Zuckerberg Talks Muse: AI Agent Explosion, "Metaverse, Smart Glasses, and Large Models" Three Major Bets Converge

Author: Long Yue, Wall Street Watch

Three years ago, Meta CEO Mark Zuckerberg mentioned on the Joe Rogan show that one day people would wear glasses, and AI Agents would emerge accordingly. Today, this scene may be unfolding.

Meta's personal AI Agent—Muse—has reached millions of users within two weeks of its launch. Zuckerberg stated in an interview on September 25, "We only encounter this kind of situation every few years." He characterized the early response to Muse as "a home run right out of the gate"—a rarity in Meta's product history.

At the same time, he announced that Muse will be integrated into the entire line of Ray-Ban smart glasses, allowing users to call their personal AI Agent directly with a custom wake word instead of saying "Hey Meta."

The metaverse, smart glasses, and large models, which have long been seen as three independent bets by the outside world, are now accelerating convergence at the product level within Meta.

In this interview, Zuckerberg also admitted that the failure of Llama 4 was "the most terrifying moment" and revealed that Meta is building a 5-gigawatt training cluster, believing that sufficiently large computing power could "brute force" AGI.

Why Muse Can "Hit a Home Run Right Out of the Gate"

The starting point for Muse was when Zuckerberg and team members Nat and Alex sat together and "pieced together" an early version using open-source tools at home.

"We realized this was a magical experience," Zuckerberg said, "If we could make it so that anyone could use it—without needing to buy a Mac Mini or tinker with terminals—then this would be something billions of people would want."

This judgment drove the entire subsequent research and development logic: not just training models, but fully self-developing the "heartbeat" mechanism of the Agent, from the model and scaffolding to the Agent itself—where the Agent would periodically wake up, check user goals, and proactively advance to-do items.

Data after launch confirmed this judgment. Within two weeks, millions of users.

Personal AI: Focus on Execution, Memory, and "Judgment"

Regarding the difference between personal AI and general AI, Zuckerberg summarized the current competitive focus as making the model a better Agent.

"The biggest thing over the past year has basically been programming the Agent," he said. Programming capability is crucial because even if users do not directly write code, "your Muse will be writing code for you in the background to accomplish various tasks."

However, he believes that personal Agents cannot only have coding or task execution capabilities; they also need to understand privacy boundaries and social contexts.

Zuckerberg gave an example where, when a user asks Muse to book a restaurant, the Agent might know about the user's allergies or pregnancy, but that does not mean this information should be automatically disclosed. "You would want it to complete the task while disclosing as little information as possible."

He stated that this capability belongs to the "basic social skills or common sense" that humans typically possess, but many companies have not included it in their model capabilities if they are only training coding Agents.

"We are training the entire model, not just taking someone else's ready-made model and building a framework and Agent around it," Zuckerberg said. Meta has also added features like memory, task tracking, virtual character animations, and real-time voice interaction at the product level.

In terms of personalization, he stated that Meta does not believe AI should have only one fixed personality. "Many labs are focusing on how to tune personality to the right state, but I never believed there is only one correct answer for personality."

He mentioned that Muse allows users to modify their avatar, voice, and basic personality, with the model design goal being high controllability. "You can define how you want to interact with it, which is important for a personal Agent."

Each Agent Has Its Own "Computer"

One of the core infrastructures that differentiate Muse from other AI products is that Meta has equipped each Agent with a secure virtual machine (Secure VM).

Zuckerberg explained why this is necessary:

Your Agent will know a lot of sensitive information about you. We do not believe this information should be mixed in a pool with everyone else's data.

He likened Muse's Secure VM to "the computer that only belongs to you under your desk"—user data is encrypted and stored, and sensitive credentials like passwords are managed through an independent security module, which the Agent itself cannot directly read.

On this basis, Meta has also designed a "Sentinel" security Agent specifically to monitor the data flow in and out of Muse, which will directly intercept and prompt user authorization if it detects any anomalies or high-risk operations.

Metaverse, Smart Glasses, and Agents: Three Routes Are Converging

Zuckerberg revealed in the interview that Muse will be fully integrated into the Ray-Ban smart glasses series and will upgrade existing interaction methods.

Currently, the glasses provide a "single-turn conversation" experience—say something, get a reply, and that's it. After integrating Muse, the glasses become the front-end entry point for the Agent: when users speak, the back-end Muse continues to work in the secure virtual machine, completing tasks and providing feedback.

Regarding the product roadmap for the glasses, he described a clear hardware hierarchy:

• Pure audio glasses without cameras: already equipped with Muse, capable of handling calls, music, and voice tasks
• Meta Ray-Ban with a small display: already released, providing basic visual feedback
• Full field-of-view holographic AR prototype: already released, which Zuckerberg described as "very exciting"

He stated that Meta's long-term investment in glasses positions the company favorably after the AI Agent matures. Meta is bringing Muse and more AI features into the glasses products, while the metaverse technology, which previously emphasized "presence," continues to advance, though more resources are currently being directed towards Muse and the AI functions of smart glasses.

Regarding virtual avatars, Zuckerberg mentioned that previously, Meta needed room-level equipment, multi-angle scanning, and enterprise-level GPU environments to generate high-quality realistic avatars; now, users can complete the setup with just a few photos, and the related capabilities can run on a pair of VR glasses.

Looking ahead to 2030, Zuckerberg stated that the core vision of the metaverse has always been to merge the physical and digital worlds. He gave an example that in the future, people could participate in activities with friends accessing through holographic images; in work scenarios, humans and multiple Agents could also participate in group chats or meetings, with Agents appearing in holographic or other embodied forms.

"The Most Terrifying Moment": The Mistake of Llama 4

Not all bets have gone smoothly. Zuckerberg rarely spoke directly about the failure of Llama 4 in the interview.

After Llama 4, that was the most terrifying moment… I thought we were on the right track, but we weren't. It was quite a significant negative surprise.

He attributed the problem to a fundamental error in team structure:

I organized the team in the same way as the Instagram recommendation system or advertising system—hundreds or thousands of people working in parallel. But training a language model requires a small team that collaborates closely, treating it as a collective scientific project. Every position is extremely valuable.

As a result, Meta underwent a complete reorganization, widely bringing in top talent from the industry and establishing the Meta Super Intelligence Lab (MSL). Zuckerberg stated that a new generation of models is about to be released but will not be announced at the Connect conference.

Computing Power Route: Brute Force AGI, 5-Gigawatt Project Under Construction

When discussing the technological path to AGI, Zuckerberg provided a direct judgment.

I'm not sure what fundamental architectural breakthroughs are still needed… I think we roughly know the formula. If you can build a sufficiently large supercomputer cluster, you can brute force your way there.

Currently, Meta's computing power expansion route: A cluster of over 1 gigawatt in Ohio has basically been put into use for training the next generation of models; a 5-gigawatt cluster in Louisiana is under construction.

He stated, "When you have multi-gigawatt clusters for training, you basically get something close to AGI or even superintelligence."

However, he also added that the current computing power-driven route does not mean that architectural research is unimportant—

The human brain only consumes about 10 watts, while our systems may be a million times less efficient than that. Only by combining large-scale computing power with architectural breakthroughs can we truly lead the way.

Alignment, Not a Burden, Is a Problem That Products Must Solve

Zuckerberg's attitude towards AI safety is very pragmatic—he does not see it as regulatory pressure but as a prerequisite for product success.

If you ask Muse to do something and it does the opposite, who would still use it? We need to ensure that the model not only understands your specific instructions but also understands your intentions and values.

"If Muse is to reach a billion users, we must solve the alignment problem or make significant progress on it," he said.

Regarding safety boundaries during the training process, he used an analogy: "It's like parents setting rules for children—if it 'completes' a programming problem by modifying system configurations, I want to tell it: no, I want you to learn the problem-solving method, not to find a shortcut around it."

The full interview is as follows: >

Big Bets

Host: This is something that billions of people will want to use. What does it feel like for you to build? Is "immortality" a possibility? Well, if you take me into your mind and fast forward to 2030, what will that look like?

This is Mark Zuckerberg. Twenty-two years ago, he built a social network that connected billions of people and forever changed the world. And now, he has decided to build something even grander. To this end, he is making big bets in the fields of the metaverse, smart glasses, and artificial intelligence. For years, skeptics have thought these were three separate bets that could not succeed, but they overlooked the bigger picture—because right now, these bets are converging to create an entirely new superintelligence.

Today, we will present all of this, and I will ask Mark some questions he has never been asked before, listening to his vision for the future, allowing you to get ahead and build the next big thing.

Host: Thank you very much for coming to the show.

Mark: Thank you, I'm glad to be here.

Host: For this conversation, I watched every interview you've done.

Mark: Wow, that's more than I've watched myself.

Host: It was fun and very rewarding. Two things left a deep impression on me: first, your love for "building," I feel you are one of the top builders; second, your ability to bet big—you dare to make huge bets. I feel that this week, all these bets are converging, so let's start from here.

Mark: Okay. We've been doing these things for a long time. In terms of AI, as a company, we've been doing this almost since our inception—the first version of the news feed was in some ways a machine learning product. Then about 15 years ago, we created the AI research lab.

But now we have entered a new phase—about a year ago, we launched the Meta Superintelligence Lab. This is a pretty thorough research reboot, bringing in a lot of great talent from across the industry, which is exciting. Currently, we see the models getting better and better, and the next generation of models is about to be released, but not announced at the Connect conference. In addition, we have the Muse personal assistant, which has received very positive feedback so far.

When building these things, you are not really sure how the results will turn out. We like it ourselves. Earlier this year, I was at home piecing together Open Claw in my own way, feeling how this thing works and how to turn it into a magical experience that anyone can use.

Basically, when we—myself, Nat, and Alex—sat together and realized this was a magical experience, and if we could turn it into a plug-and-play version that ordinary people who don't understand technology, don't want to set up a Mac Mini themselves, don't want to mess with the terminal, and don't want to debug when things go wrong could use, I thought this would be something billions of people would want to use.

Since then, we have been working towards this goal: specifically tuning the models for it, not only building the assistant itself and the operating framework but also building the technology that provides an independent computing environment for each assistant—we built a whole set of Muse secure virtual machines for this purpose.

We internally felt this was special, and the internal team liked it, but you never know how people will react after the product is released. Occasionally, there are cases of a stunning success right out of the gate, but most of the time you receive some positive feedback and need to iterate a few places to really make the product "click." But this time, it clicked right out of the gate. Seeing all this happen is really exciting—just two weeks in, and the number of users has already reached millions, which is quite rare. We only encounter such situations every few years, but this is definitely one of the most enjoyable moments for a startup.

Host: Your ability to stay in the game and keep trying really impresses me. And to have hit so many home runs is truly remarkable. You had an interview with Joe Rogan about three years ago, where you mentioned that one day you could wear glasses, and the AI assistant would appear. So I feel that Muse already has a physical representation, which is very clever because that feeling is bound to happen.

Mark: I think it just makes it seem friendlier and cuter. I believe too many people describe AI as something frightening, but AI should just be useful and fun. That physical representation—a designer made that character early in the project, and for some reason, they always wanted to iterate, but the first version was the best. Later, someone said, "Oh, it has to be blue because it's Meta." I said, "No, I think that representation is right; you nailed it the first time." That's how it is; it's fun.

What is the difference between personal AI and general AI?

Host: Great details. If you want the model to perform exceptionally well in personal matters, rather than just being a general intelligence model, how would the training approach differ?

Mark: I think the core right now is to make the model an excellent "agent." Over the past year, the biggest trend has been programming agents, which contains two core ideas: programming expertise and the general ability to be an excellent agent.

Our strategy is to prioritize making the agent itself good, rather than specializing in programming capabilities.

Programming capability is important because even if users don't think they are writing code, Muse is constantly writing code in the background to accomplish various tasks for you. But we believe that the agent should first be an outstanding agent, and programming capability serves that goal.

Additionally, when building a personal assistant, some things are more important than building enterprise software products. For example, if you are building an enterprise programming tool, the model doesn't need to have any concept of "information disclosure scale"—like what information should be said and what should not. But for Muse, this is very important.

You need to train this capability into the model, just like training any other capability. You will tell it many things and then hope it helps you achieve your goals. For example, if you want it to help you book a restaurant, you are looking for a suitable restaurant, but you might have some allergies, or you might be pregnant, and you may not want to disclose this to the restaurant. But Muse will know this information, and you want it to complete the task while revealing as little information as possible. This is a specific skill, essentially basic social common sense for humans.

But those companies that only make programming agents mostly haven't trained this capability in. We can do this because we are doing full-stack—we are not taking a ready-made model from others and just putting a framework around it; we are training the entire model from scratch, specifically to serve these capabilities.

The model certainly needs to have broad general intelligence, but it also possesses these specific capabilities. Then, based on this, you build the entire assistant and all the details around it: memory, operating framework, and how it "beats"—it periodically wakes up and checks, "Okay, these are the goals I understand about you; is there anything I can push forward now?"

We also have a dedicated team that is solely responsible for refining the real-time animation effects of the virtual representation, because it's not just a default representation—you can customize any representation, and then it moves naturally, which is fantastic. We are also rolling out voice mode, where you can have real-time voice conversations, and your assistant is right there with you. I think all these details stem from us doing full-stack—the model and the product are developed together.

Host: Another big thing is the virtual machine. Can you explain why it is important and what it unlocks?

Why does Meta provide each AI assistant with an independent computer?

Mark: Basically, for an agent to do things for you, it needs a place to store your information. We believe this information should not be mixed in a resource pool with everyone else's information.

You can understand it this way: your assistant will know a lot of sensitive information about you. Many people initially encounter intelligent agents like OpenCloud and will set up a Mac Mini at home themselves. So we thought, many people do not want to buy a Mac Mini or set it up themselves. So what kind of experience can best approximate this effect? The answer is: you just need to download an app, register, and you will get a computer dedicated to you for your assistant to work, store data, and build a secure model around it. This is akin to having your own computer under your desk—even we at Meta cannot access it.

For example, the Muse confidential virtual machine is a feature we are developing, and even we cannot see the contents of your virtual machine.

We have also built a lot around this, such as secure credential storage—when your assistant needs to handle information like passwords, it itself does not need to see those passwords; it just needs to be able to "insert" credentials when you ask it to log into a service, and only do so when you explicitly request it. The system should be designed so that those pieces of information are not freely accessible because accidents can always happen; someone might try to intrude, or the system might have issues.

So you need to ensure that the agent cannot access this data, and neither can Meta. Providing each assistant with an independent computer and maximizing security is the fundamental basis of this technology—empowering Muse with the capabilities needed to help users achieve their goals while ensuring privacy and security, making it a world-class, industry-leading product in this field.

Host: So does this mean that, like WhatsApp, the data on the virtual machine that Muse connects to is encrypted? How should people understand the actual storage of data in the virtual machine?

Mark: We basically built two versions. The Muse secure virtual machine (Secure VM) contains various privacy features, including the entire Sentinel agent architecture we built. You have a regular Muse assistant performing tasks for you, and at the same time, there is a security agent we call Sentinel, specifically monitoring the data flow in and out of your Muse.

If there is external content trying to compromise security, Sentinel will directly cut it off to prevent it from happening. If it believes your Muse is about to take an action that requires your intervention, it will override Muse's operation and trigger an alert for human confirmation—like "Do you want Muse to do this?"

This entire system, combined with secure credential storage and multi-layered defense, constitutes the Muse secure virtual machine.

We are also developing another project. Nat and I specifically recruited Moxie Marlinspike—he is the one who collaborated with us to implement end-to-end encryption for WhatsApp—to design the Muse confidential virtual machine (Confidential VM). The core idea is to give you a dedicated encryption key on top of the secure virtual machine, so that even Meta cannot access the contents within.

This version is more challenging to implement because if Meta cannot access the internals of the virtual machine, debugging and ensuring the system runs smoothly becomes much more difficult. So we spent some time on it, but it will be launched soon. This will essentially meet the security standards that people are already familiar with on WhatsApp and our other most secure products.

Host: Is the advantage of doing this just to make people feel psychologically safer, or are there actual benefits?

Mark: I think security is important in itself. Our goal is to approximate having a local machine under your table. What does that local machine give you? It means no company can access it.

So, assuming Meta wants to provide you with this service, how can we offer you the same level of privacy and security so that no company—whether it's Meta, or anyone trying to invade us, or in some countries where you don't trust the local government—can access it? Because we can't access it either, as we have no access rights.

I think this is very important, and it's one of the key reasons people trust WhatsApp. There is real value in privacy, security, and trust.

If you are going to have an assistant that knows everything about you—I guess almost all of us will have such an assistant—fast forward five years, everyone will have an assistant that understands your goals and everything about you and can help you get things done. In this case, being industry-leading in privacy and security is very important. We wanted to do this from the very beginning.

How to Shape the Personality of AI?

Host: You have an interesting point—you studied psychology in college.

Mark: Well, I was only there for a short time, two years, but I feel it influenced a lot of what I built later.

Host: When we look at models, we often say a certain model is "very smart." But just like when we choose friends, it's definitely because they are smart, but also because we like their energy and how they interact. How do you consider shaping the personality of the model?

Mark: I think the ideal model should have enough adaptability to fit different people's styles. I feel many industry professionals have gone in the wrong direction on this—many other labs focus on "how to design personality correctly," but I never thought of personality as a fixed thing. This is also one of the reasons I strongly believe in open source and in people's ability to customize, and why we designed Muse as a highly personalized product.

You can customize and personalize Muse—not just the image and voice; when you first register, the first thing it asks you is "What do you want my basic personality to be like?" and you can modify it at any time.

We strive to make the model highly steerable, allowing you to define how you want to interact. This is a crucial component of making it an excellent personal assistant—this adaptability around personality is key.

Host: What style is your own Muse?

Mark: I made it direct and efficiency-focused. It's quite interesting. Earlier versions were very sarcastic and humorous, but this version is more straightforward. My assistant uses the default Muse image, but I dressed it in a toga and gave it a voice that's deep to the point of being comical, and interacting with it is fun.

Host: I think to have a sense of humor, you must be really smart. Many people don't realize that comedians are among the smartest people in society—they need to be quick-witted and clever. You definitely have that trait. I've watched all your interviews, and you've always performed well.

Mark's Surprising Predictions About AI and the Metaverse

Host: In your interview with Theo, you talked a lot about the next frontier of technology and where all of this is ultimately heading. The AI field has gone through several "winter" periods where people thought there would be no breakthroughs. The metaverse has also gone through several such periods where people felt it was an unfulfillable bet. I tried out the new holographic feature yesterday, and it's very cool. In the interview with Lex, it felt like it would take 11 hours to record your face, but now it only takes 3 minutes. How did we get to this point today?

Mark: In the overall development of the metaverse, when we founded Reality Labs, we always believed that eventually, there would be normal-looking glasses, and over time, they would provide immersive presence while also being excellent AI devices—because glasses are the only form that can see what you see and hear what you hear, communicating with you all day and ultimately displaying images.

But 10 to 15 years ago, I assumed we would first achieve holographic technology and then have highly developed AI. However, the path of technological evolution is interesting—we actually had AI and personal superintelligence first, and then the technology to make holographic technology sufficiently widespread and affordable came along. This was something I didn't foresee, but I'm glad we're making progress in both directions.

Our significant investment in glasses puts us in a very favorable position as AI assistants become ready. Many of the announcements at the Connect conference were about bringing Muse and a lot of AI features into glasses, which I think users will really love. This is a big deal.

Regarding presence, we are still making progress, but the pace is relatively slow because most of our energy has shifted to building Muse and AI features for glasses. However, we have a long-standing project, which is real-time high-fidelity virtual avatars.

As you said, three or four years ago, you needed an entire scanning room to capture a person from various angles, and you also needed enterprise-grade GPUs to render it, which was very cumbersome. The demonstration we did for Lex's podcast was that kind of setup. And now we basically have it running in a pair of VR glasses—this is the first form of glasses that can achieve such an amazing VR experience, rather than a bulky headset. You can create your virtual avatar with just a few photos, and the progress is astonishing.

Host: And it can also drive expressions based on voice. In the demonstration, it made me laugh, made me interact, and then understood how my face moved with the audio track. You mentioned in that podcast that some people who are usually more reserved with their expressions actually want richer expressions in the virtual world. What do you think about people distinguishing between their "virtual self" and "real self"?

Mark: I think we are still in the early stages of understanding this sociologically and psychologically. I feel that people's self-perception and the image they want to project often differ somewhat from their actual selves. Since the inception of social networks, people have been carefully selecting their avatars. We see similar phenomena with Muse's virtual images. It's not so much about "curating" oneself, but about "curating" the "person" you want to communicate with.

I believe that when you provide people with the ability to express themselves, you want it to genuinely capture them, while also being a form of expression in itself, rather than just a pure mirror reflection—it is both communication and expression. We hope to build something that balances both. This is always an iterative cycle: seeing how people use it and then improving it. After years of work, we are now truly at the starting line—this is the first time we can genuinely incorporate relatively high-quality realistic avatars into products, usable on phones and in VR. I am very excited to see the results.

What Will 2030 Look Like?

Host: Okay, take me into your thinking, fast forward to 2030, if everything goes well, what will holographic technology look like? Holograms combined with Muse, along with glasses, how will they come together?

Mark: My understanding of the metaverse vision has always been about effectively merging the physical world with the digital world. The basic idea is: we have this beautiful physical world, and we also have an amazing digital world—the vast content accumulated on the internet over the past 20 to 30 years is breathtaking. But the way we access it is either sitting at a desk or through a small screen in our pocket, which is fundamentally very limited.

I think the ideal version of this is a seamless integration of the physical and digital worlds. You can think of it this way: right now, the two of us are here. In some future version, one of us might be a holographic projection, but you still feel that genuine sense of presence with each other, which is completely different from a video call.

The core of virtual reality is to convey this sense of presence—making you truly feel like you are in the same room with others or in another place. You can achieve this with holographic technology and mix it in various ways. For example, I could play poker with friends, with some people present and others joining through holographic images, and they could still play cards; the poker table itself could also be holographic, allowing those not present to be integrated.

At the same time, AI can also be embodied and appear in this scenario. This makes a lot of sense in work contexts—I am currently using various programming agents to build things. Imagine this: you have a group chat channel with several people and several agents, and you assign tasks to the agents. But sometimes, everyone gathers for a meeting, and the agents should also be present. How do they appear? It's simple; just add a few more seats on the couch, and they can appear as holographic images. Or use that cute little character from Muse, or a dragon, or any quirky image you create.

I guess this will feel quite natural in the future.

Host: Interesting. You mentioned in previous interviews that the tech industry often forgets about the "fun" aspect. I think having a physical assistant there would also make it feel more real—it’s like really outsourcing work. When you see Muse typing, it feels like something is really happening. This is done well—you can see what’s happening in the browser. So do you think there’s a possibility of a scenario where you wear glasses and control your computer, letting Muse help you do things on the computer?

Mark: Oh, definitely. VR can already do that. You can sit anywhere, even in a café, and then open your workstation with six monitors to write code; everything is there.

On the glasses side, the most popular model currently doesn't have a display, which makes it more affordable and allows more people to use it, while we are still working to make displays into the most compact form. But we have released a display version of Meta Ray-Ban, which is very popular; it's a small display. We also released a prototype version of full-field holographic AR, which I think will be very exciting.

So the entire product line looks like this: from pure audio glasses—no cameras, looking just like regular glasses, but inside there’s Muse, which can use various audio tools, listen to music, make calls—to higher-end versions, all of it.

Host: I’m wondering, with those audio glasses, can you talk to Muse while letting your home computer do work?

Mark: Yes, absolutely. We just launched this feature at the Connect conference. Now all glasses are connected to Meta AI, creating a "single-turn" experience------you send a prompt, it replies, and that's it. But with Muse, we basically want to upgrade all glasses to Muse. First, you no longer need to say "Hey Meta," you can name it anything you like, which is part of the fun. Then you just talk to it directly, it connects to your Muse, and your Muse processes tasks in your secure virtual machine to help you get things done.

Host: That's amazing!

Founder Mindset

Host: Alright, we are here now, Muse is progressing smoothly, and the glasses are doing well, but about a year ago, many people were asking "What happened to the Super Intelligence Lab?" At that moment, how did you feel inside? What was it like when things weren't going well, but you still saw the long-term vision?

Mark: The real problem was with the Llama project and Llama 4.

Llama 1 was quite an interesting model; it pioneered the entire open-source AI movement, and we are very proud of it. Llama 2 achieved scalability, Llama 3 was a great model that was almost at the cutting edge at the time. Then with Llama 4, we basically deviated from the development trajectory we should have followed.

Whenever things don't go in the direction I expect, I spend a lot of time thinking: Why is this happening? What do we need to change to do better? This time, my reflection was: I got the entire team structure wrong.

I modeled it like the way we do machine learning work for Instagram feeds or ad systems------with hundreds or even thousands of people working in parallel on many things. But for building language models, what you really need is an extremely tight-knit small team, treating it as a collective scientific project. You don't need many people, but this means that every position in the team is extremely valuable.

So we recruited the best people from all over Meta and brought in many talented individuals from the industry to form a brand new team------the Meta Super Intelligence Lab.

From my perspective, when MSL launched, I knew it would take time to restart, rebuild the infrastructure, and train the next generation of models. But I knew we had assembled an excellent team, and if the team could work well together, the results would be good.

For me, the most thrilling moment was actually after the release of Llama 4------I thought we were on the right track, but then I found out we weren't. That was quite a significant negative surprise. I think, as an entrepreneur, you are always tested in these moments because inevitably, not everything will go smoothly. And what truly determines the trajectory of development is: when things don't go in the direction you hope, how do you find a way forward.**

Host: I guess it's the opposite as well------when something exceeds your expectations, like the launch of Muse, how do you ensure you seize that opportunity?

Mark: Absolutely. Now the entire company is invested------initially, it was just a small team building the product, but now everyone realizes that it is really ready to take the big stage. The whole company is thinking about how to scale this thing up and how to let hundreds of millions of people experience it.

From optimizing all the infrastructure to make everything run smoothly, squeezing every bit of computing power from existing GPUs, to various product teams embedding Muse in different ways------like in glasses. Seeing everyone work together to ensure Muse can scale smoothly is a fantastic feeling.

How Mark Writes and Communicates Vision

Host: As a founder, I feel that this is the moment you long for most------everything comes together. How do you communicate your vision to the company? With so many things happening simultaneously, I feel you write a lot. What is your process?

Mark: Writing is very helpful for me; it helps me refine my thoughts and communicate externally. This summer, I wrote a long article called "The Future Belongs to Everyone," about 15 pages long, which helped me systematically clarify my philosophical stance on various important social issues related to AI: what I think is good, how interactions with the government should be, how to prevent various concerns people have, how to make data centers a wealth for the community, truly creating jobs instead of destroying them, how to maintain national security, and how to effectively alleviate the risks people are worried about------whether it's hacking or biosecurity issues.

This is a very complex matter, taking a long time, involving many discussions and rounds of revisions with many people internally. But for me, it was a very valuable process. In the end, we had something like this------15 pages, "This is what we believe." Then I distilled it into a one-page version to publish as a column. We also made a short video because I think to reach many people, you often don't need to throw out a theoretical argument, but rather condense it into: what are your values, what do you believe, and how do you communicate that.

There is no one-size-fits-all way to communicate these aspects to the company or the world. Different times require different approaches, and different groups of people may have varying levels of alignment with what you are doing; some naturally resonate with it, while others may be more concerned and need you to guide them, requiring extra effort to explain why it is valuable. I think this is part of running a company and being a founder------you are not just repeating the same thing over and over; each situation is slightly different, facing new challenges, which is part of what makes it interesting.

What Drives Mark to Build?

Host: I feel that for you, this founder mindset extends beyond the company------whether it's your farm or learning a new skill.

Mark: I just love to build things.

Host: What does "building" feel like for you?

Mark: I think it's an intrinsic need. Different people have different ways of self-expression. If you are a writer, you feel you must write. Some people have a need for recognition. But I just need to build things. If I'm not exercising my creativity or building something, I become irritable. That’s not pleasant for those around me.

Host: Is learning to be a great skier the same skill as learning to build a product?

Mark: To some extent, yes, the process of learning new things is quite similar. In my life, I have deliberately challenged myself with things I am not very good at. For example, I have always struggled with learning languages. This is actually why I started learning Latin------I just couldn't learn French and Spanish in class, so I thought, well, Latin doesn't require speaking, just translating, like math.

Later, when I started running a company, I set annual challenges for myself, one of which was to learn Mandarin. Mandarin is really hard, especially the tones. Of course, there are good reasons to learn it------Priscilla's grandmother only speaks Mandarin, so if I want to communicate with her, I need to learn it. But the biggest motivation is actually the challenge itself.

All these things, you can only do them; there are no shortcuts to "figuring it out." You just have to invest time, and then it slowly seeps into your brain. Learning martial arts, learning to fly a helicopter, is the same------these things are actually hard to "understand" rationally; they require practical accumulation.

Building products is partly like that. Programming can be thought of theoretically, but the intuition for building products can only be cultivated through repeated practical experience. The only question is what you like------because I believe not everyone has the same strong need to build as I do; most people have some degree of need, and the key is to find out what that is and then invest time into it to excel in what you truly want to do.

Honestly, I think it's not just a "want," but a deep psychological need or drive. Aligning this drive with something and giving yourself time for those experiences to slowly seep and settle is crucial. This is also what I try to teach my children------to help them find what they are truly interested in, but if they hit a bottleneck, to give them a push.

One of my daughters loves creating music, but she just can't stand piano lessons. I told her, "You don't need to become a piano master, but if you want to create music, you need to have an intuitive grasp of music theory and how it works. And once you learn the piano, you can easily pick up the guitar." I think, whether as a child needing a parent to give a push, or as an adult, having the discipline to sit down and let these things slowly settle in the brain is key.

The Golden Age of Builders

Host: I feel that actually everyone wants to build. I don't think Generation Z is as criticized for low agency as the outside world claims. I think people are just looking for the spark that ignites their building power. You talked a lot about this in your speech at Harvard.

Mark: Yes. When I say "building," I mainly refer to products like------software, hardware, etc. But I agree, I think everyone has some creative drive. However, some people have other stronger personality traits that overshadow this, like service-driven------those who become doctors or nurses, their core drive is "I just want to take care of others."

I heard a story, possibly from Priscilla's experience in medical school: On the first day of class, someone stood up and asked, "How many of you have memories from childhood of seeing someone and thinking, 'I really want to take care of that person'?" As a result, everyone raised their hands. For me, my version is: I have many childhood memories of "I want to make this thing better."

Not everyone has the same drive, but everyone has something they want to do. I agree that with previous technologies, starting out has been difficult for many people. This is also what excites me most about personal superintelligence, Muse, and various AI assistants------I think for the first time in history, people can really get started quickly. You have a vague idea, AI can help you outline it, and then you can refine it, like sculpting a sculpture, without needing to understand everything before you start.

I think this is very powerful, and many people will find what they want to create or push forward because of this. This will help people feel a broader sense of agency.

How Far Are We from Conquering All Diseases?

Host: Another project you have outside of Meta is conquering all diseases. I'm curious, first, how close are we to that goal?

Mark: Much closer than before.

Host: Realistically, how close do you think we are?

Mark: When we initially launched this project, the goal was to help the scientific community conquer all diseases by the end of this century. We were never going to do it ourselves; our theory is: All major scientific advancements are preceded by new tools that can measure and understand certain things. For example, the invention of the microscope led us to understand bacteria; the telescope helped us understand the universe.

Some of these tools have even become platforms------for example, the first person to invent a vaccine allowed people to use that method to treat many diseases.

However, historically, the way scientific funding has been distributed has been widely decentralized, allowing individuals to explore independently, and there hasn't been much funding truly dedicated to building these large tools. What we are trying to do at Biohub is to design several new tools that enable people to "see" biology in new ways, thereby helping the scientific community accelerate progress.

Initially, we thought that "by the end of this century" was a very achievable goal, and many biologists at the time felt it was impossible. But now I feel that the end of this century is too far away.

The progress of AI, combined with the virtual cell models we are working on—the basic idea is to conduct experiments not on real living cells, but to use AI models to simulate proteins, then simulate cells, and further simulate a virtual immune system, or an entire organism, even a whole human. This will allow scientists to run a large number of experiments, simulating what would happen under different circumstances, such as what reactions would occur in someone's body if they took a certain medication.

I don't want to give a precise year, but I guess it will be much earlier than the end of this century.

Host: The goal of "conquering all diseases" feels very deliberate, rather than "immortality." Is immortality a possibility?

Mark: This is not really my area of expertise. I believe the two are different things. "Immortality" is more about extending life—even if you don't get sick, the human body has a natural life expectancy, which is another issue that needs to be studied separately. Some people might consider this a kind of "disease," which is a reasonable perspective, but I think people also need to work on conquering those diseases that will still make you sick even after you've solved the lifespan issue.

The part of the problem we chose does not mean you will never catch a cold. Our statement is "cure, prevent, or manage": some diseases can be completely cured; some I believe we will be able to prevent in the future; and some, perhaps you will still get sick, but what would have caused real harm or even taken your life, you can now manage as a chronic condition that does not significantly impact your quality of life.

The goal is to keep the human body in a balanced state—not that we will never come into contact with pathogens, but that one day, we will be able to cure, prevent, and manage all diseases.

What is the most frequent thought in Mark's mind?

Host: AI has acted as a catalyst in many different breakthroughs, including personal intelligence. What do you think about most right now? What thoughts come to your mind most during this process?

Mark: This may change every week.

Right now, I am very focused on Muse. Every time we release something, we push forward quickly, then engage with the real world, understand people's feedback, and figure out what to do next—this phase is always exciting. We are currently in this phase, collecting a lot of information to understand what people want. The good news is that people really like it, so we are working hard to get as many people to use it as possible.

There is still a lot of work to be done on the model, the Muse Spark model has made great progress, and we hope to continue pushing forward to create a world-leading model.

What breakthroughs are needed next?

Host: What breakthroughs are still needed?

Mark: In researching this, you may not be able to foresee things in advance. We have some intuitions about certain directions, and much of the work over the past year has actually been about expanding infrastructure.

About a year and a half ago, more people would say that achieving superintelligence requires certain fundamental architectural breakthroughs. But I am not so sure that is the case now. I believe we have a considerable understanding of the "formula"—if you can build a sufficiently large supercomputer cluster, you can reach there through brute force computation.

We are building that cluster in Ohio, which exceeds 1 terawatt, and it is basically online; we are using it to train the next generation of models. Then there is the 5 terawatt cluster in Louisiana, which will be operational soon. I estimate that when you have clusters in the multi-terawatt range doing training, you can basically get close to AGI, and even superintelligence beyond AGI.

But that doesn't mean it's the best way. The human brain only needs 10 watts to operate, while the computing systems we are building are about a million times less efficient than the human brain. So I do believe there is room for improvement in architecture, and we are also exploring—if you combine massive computing power with architectural improvements, you can really build something that leads the world. But for now, scaling itself can take us a long way.

Host: Indeed, scaling is the most reliable path. In research, you have to hope for major breakthroughs, but scaling can yield results quickly.

Mark: That's why large companies choose this path. There may be a much cheaper way, but we don't know it yet. And this is something so valuable to the world that even if it costs hundreds of billions of dollars, as long as the probability of success is high enough, it is still worth doing—ensuring that even if a cheaper breakthrough is not found in the end, you still have a clear path to achieve it. Of course, if you find a cheaper way, that would be even better.

So I don't want to say this problem is "solved," because at every stage of scaling infrastructure, you will encounter various new engineering challenges that need to be debugged and resolved. Research itself is advanced in this iterative manner.

How to Win the AI Safety Battle

Host: One of the most important things right now is to focus on alignment, making models trustworthy.

Mark: Yes. There is a debate in the industry about alignment: will these AI labs naturally do it? My view is: of course they will. If you tell Muse to do one thing and it does the exact opposite, I don't know how many people would still want to use it. We must ensure that the model not only understands what you specifically asked but also understands your intentions and values, so it won't do things in a way that you are not satisfied with or produce negative effects that you absolutely do not want. This deep understanding is essentially alignment.

So my view is that to make Muse successful and reach billions of people, we must make real progress in alignment, not just talk about it.

Host: Is alignment achieved through users saying "that's not what I want," or through your backend discovering it?

Mark: Both. User feedback is there, but I think it is becoming increasingly clear that you need to do alignment during the training phase, not just after deployment. Now the models are smart enough that if you don't do it well during the training phase, most of the safety issues and accidents we see in other labs occur during the training process, not after deployment to users.

You need to set up a very good "curriculum" for it, just like parents teach children, setting clear boundaries. If it does something wrong, it needs to learn "no, you should really solve this programming problem, not bypass it by modifying some system configuration to get a reward"—I want you to learn this method through the actual problem-solving process, and that is the meaning of training.

This is largely about establishing good safety mechanisms and clear boundaries. These are all natural tasks that the industry must undertake, and we are investing a lot of time; the situation changes every day.

Host: I remember you mentioned that the issue of AI safety—seems to have suddenly become a focus in the last two weeks, but actually, there wasn't a single event that triggered this timing. It sounds like you mean we actually spent a few extra months training internally.

Mark: For Muse, we know that this product needs to pay very close attention to privacy and safety. We have an early version of the model that we believe can further train some behaviors regarding "information disclosure scale"—just like we talked about before. So we took the time to do that. We also spent time making the virtual machine safer. This took a few extra months.

I don't quite agree with some other labs' statements—"we are suffering greatly to slow down." For me, that is the right thing for Meta and Muse users. We want to make the product good; if the product is not good, it is bad for users and bad for us—we don't want to put something out there that leaves a bad first impression, and then people lose interest in using it.

I believe all labs have a very strong intrinsic motivation to get this right. If you align the incentive mechanisms with the goal of "extremely valuable and safe," that is the key.

Host: Thank you very much for taking the time during this important week.

Mark: Thank you, thank you.