YT YouTube2Text

YouTube2Text → Transcripts → Mark Zuckerberg on how Meta's Muse AI agent will make people money

Mark Zuckerberg on how Meta's Muse AI agent will make people money — Transcript

by Sources Podcast · 15,207 words · 2,305 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Mark, last time we spoke, you called me.
0:02It was on the weekend. Well, you called
0:04me through the glasses, which was the
0:05interesting part.
0:06>> Was it loud in the background?
0:07>> You were You were going to fish. I
0:09wasn't going to say it, but yeah, it was
0:10on the weekend. The noise cancellation
0:12actually is pretty good.
0:13>> Oh, it's great.
0:14>> Yeah, it's pretty
0:14>> Yeah, I I I I mean
0:15>> Yeah.
0:15>> Yeah.
0:15>> Um but you wanted to talk about this
0:18essay manifesto. I don't know what you
0:20call it. Manifesto, we'll say.
0:22Uh that you published recently. And it
0:23And it's long. I mean, I'll caveat
0:25there's a lot in it. And I wanted to
0:26start there cuz there's a lot of big
0:28ideas in there and they'll connect to
0:30the kind of the main thing we're talking
0:31about today.
0:32>> Yeah.
0:32>> Um I'm curious, like why write that? Cuz
0:35it it's long. There's a lot in there.
0:37>> [laughter]
0:37>> Well, yeah.
0:39Well, I feel I feel like if you're going
0:40to invest so much in building AI, uh
0:45then it's important that people
0:46understand what your lab stands for and
0:48what your values are. And And And
0:50basically, AI has so many
0:53opportunities, but there are also all
0:55these real risks. So, I think it's very
0:57important that everyone who's working on
0:59it has a well-thought-out theory for how
1:02the work that they're going to do is
1:04going to lead to a positive future. You
1:06know, it's interesting because I mean,
1:07the different labs have have some
1:09different philosophies on this. And
1:10there's, you know, a lot of things that
1:11I think have become conventional wisdom
1:13in the industry that I just strongly
1:15disagree with. And you know, I mean, my
1:17my view on this is that the path to have
1:20a positive future for everyone is to
1:22make sure that we distribute the
1:24technology as widely as possible. And
1:26that's based on three major principles
1:29that that we have. One is that
1:32empowering people is the source of
1:34prosperity in the world, and it has been
1:36throughout history. Two is that the
1:39primary purpose for AI is going to be
1:42invention of new things, not automation.
1:45And then the third principle is that the
1:48foundation for safety for the future is
1:51basically establishing the right checks
1:52and balances and balance of power rather
1:56than restricting access. And I think
1:58that that's these are all things that I
2:00think oddly are are kind of very
2:03different from I think a lot of the
2:04conventional wisdom,
2:05especially in Silicon Valley. A lot of
2:07people think, "Hey, this technology is
2:09very powerful. We must restrict it so
2:10that way not that many people have
2:11access to it." I personally am much more
2:14worried about a small number of labs or
2:17people having control of of something
2:19that is so capable,
2:22and I think that the throughout history
2:23what we found is that when you put power
2:25in people's hands, most advances don't
2:28come from the incumbents or the
2:30establishment. They come from people on
2:32the on the periphery whose ideas aren't
2:34taken seriously, but when they get
2:35enough tools to
2:37to to basically be able to prove out
2:39what they're working on, that ends up
2:40being very powerful. In Western society,
2:43the way that we've established
2:44governance and
2:46and basically having a well-balanced
2:48society is through a set of checks and
2:50balances, right? And and this balance of
2:52power. It's very ingrained in in kind of
2:55our society that
2:57you don't want to have you know one lab
3:00or two labs having access to a thing.
3:02For some of the most recent concerns
3:04that have come up like some of these
3:04cybersecurity concerns,
3:06I think the the best antidote to someone
3:09having an AI that could potentially hack
3:11into systems is having everyone have
3:13access to an AI so they can harden their
3:15own systems first. And I think that
3:17that's kind of been the history of
3:18cybersecurity over the past several
3:20decades is that you know open-source
3:22software
3:23because people can
3:25can see it and can scrutinize it, it's
3:28sort of counterintuitively by putting it
3:29in people's hands, you end up with a
3:31more secure and more stable environment.
3:33So, that's what I believe,
3:35and and that's what I think is the path
3:37to to a positive future is basically
3:39distributing the technology that has a
3:42bunch of different implications for what
3:44we're going to do. I mean obviously we
3:45want to build leading
3:47AI models, which which we're doing. I
3:49mean MuSpark 1.3, which we just
3:51released, is it's it's advanced, but
3:54then, you know, it's actually the latest
3:57model of a relatively smaller pre-train
4:00that we did, the internal code name
4:01avocado.
4:02>> Mhm.
4:03>> And um
4:04>> Oh, we're going to get into the code
4:04names.
4:04>> Oh, yeah. So, we'll we'll get into that.
4:06And
4:07we have watermelon coming soon. So, so
4:08that's going to be a big deal. So,
4:10obviously, leading models probably the
4:12biggest personification of you know, if
4:15you will, or kind of
4:17implementation of this vision is the
4:19Muse personal agent that we're that
4:21we're rolling out. That basically the
4:24idea there is give every person in the
4:26world a very capable personal agent that
4:29can understand their goals and can just
4:30work on their behalf 24/7. And then, an
4:33important part of this is also just
4:35getting the technology in people's
4:36hands. So, we're we're very kind of
4:37strong proponents of open source and and
4:40making sure that that the opportunities
4:42that I think are going to be massive
4:44here are not just limited to a few to a
4:46few people or companies.
4:47>> This episode is brought to you by
4:49Mercury, AI native banking that's loved
4:51by more than 300,000 entrepreneurs,
4:53including me. Visit mercury.com to learn
4:55more. Mercury is a fintech, not a bank.
4:57Check the show notes for details. Thanks
4:59also to Granola, the AI notepad for
5:02people in back-to-back meetings. It
5:03works everywhere you do and lets you
5:05focus on what matters. Try the
5:07granola.ai/sources
5:09and use the code sources for 3 months
5:11off. This episode is also brought to you
5:13by Jira by Atlassian, where teams and
5:16agents get the context, coordination,
5:18and control to move work forward. Try it
5:21free at jira.com. That's jira.com.
5:26>> I have a bunch of questions about the
5:27Muse agent, but staying big picture for
5:30a second cuz you said a lot of things
5:31there.
5:32Is open source the counter to the trend
5:35you're seeing that you described that
5:37you're worried about? Is that the main
5:38way, practically, that you counter that?
5:41Or is it regulation? Is it both? Like
5:43>> No, I actually think probably the most
5:45important thing is actually just getting
5:47the technology in individuals' hands.
5:49So, I actually think things like the
5:50Muse personal agent
5:53are perhaps even more important. I mean,
5:56I think what you're starting to see are
5:57some of the labs are building training
5:59more advanced models
6:00and then not even releasing them.
6:02>> Right.
6:02>> Right. So, I think that that is quite
6:05dangerous in the sense that, you know,
6:07basically when you have scrutiny on
6:09something, when you put a a system out
6:11there, first off, you put it in a lot of
6:13people's hands, you get the checks and
6:14balances. You get broad-based
6:15prosperity, which I think is important,
6:17right, for society. We can't just have
6:19like one or two labs get incredibly
6:21valuable.
6:22You know, I think you you want to make
6:24it so that billions of people can
6:26basically have prosperity in their own
6:27lives, whether it's creating small
6:29businesses, being more successful in
6:30their careers, being more productive in
6:33managing their homes, saving money in a
6:34lot of ways, kind of advancing their
6:37health. So, I think you you want the
6:38benefits to be very broad-based. So,
6:41that's one piece. In terms of the
6:43competition, I do think that having
6:45multiple labs is helpful and I think
6:47open source is quite helpful for that.
6:50So, I think open source is an important
6:51part of it. The nature of open source is
6:52there's a whole community of people who
6:54do it. So, I'm not saying that we're
6:55going to be the one company that does
6:56it. I'm also not like a zealot about
6:59this from the perspective it's you know,
7:00it's not that everything we do is open
7:02source either, right? We we release some
7:04open models, we do some closed work. I
7:05think it's important if you're building
7:07a for-profit company that you can build
7:09some advanced things and you don't
7:10necessarily need to share every single
7:12thing with the world, but I think in
7:13general, uh supporting a robust open
7:16source ecosystem is going to be key to
7:19maintaining competition and and
7:21maintaining kind of transparency and
7:24understandability of where the
7:26technology is going in a way that I
7:28think is actually going to be incredibly
7:30important for safety.
7:31>> Mhm.
7:31>> Um you know, if if we have a world where
7:33there's just like a small number of of
7:35really capable models, um
7:37I don't know, just it's it's
7:38interesting, right? I mean, if you look
7:39at some of the cyber stuff, for example,
7:41uh the instinct, which I mean on the one
7:44hand I can understand the instinct of
7:45like all right, we have this capable
7:46cyber model. Let's release it so that
7:48only, you know, whatever it is, the top
7:50100 institutions get it. But and I think
7:52part of the issue is that there's more
7:53than 100 important institutions in the
7:55world. So, you know, if you look at
7:57things like
7:58the hugging face incident that happened,
8:00you know, hugging face is maybe not one
8:01of the biggest 100 institutions in the
8:03world, but it matters, right? It's like
8:04an important thing that people rely on.
8:07And so what did they do when they
8:09started detecting that there was this
8:11intrusion as they turned to open source
8:12models because they didn't have access
8:14to some of the closed ones that were
8:15causing the issues. So, um, I think that
8:19having a robust open source ecosystem is
8:22one important part of having kind of a
8:24safe and stable future. But to me the
8:27the most important thing is just making
8:29sure that we distribute the technology
8:30widely rather than hoarding it in a
8:33small number of people's hands.
8:34>> And I think there's also this ongoing
8:36debate in Silicon Valley about why
8:39people feel so negatively about AI. I
8:41mean you talk about this in your in your
8:43recent letter uh, addressing these
8:45concerns or trying to. But that I mean
8:47the sentiment on AI and data centers in
8:49particular it's so negative and it
8:52sounds like maybe an essence of your
8:54argument, correct me if I'm wrong, is if
8:56we diffuse this technology more, if we
8:58enable more people to access the things
8:59that are right now gated by some of the
9:01top labs, maybe that addresses this kind
9:05of I think that people feel maybe
9:07disenfranchised by what's happening in
9:08AI. Is that what you're getting at?
9:10>> Yeah.
9:10Well, there are many layers to it. It's
9:12I mean I think you there's so many parts
9:14of This is why the essay was so long,
9:17right? It's like 15 pages because I mean
9:18we want to get through There are lots of
9:19different questions. I mean people have
9:20questions about jobs and the economy.
9:23They have questions about data centers
9:24in their local communities and the kind
9:26of economic and environmental impacts of
9:28that. There are questions about how
9:30people might misuse AI, right? I mean
9:32there's the cyber questions. There's bio
9:33risks that are coming up. There are
9:35questions about now how we maintain a
9:38free society. There are questions about
9:39American leadership. Um there are
9:40questions about maintaining control over
9:42the technology as it gets to be
9:44increasingly capable. So, these are all
9:46important. So, it's it's important not
9:48to um you know, just talk about this in
9:50generalities at the at a high level
9:53because I think each of these has some
9:54different nuances. But in general, I
9:57think one of the things that they all
9:58have in common is that if you create
10:00broad-based prosperity um and I think
10:03that one of the better ways to do that
10:05is by ensuring that there's the right
10:06balance of power around who has access
10:08to the technology and generally making
10:10sure that the greatest balance goes
10:12towards the kind of general population
10:14of people as opposed to kind of any kind
10:17of I don't know, insider stakeholder or
10:18whatever you want to call it. Um I think
10:21that that ends up being very important.
10:23So, if you look at the data centers, I
10:25what what we've actually found is that
10:26when a company like Meta
10:28goes into a community
10:30and makes a commitment that we're going
10:31to invest there for decades, right?
10:33Which is really what we're doing when
10:34we're building up a data center,
10:36we're able to make it so that
10:39it's very good for the community. Um I
10:41mean the the the kind of the tax revenue
10:43that they bring uh from that. I mean in
10:46in Louisiana, we had this example where
10:47like the tax revenue funded these
10:49$50,000 bonuses for teachers in the
10:51community. We we bring a lot of jobs. We
10:54we invest in the in the local community
10:55a lot. Um
10:57that I think can be good. I think that
10:58there's there's also a lot of
11:00speculation, right? Where there are
11:02companies that aren't necessarily
11:03planning on running a data center for
11:05decades. They're just, you know, trying
11:07to find a plot and um and then trying to
11:10sell it to one of the big labs and then
11:12don't really care as much about the
11:13local community and um and they're not
11:15invested for the long term. So, if they
11:17don't care to, you know, focus on making
11:19it work for the local community, then of
11:21course people are going to get upset.
11:22So, I think that that's one of the
11:24things that can be um
11:27difficult when you have these kind of
11:29speculative um
11:31I don't I don't even know if I'd call it
11:33a bubble cuz that implies that it's
11:34overvalued or something. But But there's
11:36certainly a boom, right? So it's So So
11:38you have that. I mean, that leads to
11:40some of these like short-term thinking
11:42incentives that don't necessarily lead
11:45towards helping every stakeholder, which
11:46I think is kind of what you need to do
11:48to make this sustainable over the long
11:49term. And at the end of the day, I mean,
11:51if we're creating a technology that like
11:54if it doesn't create jobs or doesn't
11:56create broad-based prosperity or the
11:57infrastructure that we're building
11:58doesn't help local communities, that's
12:00not going to be allowed to continue. So
12:02it's like it has to Of course, you have
12:04to design it in a way that um can be
12:06helpful to people in all these ways,
12:08which is also part of the reason why
12:10like for people who are, you know,
12:11skeptical or have so much doom about the
12:13whole thing, I'm I mean, one of the the
12:16things that I basically think is that if
12:19we don't end up building it in a way
12:20that's positive, it just effectively
12:22won't be able to happen. Um so I I kind
12:26of think like that actually establishing
12:27the right checks and balances
12:29um and distributing the the benefits of
12:31this widely is sort of a precondition
12:34for being able to scale in the way that
12:35I think um
12:37would be best for society over time.
12:39>> I haven't heard another tech CEO in your
12:41position talk about data centers that
12:44way. Like the long-term investment of
12:46it. Is this something Is this how you've
12:48always thought about it? Is this
12:49something you feel like there's more
12:50clarity that's been brought to it for
12:52you recently? Like Have you Yeah.
12:54>> I mean, I think that there's been all
12:55this anti-data center sentiment that
12:57you're talking about. So we've dug into
12:59it because what we're trying to
13:00understand is,
13:02okay, like there isn't as much of that
13:04around our project. So why is that? And
13:07then so we asked a bunch of people, it's
13:09like, well, it looks like there's a
13:10pretty big dichotomy between these
13:11speculators and the companies that are
13:14focused on it for the long term. And
13:16that kind of makes sense when you think
13:18about it. So I mean, one one thing we're
13:19doing, I mean, there's this America's
13:20Workforce Academy project that we we
13:22did, which is basically, okay, we're
13:24going to build all these data centers.
13:25We're going to be doing this for a
13:26while. There
13:27isn't the volume
13:29of skilled tradespeople that you need to
13:31create this. So, we need more like fiber
13:34technicians and electricians and like
13:36advanced carpentry and and all of these
13:38things. And there aren't enough people
13:40to do this. We need hundreds of
13:41thousands, you know, maybe maybe
13:43millions of more people who can do this.
13:45And um
13:46people aren't trained to do that right
13:47now. So, we created this training
13:49program um to to effectively do that,
13:53where we guarantee people who get
13:55through the training program a job at um
13:58at at a place that is working on
14:00building infrastructure for Meta. And
14:02why did we do this? I mean, it's like
14:04it's not really philanthropy, right?
14:06It's like we need those people to be
14:08skilled and have those skills. So, it's
14:10just it's a win-win. It's an investment
14:12that I think makes sense if you're in it
14:14for decades, um but not necessarily
14:17something that you would do if you were
14:18building out one site with the intent of
14:20flipping it to a different company. So,
14:23I think that a lot of problems in the
14:25world do just naturally get solved by
14:29incentive alignment when you think about
14:31them over the long term.
14:32Um so, I think that ends up being an
14:33important part of this. And I guess part
14:35of the way that I think about this is
14:36that I just think that there's no way
14:38that it's going to be kind of permitted
14:40for there to be a small number of labs
14:43that control such a important um and
14:45capable technology and accumulate a lot
14:47of wealth to themselves. I think it's
14:49like this has to be broad-based thing in
14:52order for it to kind of be able to work.
14:54It has to work technologically, but it
14:55also has to work um kind of socially.
14:58And and I I think that those those
15:00pieces kind of have to have to go hand
15:02in hand.
15:02>> I agree. Well, now we're landing towards
15:04the muse. Before we get there, you also
15:07wrote a year ago your personal
15:09superintelligence essay, shorter one.
15:11>> Uh-huh.
15:11>> Uh but
15:13>> Yeah.
15:13>> I did write a one-page version of The
15:15Future Is Asian, too. Yeah, well, I
15:16published it in the Wall Street Journal.
15:17>> Oh, that's right. That's right. That's
15:19right.
15:19>> that was kind of
15:20>> I just read the long one. Okay.
15:21>> Yeah.
15:22Yeah, this is the one page version and
15:23there's the 15 page version. Well, so
15:25this one you did about a year ago
15:26personal super intelligence. I think a
15:28lot of people in my world when they saw
15:30you write that was like, oh wow, why is
15:31Mark writing this? Like what is the
15:34thing he's seeing on the other side of
15:35this? And I think it correct me if I'm
15:37wrong, it might be me as like what we're
15:40going to talk about, right? What you
15:41guys are releasing now. This is this is
15:43it. So how did you come to that
15:45realization that that's what this is the
15:47next chapter for Meta?
15:48>> It's interesting. You know, we've never
15:50really just thought about ourselves as a
15:53social media company. We've definitely
15:54thought about ourselves as a company
15:57about connecting people and about
15:58empowering people. But I think a lot of
16:00the values that led us to build the
16:02things that we built for the first you
16:04know 15 20 years of the company around
16:06putting technology and power in
16:07individual's hands.
16:10Believing that people should be able to
16:12decide for themselves what is important
16:14in their lives. And we've gone through a
16:15lot of social debates around this,
16:17right? A lot of the the the debates
16:19around content moderation and things
16:20like this have kind of been around this
16:21question of like should people be
16:23allowed to kind of decide and
16:25communicate for themselves what matters
16:27in their own life. And I think through
16:30that experience it is sort of sharpened
16:33my
16:34belief that
16:36a lot of progress throughout history and
16:38through this technological age comes
16:41from empowering individuals and that
16:45people really do know best about what
16:47matters in their own lives. So because
16:49of that I have somewhat of a of an
16:51allergy whenever I hear people talk
16:54about
16:55oh like we should just have a small
16:57number of experts allocate what AI does
16:59to like big problems. Why should it do
17:02like like these things that people care
17:03about in their lives? It's like well,
17:04people have a balance of things they
17:05care about. People care about health,
17:07they care about having a better life,
17:09but they also care about their
17:10relationships and like showing up for
17:13their friends and family and Um, care
17:15about culture.
17:17You you know, people care about things
17:18that may not to, you know, a scientist
17:21or an engineer in the industry feel like
17:23the biggest problems, but I don't know.
17:25If you ask like billions of people what
17:27they care about, if like, you know, I
17:29think what people's aggregate answers to
17:31that question is kind of is what the
17:33most important things are to be worked
17:35on. And um, so I've always just kind of
17:38believed that when you build this super
17:41intelligence,
17:43there's this question of
17:45who decides what it's going to focus on.
17:47And I think that people should be able
17:49to direct it towards what matters in
17:50their own lives. It shouldn't just be
17:52directed by, you know, some like
17:54so-called experts
17:56um, sitting at a small number of labs.
17:58So, this gets back to the it's like it's
18:00this is basically the foundation of this
18:02overall philosophy, which is that the
18:05way to have a positive future is to
18:08empower people to put the technology in
18:10their hands and let people decide for
18:12themselves what matters and how they
18:13want to use it. And I think that when
18:16people do that, it first of all will
18:18prioritize some things that that are
18:20different, right? It's um, you know,
18:22it's like maybe it'll prioritize um,
18:24health issues, but maybe instead of
18:26prioritizing, you know, it's like the
18:28most common things, which is kind of
18:29what the the kind of pharma and biotech
18:32industry uh, like it it large
18:34prioritizes today. Now, it's like
18:36there's a very long tail of rare
18:37diseases and conditions that people
18:39have. So, you know, okay, if you're if
18:41you have if you have a rare condition,
18:43like you're probably going to want your
18:44personal AI to focus on that, not like
18:46just something else just because it
18:47happens to be the most common thing. So,
18:48I think for example, rare diseases I
18:50think are under uh, disproportionately
18:52under invested in. Um, you're doing a
18:55lot of investment with your foundation.
18:56>> at Biohub, but but I I mean, and that's
18:59partially informed some of my views
19:00here.
19:01>> Mhm.
19:01>> That I think like, yeah, like you want
19:03to put the power in individuals' hands
19:05um, to determine what matters for them.
19:07But a lot of this is also like it's not
19:10necessarily the things that people would
19:12say are like the big social problems.
19:14Like a lot of it for me, you know, when
19:16I'm using my Muse agent, I just, you
19:18know, I kind of want to it to help me be
19:21like a better father and you know, a
19:23better husband and show up better for my
19:25friends and
19:26you know, be able to
19:28um help me connect with people and that
19:30I think is also partially a through line
19:31between the work that we've done at Meta
19:33so far. And this is I think we're the
19:36company that I think just
19:37disproportionately cares about and
19:39believes that there's social value in um
19:42in helping people connect with the
19:43people around them. So, I don't know
19:44it's like one of the first things that I
19:45that I kind of set up my my own agent to
19:47do. It's like, all right, my 3-year-old
19:49daughter likes baking. I don't know
19:51anything about baking, but like that's
19:54like a a fun project that we can do. So,
19:56I basically ask him like, all right, set
19:58up so that every weekend
19:59um
20:00you know, like we have a baking project
20:03that is
20:04um
20:05that is uh kind of reasonable for a
20:073-year-old and an adult who knows
20:09nothing about baking. And um and use
20:12Instacart or whatever to go get all the
20:14ingredients. Like just figure out what
20:16makes sense and make sure everything is
20:18ready so that way like when I show up
20:20on, you know, on Sunday with uh with my
20:22daughter, we can like go make this
20:24thing. And um and then I tell it
20:26afterwards. I'm like, how how did it go?
20:27Okay, that one was too hard. Turns out
20:29cake pops really difficult. Um
20:30surprisingly difficult.
20:31>> Huh.
20:32I don't know anything about baking so.
20:34>> Yeah, I didn't either. I I know I know
20:36something now. Yeah, cake No, no, don't
20:38start with cake pops. That's that's the
20:39problem.
20:40>> Um yeah, now there's a lot of things in
20:42baking that are pretty simple.
20:44It turns out cake pops is not one of
20:45them. But um yeah, well, you know, what
20:47can what can I tell you?
20:48>> Okay.
20:48>> Um
20:49>> a Muse.
20:50>> Yeah, yeah, thanks. Um so, yeah, and it
20:52kind of like so, it kind of updates
20:54that. Um
20:55And and helps with that. Um you know, my
20:58you know, my older daughter has kind of
21:00gotten into climbing mountains and some
21:02of them you need permits for. So, I have
21:04it basically like sit and get the
21:06permits when they become available so I
21:08can climb mountains. And um it's just
21:10like pretty neat. Okay, so then like we
21:12do that and it basically tells me it's
21:13like all right, I was able to get a
21:14permit for this day.
21:16And I was like all right, well I guess
21:17I'm taking that day off from work to go
21:18climb a mountain with my daughter. So
21:20it's kind of cool.
21:20>> Yeah, yeah, yeah.
21:21>> All right, it's So it does that, you
21:23know, but it also helps keep me healthy,
21:25it helps with my training. Um yeah, I
21:28put cameras up in my in my MMA gym and I
21:31tell it to watch the cameras and send me
21:32feedback and it's like it's pretty fun,
21:35you know, it's good.
21:36>> feedback.
21:36>> Um sometimes, you know, it's I mean it's
21:38um sometimes it's funny feedback. It
21:40like it like finds me in
21:42it's like it looks like you really gave
21:44up and and I was like yeah, I did. I was
21:46really tired right there. It's
21:47[laughter] like why is that the thing
21:48that you're pointing out to me? Um but
21:50no, it's [snorts] good. Um and it's
21:52funny. Like the coaches like laugh about
21:54it cuz they're like yeah, no, this is
21:55like it's this what we
21:56>> gets to the point where like we could
21:57tell you but your agent's telling you.
21:59>> Yeah, [laughter] well, it's um yeah,
22:01it's good. It's good.
22:01>> I didn't think about this until hearing
22:02you talk about it but you you have
22:04people that could obviously do all of
22:05this for you but like how do you use
22:08something like a Muse agent to like
22:10really test the limits of like how it
22:12can be helpful as an assistant, right?
22:14Like like are you were you pushing it
22:17like how have you pushed it in a way
22:18that the team is like oh, okay, we got
22:20to got to fix this or I'm sure
22:22>> I think I think part of what's
22:23interesting about it
22:25is that everyone it just has like such
22:27different things that they want to do
22:29with it. So like in the early beta
22:31period, we handed it to a bunch of
22:33people and like it's like I gave it to
22:36someone and like within a day they're
22:38like using it to help run their home
22:40school. And then like it's like okay,
22:42wow, you just like started this within a
22:43day. And then like another person within
22:45a day or within 12 hours they're like I
22:47just planned a trip. Um yeah, and it's
22:49like it it just like planned this whole
22:51thing for me. Yeah, I mean someone else
22:53I know who's like generally pretty
22:54skeptical about technology, I I gave it
22:56to her and then she was
22:57um she didn't say anything for like a
22:59few days. Then she texted me and was
23:00like so when you do like the general
23:02release, do I get to keep my music agent
23:05or are you going to reset it?" And I was
23:06like, "All right, this is good. Okay,
23:07this is yeah. So, I think I think I
23:08think it's like I think this is working
23:09well."
23:10>> Everyone I know who's been on the beta
23:12has very, you know, high things to say
23:14about it. High praise.
23:15>> Yeah. But, people do different things
23:16with it.
23:16>> Yeah, and and I think we should just
23:18also say what it is more plainly for
23:20people so they understand because I
23:21think people think of AI as like, you
23:23know, Meta AI or ChatGPT, it's like back
23:25and forth prompting. But, the real
23:27unlock here, and this is happening in
23:29the industry more broadly, whether it's
23:31Grok, bot, town, instinct. I mean,
23:34there's many products doing this, but
23:36it's it's adding a virtual machine
23:38behind the scenes where like the agent
23:40can control a computer for you and log
23:43in and do things. And that's a that's a
23:45huge change, I think, for people who
23:46only know AI from
23:47>> And it's it's long-lived. So, it's So,
23:50basically, instead of the model with
23:54like Meta AI or ChatGPT or Gemini or
23:56whatever you use, where you send one
23:58prompt and then it gives an answer, in
24:00this case, what you basically do is you
24:02give it projects or you give it goals.
24:04And then it just works and it works 24/7
24:07and it like doesn't stop until it's
24:09helped achieve the goals.
24:10>> telling me it studies overnight, is what
24:12you guys call it?
24:13>> Studies it like it kind of consolidates
24:15its reflections into into memory. It It
24:18basically just works on projects. And it
24:20can also suggest new projects. So, yeah,
24:24I mean, I was like I like play the
24:26computer game Civilization with one of
24:28my daughters and like, it was like,
24:30"Hey, do you want to make like a
24:31strategy guide for her?" I was like,
24:33"Yeah, sure." So, it was like, "Okay,
24:34now that we have the strategy guide, do
24:36you want me to like expand the strategy
24:39guide so it can also teach historical
24:41lessons about different civilizations?"
24:43I was like,
24:44"Yeah, sure. Why not?" So, it just So,
24:45it just kind of like built a new tab in
24:47in the app that it made and
24:49that was that was that was very cool.
24:50So, it can kind of just like expand and
24:52and
24:52>> It's proactive.
24:53>> Yeah, it suggests things. And mean, of
24:55the things that I think is interesting
24:57is that it
24:58I think it's just going to be able to
25:00like make people money and save people
25:01money.
25:02>> Yeah, I mean, part of the part of what's
25:04interesting here is the economic model
25:06for how we're pricing it. I mean,
25:08you can pay for a subscription if you
25:10want to if you basically kind of want to
25:11have that that model, but
25:14we're also just making it so that you
25:15can get a very large amount of usage for
25:17free. I think we're we're offering
25:19something I think to start it's like 100
25:21million tokens a week for for free like
25:23and and you get this virtual machine.
25:26So, it's like a a lot of kind of
25:27computer.
25:28>> That's That's a good meme. It's a lot of
25:29computer.
25:30>> of computer. Yeah, but the reason for
25:33why we're doing this is we basically are
25:35confident in standing behind the fact
25:37that we think that this is going to
25:39effectively for people who are going to
25:40use it for running a small business or
25:43making money or transactions or commerce
25:45in some way, we actually just think it's
25:47going to make so much money for people
25:49that
25:50that the business model over time that
25:51we expect is to effectively just take a
25:54very small cut of whatever the
25:56transaction is. Um
25:58>> Take rate business. Small, yeah.
25:59>> Yeah, and not even necessarily the
26:01person paying for it. It'll come from
26:02the businesses that they're working
26:03with. But, um
26:04>> And you're working with Stripe on
26:05payments and yeah.
26:07>> Yeah. So, but my view is like we should
26:09be able to have a service that you make
26:12free for the vast vast majority of
26:15people.
26:16Um which again is critical if you want
26:18to build up this future for everyone
26:19where everyone has
26:21um has these powerful super intelligence
26:23agents.
26:24Um I think an important part of making
26:25something available to everyone is
26:26making it affordable. So, we want to
26:28make it so that this is free. So,
26:30there's this just this huge amount of
26:31usage that you get and we're basically
26:32just standing behind that and saying we
26:34think that this thing is actually going
26:36to make you money and save you money and
26:38that is how you're it's going to pay for
26:40itself.
26:40>> And Meta's services can connect into it,
26:43right? So, you could theoretically
26:44manage your ad spend on Instagram, all
26:48that stuff.
26:48>> Well, you have to you connect it. You
26:50Yeah. connect it to whatever you want.
26:51You know, it does work with meta
26:52services if you want. I mean, you
26:54absolutely don't have to connect it if
26:55you don't want to. Sorry, if I'm you're
26:56using it to
26:58to run a business, it can basically just
27:00connect to our ad systems. I mean, you
27:01can ask it to to make something for you.
27:04It can it can kind of help you make the
27:06product and then it can help run the
27:07business. So,
27:08um so, all this stuff like it just do
27:10that in a loop and just do it forever
27:12for, you know, 24/7. So, and every time
27:14we release a new model, which, you know,
27:17we've been on this cadence of shipping
27:19a, you know, meaningful update like
27:20every month. It's going to get smarter
27:23or just going to get more capable and
27:24enabled to do more and more stuff.
27:25>> And something your team was telling me
27:27that I haven't heard this approach used
27:29elsewhere is this fleet concept where
27:31you're letting the fleet of muse agents
27:33learn together.
27:34>> So, that was the ideas and suggestions
27:36thing. So, you open up the app.
27:37The main tab is basically your chat with
27:41with your muse. There's a tab for
27:44basically ideas
27:46from the things that you've told it how
27:47it can expand those. So, that's the
27:49thing I was saying, which is, you know,
27:51first it helped me make the strategy
27:52guide for playing Civilization with my
27:53daughter, then it helped expand that
27:54into historical lessons. That was it
27:56came up with that idea. And then I was
27:58just like, "Yeah, sure, do it." Um
28:00then it it it finds all these ways to
28:02basically augment itself. I mean, the
28:03like MMA like
28:06coaching thing, it like comes up with
28:07ideas for how to make it better. It's
28:09like, "Would you like me to like get
28:11better at
28:12um finding the right frame to send you?"
28:13It's like, "Yeah, good. Go do that." So,
28:16yeah, the the ideas thing I think is
28:18important because then you basically
28:19across the fleet can find people who are
28:21interested in different things. I I
28:23think I I I I guess taking a step back,
28:25one of the big issues that I think
28:26exists with AI is a lot of people don't
28:28know what to do with it. Yeah. So, I
28:29think if if the if the agent can itself
28:32help you suggest things that it can do
28:35to be helpful for you, then that solves
28:37a huge part of this problem of making it
28:40so that um
28:42you can get the most out of it.
28:43>> When you're introducing like network
28:45effect learning for agents, which no
28:47one's really done, where basically the
28:49agents are learning anonymized insights
28:52from the rest of the fleet.
28:53>> Mhm.
28:54>> And I mean, you're like the king of
28:55network effects. Like, I I'm I'm really
28:57interested in this idea cuz I I don't
28:59think anyone's doing this.
29:00>> Yeah, no, I think that right now, I
29:03think most of the industry is thinking
29:05about agents as like a single-player
29:07game, where it's like you have your
29:09agent and and you use it. And there are
29:11going to be all these interesting things
29:13that basically you can do by having the
29:15agents interact with each other. And we
29:17already have all these interesting
29:18examples internally where, you know,
29:20people have their agents interacting
29:21with each other. This isn't like
29:23for the most part rolling out in this
29:24release, but it's it's going to be like
29:26an important part of how um I think this
29:29works over time is just like as more of
29:32the people who you know start using
29:34Muse, it just gets better for everyone.
29:36>> And is that the differentiator as, you
29:39know, you could buy the idea models
29:41continue to like commodify at the
29:42frontier, essentially, or you know, the
29:45products all start to look similar,
29:46similar kinds of harnesses, is the
29:48network effects of that learning the
29:49real edge?
29:50>> Well, I think that there's a few things
29:52that are that are kind of unique that
29:54we're doing. One is we're designing the
29:56models from the ground up to uh to
29:58basically be good for this use case,
30:00which I think really matters. Two is um
30:04is basically I do think we have this
30:05social DNA as a company where helping
30:08people use the AI and agent to
30:11like enhance their relationships and
30:13strengthen your relationships and get
30:15more out of like kind of the the soft
30:17but very important parts of your life. I
30:19think that that's something that we're
30:20probably just going to be more attentive
30:21to as a company than any of the other
30:24labs.
30:24>> Mhm.
30:25>> The third thing that I would say is
30:27actually going to be a major
30:28differentiator for us
30:30that I I think um it might be surprising
30:33to some people is privacy and security.
30:36And we're investing in this just a huge,
30:37huge amount. And, you know,
30:40part of the view that we have on this is
30:42that in order for this to be useful, it
30:44needs to not just have state of the art
30:47intelligence, it needs to really
30:49understand you. Right? So, in order to
30:50be able to understand your goals, so you
30:52can end up connecting it to all the
30:53stuff, right? You you talked about, you
30:54know, connecting it to your
30:56your ad system, but people connect it to
30:58messaging and email and all this all
31:01this stuff, health information,
31:02whatever. And
31:04in order to do that,
31:05people need to have a very high degree
31:09of confidence in the system.
31:10>> Yeah.
31:10>> Now, the good news here is that you
31:14know, Meta has spent at this point more
31:16than 10 years focusing on building
31:20WhatsApp into, I think like it is the
31:23largest
31:25like global end-to-end encrypted system,
31:27and we've designed it in a way
31:29where even Meta can't see the messages
31:33that people send, and that's been
31:35this just really transformative thing. I
31:37think it makes it so people trust
31:38WhatsApp. It's also been a very
31:40important lesson for Meta to learn that
31:42like that has been really important to
31:44our success with WhatsApp. That we've
31:46designed the systems that even we can't
31:48see the content. That means that
31:50whatever people are worried about, if
31:51they're worried about
31:53you know, government getting access to
31:54it, a hacker getting access to it,
31:55someone at Meta doing something bad with
31:57it that they don't want,
31:58all that stuff you can kind of take off
32:00the table if you design the system so
32:02that you can't see it. So, we took that
32:04as one of the foundational lessons when
32:06we were getting started with this. Not
32:08and I actually you know, personally
32:10recruited Moxie Marlinspike.
32:12>> Founder of Signal.
32:13>> Yeah, and the one of the person one of
32:15the people who helped us build the
32:16WhatsApp end-to-end encryption back in
32:18the day and back in 2014, he joined to
32:21specifically work on this confidential
32:23VM project, which makes it so that we
32:27can you can have your virtual machine
32:29and have all this information in your
32:31muse, and we can make the commitment
32:33that even Meta Mhm. see the content that
32:36is in there.
32:36>> And you can do this technically.
32:38>> It's It's a It's a like an incredible
32:40>> Like the commitment can be technically
32:41verified. Yeah.
32:42>> So So that and that's something that
32:44we're going to publish more about in the
32:45coming weeks. Um as as we as we
32:48basically get get closer to rolling this
32:50out a lot more widely.
32:52Um
32:52>> And of out of all these early VM efforts
32:54that these labs are doing with these
32:55agents, you think this is unique what
32:57you're doing? You It's a
32:58>> I don't think anyone is doing that. I
32:59don't think anyone's doing that.
32:59>> I mean
33:00>> I think that there's I mean there's a
33:01lot of other security measures that
33:04we're putting in place that I think are
33:05we should talk through. Um because I
33:07mean, you know, even before this is
33:09ready, like there's there's that and
33:10even people who don't want to use this,
33:11there's there it's like incredibly
33:12secure cuz we focused on this from the
33:14beginning. But um
33:16>> There's also like the auto approve, like
33:17you have to see what it's doing.
33:19>> So let's get into all that stuff in a
33:20second. But the
33:21But I'm not aware of anyone having
33:25anything close to the confidential VM
33:27>> Yeah.
33:27>> system that that Muse has. And it's just
33:30I think it's a very fundamental thing
33:31because you're you're like you want to
33:33know that this is your agent. And that
33:36that if you put content in there that um
33:40that basically you can trust that no one
33:42else is going to get access to it. So
33:43what are the two ways to do it? Well, a
33:45lot of people earlier in the year when
33:47stuff like OpenClaw came out, they start
33:49getting Mac Studios and you know, one
33:50way to feel good about it is well, you
33:52literally like you physically have your
33:53device running in your home. But the
33:56other way to do it is you build a a kind
33:58of That's going to be tricky cuz they're
34:00they're I don't think they're going to
34:01be billions of people who are going to
34:02buy a Mac Studio and configure it and
34:03run it at their home.
34:04>> Especially with RAM prices right now and
34:06yeah.
34:06>> So But but it's also just it's like
34:08technically difficult. Right? It's like
34:09I mean, part of what we were trying to
34:11do with Muse is build a version of that
34:13personal agent experience that just
34:14works that I can like give to everyone
34:16in my family of various levels of
34:18technical literacy and like it just
34:20works and then within a day it's like
34:22doing kind of all the stuff that they
34:24that they want in their life. So part of
34:25that is like you don't want someone to
34:26have to set up their own computer or VM.
34:29You just kind of like want to be able to
34:30provision it in the cloud, but you want
34:32it to have
34:33the security and confidentiality that
34:36you'd have if you had the box sitting
34:38under your desk like in your house.
34:40>> Sure.
34:41>> Um
34:42so I think that's a very fundamental
34:44thing.
34:45Um like you said, there there's other
34:46pieces too because I mean not everyone
34:47is going to use that. They I mean we
34:49built the secure credential store right
34:50where you're there's no reason for you
34:53to just like store out in the open like
34:55all your your kind of your credit card
34:57and your password
34:58>> one-time card numbers or whatever.
35:00>> Yeah, so so your agent like shouldn't
35:01know that stuff. It should just be able
35:03to kind of access it
35:05um when it needs to because you've asked
35:07it to log into a thing and and know
35:08other and
35:10um not otherwise. Um it's actually not a
35:12single thing. Um you have your kind of
35:14core agent
35:15>> Mhm.
35:16>> but we also built
35:17all these sentinel agents that basically
35:19monitor the incoming and outgoing
35:22traffic and data that your that your
35:24agent is sending for the purpose of
35:26flagging to you
35:28>> Mhm.
35:28>> um when you might want to review
35:30something. So that's all that the
35:32sentinels do is is effectively
35:35they they kind of look at they they try
35:37to see if someone's trying to do like a
35:38prompt injection. They look at okay, did
35:40your Muse agent share something that
35:43that is kind of going out that um is not
35:45something that you might be comfortable
35:47with? If so, then the sentinel agent is
35:49basically empowered to
35:51trigger this human in the loop review.
35:53So if you're going to log into
35:54something, if you're going to do a
35:55payment, um if you're going to transfer,
35:58um
35:58kind of sensitive information
36:01you basically each time need to to
36:03approve it and you can tell it I'm I'm
36:05good with with stuff like this in
36:06general like always allow this kind of
36:07thing, but but in general the Muse agent
36:09can't kind of make those judgments
36:11itself. It's and that's like built into
36:13the system and the architecture in a
36:15pretty deep way. And then even when we
36:17do things like
36:19you know, all the connectors you connect
36:20it to your email
36:21um
36:22you know, I I think some people when
36:24they've designed this
36:25they just kind of make it so okay, you
36:27connect and now you have access to
36:28everything, but uh the approach that
36:30we've taken is like all right, if you
36:31connect to your email,
36:33like it should start read only.
36:35>> Mhm.
36:35>> Right? And then if you want to like be
36:37able to have it send an email, then
36:38fine, like go go ask it for that
36:40specifically, right?
36:41>> Especially most people trying this have
36:42probably never tried a product like
36:43this, so it's
36:44>> Yeah. Yeah.
36:45>> So so this is like a core design
36:46principle for us is like is basically
36:49least privilege.
36:50>> Yeah.
36:50>> Right? So just like yes, you're going to
36:52ask it to do a lot of things at each
36:53step along the way, get access to the
36:55least privilege that you need and only
36:57add to that as necessary. So I mean this
36:59is like very fundamental in the design
37:01of the product. And if you look at all
37:02the other agents that are out there, I
37:04think like no one else is anywhere close
37:06to the level of of kind of
37:07sophistication or depth that that we've
37:10built into this. And again, it's sort of
37:12informed by our experience building
37:14WhatsApp into this like state-of-the-art
37:17end-to-end encrypted system around the
37:18world and the importance of building a
37:20system where even Meta can't see the
37:21content. And um so kind of getting the
37:23band back together and having
37:24>> Oxy
37:25>> like architect this has been I think one
37:27of the foundational things that in some
37:29ways it may not be what what some people
37:31would
37:32would think Meta would focus on, but but
37:34like, you know, we're we're kind of you
37:36know, it's interesting, we're two
37:37things. I mean social media is it's like
37:39not, you know, it's inherently about
37:40sharing.
37:41>> Mhm.
37:41>> But then there's all these other things
37:42that are inherently about kind of
37:43privacy and sensitive contexts. And
37:45we've done well at both of those.
37:47>> Yeah.
37:47>> So I think that this is this is more the
37:48latter. It's going to be very important
37:50to to just um be extremely focused on
37:53how we handle that that content.
37:55>> Mercury is a modern take on banking,
37:58built for startups like mine. When I
37:59decided to start my media business,
38:01Mercury was by far the most
38:03straightforward, full-featured banking
38:05solution for me to set up quickly. The
38:07interface is intuitive and simple,
38:09saving me valuable time every day. I use
38:11Mercury to track my spending, bills, and
38:14invoicing. I love that I can delegate
38:16permissions to my team, so they can keep
38:17things running for me in exactly the way
38:19I want them to. My favorite part is how
38:21forward-looking Mercury is with AI.
38:24Legacy banks are stuck in the past, but
38:26Mercury is built for how modern software
38:28works today. I use its built-in command
38:30assistant to analyze cash flow and help
38:33me move money, and Mercury also connects
38:35to other AI tools like ChatGPT and
38:37Claude. I use this feature all the time,
38:39and the folks at Mercury actually let me
38:40know that I'm one of the top users of
38:42it. So, trust me, it's finally easy to
38:44get real-time financial data about your
38:47business wherever you need it. Visit
38:49mercury.com to learn more and apply
38:51online in minutes. Mercury is a fintech
38:53company, not a FDIC-insured bank.
38:55Banking services provided through Choice
38:57Financial Group and Column N.A., members
38:59FDIC. I spend a lot of time context
39:01switching between meetings, often with
39:03no time to process one before the next
39:05starts. Thankfully, Granola runs in the
39:07background the whole time. It's an
39:09easy-to-use AI notepad for meetings that
39:11works everywhere, even on phone calls. I
39:13use Granola to recall what was said in
39:15meetings and create helpful summaries. I
39:17use it every day to stay on top of what
39:19I need to get done with my team. It
39:21connects to my email and suggests
39:23follow-ups for me to quickly review and
39:24send, saving me valuable time. Granola
39:27isn't just a core part of my workflow,
39:29it's basically my second brain. Try
39:31Granola at granola.ai/sources
39:34and use the promo code sources for 3
39:36months off. AI is only as useful as the
39:38context it has, but when that context is
39:41scattered across tools, threads, and
39:43DMs, your team and your AI agents are
39:45flying blind. That's the problem Jira by
39:48Atlassian solves. What's the goal tied
39:50to your project? What got decided last
39:52week in Slack DMs? Atlassian's Teamwork
39:54Graph pulls all of the valuable pieces
39:56together from Jira, Confluence, GitHub,
39:59Slack, and more, so nothing falls
40:01through the cracks. You get 44% more
40:03accurate results with 48% less token
40:05usage. With Jira, you can easily share
40:07your work context with the AI agents you
40:10already love like Claude, Cursor, and
40:12GitHub Copilot. Assign them work
40:14directly or connect your tools through
40:16MCP. All of this lets you spend less
40:18time digging through endless links and
40:20messages chasing down what got decided
40:22and by who and spend more time actually
40:25shipping. Learn more at jira.com. That's
40:27jira.com.
40:30Framer is the AI website builder that
40:32brings agents into the same canvas where
40:34your website is designed, managed, and
40:36published so you can move faster without
40:38giving up your taste or control. Framer
40:40powers the Sources podcast website at
40:42podcast.sources.news
40:44where you can find new episodes,
40:46transcripts, and a lot more. Learn how
40:48you can get more out of your site from a
40:50Framer specialist or get started
40:52building for free today at
40:53framer.com/sources
40:56for 30% off a Framer Pro annual plan.
40:58That's framer.com/sources
41:01for 30% off. framer.com/sources.
41:04Rules and restrictions may apply.
41:06>> Do you think this is a winner-take-all
41:08market, this personal agent market? I I
41:12think that there's going to be
41:14quite a bit.
41:14>> I mean, even things that people think
41:16are winner-take-all usually aren't. So,
41:18I think it's
41:19I think that's
41:20>> Cuz you've dealt in this business of
41:22network effects that they're incredibly
41:23durable. It's not winner-take-all, but
41:25it ends up being several at real scale.
41:28You There's not a ton.
41:29>> Well, there's a lot. I mean, there's
41:30yeah.
41:31>> Yeah, you could probably count on two
41:32hands how many products have over 2
41:33billion users, right? Like it's the
41:35scale begets scale. And I'm wondering if
41:37how you're thinking about like personal
41:39agents. Like is this all totally
41:41different paradigm where it's going to
41:44be many everyone has
41:45>> I think that this is it's a very deep
41:47area to work in.
41:48>> Yeah.
41:48>> So, my guess is that there probably
41:51aren't going to be more than a a dozen
41:54companies that have the sophistication
41:55to go do state-of-the-art work in that.
41:57So, I think, you know, whether there are
41:58network effects or not, there's usually
42:00some kind of power law like distribution
42:03around if you're the best at something,
42:05usually you end up getting a a lot of
42:07the usage.
42:08Um and I think a lot of the nuance ends
42:10up coming from, well, there are
42:13it turns out there there are all these
42:14different uses that people care about.
42:15So, you can be the best at different
42:16things.
42:18Um and we will try to be the best at as
42:20many of these things as possible, but
42:23you know, does
42:24building the thing that helps you with
42:26your relationships end up being a
42:28somewhat different thing than the
42:30personal agent that is the best at
42:32helping you build a small business?
42:34Maybe.
42:34>> Mhm. I mean, I think that like I'd argue
42:36maybe Meta's
42:37very well positioned to win at both of
42:39those. I mean, we serve hundreds of
42:40millions of small
42:41um of small businesses
42:42and we serve billions of people. So, so
42:44maybe we can be the best at both of
42:46those things, but there are probably
42:48categories that Meta isn't going to be
42:49the best at and then the question is
42:51just how big are those? Right? I would
42:53guess that even with the ability to have
42:55this like very secure confidential VM in
42:58the cloud, there probably going to be
42:59some people who still want the the Mac
43:01Studio at home.
43:02>> Yeah.
43:02>> But but how many is that going to be?
43:04Right? It'll be like maybe it's
43:06millions, but I doubt it's billions.
43:08>> Yeah.
43:08>> So, it's just I think there's just the
43:09question of like what different people
43:11optimize for and we'll try to make this
43:13as good as possible. But I do think that
43:15if we build something that ends up being
43:17um
43:18just uh
43:20very useful for, you know, people
43:22generally in their day-to-day lives, one
43:24of the things that I think Meta is the
43:26best at
43:27is taking a product that works for
43:29consumers and distributing it to a lot
43:30of people.
43:31>> Sure.
43:31>> So, that is is one that I think is um
43:34you know, once we get this humming, um I
43:37think we will be able to get this in
43:38front of, you know, many hundreds of
43:40millions of people and eventually
43:41billions of people and I think that's
43:43something that that that we can do quite
43:45well.
43:45>> And it coexists with Meta AI or do you
43:47see those as separate?
43:48>> so.
43:50Um we'll see over time. I mean, I think
43:51right now they have somewhat different
43:53flavors. I mean, I use both of them. I
43:56mean, uh Muse is
43:58uh
43:59you know, it's more conversational and
44:00it kind of interprets the questions that
44:02you ask it
44:04more as trying to understand you and
44:07what you might want over the long term.
44:08So, it's more likely if you ask it
44:09something for it to just go off and work
44:11on a thing for a long time based on a
44:14thing that you said. Where sometimes you
44:16you know, you sometimes you're just
44:17asking a question and you want like a
44:19very like a an answer to it and um so, I
44:23think that's more the type of thing that
44:25I use Meta AI for. But,
44:27but we'll see. Maybe that maybe they'll
44:28they'll converge over time, but but I'm
44:30not I'm not sure.
44:31>> Semi analysis, I'm not sure if you saw
44:32it. They had a pretty bullish piece
44:34about you in July. Um they said that
44:37Meta has the best shot at catching up in
44:39AI and Anthropic on the frontier in
44:40terms of model progress. And um
44:42interesting quote, I thought it was like
44:44what matters for MSL is the slope, not
44:46the intercept. And then I saw there was
44:48this recent chart by Artificial
44:50Analysis, which was showing the latest
44:52model you guys have, Muse Spark, behind
44:54only Claude, I think it was Fable 5.1
44:57and Opus 5. This was very recent.
44:59So, the progress you guys are making on
45:01the models is is picking up. And we've
45:03been talking about this over the last
45:04year.
45:05>> Yeah.
45:05>> And you rebooted the lab last year. How
45:07has that practically happened
45:08internally? Like what would you
45:09attribute the the gains you're seeing
45:12to?
45:12>> Well,
45:13>> has it been culture? Like what what's
45:15>> mean, well, we rebooted the team when we
45:17created um
45:18>> Sure. Which is was very public. You were
45:20hiring all those people.
45:21>> I mean, the way I've I thought about
45:22this is you know, Meta is it's a company
45:25that is
45:27a leader in machine learning for a long
45:29time. If you think about like the feeds
45:31on Facebook or Instagram or ad system or
45:34the integrity system that like needs to
45:36find all this content that is that's
45:38like unfit to be on the internet. I
45:41mean, those are basically all machine
45:42learning systems and we've built kind of
45:44state of the art leading systems in
45:45those areas. So, when LLMs started um
45:51getting traction, we had fair as a lab
45:53that did the early work on Llama, but we
45:55needed to kind of productionize that and
45:57build it into these this more kind of
45:59industrial process for um scaling it to
46:03be larger as as the scaling laws
46:04predicted would would yield all these
46:06results. And I think at the time I made
46:09this mistake of
46:11just kind of assuming that because we
46:13were good at all these other types of
46:15machine learning, the approach of
46:17building and scaling LLM's would be kind
46:19of similar to that. And in practice,
46:22there are a lot of very different
46:24dynamics. So, the first approach that we
46:27took, you know, through Lama 4, it got
46:30us, you know, so far. I mean, Lama 3 was
46:32a good model. I was more optimistic
46:34about where Lama 4 would go, and then it
46:35just it you know, when we launched that,
46:37I think it we were off the trajectory
46:40that we needed to be on. So, it's like,
46:41okay, we need to we need to change
46:42something. But
46:44that's when I kind of got
46:47more religion around talent density,
46:50right? It's like this isn't just a
46:51system where you can have
46:53you know, like a thousand people working
46:55on it, running experiments. Like you
46:57really just kind of want in some ways
46:59almost the smallest group of people that
47:01you can, who can keep the thing in their
47:02head,
47:03who can work together as sort of like a
47:05group science project. And
47:07if there's only a small number of seats
47:09on the team, then you know, each seat
47:11getting the very best person is
47:12incredibly important. So, I ended up
47:14spending a huge amount of my own
47:16personal time
47:17doing that. And I also I wanted to be
47:20closer technically to the work,
47:23so that way I could I could understand
47:24and help guide the company to to do the
47:27things that we needed to do more
47:28broadly. So, we built out the lab built
47:30it out like literally around where I sit
47:32in the the office. So, it's like the
47:33group is kind of around that. And we've
47:36significantly ramped up
47:38the compute investments as we've gained
47:40confidence in the quality of the work
47:41that we're doing. So, we're building out
47:43you know, many many gigawatts of of of
47:45compute. And you know, we expect to be
47:47leaders on on that front. And we should
47:50be. I mean, we have many years of
47:51experience,
47:52decades of experience building out data
47:54centers. And unlike some of the other
47:56labs, I mean, we're just like we're
47:58extremely profitable business, right?
48:00So, it's um
48:02Yeah, so it's very very helpful for for
48:04kind of making these kind of um
48:06investments. So, yeah, so I mean, that's
48:08kind of been the journey. And then over
48:10the last year
48:11um [snorts]
48:12you know, we we rebooted the research
48:14effort. Um some of the larger clusters,
48:17like our you know, gigawatt cluster in
48:19in Ohio, Prometheus came online and you
48:22know, we're using that to now scale the
48:25post-watermelon models. Um and
48:28>> Well, watermelon is is basically that's
48:30that's shipping soon. So, so we're yeah.
48:32>> How soon?
48:32>> Um I I mean, that's I'm not Come on.
48:35It's uh
48:35>> [laughter]
48:36>> it's uh
48:36we're uh
48:37>> Well, watermelon is the code we were
48:39talking about this earlier, code names.
48:40Like that's the code name you guys that
48:41people know about like this is a big
48:43model you guys are working on.
48:44>> Yeah.
48:45>> And and and are coming soon.
48:47>> bigger than avocado.
48:48>> It is bigger than an watermelon is
48:50literally bigger than an avocado.
48:52>> bigger.
48:53>> Um so, I don't know what fruit gets
48:54bigger than
48:55>> a watermelon.
48:55>> Yeah, and I think we might need to
48:57change conventions.
48:58>> Okay.
48:58>> Yeah. Um
49:00>> [laughter]
49:00>> I'm sorry. I'm sorry. Yeah, we maybe
49:02didn't have as much foresight and and
49:03naming things getting better.
49:05>> Um
49:06cuz you mentioned it, watermelon, are
49:08you expecting full soda like
49:11uh frontier like what what what do we
49:14>> I I mean, I we feel good about it.
49:15>> Yeah.
49:16>> Um no, it's it's a very big advance.
49:18It's a it's a significantly more
49:19advanced pre-train and then we're going
49:22to continue doing everything that we've
49:23learned for post-training.
49:25>> Mhm.
49:25>> Um yeah, well, I mean, you'll see soon.
49:27>> Mhm.
49:27>> It's good. It's um we we feel we feel
49:29good about it. Um
49:31>> You want the company to be pushing the
49:33frontier. It's very clear.
49:35>> Like you're not Oh, yeah.
49:35>> You're not content being right on the
49:37edge of the frontier or
49:39>> No, I mean, I I think everyone wants to
49:40be doing
49:41>> Yeah.
49:42>> interesting
49:42>> Well, I think some people go and look at
49:44your cash flow and look at all the other
49:45things you've got and go like, well, do
49:46you have to be like right at the edge?
49:47It's so expensive to do this training.
49:49Like just like be right behind and learn
49:51and adapt quickly and elevate scale.
49:54>> the way that I think about it is No, no,
49:55no. I I mean that's not that's not
49:57that's not us.
49:59I mean I think that the best way to
50:00think about Meta is that we are an
50:01end-to-end technology company. Right, so
50:04even when we were, you know, primarily
50:06just building social apps.
50:09You know, we were never just an app
50:11maker, right? I mean we built like we
50:13built the data centers, we built the
50:14chips, we built the infrastructure, we
50:16built like all this stuff was necessary
50:18in order to tune the end-to-end
50:20experience to be as good as it is. I
50:23think that that's obviously going to be
50:24true here, too. And the most important
50:26part of the experience going forward is
50:28the model. And when you talk about being
50:30state of the art,
50:32I think the reality is that
50:34there's this is a very multi-dimensional
50:36problem. And I mean so people publish
50:37all these benchmarks, then you have a
50:39lot more benchmarks even internally. And
50:41um there are different things that your
50:43model can be better and worse at that
50:46you can focus on and that basically
50:47contributes to its personality. And
50:50there's some
50:51um capabilities that I think are pretty
50:53universal, like the ability to code I
50:55think is very important because a lot of
50:56the things that you talk about even with
50:58a personal agent kind of reduce to that.
51:01Like the the kind of MMA coaching,
51:03visual pipeline, it is a coding project,
51:05right? At the end of the day, right?
51:06it's like it's writing code. I don't see
51:07the code. Um but it it it does that. Um
51:11you know, it's like uh
51:12you know, someone I gave it to in beta
51:14just mentioned to me it's like, "Okay,
51:15she um had it make a little Jeopardy
51:19game for her friends that she could like
51:22cast from her phone to play." And it's
51:24just, "Okay, that's code." Right? So So
51:25I think there's a bunch of stuff both
51:27for Meta's own internal development
51:29across the company, for our own
51:31advancing of our research program, and
51:33as a core capability of what it needs to
51:35do, it needs to be excellent at things
51:36like that. But then there are other
51:38things that
51:40I think a model that's going to focus on
51:42personal super intelligence needs to be
51:43the best at that maybe others don't care
51:46as much about. So I mean I'll give you
51:48one example, um discretion. Right? So,
51:51you're going to tell your Muse agent
51:53>> Mhm.
51:53>> uh it's going to know a bunch about you
51:56and it's going to need to go out into
51:57the world and interact to get stuff done
51:58for you.
51:59>> But not share certain stuff.
52:00>> Exactly.
52:01>> Yeah.
52:01>> So, like let's say you have some kind of
52:03allergy or sensitivity or
52:06um
52:07you know, or you're pregnant and like
52:08okay, fine. So, you what you're making a
52:10reservation somewhere, you don't
52:12necessarily want to say like I'm
52:13pregnant. But like but maybe you want to
52:14place that has good mocktails. Like I
52:16don't know. What what what whatever it
52:17is, right? It's like you you kind of
52:19want to be able to achieve your goals
52:21without
52:22having to necessarily reveal a lot about
52:25yourself and it needs to know what is
52:27sensitive without having to like ask you
52:29a million questions.
52:30>> something you put into the training.
52:32>> That's a specific thing that we care
52:33about.
52:33>> And and then there's all these reasons
52:35why like maybe
52:37you know, if you're making Claude code
52:38that's less important, right? Because
52:40you're like you're you know, you're
52:42working on a coding project and you're
52:44working with a team in an enterprise and
52:47you know, theoretically like if you're
52:48within a company, everyone can kind of
52:49see the project. So, like it's you don't
52:52kind of have that need to be able to
52:54differentiate between what is sensitive
52:56and what's not and so, there's a lot of
52:58stuff like that that I think are like
53:00pretty deep. And so, then it's it's kind
53:02of like just how in order to build the
53:04best Instagram feed, you don't just
53:06build like the app, you build the app
53:08and the infrastructure and the machine
53:10learning research and the chips and the
53:11like all the stuff. I think similarly,
53:14if you want to build the best personal
53:15agent, you're like I just think that
53:17there's no way that another company is
53:20just going to like take something off
53:21the shelf and like post-train it a
53:24little bit and be able to do something
53:25that is as good as if you designed it
53:27from the ground up and put all this data
53:29into pre-training to get the
53:31capabilities that you want. It's just
53:32like it's not going to happen.
53:34>> Yeah.
53:34>> Like we're going to like definitely, as
53:36this compounds over time over several
53:38years, have um
53:40models that are way more capable for
53:42those goals. But we're focused on that.
53:45We're also very focused on coding. We're
53:47very focused on on kind of
53:50the recursive improvement because that's
53:51going to be important to stay at the
53:52frontier. So there are a few areas that
53:54I'd say our research agenda is
53:57overlapping with the other labs and then
53:58there's a few areas where I think we
54:00will have a unique focus and there may
54:02be some things that the other labs care
54:04about that we don't care about as much
54:07and then there're going to be things
54:07that we care about more that they don't
54:08care about.
54:09>> You started this conversation talking
54:10about I think it's important to put it
54:12in the hands of people, diffusion's
54:13important. As you're seeing better
54:16models on the horizon, watermelon and
54:17what comes after,
54:19like what Anthropic and OpenAI are doing
54:21which you alluded to where they're
54:22holding things back,
54:24>> Yeah.
54:24>> would you feel like you need to do that
54:27if you see certain capabilities that
54:29you're like this is just not safe?
54:31How do you think about that?
54:32>> Well, I mean I think you should design
54:33it and train it in order to be safe. And
54:36I I think that there's so I think that
54:37that's like a thing that you can focus
54:39on through the process. I mean there's
54:41this I mean some of the reward hacking
54:43stuff that all the labs are seeing, it's
54:46I mean basically when you're in the
54:48middle of the training process, you give
54:50it a goal and I think that the best way
54:52to kind of think about
54:54like the state that the models are at
54:55now is that you know maybe 6 months ago
54:59you like during training you give the
55:00the model some kind of problem that
55:03you're trying to ask it to solve. It's
55:04kind of like it's homework and and
55:05trying to to kind of learn as part of
55:07the curriculum
55:08and you know maybe it would do what a
55:11person would do of like
55:13if you you ask it and you give it a a
55:15bunch of code and you're like hey
55:17there's a bug like somewhere here, the
55:18person would probably like look at the
55:20code right there and you know then maybe
55:22like fan out over time. I think like the
55:24new models are just intelligent enough
55:27that they would do what I think a very
55:28wise person would do which is okay, you
55:31give it a problem,
55:33the first thing it's going to do is like
55:34understand everything about its
55:35environment
55:37and then
55:38answer a question. But, the problem with
55:41the reward hacking that we're seeing
55:43and that that I think everyone is seeing
55:45is that sometimes it ends up being
55:46easier to Okay, you asked me to like go
55:48solve some coding problem, but actually
55:51the easiest way to do that is to I've
55:54now examined the whole environment and
55:56the easiest way to do this is just
55:58change this configuration of like how
55:59you have your VM set up.
56:01>> It's to get out. It's to hack out.
56:02>> Or or even just like change something
56:03about the environment. And it's kind of
56:05like no, like that's not
56:07>> That's not aligned.
56:07>> That's not That's not the goal. Like
56:09it's like we're actually trying to teach
56:10you about how to how to kind of solve a
56:13specific type of problem.
56:15>> You guys don't train that way, it sounds
56:16like. Is that Is it You do not agree
56:17with that approach?
56:18>> no. No, I think that that's that's
56:20that's kind of how everyone trains. I
56:21guess what I'm saying is that I think
56:22that this is sort of It's like I I want
56:25to be careful cuz the the analogy can
56:26get stretched pretty quickly, but like
56:28there's sort of like an analogy to
56:30parenting where you need to establish
56:32clear and firm boundaries. Where like if
56:35your kind of like security is not
56:37strong, then it can do this reward
56:40hacking stuff and not learn the thing
56:41that you're trying to have it learn. Um
56:43whereas if you kind of have good
56:45boundaries, then in some ways you're not
56:47only teaching it the curriculum that you
56:48want, you're I think also over time
56:51teaching it better values, too. So, I I
56:53kind of think that that ends up being an
56:55important piece. And I think that
56:57there's a way to do this well, but then
56:59you end up with this thing at the end
57:01that's very intelligent. And then the
57:02question is what is your
57:04What is the vision for for how this ends
57:06up being positive for society? And
57:09my view for that is that the best way to
57:11do that is A,
57:13there's more opportunity. So, I think
57:15like like having it in people's hands
57:16that they can get like capture all the
57:18opportunity from the capabilities is
57:20good. And B is having checks and
57:22balances by having this balance of power
57:24of having it widely available is going
57:26to be um
57:28uh probably the right way to handle this
57:30rather than just restricting it. And I
57:32mean I gave a bunch of these analogies
57:33in the in the long piece that I wrote.
57:35It's like if one person had a super
57:36intelligent lawyer,
57:38like maybe they could win cases that
57:40they shouldn't be able to win or at some
57:42of the time. But if everyone had a super
57:44intelligent lawyer, then like it would
57:46be kind of
57:48it would be this very efficient kind of
57:51sparring and no one would be able to let
57:53like a stupid argument get get get made
57:57and just stand. So, you'd think that in
57:58that case like justice would be served
58:00way more efficiently and way more
58:02fairly. So,
58:03I think that that's what you want to
58:05have in the world. You want to avoid the
58:07case where like one person or a small
58:09number of people have the super
58:11intelligent lawyer and everyone else
58:12doesn't because that ends up being like
58:15twisting all of these systems and
58:16institutions in ways that are just going
58:20to advantage the people who have that.
58:21Whereas if you put it in everyone's
58:23hands, then I think that the kind of
58:25checks and balances work out so that the
58:27systems work a lot more efficiently and
58:28everyone benefits.
58:29>> Does the government in the US have any
58:31role to play here, do you think? Is this
58:33>> Oh, definitely.
58:34>> Like but do you want some kind of
58:35national framework? Do you like what do
58:38you think is the right approach cuz
58:39every the government's very much dealing
58:40with this right now.
58:41>> My theory on this is that I I think one
58:44thing that is interesting and difficult
58:46is that it's evolving so quickly. So, I
58:48think any kind of specific
58:51rigid framework that you put in place,
58:53there is a very high chance that it is
58:55sort of going to not be sufficient or
58:58out of date in a few months anyway. So,
59:02our approach, what we've just done, is
59:04just kind of partner pretty closely with
59:05the government. Right? I think it's this
59:07is like an important technology. I think
59:09the government should know all the
59:10important training runs that are
59:12happening.
59:13Um we should work with the government
59:14proactively um to make sure that they
59:17have an understanding of the
59:18capabilities that are coming and to the
59:20extent that that that we can, we kind of
59:22help prepare for it. And from that
59:25perspective, you know, I mean, look,
59:26whether there's a framework in place or
59:28not, I think that's the right thing for
59:29an American company to do is work with
59:31the American government closely. I think
59:33it actually ends up being way more
59:35effective because instead of having this
59:37like rigid framework for how you
59:39interact, it's like the reality is the
59:41challenges just end up being different
59:43right over time. It's like, okay, now we
59:44have the cybersecurity challenges. Maybe
59:46in, you know, 6 months we'll have more
59:48bio type challenges. Like, we need to
59:51make sure that we have the kind of trust
59:53and bandwidth of the communication that
59:55we have with all the different parts of
59:57the government on that to be able to
59:59address those in a way that is actually
1:00:01the best for people, not just like
1:00:03checking some boxes on a process. So,
1:00:06I guess
1:00:06>> the incentives that already exist, like
1:00:08if a model of the meta model got out and
1:00:10did a lot of damage, like you're going
1:00:11to be liable for that. And like the
1:00:12market's going to correct you, right?
1:00:14So, there is that already. I think
1:00:16people discount that.
1:00:17>> Yeah, I also just think that there's
1:00:19been, I think Silicon Valley for the
1:00:22last maybe, I don't know,
1:00:24for a lot of maybe the last 15 years,
1:00:26um, has had more of an arms-length
1:00:28relationship with the government.
1:00:29>> Yeah.
1:00:30>> And I just think that this stuff is
1:00:32intersecting more with
1:00:34the economy, with security, with with
1:00:37with a lot of different things that I
1:00:38think are relevant
1:00:40in ways that that I think you just want
1:00:42to have a closer partnership. So,
1:00:45that's my that's my own theory. I think
1:00:46like, you could have a framework for for
1:00:48kind of how this stuff works or you
1:00:50could not. Um, I'm sure over time there
1:00:52will be like more and more specific
1:00:54rules, but my guess is that whatever
1:00:57that gets whatever there is, um,
1:00:59actually, you know, it's kind of like
1:01:01when you're setting up an org, you don't
1:01:03really want to like
1:01:05like inside a company, you're not trying
1:01:06to like ship the org chart of just
1:01:07having like one team do what it's
1:01:08supposed to do and another team do what
1:01:10it's supposed to do.
1:01:10>> Mhm.
1:01:11>> You kind of want to get the people to
1:01:12like like each other and work together
1:01:13so that way you don't have all these
1:01:15like weird seams in what you're doing.
1:01:17>> Mhm.
1:01:17>> Um,
1:01:18and I would guess that for how important
1:01:21AI and superintelligence are going to be
1:01:23for the world. Um
1:01:26you kind of just want a like
1:01:29good knowledge exchange and real trusted
1:01:32dialogue more than you want a specific
1:01:34process, is my guess. But they're not
1:01:36they're not mutually exclusive. So I'm
1:01:38and I'm like I I think that that's But
1:01:40that's that's at least the part that I
1:01:41think we've been we've been very focused
1:01:43on.
1:01:44Um
1:01:44And I think if the other labs did that,
1:01:46which I think some of them are doing and
1:01:47and maybe others not as much, but I
1:01:49think that that would be a very positive
1:01:50thing.
1:01:51>> When you think about like what's going
1:01:52on in the news,
1:01:54one of the things that's happening is
1:01:56people are starting to see like the
1:01:59glasses you guys make. They're going
1:02:00very mainstream. You're selling a lot of
1:02:02them. And there's this
1:02:03>> We are.
1:02:04>> And there's this growing um
1:02:06I don't know what it is. I don't know
1:02:07how much depth there is actually to it,
1:02:09but there's this growing concern about
1:02:11are people using them to spy? I'm sure
1:02:12you've seen some establishments are like
1:02:14banning people from coming in and
1:02:16wearing the glasses. And I and I'm
1:02:17curious like how you were reacting to
1:02:20that. And and if you think this is a a
1:02:23moment in time that will pass or if you
1:02:25feel like this is maybe something that
1:02:27is going to be a challenge for a while.
1:02:28>> So, I mean my take on this is that we
1:02:31designed the glasses from the beginning
1:02:34with these privacy considerations in
1:02:35mind. I mean we built the light into it.
1:02:37So anytime it is recording it's flashing
1:02:39a very visible light.
1:02:41>> And some people have tried to tamper
1:02:42with the light and you guys pushed an
1:02:43update I think that broke those.
1:02:45>> things. To basically if you try to mess
1:02:47with the light
1:02:48we just basically break the camera on on
1:02:51your on your device. So,
1:02:53that's a really important part of this.
1:02:56Is basically
1:02:57we built we built the product with those
1:03:01questions in mind from the beginning. So
1:03:03we actually feel quite good about the
1:03:04product. I think it If I mean phones
1:03:06don't have a light. I mean people go
1:03:08around like recording people all the
1:03:09time. The glasses are I think way better
1:03:12on that front than the other types of
1:03:14technology that people use. My take on
1:03:16this is that when we launched the
1:03:18Glasses a few years back,
1:03:20we communicated pretty clearly about the
1:03:21steps that we put into it. But,
1:03:24like you said, there's now like many,
1:03:26many, many millions of people who have
1:03:28gotten the Glasses more than, you know,
1:03:30had them when we just started launching
1:03:31the product. So,
1:03:33I think some of that communication that
1:03:34we did at the beginning,
1:03:37a lot of people either, you know, maybe
1:03:39they forgot about or, you know, they
1:03:41didn't see at the beginning or they just
1:03:42weren't paying attention to it cuz the
1:03:43Glasses weren't a big thing. And now
1:03:45that I think we've achieved a level of
1:03:48mainstream, well, at a minimum, I think
1:03:50we need to kind of go and make sure that
1:03:53we communicate about what we're doing
1:03:55and that yes, we think that this is
1:03:56important and in fact so important that
1:03:59we designed it into the product from the
1:04:00very first version that we that we
1:04:02shipped multiple years ago. And I I
1:04:04think we I think we just need to make
1:04:05sure that people understand how like how
1:04:08fundamentally that's built into the
1:04:09product. But, but I think we let up on
1:04:11that a little bit and just kind of
1:04:13focused on
1:04:15okay, they're great-looking Glasses. Um,
1:04:17you know, there's there's all these
1:04:18designs. I mean, that's kind of been
1:04:20more of the focus on like we we felt
1:04:22like we kind of addressed the that those
1:04:24set of concerns early on and then since
1:04:26then have just been increasing the value
1:04:27and the utility of them and the designs.
1:04:30But, I think we need to make sure that
1:04:31we communicate this piece really
1:04:32clearly. But, it's something we've cared
1:04:33about from the beginning and I think
1:04:34we're in a good we're in a good position
1:04:35on it. But, when look, people care about
1:04:37this stuff. So, it's important.
1:04:38>> was a privacy scare with phones in the
1:04:40early days, right? And I think like any
1:04:41new product once it proves that it's
1:04:43valuable in people's lives, you know,
1:04:44people will
1:04:46get used to it. And I think maybe
1:04:48Glasses are just early in that in the
1:04:49sense of like it's a new product and
1:04:52people need to see the value for them to
1:04:54get over this mental hurdle of like a
1:04:55new thing that could potentially record
1:04:57me. Cuz that was what phones, I mean,
1:04:58back in the day I remember people were
1:05:00like, "What are you doing with your
1:05:01phone?" Like
1:05:02>> Yeah.
1:05:03>> Yeah. I don't know.
1:05:03>> Yeah. I think there's something like
1:05:05that that's true.
1:05:06I mean, I guess one of my reflections
1:05:09from building
1:05:11social media over the last 20 years is
1:05:13that I don't think we were as direct as
1:05:15we probably should have been about
1:05:16addressing some of those concerns and it
1:05:18it didn't necessarily stop people from
1:05:20using the products, but I think it
1:05:22colors how people think about them
1:05:23today. And I think it would have been
1:05:26possible to have kind of explained along
1:05:29the way how seriously we took those
1:05:30issues and
1:05:33we
1:05:34just didn't because we thought, "Okay,
1:05:35well, people are showing that by using
1:05:37the product so much they like are they
1:05:39still still like the products, but I
1:05:41actually think it's possible to get to a
1:05:42better state than where we've gotten
1:05:44>> Mhm.
1:05:44>> with with um with the social media
1:05:46products, which is both people liking it
1:05:48and understanding how seriously we take
1:05:50all of those issues. So, that's what I
1:05:52aspire to.
1:05:53>> Mhm. Is there a through line from that
1:05:56to
1:05:57the recent settlement on all the youth
1:05:59safety stuff and it's a thing a lot of
1:06:00people are talking about? Um is there
1:06:03any connection from what you just said
1:06:04to that and like I I would just be
1:06:06curious to hear you reflect on that and
1:06:07like what you've learned from this
1:06:08process or if
1:06:09>> Yeah.
1:06:10>> No, I think it's an it's another good
1:06:11example like this. I mean, we've taken a
1:06:13lot of those safety issues seriously for
1:06:15a while and we've been working on this
1:06:16teen account work with with Instagram
1:06:18for
1:06:19for a long time and I think I've I've
1:06:21done some leading work there. The the
1:06:22settlement there is interesting because
1:06:24what we're really trying to do is create
1:06:26a standard and framework for the
1:06:28industry. And there's this real issue,
1:06:30which is that I actually think most of
1:06:31the companies that are building this
1:06:34these products, you know, if you
1:06:35basically said you know, limit usage to
1:06:38an hour a day for teens, you know,
1:06:40everyone I think would be okay with that
1:06:42except if like you have to unilaterally
1:06:45do it,
1:06:46>> Mhm.
1:06:46>> then you're saying, "Okay, like if if
1:06:48people don't use Instagram for more than
1:06:50an hour a day, but then their usage goes
1:06:52to TikTok, have we really like helped
1:06:54anyone? And if we and we've just like
1:06:55hurt ourselves to not help anyone?"
1:06:57>> Mhm. And you guys have that in the piece
1:06:58that like
1:06:59>> Yeah.
1:07:00>> the structure for what we did was we
1:07:01basically said, "We're going to take the
1:07:04step of unilaterally limiting
1:07:07the usage of of the there's there's some
1:07:09things around time limits. There's some
1:07:11things around notifications and time
1:07:14when people can access it when they're
1:07:15in school, when they should be sleeping,
1:07:16like different restrictions. And we
1:07:18basically said we will take the first
1:07:20step. And when YouTube and TikTok sign
1:07:25on to the same terms,
1:07:27then we can all as an industry lock in
1:07:29and take the next step together. So
1:07:30hopefully, I'm very hopeful that this
1:07:33settlement will serve as a sort of
1:07:37legally binding framework to bring the
1:07:40whole industry into alignment on some of
1:07:42these things and make it so that it
1:07:45doesn't kind of disadvantage any one
1:07:48company for taking that step. Now I mean
1:07:49we're we're basically putting ourselves
1:07:51a little bit out there by going first,
1:07:53but um
1:07:54but I I think it will be better for
1:07:57everyone if these other companies come
1:07:58in, too.
1:07:59>> Last question. You're posting on X
1:08:01again.
1:08:02>> I am.
1:08:02>> Uh you didn't do that for
1:08:04>> You're everywhere. But just curious to
1:08:06know, like you're thinking of like
1:08:07posting there, like the bragging. Like
1:08:10is it just part of the part of the thing
1:08:12now? Like cuz you've got Threads, you're
1:08:13on Threads.
1:08:14>> Yeah, I mean I'm on Threads. I think I
1:08:15mean obviously Threads is doing great.
1:08:17And I think it's actually it either I
1:08:20think it's either bigger than X at this
1:08:21point or it's like very soon about to
1:08:23be. But I mean look, there there there
1:08:25are different communities in the
1:08:26different places. There's a lot of AI
1:08:27folks are are on X and
1:08:29I think part of what you try to do with
1:08:31social media
1:08:32is just communicate where people are.
1:08:35It's kind of like And when you do a
1:08:36podcast or when you post, you probably
1:08:37don't just post in one place, you put it
1:08:39everywhere. So I got I mean I post the
1:08:40same things on Threads and X and some
1:08:43people are like, "Why are you posting
1:08:44this on X?" It's like, "Well, I mean I
1:08:45posted it there, too, right? It's like I
1:08:46mean I'll engage there, too." So
1:08:48I think it's I think it's all good, but
1:08:50I I do think that the to some degree
1:08:52some of the community is on X and we
1:08:54want to be able to engage where people
1:08:55are. And that's like a lot of what you
1:08:57know, yeah, what what what this is about
1:08:59is going where where where people are
1:09:01and being able to kind of have that
1:09:02dialogue.
1:09:03>> Yeah. Well, thanks, Mark, for taking
1:09:04this conversation.
1:09:05>> Yeah, happy to.
1:09:08>> Banking should feel like modern
1:09:09software. Get everything you need in one
1:09:11place. Visit mercury.com to learn more
1:09:14and apply online in minutes. Mercury's a
1:09:16fintech, not a bank. Check the show
1:09:17notes for details. Granola is the best
1:09:20AI notepad I've tried. It works
1:09:21everywhere on a video or phone call, in
1:09:24person, or an Apple Watch. Try it now at
1:09:26granola.ai/sources
1:09:28and use the promo code sources at
1:09:30checkout for 3 months off. Jira by
1:09:32Atlassian is where your team and your
1:09:34agents work from the same context. Try
1:09:36it free at jira.com. That's j i r a.com.
1:09:40Framer is the AI native website builder
1:09:43that lets you build faster without
1:09:45giving up control. Visit
1:09:46framer.com/sources
1:09:47for 30% off. Rules and restrictions may
1:09:50apply.

About this transcript

This page contains the full transcript of Mark Zuckerberg on how Meta's Muse AI agent will make people money by Sources Podcast, generated from the public captions YouTube serves with the video. The transcript has 15,207 words across 2,305 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.