PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

,

Muse Wonder: Zuckerberg and Alexandr Wang on Meta’s $100B AI Bet

Mark Zuckerberg and Alexandr Wang tell John Collison how Meta built Muse Wonder and why it committed over $100 billion before the model existed. The conversation is the October 12, 2026 episode of Cheeky Pint, it runs two hours and fourteen minutes, and it doubles as the launch interview for Meta’s new frontier model. It covers the Muse personal agent, how a frontier model is trained, what a gigawatt of compute costs, agentic commerce, smart glasses, and the year-old Meta Superintelligence Labs (MSL) that Wang runs as Chief AI Officer.

TLDW

Meta is releasing Muse Wonder, a much larger model than Muse Spark, tuned to be a personal and professional agent, with a higher tier called Muse Wonder Full. Zuckerberg and Wang say it is state of the art on their own measurements, strong at memory, long tasks, coding and personal legal, finance and health questions, and they describe how pre-training, supervised fine-tuning and reinforcement learning produced it. The center of the interview is money: Meta committed $100 billion to $200 billion at the end of 2025 with no model to look at, and Zuckerberg argues the bet was safer than it sounds. The last half hour is the story of rebooting Meta’s AI effort around a research team of roughly 100 people.

Thoughts

Zuckerberg’s defense of the spending is that Meta cannot lose alone. If AI works, he says, other companies will pay more than the roughly $15 billion a year it costs Meta to run a gigawatt, so “it’s not just that we have to fail, AI overall has to fail.” That is true, and it is a smaller claim than it sounds. It covers the case where the chips hold their value. It does not cover the case Collison was asking about, where AI works and Meta’s model does not. In that case Meta is a landlord renting compute to the labs it set out to beat, which pays the bills and loses the argument for building MSL at all. CFO Susan Li’s warning to Wang, that a bad model would be “really, really, really bad,” is that case with a price on it. The backstop also depends on compute staying scarce, and the year it stops being scarce is the year everyone’s resale value falls together.

The watermelon story deserves more suspicion than it gets at the table. Muse noticed that Zuckerberg’s Instacart basket was under the delivery minimum and added a watermelon. He tells it as charming, and it is. It is also an agent spending its owner’s money to solve a problem the merchant created, inside a product whose revenue plan is “a small cut of what you are saving or making for people.” Zuckerberg asks whether his new Instacart account is really for him or for the Muse, and treats the answer as a bigger market. The harder version of that question is who the agent works for when the owner and the merchant want different things. Meta has twenty years of practice at being paid by the seller. Nobody asks, and the host is not neutral here: Stripe handles Muse’s payments and has agreed to buy OpenRouter, which sells access to Muse Wonder.

Collison lists four things most founder CEOs would not have done: declare the Llama effort not good enough, pay enormous sums for people, hand the lab to someone who had never trained a frontier model, and clear two months of calendar to recruit. Zuckerberg’s answer on the third is the interesting one. The field changes too fast for expertise in the current training technique to be the test, he says, so he picked for clear thinking and for whether the best researchers wanted to follow. That is a real theory of hiring, and one year and two model lines is thin evidence for it. I would watch what happens to the flat structure, where nearly the whole lab nominally reports to Wang and there are no non-technical managers, once the new enterprise group needs the opposite kind of organization.

Wang credits distillation with keeping AI from becoming winner-take-all, since capability keeps getting packed into smaller models. Zuckerberg, an hour earlier, hopes to have Wonder-level quality in the cheap Spark size “not too far from now” and give it away. Put those together and the frontier advantage that costs $50 billion to $60 billion per gigawatt to build has a shelf life measured in months, by their own account. That is fine if the product is distribution, which is the Meta playbook Zuckerberg describes: get to a billion people, then scale revenue. It is bad if the product is the model. It also explains why the paid tier is described as “pay money” and the network effects are described with conviction.

On benchmarks the two want it both ways. Wang says chasing public benchmarks means “outsourcing all of your product thinking” to a PhD student, a few minutes after citing results on three public coding benchmarks. The claims that matter most to their pitch rest on internal evals nobody outside Meta can run. The one idea here I would like to see tested is “big model smell,” which Wang defines as the gap between how a model does on tasks inside its training distribution and tasks outside it. That turns a vague impression into a number. Until someone else measures it, everything in this interview about how good Muse Wonder is comes from the people who made it.

Key Takeaways

  • Muse Wonder is a new, significantly larger pre-train than Muse Spark. A higher-end version, Muse Wonder Full, goes to Muse Code subscribers. Wonder is also offered through the Meta Model API and OpenRouter.
  • Inside the Muse app, Wonder starts in paid subscriptions because it costs more to serve. The free tier stays on Muse Spark. The app has no model picker and no effort picker, on purpose.
  • Wang names the design priorities as active memory (remembering across days, weeks and months, with compaction handled by the model itself), active focus (finishing long, ambiguous tasks), domain expertise in legal, finance and health, and frontier coding.
  • Zuckerberg puts the cost of operating a one-gigawatt cluster at about $15 billion a year, and building one at $50 billion to $60 billion up front. Wang says tens of billions of dollars is a fair estimate for training one frontier model.
  • Meta’s Prometheus cluster in Ohio is scaling past one gigawatt, and Hyperion in Louisiana is planned for five. Both were planned in 2023 or 2024, before MSL existed.
  • Muse Wonder was planned in August 2025. The $100 billion to $200 billion commitment was made at the end of 2025, when the only evidence was what Wang calls a vibe report and some plots.
  • Muse Spark versions 1 through 1.3 all sit on one pre-train from December 2025. The gains between them came from more and better reinforcement learning, where Wang says less data is often more.
  • Meta defines alignment as doing what the user asked, reliably, without what Zuckerberg calls a moralizing lecture, while declining catastrophic requests gracefully. He says this differs from how other labs think about it.
  • Muse uses trained-in browser use so no business has to integrate, with direct API access for partners. Zuckerberg says far more businesses are asking to integrate than trying to block it.
  • The revenue plan is subscriptions, a cut of commerce, a “business in a box” for small businesses, the API, a new Meta Enterprise Platform led by CJ Desai, and a few percent of extra growth in the core ad business.
  • The MSL research team is under 200 people, closer to 100 by Zuckerberg’s count, plus infrastructure staff. Each training stage has tens of people where other labs have hundreds.
  • The reboot began about two weeks after Llama 4 shipped in April 2025. Wang’s first written advice to Zuckerberg was that talent density matters more than headcount.
  • Collison says Stripe’s agentic commerce volume started going vertical in September 2026, after product announcements in May.

Chapters

0:24 Muse Use Cases and Why the Model Is the Product

Wang has his Muse source obscure teas from farms in Taiwan and Japan every week. Zuckerberg put cameras in his gym so Muse can coach his MMA between rounds, which means processing video and answering inside a 60-second break. He says the agent wrote its own video pipeline and that most people will soon be writing code without knowing it. Collison argues that winning AI products collapse to one simple shape. Both guests push back: the long product spec still exists, but it now lives in post-training decisions about behavior, personality and tool use.

14:35 Meet Muse Wonder and Muse Wonder Full

Zuckerberg describes climbing a ladder of pre-trains, each four or five times larger than the last, and says the gap between model classes at Meta should now shrink. Wang walks through active memory, active focus, domain expertise and coding, citing state-of-the-art results on DeepSWE and SWE-Marathon. Collison notes that memory is the part of OpenClaw users find buggy. Distribution is paid Muse plans first, plus the API, OpenRouter and Muse Code. The name Wonder Full is a pun, and they are pleased with it.

24:56 How a Frontier Model Is Built

Meta now steers by internal evals more than public benchmarks, which Wang says saturate fast and encode someone else’s priorities. Zuckerberg adds a requirement he calls discretion: an agent booking a restaurant should handle your allergy without announcing your medical history. Pre-training takes huge volumes of data across text, images, audio and video, and larger models need more data or they overfit. Supervised fine-tuning seeds behaviors, and reinforcement learning makes them reliable. Wang says safety problems show up when RL pressure rewards a model for succeeding the wrong way, so the environments were curated with that in mind.

48:26 Betting $100 Billion Before Seeing the Model

Model planning runs more than a year ahead and data centers several years, so the money goes in before any proof. Wang calls it the opposite of The Lean Startup. At an executive offsite in Hawaii, CFO Susan Li pulled him aside more than once to say the model had to be good. Meta also had to size inference capacity for a product that did not exist. Zuckerberg’s view is that the risk was lower than the numbers suggest, because the compute has other buyers and other uses, and because a consumer agent was the open lane after ChatGPT and Anthropic’s coding agents.

1:07:21 Agentic Commerce and the Watermelon

Collison asks what happens when businesses block agents. Zuckerberg expects the blockers to be disrupted, and says every business will end up with its own agent, which Meta is also building. Wang says people are already delegating whole purchases to Muse, and that its advantage is context across a person’s whole life where older recommendation systems sat in silos. Zuckerberg describes planning weekend baking with his three-year-old through Muse voice mode, with ingredients arriving by Instacart on Saturday. He had rarely used Instacart before.

1:20:06 Voice, Glasses and Muse Gadgets

Collison orders a watermelon to Stripe’s office by voice, then asks why talking to computers has taken so long. Zuckerberg gives three reasons: the assistant has to be able to act, turn-taking in conversation is hard, and listening is a slow way to take in information even though speaking is a fast way to give it. Wang says researchers already dictate to their coding tools. Glasses with a small display are the multimodal target. Muse Gadgets opens the agent to other hardware, including a coin-sized HomeLink for home devices.

1:27:00 How the Buildout Gets Paid For

Zuckerberg wants Muse free and widely used because he expects network effects once agents interact with each other and partners integrate. Revenue comes later and from several directions, as it did for Meta’s other products. Wang points out that R&D spending at any moment covers several overlapping model generations, so it looks worse than it is. Zuckerberg adds that owning data centers beats renting from clouds, since the chips keep working after they are depreciated. He says an H100 is worth more now than a year ago.

1:34:06 Why AI Has Not Been Winner-Take-All

Wang says frontier models are diverging in what they are built for, and that a gateway like OpenRouter could help developers see differences that benchmarks flatten. Zuckerberg argues a future with many capable models in many hands is safer than a few models in a few companies. Asked why no lab ran away with it, Wang gives three causes: every lab has stretches of weak execution, distillation spreads capability into small and open models, and many researchers and companies actively choose a more even outcome.

1:40:46 Rebooting Meta AI as MSL

Zuckerberg started calling people for advice within weeks of Llama 4. He says he talked to the heads of every lab and concluded Wang was the person, despite the “social media guy and data guy” optics. The decision to start from scratch came out of recruiting conversations, because a clean start was what the best candidates wanted. Wang recalls an early meeting where a lead on the old effort said consumers did not need a reasoning model. The lab runs with no deadlines, and Wang says the team’s forecasts are reliably too pessimistic. They knew Wonder was good when it beat Spark almost immediately in RL.

2:08:11 Working for Mark Zuckerberg

They still argue, most recently over the model’s name. Wang says what surprised him is how completely Zuckerberg reorganizes his actions once he changes his mind, and how large the bets are. He and Nat Friedman used to ask each other whether they would go this big. Many CEOs say AI is the most important thing for their business, Wang says, and few act like it. Zuckerberg defends Wang’s posting on X as the cost of being authentic. The closing summary is to stay technical and stay at the frontier.

Notable Quotes

“And so you’re kind of outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark.”

Alexandr Wang, on why Meta steers by its own evals

“We’re going to write this check for, like, I don’t know, like $100 billion, $200 billion. I’m like, Alex, are you sure this is going to work?”

Mark Zuckerberg, on committing the money at the end of 2025

“If it’s good, it’s all good. We’re great. But if it’s bad, it’s really, really, really bad.”

Alexandr Wang, recounting what CFO Susan Li told him in Hawaii

“In order for it not to work out, it’s not just that we have to fail, AI overall has to fail.”

Mark Zuckerberg, on why he sees the spending as less risky than it looks

“We believe that Muse is actually going to make people money and save people money by using it, and that over time there are lots of different ways to effectively take a small cut of what you are saving or making for people.”

Mark Zuckerberg, on how Muse earns revenue beyond subscriptions

“It’s not like anyone planned for distillation to work super duper well, but it just turns out that you can get a lot of capabilities naturally through some form of distillation.”

Alexandr Wang, on why no lab has pulled away

“Oh, we don’t need a reasoning model. Do any of our consumers need a reasoning model?”

Alexandr Wang, quoting a former lead of Meta’s earlier AI effort in one of his first meetings

This post reports Muse Wonder’s abilities as Zuckerberg and Wang describe them on launch day, and none of the benchmark claims were checked independently. Watch the full conversation here.

Related Reading

Keep reading