Mark Zuckerberg's Plan for a Personal AI That Works For You | Meet Muse

Vídeo em Inglês, publicado pelo canal Tiff In Tech.

Video transcript

Transcript from available captions; it may contain errors.

It's very novel. I don't think anyone else has anything like this. Meta can't even see the content of what you're putting into your Muse agent. >> Can you share some specific examples that you're personally using Muse for that made you really stop and think this is different. This is so unique. >> You know, just this huge amount of of compute. Basically like a hundred million tokens a a week, which is like a huge amount that we're basically giving to people for free. Oh, you like really we're kind of lazy that round. The shot that it sends me is like me just like on my back on the ground. One of our goals for this is to be able to enable what we call like a business in a box. I literally think that people are going to have personal agents that can effectively help them with their health, which is ultimately just a very personalized thing. >> Mark, thanks for joining me today. >> Yeah, great to do this. >> You know, Mark, I was thinking back to the last time we sat down, which was actually almost exactly 2 years ago at Meta Connect. And during that conversation, we were speaking about some other really big announcements you were making at the time. Now fast forward to 2 years later, here we are again sitting down about something that I'm really excited to speak with you on, which is Muse, the new personal AI agent. >> Yeah, I mean this is a big one. We've been focused on this for I don't know, maybe a year. Basically since we started Meta Super Intelligence Labs and had the the this whole vision of trying to deliver personal super intelligence to to everyone in the world. This is kind of what we had in mind. The the idea is giving everyone an agent that can understand your goals and can just work on your behalf 24/7 without you even needing to to to go do a lot more. And it can just help you in all different parts of your life from you know, helping you um strengthen your relationships and be present for your family and friends to you helping you with your career, helping run your home. I basically have it like hooked up to all my home systems and it like runs everything for me. Basically like helps uh helps with health. Um I like connect all my like my my aura ring and and like all my training and and stuff that I do in there. Um yeah, so I mean it's just everything that you do in like in the different parts of of your life, you can basically tell it what you're trying to do or it can even give you suggestions for things it can help with and um then it just goes and does projects for you. So it's uh it's it's it's pretty cool. And it's been fun um giving it to family and friends in in the early beta period because we basically just see like how quickly it kind of takes on like very big projects in people's lives. So it's uh it's been fun so far and I'm looking forward to getting this in more people's hands, but this is this is a big part of our focus. >> Absolutely. And I wanted to dig into that more. What are Can you share some specific examples that you're personally using Muse for that made you really stop and think, "This is different. This is so unique." >> Yeah, so I mean one thing that I do is um it helps me organize activities that I do with my kids. So, you know, my 3-year-old daughter, Aurelia, she likes baking and I don't know anything about baking, but but basically every weekend I I basically like it has the standing project which is I'm going to do a baking project with Aurelia and go find a a a recipe that is appropriate for a 3-year-old and an adult who knows nothing about baking and go get all the ingredients from Instacart or you know, whatever whatever you're going to use to to make sure that we have everything that we need and I just like want all that stuff to show up so that way when I like go sit down to you know, start baking with um with Aurelia, it's all ready. So, that's been a fun one because again, like I don't know anything about baking and uh we give the Muse agents um uh some feedback on which projects we want, which ones we we we think are are not good and and and Aurelia like has some ideas about what she thinks she wants to do next, but it's it's good. I mean it basically like like I'm busy and you know, it's great to just like have everything be prepared. So that way I can like spend some quality time with with my daughter just kind of doing this thing and learning along the way. Some some things are easy to bake, some things are are are hard, it turns out. But that's one. >> I hear you. >> I mean I play it one of my other daughters is really into building robots and and it made this whole plan over like a multi-week period to help us make a robot together. So that was that was a pretty cool one. And then my other daughter has gotten really into climbing mountains and like summoning mountains. So some of them you need a permit to be able to go get access to. So I basically asked Muse to um go find like figure out when we when we can get a permit and some of these is like it's pretty hard to get a permit. When you get the permit it's like you just drop stuff and you go and do it. So that's that's been pretty cool. It's like okay, here's the day when I can do it. It's like all right, maybe I'll just like I'll you know, take a day off from work and go go climb a mountain with my daughter. So that's that's kind of cool. So I mean that stuff is pretty fun. Then there's all this health stuff. I use it for training. You know, I'm really into MMA. I I put cameras in my MMA gym and I ask it to just kind of watch the cameras when I'm training and in between rounds give me feedback so I can so I can do better. So that's that's pretty cool. It kind of figures out what I'm doing if I'm like boxing or kickboxing or just doing grappling and kind of like gives me feedback and sometimes the feedback is pretty funny and even kind of brutal. It's like oh, you like really were kind of lazy that round. It's like and then the the shot that it sends me is like me just like like on my back on the ground. Like it's like come on. It's like can't you have you picked something more uh motivational to to to kind of to to highlight for me, but that's pretty cool and you it kind of connects to in my like aura ring data the other the other kind of health tracking data other like lifting and stuff that I do so that way it just kind of helps me balance all of the the the stuff to keep stay healthy, which is kind of like just important to in my life overall cuz I feel like if I don't if I don't get to a chance to burn the energy and stay and then you know, stay active then um you know, everything else suffers. >> Absolutely. And and I think through your examples, one thing that really hit home to me and what I love about it so much is I might be asking my Muse a question about I don't know, say say something around my health or whatnot. It will then if there's say an upcoming appointment for my son or something else completely outside of what we're speaking about, but it has that context, it will be like, "Reminder, Tiff, you know, your son's appointment is tomorrow." And that to me was is so useful and so helpful. But I have to admit, one of the hardest parts when I was setting up Muse was coming up with the name. So I'm curious to hear, what is your Muse's name? >> Oh. Yeah, I know, I think you just kind of need to do what the first thing is that comes to mind and then you can change it later. But I mean, I've had I've had that feedback from a lot of people that it's like, "Ah, this is like going to be a hard thing." Um mine is um it's Agrippa. He's the it's this uh historical figure. He was a he was a Roman general. I I think it's it's kind of funny cuz he just gets paired with this like cute teddy bear like um like avatar and it's uh it's cute. So I I think that this is something that people have really enjoyed internally. Is like, "Yeah, you you have a name, you get like an avatar. It's like you get to see your buddy working, all right? He's just like like you get to see him like typing away or something when he's when when it's like when it's doing work for you. So yeah, I I think people really enjoy that. What's yours? >> Um my Okay, so mine is Pip. And I'm not sure if I'm committed to it or not. It's been Pip for a while, so I think I'm going to run with it, but you know, we'll see. >> Yeah, I know, that's how I feel. >> Right? Yeah, but the good thing to your point is you can change it. So but I still felt that pressure. I was like, "I'm going to is going to be speaking to you all the time. But, shifting gears a little bit, recently you wrote in the Wall Street Journal that you think personal super intelligence should be something everyone has access to. And I'm curious to get your thoughts around, how do you see news fitting into that vision? >> Yeah, I mean, I think that news is it's sort of the clearest kind of implementation of of of this vision overall. I mean, my my overall view is that in order to have a positive future, we need to make sure that everyone broadly gets the benefits and opportunities of AI and that it isn't just concentrated with a small number of labs that keep the best models for themselves. Um and just kind of grow and grow and grow while while everyone else doesn't have access to the same level of intelligence. So, my my view is that the kind of the antidote, the thing that you want to do is build good technology and put that power in people's hands. That's what we've tried to do for a long time with social media, giving everyone a voice, giving people the ability to connect in the ways that they want. And it's how we think we should approach this with AI, too. And the best possible way that I can think of to put the power of of this new technology in people's hands is by giving billions of people um a personal agent that they can basically share what their goals are and it could just work and advance their goals on their behalf. And of course, if you want to do that, it's very important that it is affordable for everyone to use. Right? So, a lot of AI products are very expensive. But, we're designing this that way. Um there is a subscription if you want to use like a a like extra um capacity, but it's designed so that the vast majority of people can use it for free and you get, you know, just this huge amount of of compute, basically like 100 million tokens a a week, which is like a huge amount that we're basically giving to people for free. And the reason why we're doing it is that we believe that news is actually going to make people money and save a lot of money and that the the real long-term model for this is that we will it'll basically be this commerce model where over time we will take a very small cut of the value that we create in in either commerce or um or saving money and that will be how this gets paid for but this will be something that I think billions of people will have access to that will help them advance their goals and it will be something that you don't even need to pay to to really use. So um so I think that that's like an important part of getting a positive future for everyone is making it so that the power of the this kind of most advanced AI technology can be something that is affordable and in everyone's hands. >> Do you think the the end state is that we're spending less time actually interacting with AI directly because our agents are increasingly just getting these things done then in the background? >> I don't know. It's hard to tell. I mean I think this will probably replace a bunch of our other interaction with computers but I I certainly hope that I mean I think that the way that this is going to work is that people are probably going to spend less time having to kind of work and do productivity type stuff over time and they'll get to spend more time on the things that I think matter to us as people like getting to spend time with our family and our friends and creating things and on culture and understanding the world and that that type of stuff I think is just like a lot of time you know people people's days are are consumed with somewhat mundane tasks that they have to do and hopefully the agents will um make it so that that stuff can get taken off your plate. I mean I do think that if people find this useful I mean they'll interact with their agent a fair amount just because each time you tell it to do something you're basically getting the superpower of it will you know now go do what what would have been hours of work for you. So I mean it's like you know if you if you have you know you people have a lot of stuff that that I think that they would like to get done if they didn't have to go do those hours of work themselves. So I think that that'll be a big thing but our goal is that this should also save people a lot of time. Just make it so that you take all these tasks off their plates that they can just You can just do the stuff that you care about, whether it's hanging out with the people you care about, connecting with them, your hobbies. Um yeah, all that. >> You know, there's the time aspect, and for me there's also been the actually the inspiration aspect. So, as with my Muse, obviously as it continues to get to know me better, it sets different ideas, or I can set goals with it, and some of those ideas that it's sharing with me are things that, you know, when you're so focused on one project or a few projects, you kind of don't see maybe a certain picture or other side of things. And then when uh my Muse pip is is giving me ideas around, you know, the context that I've shared with it, that's that's sparked a lot of creativity uh for me. >> Yeah, I mean, I think one of the big issues that a lot of people have with AI is they're not quite sure what to do with it or how to get the most out of it. So, one of the specific things that we built into the product is this ideas feed, where it can generate a bunch of ideas for things that it can do to make your life better. And some of them are general things, like, okay, it can help look through all of the services that you're subscribed to, and it can save you money by figuring out where there are duplicates, and it can just kind of go cancel them with your permission, and then go save you a bunch of money. That's probably applicable to most people. But then as it gets to know you and the things that you care about, it comes up with ideas that are personalized to you. So, you know, like I play computer games with my daughters, and you know, so it came up with this idea of, "Hey, do you want me to build a a website for your daughter that goes through the strategy of the game Civilization." So, it's like, "Yeah, sure, go for it." And then So, we started using that, and then after that, basically it suggested, "Hey, do you think would you want me to um add to that website that I made, so that way you can also help teach her more lessons about history, about the civilizations that are in the game as you're playing with it. So, it's like, "Yeah, go do that." So, then I just added a tab to to the website that I made. And I think it's just like one of the more interesting and powerful parts of this is that it can suggest ways to just like improve your life and suggest projects that it can take on either for you or that you can do with the people you care about to make it so that you can just do more together. So, I I I think that that's that's something we've put a lot of time into making that so that it can it can help you with that. >> Absolutely. And you know, it is a very consumer-facing agentic AI product. What are some of the differences that you experience when building this product versus something that was more developer-facing in the past? >> Well, one of the things that I think is most important about how we designed this is that the the element about trust and security is is huge, right? It's like the In order for this to work well, it needs to have a good amount of intelligence, but it also needs to understand you, right? So, at a minimum, it kind of needs to understand your goals, but I think people as they use it, they plug in a lot of information into it. And in order to feel comfortable with that, we've done a lot of specific security and architecture work to make it so that all of that is is like really secure. So, we've learned a lot from the work that we did with WhatsApp, where we architected WhatsApp so that it's end-to-end encrypted, which means that as you're messaging people, Meta can't see the messages that you're sending. Like, even if we wanted to, it just it's encrypted as it flows through Meta's servers. And we took that principle, and we wanted to design an architecture here that we call confidential VM or virtual machine. So, that way we're we're building this technology. It's it's very novel. I don't think anyone else has anything like this as far as I'm aware that will make it so that if you're using a confidential VM, even it meta can't even see the content of what you're putting into your muse agent. So, and that's very powerful and it creates a a level of security and privacy and trust that I think is is just kind of unmatched in the industry. But even beyond that, we built a a lot of other measures. So, we built this secure um credential store, so you know, you your your agent doesn't need to know um you know, your login and password to different sites, your credit card number. That should go in a vault and when it wants to access those credentials to log into something or use your credit card, it just ask you for permission and and you could give it permission to to use stuff if if if you want, but there's no real reason why it should have that stored in a place where it can have access to that. Designing the system so that basically the you is the person staying in the loop on the key decisions that it's making is really important. You don't want an agent to just go off and like and make a bunch of decisions for you. So, it's basically we have this whole security architecture that we've designed so that way you have a second agent that we call the sentinel that it like basically can monitor what your main agent is doing and anytime it detects that there might be something that it that that you would want to be in the loop on, it will flag it for you. And um that's like a core part of the whole security architecture that we've created around this. But anyway, I think that the the the bottom line on this is this is I think one of the most important things is in order for this type of an agent to to work and be successful, people are going to need to you know, feel confident in its security and its trust to be able to give it a lot of context about their life and that's been probably one of the main things that we've spent a lot of time working on to to make sure we get right from the foundation up here. >> To me when I I first saw my own my Muse's a VM or private computer, I to your point, I mean, I I've never seen anything like that before, and it was really interesting to me cuz I always started digging into it more. I know it stores things like memories, our data. It's all right there, and it's just for our agents, which I think is such a huge unlock when it comes to security and trust. You you mentioned around keeping the human in the loop aspect. How did you decide where Muse should be able to do things that are autonomous versus where it has to come back and ask me? How did you make that decision? >> Well, I think basically anytime it's accessing a new service or if it's going to do a transaction, I mean, for the first time, it should it should basically get your your permission. And the way it works is that the agent itself doesn't actually get to decide what it can do on its own. There's this whole separate system that I mentioned that we call the the sentinel agent that monitors the data that's going out from your virtual machine, and that if it notices that there's anything sensitive, then it will flag that for you to review before taking the action. So, I think it's actually the security system you basically have multiple agents as part of the Muse system. There's the system that does things for you, and there's a whole system that is just focused on making sure that that your data is secure and that anything that should be reviewed by you is. >> That's really interesting, and and I think I can see a lot of people well, I'll speak for myself using this for for shopping, for booking travel. And that's really where the payment side comes in, and I kind of want to dig into that a bit more because I found it really interesting how I with my credit card, yes, it's saved in a secure you know, payment system, but when the transaction's actually made, Muse will then pay with a one-time virtual card number. So, the merchant will never actually see my real number, which I found really fascinating and and kind of brilliant cuz it makes me feel a lot better about making those transactions that way. >> Yeah, I mean, we designed the system so that way it is just it's just designed to be very private from the ground up. And when I'm doing transactions myself, I mean, sometimes I think about, okay, should I use my credit card or should I create one of these virtual credit cards? But it takes extra effort to go create the virtual credit card. So a lot of time I'll just use my normal credit card. And I mean, this is part of the value of having an agent that can take on these tasks for you is it doesn't mind doing doing that extra work. So just having the default be that it's you know, it's it's disclosing as little information as possible when you connect it to a new system, whether it's your email or some some system. It defaults to having the least privilege possible. So if you want to connect to something you don't necessarily need to give it right access to go edit your content just to you know, if like the most thing that that people want to do most commonly is is just be able to you know, have your agent have contacts to different parts of your life. So yeah, I mean, a lot this is like all very foundational to the architecture and design of it is is basically granting it the least privilege possible and disclosing the least information possible to complete whatever task you're trying to do. >> How did you decide what belongs in long-term memory versus what should say disappear after the task is done? >> Well, I think it decides that for the most part. I mean, there there are some kind of security-oriented questions like so stuff like credentials we keep in the in the secure credential store rather than keeping in in the Muse agent's database or memory. But in general, it um you know, every evening it will look at everything that it did that day and it will compact it into memories. And I mean, this is kind of how people think and and compact when we sleep. And I think that there's um that just like other parts of the the underlying AI model will improve over time. Like every month we're shipping a new model that gets more capable and I don't know if that's going to be part of what is what is interesting about this. I mean, I mean actually if you've been playing with this for a while the um and we just shipped Muse Spark 1.3 and your your uh Muse agent will probably be upgrading to that model like in the next couple of days and you'll notice it'll just get smarter and able to do more things automatically and it'll also just get better at kind of compacting its memory and and being more efficient about remembering the right things and not remembering the things that are relevant. >> Oh, well, I'm very excited for I mean, I can't even imagine cuz it's already uh my day-to-day is just it's so much it's honestly so much more seamless. Final question I have for you Mark is if we're sitting here in 2 years, which honestly seems like such a a long time from now, but we're still going to go with that. What do you think people will be letting their personal agents do then that would be hard for them to maybe imagine today? >> Well, I think that a lot of people are going to use it to run businesses. So I mean, I basically what one of our goals for this is to be able to enable what we call like a business in a box, right? So like you just have an idea, you wanted to create something, it helps you create the thing, it sets up the presence online, it connects to all the Meta services and whatever whatever else you want. It can help you run advertising so that way you can get the product or whatever you're doing into people's hands. Um it can handle a lot of the back end and I mean, in theory you should be able to just like have an idea, it should help you do that and then it should just be able to go like carry on the business and make money for you while you just go do the things that you want to go do. Now, I mean how close we are to having to to having that, I think we'll we'll see, but the models are getting good at a pretty fast rate and I think the idea is I think people like the creative aspects of that or different people like different parts of the process, but to be able to have an agent that can do the other things is going to be pretty valuable. I think we see a lot of people using it for health-related things, right? So, there's um a lot of the simpler things around wellness and training, which you know, we we talked about earlier like exercise sleep but I would guess that in the next couple of years the AI models are going to get good at starting to design different treatments for different conditions that people have. And I think one of the interesting things, I mean, just for I mean, outside of Meta, my my main philanthropic effort is working on Biohub, which I I started with Priscilla. And one of the main takeaways that we've had from that is there is a very long tail of rare diseases and conditions that people have, right? I mean, it's you know, the the industry largely focuses on the common things because that's where there's a market. You know, if you make a a drug for something then that a lot of people have, then then you can you can sell that to a lot of people, but there are a lot of conditions that, you know, maybe like a couple dozen people have or, you know, a few hundred people have and there's not like a big um market for, you know, a biotech company to go invent a cure for that, but I think what we're going to start to see is a model that can work on your behalf can potentially come up with personalized treatments if it has information around or around like the different things that are important for that that both make up your condition and who you are and your medical history and all of that. So, I I would guess that within a couple of years, if we don't have that yet, I would guess it'll be kind of close on the horizon, but it it gets to how kind of intimate and and and personal, but also serious and important um I think people are some of the things are that people are going to be doing with this. I do I literally think that people are going to have personal agents or that that that can effectively help them with with their health, which is ultimately just a very personalized thing. I mean, everyone is different. >> Absolutely. Well, I think that's a great great note to end on. Mark, thank you so much for your time. >> Yeah, thank you. I'm excited for this.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Muse. Publicado no YouTube em 2026-09-08.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.