Mark Zuckerberg's Cheeky Pint Interview: Summary and Timestamps

A summary of Mark Zuckerberg and Alexandr Wang's Cheeky Pint interview with John Collison: Muse Wonder, the $100 billion compute bet, agentic commerce and how Meta rebuilt its AI lab.

▶ Watch the full interview on YouTubeCheeky Pint · 12 October 2026

Mark Zuckerberg, the founder and CEO of Meta, was a guest on Cheeky Pint, the podcast hosted by Stripe co-founder John Collison^[2]. Alexandr Wang, the Chief AI Officer of Meta Superintelligence Labs (MSL), joined him. The episode was published on YouTube on 12 October 2026 as "Zuckerberg's $100 Billion Bet" and runs 2 hours 14 minutes^[1]. It had more than 1.2 million views by 10 November 2026. The conversation covers Meta's new Muse Wonder model, the money Meta committed before the model existed, agent shopping and how MSL was set up. Cheeky Pint also published a full transcript on Substack^[2]. This page summarises what each guest said, grouped by topic, with timestamps.

We also have a profile of Alexandr Wang.

Topics and timestamps

The chapter names come from the episode description^[1]. The OpenRouter chapter is a sponsor message.

How the guests use Muse

Muse is Meta's personal agent app. According to third-party reports mentioned at the start of the episode, it might be the fastest-growing agent or AI app yet^[1]. Wang's favourite use is having his Muse source increasingly obscure teas. Every week it buys tea from an obscure vendor, such as a farm in Taiwan or Japan.

Zuckerberg uses Muse to coach him in MMA, and he called this both his favourite use case and an eval for the product. He put cameras in his gym. Muse watches the video and gives him feedback in the one-minute break between rounds. It pulls still frames from the video by itself, and its feedback covers balance and hip position. He said he first did this with OpenClaw, before Muse was working.

Zuckerberg said Muse is writing a whole video and image processing pipeline for this, though he never sees the code. When it took 30 seconds to answer, he told it to optimise itself. He thinks this shows where software is going: "everyone is actually going to be writing tons of code. They're just not going to know that they're writing code."

Why Meta says the model is the product

Zuckerberg said Meta trains its models end to end to be good personal agents. He said Wang focused on the multimodal program from the start, because it matters most for the things people use these services for. Wang said most people experience the world in a highly multimodal way, by talking, listening and watching video. Earlier in the year, he said, much of the AI industry thought coding was all that mattered, while a consumer product like Muse depends on strong multimodal skills. Zuckerberg added that most people do not like reading that much, and Meta's glasses will eventually be a fully multimodal product.

Collison argued that winning AI products quickly converge on one simple design. Zuckerberg disagreed. He said an agent that gets things done is different from the text assistants every company has, including Meta AI. Giving each agent a dedicated virtual machine (VM) matters for the security and privacy of your data, he said.

Zuckerberg said products are getting thinner and "the model is the product". Much of Muse's product work went into the privacy and security model for your data. Capabilities and personality are trained into the model. Zuckerberg said that when the person whose team works on the product thinks about making Muse feel good, he talks about post-training the model. Wang compared the moment to early mobile, when everything was "just an app". He said today's product differences show up in how proactive an agent is and how it talks to you, and those come from model design, training data and evals.

Muse has no model selector and no effort selector. Wang said many AI products were designed by technical teams who do not picture an ordinary user, and some model names are far too confusing for people who do not read API specs.

What Muse Wonder adds

Zuckerberg described Muse Wonder as a new class of frontier model. Meta trains models up a ladder, each maybe four or five times bigger than the last. The Muse Spark models came from a pre-training run at the end of last year, and Muse Wonder is significantly bigger. Internally, the Spark line was code-named Avocado and the Wonder line Watermelon. Zuckerberg said Meta was not happy with the trajectory it was on with Llama 4 and rebuilt the whole team, and that from now on its work will push the frontier.

Wang said there are two models in the series, Muse Wonder and Muse Wonder Full, a much higher-end version. He said Muse Wonder is state of the art on Meta's own measurement of personal and professional agent tasks. He named four strengths:

• Active memory. How well the model remembers what you told it over days, weeks or months and uses that in later tasks. Wang said Meta designed both the model and a memory system for it to use, and skills such as compacting memory are in the model itself. He said there is still plenty of room to improve.
• Active focus. The ability to carry long, complex and ambiguous tasks through to the end, and to find another way when something goes wrong.
• Domain expertise. Wang wants the best personal agent to "feel like a team of experts in your corner". He said Muse Wonder is state of the art on many finance and legal evals.
• Coding. Wang said it is state of the art on DeepSWE and SWE-Marathon^[2] and does well on Terminal-Bench 4. He expects coders to be a minority of its users.

Zuckerberg said Muse will stay free, with free tokens on Muse Spark. Muse Wonder is bigger and more expensive to serve, so in the Muse app it will start mainly in paid subscriptions. He hopes Meta will soon have a model as good as Wonder at Spark's size, which could serve hundreds of millions of people. Wang said Muse Wonder will also be available through Meta's model API, OpenRouter and Muse Code, and Muse Wonder Full will go to Muse Code subscribers.

Benchmarks and big model smell

Zuckerberg said public benchmarks get saturated quickly. Early on, Meta used external benchmarks to check it was on the right track. With each release, a larger share of what it measures comes from its own internal benchmarks.

Wang said Meta's research blog post on Muse Wonder measured "big model smell", the sense that you are using a very large model. Meta measures it as the gap between performance on tasks inside its training distribution and tasks outside it. He said this gap showed a big step up from Muse Spark to Muse Wonder. He also said that chasing public benchmarks means "outsourcing all of your product thinking" to the PhD student who made a viral benchmark.

Zuckerberg said a personal agent needs things a coding agent does not. One is discretion. If you have an allergy or are pregnant, the agent should be able to get you a table at a suitable restaurant without telling the restaurant about your medical condition.

How a frontier model is built

Zuckerberg said the MSL research team is less than 200 people, closer to 100, plus a group of infrastructure staff. Wang said each stage of training has tens of people working on it at Meta. Zuckerberg said Meta gives a lot of compute to each researcher and does not spread it across many competing efforts.

In pre-training, Zuckerberg said, you add large amounts of each modality you care about, such as images, audio and video for perception. Domain data matters from the start too. He named shopping, personal advice, and personal legal, health and tax questions as areas people care about. A bigger model needs more data, or it overfits. Since the internet is not growing that quickly, he said, finding enough data is one of the hard problems of scaling.

Wang said Meta invested a huge percentage of its compute in pre-training Muse Wonder, which made it the highest-stakes part of the research. He described it as scientific in an unusual way, because a team has very few samples to learn from.

Post-training has two stages. Zuckerberg said supervised fine-tuning (SFT) shows the model questions and the answers you want, and it is important for shaping personality, but most recent advances have come from reinforcement learning (RL). Wang said SFT might teach a behaviour the model shows 10% of the time, and RL raises that rate sharply. In RL, he said, less data is often more. RL also magnifies what the model learns. If heavy RL pressure is applied where the only way to succeed is unsafe or misaligned, the model learns that behaviour. So Meta curated the RL environments for Muse Wonder to drive the model toward aligned behaviour.

On alignment, Zuckerberg said Meta wants the agent to do what its user asks with high reliability. "You don't want a moralizing lecture" for asking, he said, but the agent should refuse something catastrophic without annoying the user. He said this differs from how some other labs think about alignment.

Wang said Meta has planned Muse Wonder since last August and saw Muse Spark as a stepping stone to it. Zuckerberg said that after pre-training, the pre-training team warned it could take a while in RL before Wonder beat Muse Spark, and it happened almost immediately.

The $100 billion bet

The episode opens with a clip of Zuckerberg describing the end of last year. Meta had not seen the models yet, and it was about to write a check for "$100 billion or $200 billion". He said: "I'm like, Alex, you sure this is going to work?"^[1]

Wang said AI is the opposite of the lean startup method, because a lab cannot wait for proof that something works before investing. Meta had invested in data centres for years. At an executive meeting in Hawaii at the end of last year, Wang said, CFO Susan Li pulled him aside many times to say that if the model did not work, it would be "very very bad in financial terms". He said that at that point he could show only some research plots, and it was not clear the model would work. He called it a lot of risk, even for a founder who has taken risks before.

Zuckerberg put the cost of operating a one-gigawatt cluster at about $15 billion a year, and the cost of building one at about $50 billion or $60 billion up front. Research costs about as much again, and serving the product adds more. Wang agreed that tens of billions of dollars is a good estimate for training a frontier model.

Meta also had to build serving capacity years before Muse existed. Zuckerberg said some people wanted to be more conservative, and he decided to go ahead. He argued the bet is less risky than it looks. If AI works, other companies would pay for Meta's compute, so "it's not just that we have to fail, AI overall has to fail", and Meta has other uses for compute anyway.

Zuckerberg said Meta measures a lab's compute strength by its largest single cluster. Meta's Prometheus cluster in Ohio scales to more than a gigawatt, and the Hyperion cluster in Louisiana will scale to 5 gigawatts. Meta made those plans in 2023 or 2024, before MSL existed. He said Muse's launch was a "home run", although his median expectation had been that it would need a few rounds of changes.

Agentic commerce

Zuckerberg said Meta trained good browser use into the models so that businesses do not have to integrate anything for Muse to work. Partners who want it get direct API access, which is faster and more efficient. He expects every business to have its own agents, just as every business now has a website. He said far more businesses have asked Meta to integrate deeply than have tried to block Muse.

Wang expects agent shopping to grow as a share of commerce, the way e-commerce grew next to physical stores. He said Muse is one of the first products where people fully delegate purchases to an agent. The agent can apply a rule such as "I need something by Wednesday". The long-term benefit he sees is personalisation, because Muse knows about your family, friends and co-workers, while today's recommendation systems are separate from each other. Wang said that at a dinner in 2023, someone asked him what they should be doing, and he warned that AIs were going to buy things and they were not prepared for it. He later called that view correct in direction but three years early.

Zuckerberg described baking with his three-year-old daughter, Aurelia. During the week they talk about what to bake with Muse's voice mode, which is not fully rolled out yet and can tell their voices apart. Every Saturday, Muse orders the ingredients from Instacart. One week the order was too small for delivery, so Muse added a watermelon. Zuckerberg said he rarely used Instacart before. He opened an account to give to Muse and now uses it once or twice a week. He asked whether that account is really his or the Muse's, and said the market for commerce could grow because agents will have their own accounts.

Collison said agentic commerce at Stripe only started growing quickly in September. Later in the episode, Zuckerberg asked his Muse to order a watermelon to Stripe's office in South San Francisco.

Voice, glasses and devices

Asked why talking to computers is still hard, Zuckerberg named three problems. The agent needs to act on what you say. Conversation has its own timing, such as when a pause means you are done. And listening is a slow way to take in information compared with seeing it. He said the ideal interface is probably not voice in and voice out.

Wang said voice input is growing fast. Most of the researchers speak their prompts, and a common habit among college students is to talk to an AI for 10 or 15 minutes and let it organise their thoughts. He said Muse can be shown as a small character working for you only because it has persistent memory, can take actions and has some personality.

Zuckerberg said many people want glasses with no display or a small one, and Meta has started shipping display glasses that are not full augmented reality. Wang said the Muse Gadgets program lets others build devices for Muse, and several phone makers have contacted Meta. Zuckerberg described a small USB gadget, about the size of a nickel, that lets Muse control a home. He uses it to dim all the lights in his house in the evening.

How Meta plans to pay for it

Zuckerberg said Meta wants Muse to be free so it reaches large scale, which brings network effects and more partners. He listed the ways Meta expects to earn money:

• subscriptions, for example to use Muse Wonder in Muse
• commerce, by taking a small cut of the money Muse saves or makes for people
• a "business in a box" for small businesses
• selling the API
• a new Meta enterprise platform, which CJ Desai joined Meta to build

He said Meta already starts as a very profitable business with good credit. Its playbook is to get a product to a billion people and then grow revenue. Meta also uses the models to make its core business, which he described as hundreds of billions of dollars, grow a few percent faster. Wang said R&D spending at any time covers several model generations at once, because each takes so long to plan.

Zuckerberg said owning data centres is a financial advantage over companies that rent cloud capacity. He expects Meta to keep using its GPUs after they are fully depreciated. He said an H100 is worth more now than a year ago because compute is still scarce.

Why AI is not winner-take-all

Collison asked why no lab has pulled far ahead. Wang gave three reasons. Every lab has periods of strong and weak execution, and a leader that executed perfectly would probably stay ahead. Distillation lets small models gain a lot of capability, which helps open-source models, and small models today are much more powerful than the biggest models of a year or two ago. And many people in the industry do not want one company to win everything, which affects where they choose to work and who they do business with.

Zuckerberg agreed. He said that having only a few good models in a few companies' hands is "probably a riskier configuration for the future" than many people having access to different models. For OpenRouter, Wang suggested helping developers understand how models differ beyond benchmark scores, since Muse Wonder focuses on different things than rival models.

Earlier in the episode, Zuckerberg said every generation of technology has a few very large companies, and a small handful will be able to build frontier superintelligence. He said many rivals focus on automating knowledge work, while he sees "invention and creation, not automation" as the highest use of superintelligence.

How Meta Superintelligence Labs started

Zuckerberg said Meta launched Llama 4 on 4 April 2025, and around 20 April he started calling people for advice. He contacted Wang, and they talked for 44 minutes. Two days later Wang sent a long plan. Its first point was that you do not need a lot of people, and talent density is what matters. Collison first called MSL two years old, and the guests corrected him: it is a little over one year old.

Zuckerberg said he cleared his calendar for a month or two to meet researchers, and neither he nor Wang slept much in the first couple of months. He had talked with Wang every few months for years and thought he was the sharpest entrepreneur of his generation that he had met. He knew some people saw the pairing as "the social media guy" working with "the data guy". After talking to the leaders of all the labs, he said, "Alex is the guy."

Zuckerberg said they were not sure at first that MSL needed a full reboot. It became clear while recruiting: the best people found it more attractive to start from scratch, and MSL still brought in people from the Llama team and FAIR. Wang added that in one of his first meetings, a lead of the old effort said consumers did not need a reasoning model. He called it a "heart-sinking moment". He said Meta's leaders needed to be much more forward-leaning on how fast AI is moving, and that many of them adapted quickly. He named Adam Mosseri, who runs Instagram, as having one of the most effective personal setups of vibe-coded mini-apps he has seen.

Wang said Meta told MSL there were no deadlines. He described the team as "insecure overachievers" who give pessimistic forecasts. He records their predictions and shows them three months later how far they got. Zuckerberg said the point of the flat structure is "the absence of non-technical managers". People at MSL have to stay technical, use the agents and write code.

Asked whether he would hire outsiders for other projects, Zuckerberg said the experience made him somewhat more optimistic. He gave the example of Kunal Shah, who built the company CRED and now runs WhatsApp. He called MSL "a level 10 out of 10 difficulty thing" that is still going on, and said it puts strain on the rest of the company and its financial model.

Working with Zuckerberg

Wang said Zuckerberg changes everything he does once he concludes something is right. Much of Meta pivoted to support MSL, with infrastructure and data centres, and Wang thinks only a founder CEO could make that happen. He said he and a colleague often asked each other whether they would go as big as Zuckerberg did. Collison described this as taking in new information and fully updating his worldview, and Wang called him "very Bayesian". Zuckerberg replied: "Life is short. Just do the right thing."

Zuckerberg said he wants people to argue for what they believe in, not blindly execute. Asked about Wang's posts on X, he said it is most important to be authentic, even if some posts do not come out as intended, and he finds most of them hilarious.

Sources

1. Zuckerberg's $100 Billion Bet — Cheeky Pint, YouTube (12 October 2026; views checked on 10 November 2026)
2. Muse Wonder: Zuckerberg and Wang's $100B AI — Cheeky Pint, Substack (12 October 2026), full transcript and show description

More interviews with Mark Zuckerberg

• Mark Zuckerberg on the Sources Podcast: Muse AI Agent Key Points