to build frontier AI, but one year later the company is quietly replacing its own open-source Llama model with Wang's proprietary Muse Spark on its smart glasses — while publicly telling every developer on earth that Llama is the future.
Jonathan Ingles: Right. And nobody's saying it, but what Meta just did is tell its entire developer ecosystem one thing and then do the opposite thing on its own hardware. That's not a messaging problem. That's a choice.
Cole Bryant: Wait — $14.3 billion? Like, bro, that number stops me every time. That's not a hiring bonus, that's a — that's an acquisition of a person almost.
Jonathan Ingles: It's an acquisition of a narrative. Zuckerberg needed to be in the frontier AI conversation, Wang gave him that, and Muse Spark is the proof of concept — except developers are ignoring it, it's nowhere near OpenAI or Anthropic, and now there are trust and safety concerns because Meta gutted the teams who were supposed to catch those problems.
Cole Bryant: So Meta's like — open source for thee, proprietary model on the device in your face, and also we fired the safety people, good luck.
Jonathan Ingles: Frankly, that's exactly the situation. And the reason nobody wants to say it plainly is that Zuckerberg spent fourteen billion dollars and people are still being polite about it.
Cole Bryant: Meta paid fourteen point three billion dollars to get Alexandr Wang inside the building — and within twelve months they're quietly pulling their own Llama model off the Ray-Ban glasses. No press release. Just — different model now. That's not a product update. That's a confession.
Jonathan Ingles: A quiet confession, which is worse. Zuckerberg spent a year telling developers, telling the press — open-source is the future, Llama is the future. Then in May 2026 he pulls Llama off his own hardware. That's not iteration. That's the company admitting its flagship model wasn't good enough for its own products.
Cole Bryant: And the Oakley glasses too, right? Not just the Ray-Bans.
Jonathan Ingles: Both. Ray-Ban and Oakley. Both quietly swapped. Nothing public.
Cole Bryant: Bro. Okay — walk me back to June 2025 for a second. Meta acquires a 49% stake in Scale AI. Fourteen point three billion. They bring in Wang to run this brand new unit — Meta Superintelligence Labs, MSL. Zuckerberg is essentially saying: this is how we compete with OpenAI and Anthropic and Google. Frontier AI. That's the whole bet.
Jonathan Ingles: That's the stated bet. Here's what I'd push on — when you spend fourteen billion on someone, you're either buying their talent or their name. And within a year we're getting reports of clashes between Wang and Zuckerberg over MSL's direction. That doesn't happen if Wang actually has authority. That happens when you brought in a prestige hire and kept the decision-making exactly where it already was.
Cole Bryant: Wait — does that surprise you though? Zuckerberg is one of the most controlling founders in tech. He's not handing the keys to someone he's known for twelve months.
Jonathan Ingles: Then don't pay fourteen billion and frame it as handing someone the keys. The pitch was 'Wang leads frontier AI for Meta.' If Zuckerberg is overruling him in year one — that framing was always marketing.
Cole Bryant: I mean — okay, I hear that, but I also think there's a version of this where the clash is just... normal? Wang built Scale AI from scratch. Now he's inside a company with two billion users and a founder who's personally obsessed with winning the AI race. That friction might be the process working. Two smart people arguing about direction isn't automatically a failure.
Jonathan Ingles: Sure. Except the output of that productive friction is a model the Wall Street Journal reported was repeatedly delayed. Shipped to a private API preview. Weak developer adoption. That's not two smart people arguing their way to a breakthrough. That's organizational paralysis dressed up as deliberation.
Cole Bryant: Okay but — Muse Spark did ship. That's real. Meta built a multimodal proprietary model with selectable modes for different task types. Wang called it an appetizer — meaning more is coming. That's—
Jonathan Ingles: Right, but—
Cole Bryant: — no, let me finish — calling it an appetizer is either confidence or spin, and I genuinely don't know which one it is yet.
Jonathan Ingles: Here's the thing about 'appetizer.' Developers are ignoring Muse Spark. That's not my read — that's what's being reported. The broader developer ecosystem is just not engaging. So what does it mean to call something an appetizer when nobody's at the table?
Cole Bryant: That's the part I keep getting stuck on. Because developer ecosystem stickiness is real — once you're building on OpenAI's API or Anthropic's Claude, switching costs are high. But is that the problem here, or is the problem that Meta broke trust before the model even launched?
Jonathan Ingles: Both. And the trust piece has a timeline. Meta laid off thousands of employees. Zuckerberg publicly said the cuts were necessary to fund AI ambitions — to fund MSL, to fund the whole play. Then in that same restructuring wave they cut the Integrity team. The trust and safety people. And what happens after? WIRED reports face-recognition code was quietly sitting inside the Meta AI app before it got removed. That sequence — cut safety, then have a safety incident — developers notice that.
Cole Bryant: Wait — face recognition code? In the Meta AI app?
Jonathan Ingles: Quietly removed after WIRED found it. And the timing is the thing. Integrity team cut, then this surfaces. That's the pattern.
Cole Bryant: Dude. That lands different when you put the layoffs right before it. Because Zuckerberg's whole framing was 'these cuts are a reallocation, this is about investing in AI.' But if you're reallocating away from the people who catch safety problems—
Jonathan Ingles: You're not reallocating toward capability. You're hollowing out the infrastructure that makes capability trustworthy.
Cole Bryant: And then there's the Applied AI unit. This is the one that got me. Six thousand five hundred people — massive team — and engineers inside Meta are calling it 'the gulag.' Publicly. That's not a leaked Slack message. That's people willing to say it out loud.
Jonathan Ingles: Which tells you the demoralization is past the point where people are afraid to say it. When your own engineers use that word for a unit Zuckerberg just created — that's not a morale problem. That's a credibility crater.
Cole Bryant: And that's the talent problem. Because the people you need to close the gap with OpenAI and Anthropic and Google — you need them engaged, you need them believing in the mission. If the vibe inside MSL and the Applied AI unit is 'we got restructured into a punishment detail,' those are not people shipping frontier AI.
Jonathan Ingles: Google published Gemini 2.5 — Pro and Flash — hitting state-of-the-art on coding and reasoning benchmarks. Anthropic is not slowing down. OpenAI is not slowing down. Meta is not just behind; Meta is trying to close a gap while demoralizing the people who are supposed to close it.
Cole Bryant: Okay but here's where I want to push back — the $145 billion capex number. Meta is spending one hundred and forty-five billion dollars on AI infrastructure this year. Does that not buy you something? At some point compute matters.
Jonathan Ingles: Compute matters when you have a coherent strategy to use it. What Meta has right now is a $145 billion infrastructure bet running alongside a publicly championed open-source model they're quietly replacing in their own hardware, a proprietary model that's delayed and ignored, and an internal culture in revolt. You can buy a lot of GPUs. You can't buy your way out of that.
Cole Bryant: No but — the spend has to compound eventually, right? Meta isn't going away. Two billion users, the distribution, the money. Is the developer adoption problem a timing thing that fixes itself when the models get better, or is it permanent damage?
Jonathan Ingles: That's the right question. And I think it depends on whether Muse Spark 2.0 — whatever comes next — lands in a company that actually has its house in order. If Wang and Zuckerberg are still clashing over direction, if the Applied AI unit is still 'the gulag' — a better benchmark score does not move developers. Developer ecosystems are winner-take-most. And look — Llama 3 is still the most downloaded version of their own model. Llama 4 underperformed. Muse Spark is being ignored. That's a pattern, not a bad launch.
Cole Bryant: Wait — Llama 3 is still the most downloaded? Over Llama 4?
Jonathan Ingles: Llama 4 landed poorly. Llama 3 is still what developers are using. Which means Meta's open-source strategy — the thing they've been publicly banking on as the democratizing force against OpenAI's closed models — isn't even building on its own momentum.
Cole Bryant: So let me say the quiet part out loud — Meta told developers 'build on Llama, open-source is the future,' and then put a different model in their own glasses. That's not just a contradiction. That's Meta signaling to every developer watching that Llama is the public story and Muse Spark is what they actually think is better. And they can't say that publicly because they've staked their whole identity on open-source.
Jonathan Ingles: Now you've got it. And Meta cited safety concerns — bio, chemistry, cyber risks — as the reason Muse Spark stays closed-weight, behind a private API. Which might be true. But it's also a very convenient framing for 'we're not ready to let people kick the tires on this.' The developers who've been paying attention know the difference.
Cole Bryant: And Zuckerberg — literally a year after the Wang hire — admitted mistakes in the restructuring. Like, he said it out loud. That's—
Jonathan Ingles: Zuckerberg doesn't do quiet admissions. That's what makes it notable. When he says mistakes were made, that's the man who built Facebook admitting something went sideways on his flagship bet of the decade.
Cole Bryant: Which brings me back to Wang. Because — I keep thinking about what it's actually like to be him right now. You built Scale AI. It mattered. You ran it. Then you get recruited into Meta Superintelligence Labs, leading frontier AI for the world's biggest social company. That's supposed to be the job. And then year one is: reported clashes with the founder, a model the ecosystem is ignoring, your unit's culture imploding. What's he thinking?
Jonathan Ingles: Probably what anyone thinks when they take a big job and discover the organization wasn't actually ready to give them the authority they were promised. But look — that's the human version of a structural problem. The structural problem is that Meta spent fourteen point three billion to buy a founder's credibility and then kept the founder's authority exactly where it already was. With Zuckerberg.
Cole Bryant: And here's what I keep circling back to — the $145 billion. If the next model lands with the same developer indifference, same internal friction, same quiet product swaps — Meta will have spent close to two hundred billion dollars to prove you can be technically competitive in frontier AI and still lose the ecosystem war.
Jonathan Ingles: Because trust and clarity compound the same way compute does. Meta has one and not the other. Cut safety, have a safety incident, ship a delayed model your own hardware isn't running — that's not a bad year. That's a strategy that isn't working. And no amount of capex fixes the sequence.
Cole Bryant: So what does a good year two even look like for MSL? Like if Wang is still there — what actually has to change?
Jonathan Ingles: Developers building on Muse Spark instead of ignoring it. An open API instead of a private preview. And Wang staying long enough to have a second year that doesn't start with reported clashes. None of that is guaranteed. Frankly, none of it is even close to guaranteed right now.
Cole Bryant: Man, okay. Like — $14.3 billion to start, nearly $200 billion all in, and the question at the end of it might just be: do developers trust you? That's it. That's the whole thing.
Jonathan Ingles: That's always been the whole thing. Meta just thought they could buy their way past it. And the honest answer — the one nobody at Meta wants to say out loud — is that you cannot acquire trust. You cannot hire it. You either build it slowly or you don't have it.
Cole Bryant: So I guess the actual question — the one we don't have an answer to — is whether Meta even knows that's the problem they're solving. Because if they think they're still just solving for the model, they're already lost. Catch you next week.
