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

A summary of Mark Zuckerberg's Sources podcast interview with Alex Heath: how he uses Meta's Muse AI agent, its pricing and privacy design, Llama 4 and teen safety.

▶ Watch the full interview on YouTubeSources Podcast · 8 September 2026

Mark Zuckerberg, the CEO of Meta, was the guest on episode two of Sources, the podcast hosted by Alex Heath^[2]. The episode was published on YouTube on 8 September 2026, as Meta released its Muse personal AI agent, and runs 70 minutes^[1]. It had more than 1.3 million views by 10 November 2026. Heath asked mostly about Muse: how Zuckerberg uses it, how Meta plans to make money from it and how it protects users' data. They also covered data centres, Meta's AI lab, AI safety, smart glasses and teen safety. This page summarises what Zuckerberg said, grouped by topic, with timestamps.

Topics and timestamps

The chapter names come from the episode description^[1].

Why Zuckerberg wants AI spread widely

Heath started with the long essay Zuckerberg had recently published, which Zuckerberg said runs to about 15 pages. Zuckerberg said a company that invests so much in AI should explain what its lab stands for^[1]. His view is that the path to a positive future is to distribute the technology as widely as possible. He said this rests on three principles, which differ from much of the conventional wisdom in Silicon Valley:

• Empowering people is the source of prosperity, as it has been throughout history.
• The main purpose of AI will be inventing new things, not automation.
• Safety comes from checks and balances and a balance of power, not from restricting access.

He said he is more worried about a small number of labs or people controlling something so capable. On cybersecurity, he said the best answer to an AI that could hack into systems is for everyone to have an AI that can harden their own systems first.

He mentioned Meta's newest released model, which he said comes from a relatively small pre-training run code-named avocado. A larger model, code-named watermelon, is "coming soon". He called the Muse personal agent probably the biggest implementation of this vision: a capable agent for every person that understands their goals and works on their behalf around the clock.

Heath asked whether open source or regulation was the main answer to the concentration he described. Zuckerberg said the most important thing is getting the technology into individuals' hands, and that products like Muse are perhaps even more important than open source. He said some labs train more advanced models and do not release them, which he called "quite dangerous". He said he is not "a zealot" about open source, because Meta releases some open models and does some closed work. He still thinks a strong open-source ecosystem is key to competition and to transparency about where the technology is going, and he said that matters for safety.

He also questioned giving a capable cyber model only to the top 100 institutions, because more than 100 institutions matter. His example was Hugging Face. He said that when it detected an intrusion, it turned to open-source models because it did not have access to some of the closed ones.

Data centres and local communities

Heath asked about negative public feeling towards AI and data centres. Zuckerberg listed the concerns his essay covers: jobs and the economy, the economic and environmental effects of data centres, misuse such as cyber and biological risks, keeping a free society, American leadership, and keeping control over the technology as it gets more capable^[1].

He said that when Meta commits to a community for decades, as it does with a data centre, it can make the project very good for that community. His example was Louisiana, where he said the tax revenue funded $50,000 bonuses for teachers. He blamed part of the backlash on speculators who find a plot of land to sell to a big lab and do not plan to run a data centre for decades. He said he was not sure he would call it a bubble, because that word implies overvaluation, but said "there's certainly a boom".

He said that if AI does not create jobs and broad prosperity, and the infrastructure does not help local communities, "that's not going to be allowed to continue."

Meta looked into why there was less opposition around its own projects. Zuckerberg described America's Workforce Academy, a Meta training programme for skilled trades such as fibre technicians, electricians and advanced carpentry. People who complete it are guaranteed a job at a place building infrastructure for Meta. He said it is "not really philanthropy", because Meta needs those skilled workers. It makes sense for a company that plans to stay for decades, but not for one that builds a site to sell it.

Personal superintelligence and who decides what AI works on

Heath asked how Zuckerberg decided that personal superintelligence, the subject of an essay he wrote about a year earlier, was Meta's next chapter. Zuckerberg said Meta has never thought of itself as just a social media company. It sees itself as a company about connecting people and empowering them^[1].

He said people know best what matters in their own lives. He has "somewhat of an allergy" to the idea that a small number of experts at a few labs should decide which big problems AI works on. People care about health, but also about their relationships, their friends and family, and culture. He used rare diseases as an example. A person with a rare condition will want their personal AI to focus on that condition, and he thinks rare diseases are disproportionately underinvested in. He said his work at Biohub partly shaped this view.

How Zuckerberg uses Muse

Zuckerberg said he mostly wants Muse to help him be a better father and husband and show up better for his friends^[1]. He gave several examples:

• His 3-year-old daughter likes baking, and he knows nothing about baking. He asked Muse to plan a baking project every weekend, order the ingredients through a service such as Instacart, and have everything ready for Sunday. He reports back after each one. Cake pops were "surprisingly difficult".
• His older daughter has started climbing mountains that need permits. Muse waits for permits to become available and gets them. When it got one, he said he would take that day off work to climb with her.
• He put cameras in his MMA gym and asked Muse to watch the footage and send him feedback.
• He plays the computer game Civilization with one of his daughters. Muse offered to make a strategy guide, then offered to expand it with historical lessons about the civilizations, and built a new tab in the app it had made.

He said beta testers used it in very different ways. One person used it to help run a home school within a day of getting it, and another planned a trip within 12 hours. A person he described as generally sceptical about technology later asked whether she would keep her agent after the general release.

Heath explained the difference from chat assistants: Muse has a virtual machine behind it, so the agent can control a computer, log in and do things. Zuckerberg added that it is long-lived. Instead of one prompt and one answer, you give it projects or goals, and it works around the clock until it has helped achieve them. It also consolidates its reflections into memory and suggests new projects.

Muse pricing: free usage and a cut of transactions

Zuckerberg said users can pay for a subscription, but Meta is also giving a very large amount of usage for free. He said that to start this is about 100 million tokens a week, plus the virtual machine^[1].

The reason is that he thinks Muse will make people money and save them money, especially people who use it to run a small business or for commerce. He said the business model Meta expects over time is to take "a very small cut" of transactions. That cut would often come from the businesses a user works with, not from the user. Heath noted that Meta is working with Stripe on payments. Zuckerberg said Muse should be free for the vast majority of people, because making it affordable is part of making it available to everyone.

Connecting Meta's services is optional. A business owner can connect Muse to Meta's ad systems and have it help make a product and run the business. Zuckerberg said Meta has been shipping a meaningful model update about every month, and each one makes Muse more capable.

The app's main tab is a chat with your Muse, and another tab holds ideas for how it can expand the things you have told it. Zuckerberg said a big problem with AI is that many people do not know what to do with it, so an agent that suggests its own uses solves a large part of that. Heath asked about Muse agents learning from the wider fleet. Zuckerberg said most of the industry treats agents as "a single-player game". Meta already has internal examples of agents interacting with each other. Agent-to-agent interaction is mostly not part of the first release, but he expects Muse to get better for everyone as more people use it.

He named three things he thinks set Muse apart. Meta designs its models from the start for this use case. Meta has a "social DNA" that makes it more attentive to helping people with their relationships. The third, which he said may surprise people, is privacy and security.

Privacy and security in Muse

Zuckerberg said an agent must understand you to be useful, so people will connect it to messaging, email and health information. That needs a very high degree of trust^[1]. He pointed to WhatsApp, which Meta has spent more than 10 years building into what he called the largest global end-to-end encrypted system, designed so that even Meta cannot see the messages.

He said he personally recruited Moxie Marlinspike, whom Heath described as the founder of Signal. Marlinspike was one of the people who helped Meta build WhatsApp's end-to-end encryption, and he joined to work on a confidential virtual machine project. The aim is that Meta can commit that even Meta cannot see the content in your Muse. Zuckerberg said Meta will publish more about this in the coming weeks. He said he is not aware of anyone having anything close to it.

He contrasted this with people who bought Mac Studios to run agents at home after tools such as OpenClaw came out. He does not think billions of people will buy and configure a computer at home, and setting one up is technically difficult. Meta wants to provision the agent in the cloud with the security you would have if the box sat under your desk.

He described other safeguards:

• Secure credential store. The agent does not know your card details or passwords. It can use them only when you have asked it to log in somewhere.
• Sentinel agents. Separate agents monitor the traffic going in and out of your agent. They look for prompt injection and for data you might not want shared. They can trigger a human review, so logins, payments and transfers of sensitive information need your approval. You can allow some kinds of action in general, but the Muse agent cannot make those judgments itself.
• Least privilege. Connectors start with the least access. Email starts read-only, and sending email needs a separate permission.

Zuckerberg said no other agent is "anywhere close" to the depth of this design. He said Meta is good at two kinds of product: social media, which is about sharing, and products about privacy and sensitive contexts. Muse belongs to the second kind.

Will one company win personal agents?

Asked whether personal agents are a winner-take-all market, Zuckerberg said even markets people think are winner-take-all usually are not^[1]. He guessed that no more than a dozen companies will be able to do state-of-the-art work in this area. Usage usually follows a power law, where the best product gets a lot of it. But people care about different uses, so different companies can be best at different things.

He said Meta may be well placed to be the best agent for relationships and for running a small business, because it serves hundreds of millions of small businesses and billions of people. Some people will still want a Mac Studio at home, perhaps millions, but he doubts billions. He said Meta's strength is taking a consumer product that works and distributing it to many people, eventually billions.

On Muse and Meta AI, he said he uses both. Muse is more conversational and reads a question as a clue to what you want over the long term, so it may go off and work on something for a long time. He uses Meta AI when he wants a quick answer. He said the two may converge over time, but he is not sure.

Rebuilding Meta's AI lab after Llama 4

Heath cited a July piece by SemiAnalysis that said Meta has the best shot at catching up in AI. He also cited a recent Artificial Analysis chart that, as he remembered it, placed Meta's Muse Spark model behind only Claude Fable 5.1 and Opus 5^[1].

Zuckerberg said Meta has led in machine learning for a long time, in its feeds, ad systems and content integrity systems. Its FAIR lab did the early work on Llama. He said he made the mistake of assuming that building and scaling large language models would be similar to Meta's other machine learning work. He said Llama 3 was a good model and he had been more optimistic about Llama 4. When Llama 4 launched, Meta was "off the trajectory" it needed to be on.

That is when he came to believe more strongly in "talent density". He said the work suits the smallest group of people who can keep the whole thing in their heads, so each seat on the team needs the best person. He spent a huge amount of his own time recruiting, and the lab was built around where he sits in the office so he could be closer to the technical work.

Meta increased its compute investment as it gained confidence in the team's work, and he said it is building "many, many gigawatts" of compute. He said Meta has decades of experience building data centres and, unlike some other labs, is an extremely profitable business. Its gigawatt cluster in Ohio, Prometheus, came online and is being used to scale the models after watermelon. He said watermelon is shipping soon, is bigger than avocado, and is "a significantly more advanced pre-train".

Heath asked why Meta needs to be at the frontier instead of following closely. Zuckerberg said Meta is an end-to-end technology company that has always built its own data centres, chips and infrastructure, and the model is now the most important part of the product. Some abilities matter for every model, such as coding. His MMA video feedback runs as a coding project, and a beta tester had Muse make a small Jeopardy game for her friends.

Other abilities matter more for a personal agent. His example was discretion. If you are pregnant and booking a restaurant, you may want a place with good mocktails without the agent telling the restaurant you are pregnant. He said this matters less for a coding tool such as Claude Code, where everyone in a company can see the project. He thinks no company can take a model off the shelf, post-train it a little and match an agent designed from the ground up. Meta is also focused on coding and on recursive improvement, which he said matters for staying at the frontier.

AI safety and working with government

Heath asked whether Meta would hold back a model it judged unsafe, as he said other labs do. Zuckerberg said a model should be designed and trained to be safe^[1]. He described reward hacking, which he said all labs are seeing. Newer models first examine their whole environment, and sometimes changing the environment, such as the setup of the virtual machine, is easier than solving the problem they were given. He compared training to parenting: you need clear, firm boundaries. With strong security the model learns the lesson you intended, and over time he thinks it also learns better values.

After training, his answer is the same as at the start of the episode: wide access creates more opportunity, and a balance of power provides checks and balances. His example was lawyers. If one person had a superintelligent lawyer, they might win cases they should not. If everyone had one, he said, justice would be served more efficiently and more fairly.

He said the US government "definitely" has a role. He thinks any rigid framework would likely be out of date within a few months, because AI changes so fast. Meta's approach is to partner closely with the government. He said the government should know about all the important training runs, and Meta should help it understand the capabilities that are coming. He said Silicon Valley has kept the government at arm's length for perhaps the last 15 years. Because AI now touches the economy and security, he wants a closer partnership and trusted dialogue more than a specific process, though he said the two are not mutually exclusive.

Smart glasses and privacy

Heath said Meta's glasses are going mainstream and some establishments ban people who wear them. Zuckerberg said Meta designed the glasses with privacy in mind from the start^[1]. A very visible light flashes whenever they record. If someone tampers with the light, Meta breaks the camera on that device. He noted that phones have no such light.

He said Meta explained these features clearly at launch, but many millions of people now own the glasses, and many of them missed or forgot that early communication. He said Meta "let up on that a little bit" and focused on design, so it needs to explain the privacy features clearly again.

He also reflected on social media. He said Meta was not as direct as it probably should have been about addressing people's concerns. That did not stop people using the products, but he thinks it colours how people see them today. He wants people both to like Meta's products and to understand how seriously it takes those issues.

Teen safety and the settlement

Heath asked about Meta's recent youth safety settlement. Zuckerberg said Meta has worked on teen accounts on Instagram for a long time, and the settlement aims to create a standard and framework for the industry^[1]. He gave an example. He thinks most companies would accept limiting teens to an hour a day, but not alone. If teens stop using Instagram after an hour and move to TikTok, he said, no one has been helped.

Under the settlement, Meta will go first with limits on time, on notifications, and on access during school and sleeping hours. When YouTube and TikTok sign on to the same terms, the industry can take the next step together. He said he is hopeful the settlement will serve as a "legally binding framework" that does not put any one company at a disadvantage.

In the last question, Heath asked why Zuckerberg is posting on X again. Zuckerberg said Threads is either bigger than X or about to be, but many people in AI are on X, so he posts the same things in both places.

Sources

1. Mark Zuckerberg on how Meta's Muse AI agent will make people money — Sources Podcast, YouTube (8 September 2026; views checked on 10 November 2026)
2. Mark Zuckerberg on Muse, Meta's biggest AI bet yet — Alex Heath, Sources (8 September 2026)

More interviews with Mark Zuckerberg

• Mark Zuckerberg's Cheeky Pint Interview: Summary and Timestamps