Alexandr Wang on Muse's Japan Launch and Meta's Bet on Personal Superintelligence
TBS CROSS DIG with Bloomberg
Muse, Meta's personal AI agent, launched in the United States and Canada and has now arrived in Japan. In this interview, TBS CROSS DIG business editor Masahiro Nakagawa asks Alexandr Wang, Meta's Chief AI Officer, why Japan came next and how Meta rebuilt its AI effort after Llama 4. They also discuss what it takes to make an agent useful in daily life, how Meta approaches agent security, and how it plans to pay for the compute. Wang's central argument is that much of the AI industry has concentrated on coding agents and enterprise customers. That focus, in his view, left Meta an open lane: building a capable, trustworthy personal agent for ordinary consumers, at the scale of billions of people.
Early reaction in North America
Muse launched in the US and Canada a little over a month before the interview. Wang described the reaction as "beyond many of our original expectations." He said the most rewarding thing for the team has been how widely the product was adopted and how easily people "from all walks of life" have used it to save time and money.
He also said Meta's emphasis on safety, privacy and security had resonated with users. The company spent "many, many months" on these areas before launching in the US. In his telling, Muse has become something of a cultural phenomenon in the US, and its mascot, Jolly, has turned into an online meme. People have connected with the product, he said, more than the team ever expected.
Why Japan, and why a non-English market so early
Wang said Meta wants to make Muse steadily more accessible around the world. It prioritized countries where the team felt Muse would resonate with the population and where the regulatory framework is friendly to launching the product. Japan, he said, "hit many of those boxes." He added, while saying it was not the deciding factor, that he is a big fan of Japan and has visited nearly every year for many years.
Nakagawa pointed out that Japan is not an English-speaking country, unlike the US and Canada, so it seemed an unusual choice for the first expansion. Wang answered that Meta's mission is to bring personal superintelligence to billions of people regardless of what language they speak. The company has therefore invested in its models' multilingual capabilities for many years and has published much of that research. He presented that long-term investment as the reason Meta could release the product quickly in a market like Japan.
Rebuilding Meta's AI effort after Llama 4
Asked about the history of Meta Superintelligence Labs, which he leads, Wang went back to his arrival in the summer of 2025. He joined after Meta had released Llama 4, which, he noted, the company had publicly acknowledged was "not on the trajectory that the company needed."
His first priority was to lay a foundation. He wanted a strong culture of scientific discovery that could push the company to build frontier models. At the same time, he wanted a culture of product experimentation, so the team could explore what new AI products might make possible and ship them to users.
To do this, Meta ran a large effort to build a talent-dense research environment. It drew on many of the most talented people already inside Meta and ran an extensive recruiting process to bring in people from outside. Wang emphasized that the lab was built around strong scientific foundations and technical decision-making, and that the core team "has remained quite small even to this day." He said this team built the foundations of a new model series "nearly from the ground up." The series began with Muse Spark, and the most recent release, Muse Wonder, he called "a really big moment."
The product work ran in parallel. Since the start of the year, Wang said, he, Nat Friedman and Mark Zuckerberg had all been excited about personal agents built on strong agentic models. Getting the product right took a long time. The team iterated on the model for many months, and it also iterated on the product and on the privacy and security architecture needed to support it. Meta delayed the launch by many months to get those pieces in place before Muse finally shipped.
The gap Meta saw in the competition
Nakagawa asked how Wang viewed the gap between Meta and labs like OpenAI and Anthropic when he started. Wang first said he has a lot of respect for both companies, describing them as highly innovative and as having delivered impressive models.
Beyond the obvious need to build research foundations for frontier models, he said Meta saw a gap in the market. The question was how to bring powerful agentic AI into a product and form factor that resonates with all consumers. According to Wang, that gap formed because so much of the industry, including many leading labs, was focused on coding agents, the B2B market and enterprise. That left open the lane of "building the best personal agents for everyone," and he described Muse as Meta seizing that opportunity.
He believes the reaction so far has validated the bet. In his account, many users are saving real time and money, and the product is making a tangible difference in their lives. He gave examples such as letting people spend more time with loved ones, or helping them do things for their loved ones.
From founder to executive at a large company
Wang founded Scale AI at 19, while at MIT, and moved to Meta after a major deal. Asked what that transition was like personally, he pointed first to the vision. Around the time he started, Zuckerberg wrote a memo on personal superintelligence. The two had already discussed that idea before Wang joined: building AI that could reach billions of people and make everyone's life better.
He was candid that joining a very large company "as kind of the new guy" with a big mandate is complicated. In some ways it was very different from running his own company, and in others it was similar. Much of the work was a blank slate. A lot had to be defined and written from scratch, and he and Zuckerberg both did work that was "very entrepreneurial." The new part for him was doing this inside a large organization and working cross-functionally with many partners, which he described as a great experience but an unfamiliar one.
How the Muse models were built
Muse Spark and Muse Wonder are the models behind the Muse agents. Wang said a "huge percentage" of the team's overall effort went into building the right foundations. That meant pre-training, careful scaling, and getting reinforcement learning and post-training right so the models' intelligence could keep scaling. Much of the work centered on the science of scaling and on data.
The team also thought carefully about what a personal agent product specifically requires. Wang named multimodal capabilities as one critical area. He named memory and the model's ability to focus as others, which he said were highlights of the Muse Wonder family under the names "active memory" and "active focus."
The process, as he described it, was systematic. The team wrote down everything that would make a personal agent like Muse excellent, and the list grew to hundreds of areas of investment. They then built evaluations and measurement methods so that every model coming out of the lab could be scored against those areas, and they kept iterating to improve quality. He called it "a real labor of love" for the whole team.
Why Meta built its own benchmark
Nakagawa asked about the internal benchmark Meta developed, which Wang called the personal superintelligence index. Wang said the team found that many of the things that matter for a personal agent were not captured by existing benchmarks. Most benchmarks companies report on focus on coding, math or science. Those capabilities are valuable, he said, but they are not necessarily what matters when building the best personal agent. So Meta developed many separate axes of behavior and combined them into the index.
He gave several examples of what those axes cover:
- Memory and focus, as mentioned earlier.
- Chaining multiple tool calls, meaning the ability to combine many tools into one continuous process.
- Personality and how the agent communicates, which he said makes a large difference to the user experience.
- Discretion: the model had to be trained to be thoughtful about when to share certain information, and about when to reach out to the user versus when not to.
- Fault tolerance: how the model responds when something fails, whether externally or internally.
- Browser use: Muse's ability to operate a web browser. Users can watch it work and take over if needed. Wang said this was "so important for really unlocking the abilities of the model."
He said the list "goes on and on." The index compresses all of these into a single number, which he said required a lot of careful effort.
Security in a world full of agents
Nakagawa raised a series of AI-enabled cyberattacks, including reports of US banks detecting unauthorized AI agents in customer accounts, and asked how a world full of agents can be made safer. Wang said security was a major focus for Muse. The team did extensive testing, and it delayed the release by many months to get this right.
He described three layers in Muse's architecture. The first is the Muse Secure VM, a secure sandbox environment in which the agent runs. The second is a separate "sentinel agent" that observes the actions the main agent wants to take. The sentinel must approve those actions, and any sharing of information, before they happen. The third is human-in-the-loop verification: many of the agent's actions require a person's approval before the agent carries them out.
Wang said Meta believes these designs are important paradigms for scaling consumer agents to "billions and billions of people." Agents are powerful and exciting, he said, but companies have to invest in solutions that let them give users a trusted experience.
Paying for the compute
Nakagawa noted that the service is initially free of charge. He asked how Meta will cover the costs and whether its compute investments, including projects such as Prometheus and Hyperion, can keep up with Muse's growth.
Wang first described the business opportunities he sees in Muse itself. One is subscriptions: many users have signed up for a monthly plan or bought additional tokens as needed. Another is B2B. He said Meta has seen adoption and inbound interest from small businesses, and from enterprises that want to use Muse internally. The third is commerce. He described "an explosion in shopping," with people trusting their agents to do more and more of their buying.
Beyond Muse, he said Meta has many ways to put its compute to economically efficient use. He called the most important of these "maybe the most boring": using compute for the core business, meaning recommendation and ad systems, where he said Meta sees "amazing return on investment." He also mentioned broader consumer use through scaling up both Muse and Meta AI. Finally, he pointed to the recently launched Meta Enterprise platform, which CJ Desai was brought in to help lead, as a future route to monetizing Meta's technology with enterprises.
On whether compute can keep pace, Wang answered "for sure." He referred to remarks Zuckerberg has made and said there is substantial room for efficiency improvements within Muse. In his view, those gains would let Meta serve more users on the same level of compute while expanding to more countries.
What personal superintelligence means to Wang
For the final question, Nakagawa recalled that Wang had once said he was waiting to have children until brain-computer interfaces like Neuralink became available. He asked whether that is the future Wang now sees for personal superintelligence. Wang did not take up the brain-computer interface part of the question. Instead he described what excites him most.
What he looks forward to is being able to give a personal agent vague goals and vague dreams, and having it start and carry through the process of helping you achieve them. His own example was making a movie. He said he has "all sorts of crazy ideas for movies," but he has a day job and no idea how to get started. He expects that in the future the best personal superintelligence, "whether that's Muse or something else," will let him say he wants to make a movie, offer some rough ideas, and then work alongside him to make it happen. The goal, he said, is that people look back and think they might not have managed it without Muse.
Asked whether Meta can actually deliver this, Wang said this is where Meta wants Muse to go and that he thinks "the path is very clear." He framed the work ahead as continuing to improve the models and the product and to give users better tools. As the interview closed, he mentioned that he was using Muse at that moment to help organize an upcoming trip to Japan.
One-on-one.
Agents.
Agent.
Alexandr.
Superintelligence.
You can.
Agree.
Superintelligence. Personal supery.
Okay, Alex, thanks for your time today. It's a huge honor to have you on the show. Thank you very much.
Of course. Excited to be on the show.
Yeah. So, Muse started in the US and Canada over a month ago. So, I was wondering how you think about the reaction so far, and now you brought it to Japan. So, why did you pick Japan after North America?
Yeah. Well, first, the reaction has been absolutely incredible. I think it's been beyond many of our original expectations and what we thought was possible. Frankly, I think there's many things that have been very exciting about it. I think the first was just seeing how widely it's been adopted and how easy it has been for people from all walks of life to be able to use Muse and use it to actually save time and save money. That has been just one of the most rewarding things for the whole team.
Also, it's great to see, one of our big focuses in releasing Muse has been our focus on safety and privacy and security. We invested many, many months in the development process to get those things right before we ultimately launched in the United States. It's been great to see that resonate with so many of the users and so many of the people who have been adopting Muse.
And then in the US, it's been very exciting because it's become a bit of a cultural phenomenon. We've seen our mascot Jolly become a meme online. So we've seen just really exciting adoption, and it's something that's clearly struck a chord with people. People have connected with it more than I think we ever would have expected.
And we're very excited to be launching in Japan. One of our focuses is to continue making Muse more and more accessible globally. We wanted to really prioritize countries and places where we felt that Muse would really resonate with the populations of those countries, and that have had friendly regulatory frameworks for us to be able to launch our products in. So Japan hit many of those boxes. And not that this ended up making the biggest difference, but I am also just such a big fan of Japan. I have visited nearly every year for the past many, many years, and it's just an incredible country.
But Japan is a non-English country, unlike the US and Canada. So isn't it hard to choose the country whose people are not speaking English, because default language is English? So why?
Yeah. Well, because we've known for a long time that our mission as a company is to bring personal superintelligence to billions of people all around the world. We've known for many, many years now at Meta that our goal is to deliver AI that can be used by billions of people regardless of what language they speak. So we've made a lot of investments over many years into the multilingual capabilities of our models and have actually published a lot of this research on our ability to make these models very useful and multilingual. And so this has been a long-term investment of ours, which enables us to release the product very quickly in countries like Japan.
Right. So I was curious about the journey of Meta Superintelligence Labs, which you are leading. So from day one, all the way up to now, the launch of Muse agents or a bunch of Muse models, tell us about the journey.
It's been so exciting, but it's also been a real journey. I came into the company last summer, the summer of 2025, and it was after the company had released the Llama 4 model, which we publicly talked about as not being on the trajectory that the company needed. And so I came in, and immediately one of the most important things was really setting up the foundation and developing the best culture of scientific discovery to be able to propel the company to develop frontier models and develop the most advanced AI capabilities, while also developing a strong culture that enabled product experimentation and enabled us to really explore what is the realm of the possible with new AI products, and really be at the front end of exploring that to be able to deliver those to our users.
So I came in, we had this very large-scale effort to bring in the very best talent into a very talent-dense research environment. This included many of the most talented folks from within Meta, as well as a very extensive recruiting process to bring in the best talent from outside of Meta. We established a great research lab with an incredibly strong focus on a culture of scientific discovery, strong scientific foundations and technical decision-making. That team has remained quite small even to this day. And through that team, we've been able to, over time, build nearly from the ground up the foundations of this incredible model series, which started with Muse Spark. And obviously, most recently, we released Muse Wonder, which has been a really big moment.
In parallel with that, we've also been iterating on the product side. Since the start of the year, we've been extremely excited about personal agents. Many of us, Nat Friedman, myself, Mark, we all really saw and were excited about the opportunity and potential with personal agents with strong agentic models. But it took a lot of time for us to get that product right. We iterated on the model for many, many months. We iterated on the product, the privacy and security architecture to be able to support releasing this model, and delayed the launch for many months to be able to support that. And then ultimately this culminated in this incredible moment of launching Muse, which has really resonated.
Right. So when you started, what were your thoughts on the competition with AI labs like OpenAI or Anthropic? And how did you see the gap and what you needed to fill in the gap? And so how did you overhaul Meta's AI efforts like Llama?
Yeah. Well, first of all, I have a lot of respect for those companies. I think they're incredibly innovative companies that have delivered incredible models, both of them. But one of the things that we saw, in addition to obviously needing to develop the research foundations to be able to innovate at the frontier and develop the most powerful models, we saw a real gap in the market, which was focusing on the consumer and focusing on how do you bring powerful agentic AI experiences into a product and form factor that can really resonate with all consumers.
And that was a real gap that we saw that was forming because so much of the AI industry, including many of the leading labs, were so focused on coding agents and so focused on the B2B market and enterprise that it really left open this lane for building the best personal agents for everyone, for every single person to use. And I think that really was the opportunity that we seized with Muse. We've obviously been very excited and fortunate to see all the reaction, and I think it's really been validated that this powerful AI can make a really big difference for people. Lots of our users are saving real time, real money. It's making a very tangible difference in their lives, whether it's enabling them to spend more time with their loved ones or do amazing things for their loved ones. So we've just been really excited about all of that.
You've been a founder since 19 years old. When you were in MIT, you founded Scale AI, and the huge deal came up and you jumped into a huge global company like Meta. What was it like personally for you?
Yeah. Well, I think first and foremost, the thing that was exciting at Meta was the potential and the vision, right? Mark wrote this memo right around when I started around personal superintelligence, and that's what we had talked about before I joined Meta as the really exciting opportunity, which is how do we build AI that can be delivered to billions of people in the world and really make everyone's life better. And I think that was an incredibly exciting mission to work on. That was an incredibly exciting thing to work towards.
But also, I was joining Meta, which is a very large company, and there's a lot of complexity in being able to join a large company, especially as kind of the new guy with a really, really big mandate and a big thing to accomplish. And so in some ways it's very different, in some ways it's also the same as starting my own company. It was in very many ways a blank slate, and there was a lot that had to be defined and written, and stuff that Mark and I both had to do that was very entrepreneurial. And that was very exciting for me. And then I also had to learn how to do that in the context of a large organization and a larger company, and work more cross-functionally with many of my partners, which has been a great experience, but also a new experience for me.
And Muse Spark and Muse Wonder are the foundation behind the Muse agents. So what was the process like to build those models to make personal agents like Muse work in real life? Or tell us specifically about what use cases, what kind of tasks and scenarios you wanted models to be good at.
Yeah. So a huge percentage of our overall effort just went into building great foundations for these models, building the right foundations on pre-training, being very careful about scaling these models, and then building the right foundations on reinforcement learning and post-training to be able to continue scaling the intelligence of these models. So a lot of our effort went into the science, scaling and data to be able to support all these efforts.
But then we also took a lot of focus to really understand and think deeply about what were the things that were going to be very important for a personal agent product. Some of these things are multimodal capabilities. That is one area that is incredibly important for personal agent products. Other areas that we thought a lot about were memory or the ability of the model to focus. And so these were real highlights in the Muse Wonder family, which is active memory and active focus.
So we really thought deeply about what are the things that would make a personal agent like Muse really amazing. And we wrote all of those down into a big list. And this list includes hundreds of different areas that we're investing in to make sure the models are really good at. And then we developed evaluations and methods to measure, for every model that came out of the lab, how well did it perform. And then we continued iterating and iterating to continue improving the quality of these models. So it's been a real labor of love for the entire team to be able to develop these advanced models.
You even built the internal benchmark like the superintelligence index. So what did you focus on those fronts?
Yeah. So one of the things that we had found is, in particular, many of the things that mattered for developing a great personal agent were not well captured or measured by the existing benchmarks. So many of the existing benchmarks that most companies reported on are focused on coding or math or science, capabilities that are really valuable and important, but not necessarily the things that matter if you're trying to build the very best personal agent. So we had to develop a lot of that ourselves, and we developed many individual axes of behavior that all coalesced into the personal superintelligence index.
To give a few examples, like I mentioned, memory is very important. Focus is very important. The ability to chain together multiple tool calls, so to bring together many tools into one continuous process, is very important. Personality is very important, and how it communicates with you. That makes a really, really big difference in terms of the user experience. Also, discretion was very important. We really had to train the model to be very thoughtful about when it shared certain pieces of information and how it thought about that, and was it being thoughtful about when to reach out to the user versus when it didn't need to. These are things that are very, very important.
Fault tolerance was also critical. If something happened, whether some external or internal failure happened, how is the model able to respond to that? I could keep going on and on. I'll maybe mention one other area which has been really important, which is browser use. So the ability for Muse to be able to use a browser. And many of our users, you can see when it's using a browser, and you can take over if you need to. But that was so important for really unlocking the abilities of the model. And so the list goes on and on, and the personal superintelligence index captures so many of these all into just one number. But this really required a lot of thoughtful effort.
So we've seen a lot of series of AI-enabled cyberattacks where it is reported some banks in the US detected unauthorized AI agents in customers' accounts. How do you think we can make a world full of AI agents safer?
Yeah. So this was a huge area of focus for Muse specifically, and an area where we did lots of very extensive testing and were very thoughtful about it. We actually delayed the release of Muse many months to be able to ensure that we got this right. In our architecture, we have the Muse secure VM, which is a secure sandbox and an environment for the agent. Then we also have a separate sentinel agent that observes the actions that the main agent wants to make and has to approve those actions whenever the main agent wants to make one of those actions or share any kind of information. And then there's also human-in-the-loop verification. So many of the actions that the agent makes will require a human to approve before the agent will go out and make those actions.
So a lot of the effort that went into our architecture and our security architecture and our design to ensure that these models were private, safe and secure, we think are really important paradigms to be able to scale these consumer agents to billions and billions of people. So I think a lot of these things are quite important. I think obviously agents are very powerful technology that's very exciting, but also you have to make sure that you are investing into a solution that enables you to deliver them to anyone and give them a trusted experience.
Muse is initially free of charge, and I was wondering how you guys are going to cover all those costs like compute. And I was wondering, do you think your investment plans for compute, like Prometheus or Hyperion, can keep up with the Muse growth?
Yeah. So first of all, I think we're excited about Muse. We see really exciting business opportunities and business model with Muse. One is obviously subscription plans that we've released. So we've seen many of our
Users subscribe to a monthly subscription plan or buy additional tokens as they've needed to. Second, we actually think Muse is a very exciting B2B product. We've seen a lot of exciting adoption and outreach from small businesses and a lot of excited enterprises who want to be able to use Muse within their enterprise. So we think that it'll be a great business product as well.
And then lastly, there's just so much commerce happening on Muse. We've seen just an explosion in shopping where people are trusting their agents to do more and more shopping, and that's been really, really exciting. So as a result, we're quite excited about the business opportunity for Muse.
But in addition to that, Meta has many ways to be able to monetize our compute and to utilize that in a very economically efficient way. So one of the most exciting ways is maybe the most boring for most people, which is just using it for our core business and using it for our core recommendations and ad systems. And we see just amazing return on investment from being able to make those investments.
On top of that, we also just see lots of exciting uses in broader consumers. So both Muse and Meta AI and scaling those up. And then lastly, we recently started the Meta Enterprise platform and brought in CJ Desai to help lead that area. And so we see a lot of opportunities to monetize our technology with enterprises in the future, especially as we develop more and more incredible products and tools utilizing our AI investments.
Can compute keep up with the growth of Muse?
For sure. Yeah, we are very excited. Mark has talked about this, and we're very excited. We believe that there's a lot of opportunity for continued efficiency improvements within Muse, and as we're able to do that, we'll be able to continue scaling up the Muse product to more and more users using even the same level of compute. And at the same time, we'll continue offering it to more and more countries and releasing it more and more widely.
Okay, last question. What's your true personal intelligence for you? I remember that you said you waited to have kids until you get brain-machine computer interface like Neuralink. Is that the future you're seeing right now in terms of personal superintelligence?
I think what I'm very, very excited about is the point at which you can give a personal agent some very vague goals and very vague dreams, and then it's able to really begin and undergo the process of helping you accomplish all those, right? One of the things that I would love to do one day in my life is make a movie. I have all sorts of crazy ideas for movies, but I have a day job. I have no idea how to get started with actually making a movie.
But I think in the future, I really do expect the best personal superintelligence, whether that's Muse or something else, I should be able to just tell it I want to make a movie and give it some very vague ideas, and it should be this helper and work with me to make that happen. And that to me is the really exciting thing, where people have wants and ideas and dreams and things that they're really excited about, and then they're able to utilize AI as this helper to make those happen. And you would look back on that process and you would say, "Oh, clearly without Muse, I don't know if I would have been able to do that."
Do you think you can make it happen at Meta?
I believe that, yeah, this is where we want Muse to get to, and I think the path is very clear for us to get there. And so we're quite excited about it. Continuing to improve the models, continuing to improve the products, and just continuing to give better and better tools to our users.
Okay, I think the time is up. Thank you very much for joining our show.
Yeah, thank you so much, Mas.
Thank you. Hope to see you in Japan in the future.
Well, I hope so, too. I love Japan. I'm really excited. I'm using Muse right now to help me organize a trip in the coming months.
Okay. Yeah, that would be great. Thank you very much.
Thank you.
Article published · Updated