Mark Zuckerberg on the AI bubble and Meta's new display glasses | ACCESS

25:46So that's actually an important part of the technology is light leakage is a feature of some types of waveguides. And so basically you get these tradeoffs where you want them to be very efficient. There are waveguides where you have to pump just a ton of light through them in order to get anything to show up. But then some waveguides have just different artifacts. artifacts and usually in a bad way. It's like the light will catch them and you'll see all kind of like rainbowing or something. Another artifact that we think is pretty bad because it's a privacy issue is if the person who you're looking at can basically see it.

26:28Yeah. Well, the very worst version of it would be if they could see what you're seeing. But I think another version that I think is still not that socially acceptable is if they can see that you're looking at something at all yeah right um so i i think that that's one of the things that we're really proud of in the design here and that that we put a lot of energy into is you know the the uh the displays are super bright to you and to the person that you're looking at they can not even really tell that you're doing anything i mean that that's an important thing for having it to be socially acceptable right i mean we we also we design them so that way when when the display comes up it's offset slightly to the side we don't want to block what you're doing an important principle for the glasses is the technology needs to get out of the way all right it's i mean fundamentally it's like you know this is something that you're going to be wearing on your face for a lot of the day i mean we designed these specifically to be um you know both indoor and outdoor that with the transition lenses they work really well sunglasses outdoors um but the reality is is that most of the day you're not going to be using technology or at least a visual yet right so like maybe maybe you'll be listening to music or something but um but we want you know when you are interacting with something it should show up it should kind of be off to the side if you don't interact with it it needs to get out of the way really quickly that's like a really important principle of the of the whole technology you've got this wake gesture with the band where you can just tap quickly to make the display go away yeah very very subtle and again i want to get into the band the band is wild in its own right um the thing that really stood out to me from my demo from these was some things that you only could do in a form factor like this because i mean the texting is cool but like there's this live captions thing where i was in a room with a bunch of people and they all started talking really loudly and i if i just looked at someone it would live caption what they were saying yeah tune out everything else it's like super hearing um and then you're also doing that with language translation so you can do real time i mean ideally both people are wearing the glasses to get the full experience but you don't actually need to have the other you could just hear what they're saying in your language or see it yeah that's pretty wild and that speaks to like what just this form factor could do i'm curious like there's that are there other things that this form factor uniquely can do in your mind that a smartphone can't um i mean the all the things around AI where you have an AI that basically you want to have context around what's going on with you right so like if you want an AI that can see what you see hear what you hear can just kind of talk to you passively throughout the day um and then can show you information contextually that's just not something that a phone can do I mean I guess technically you could walk around holding a phone like this but you can't really do that that's no one does it yeah you can't those demos have existed forever and I'm always like I don't want to hold my phone yeah So I think that's actually going to be the main one.

29:28And I think all the live AI stuff, it's interesting. It takes on a different feel. So we have live AI in the Ray-Bans without a display too, the kind of classic Ray-Bans. And for that, it's audio only live AI. So it's really helpful for when you're doing something kind of by yourself. If you're cooking or something, then, you know, it can just give you, it can, it's, it's watching what you're doing with the video and, and you can ask questions about what you should be doing, or it can give you tips and that's all, that's all great. But it's not really useful when you're in another kind of conversation.

30:08conversation and the thing that the thing that i've observed with um the thought experiment i've run but also just wearing these is you know so we go through the world we have dozens of conversations a day um and in every conversation i usually have like five things i want to follow up on maybe there's like some it reminded me that i should you know go do a thing or or it reminded me of a person who I wanted to talk to, or maybe I'm talking to someone and maybe they like assert some assumption that doesn't quite sound right. And I want to fact check it or like gut check it. These are all things that I think with live AI, you can have this AI that's sort of running in the background and that goes and often does work for you and then can bring that context back, whether it's asynchronously kind of offline when you're done with the conversation, or sometimes like when you're in the middle of a conversation, it's just useful to have more context right then.

31:11Yeah. How have you been using these? Someone on your team was saying you text a lot through them. Yeah. Well, I'm a texter. I like run the company through text messages. So when you were asking about what can you do that you can't do on a phone, I mean, I guess you can, you can obviously text on a phone and we all do it dozens of times a day but I think that there's like a lot of times where it's just not socially acceptable to send a text message and so like let's say you're in the middle of a conversation you want to like ask someone a question or get some information I mean I have this like all the time I'm like having a one-on-one conversation with someone and I like I like wanted to ask someone this or like I wanted to ask someone else a question to like pull some context so I can ask this person what they think about it but I'm not gonna like pull out my phone in the middle of a conversation um with this it's actually just super quick you you can just like send a message in like five seconds get the context back like it actually just really improves the conversations that you're having I find it's um to me this is the one thing that I think is basically better about zoom than in-person conversations is that you can sort of multitask a little bit, right?

32:26It's like, it's worse than basically like every other way than kind of an in-person physical conversation. But the one thing that I think is useful is you can go from having a conversation to basically asking someone else a question. It's not necessarily distracting, it's additive, right? Because otherwise your option is like, all right, you have a conversation, then you go check in with someone else, then you have to go back and call the other person back and have a whole second conversation. So it just short circuits these things all the time. And now I think this kind of brings the best part of that into physical conversations.

32:58Well, you basically feel present in the conversation. You can pull in whatever information you need. It's super, super helpful. Yeah. A real like, holy shit thing about this product is the band. I thought that because it was with Orion, the demo you guys had last year too. And I thought it at the time, I was like, there's something special with this band and you're calling it's a neural band. Is that right? Yeah. The neural band. Because it's a neural interface it picks up nerve activity so like it feels like it's reading your mind it's not doing that but it's not reading your mind what you're doing is you're sending signals through your muscular nervous system that it actually picks up before you even make movements but um you it basically picks up these micro gestures and that allows you to control your interface um no matter where your hand is so it's not doing like hand tracking visually or anything like that like you can have your hand by your side you can have your hand behind your back like whatever in your jacket pocket um and it's fine and um and the gestures are really subtle right so like it's like this is all i need to do to bring it up i mean this brings up met ai so i really like the music one did did you um i didn't try that one no oh so the way when you're listening to music that you adjust the volume is you just kind of oh the dial yeah you pretend that there's a dial and you just turn the dial i did that with the zoom in something yeah you could do you could do it on on photos too yeah um that's like it feels like minority report when you do that it's like in real life it's a it's a good interface yeah i think it's what you mean but yeah it's like not i mean not the weird part of minority reports yeah it's a good part of where tom cruise is doing the hands like it's it feels like sci-fi and um yeah and i'm wondering why this band like why did you land on this band as the input for this because people have been trying to figure out input for glasses like these forever and it's usually voice or hand gestures or something but it's like i'm not going to be in the subway like gesturing out into space so okay i think that those are going to be useful too but i don't think that they're i don't think that they're complete right so voices is obviously going to be important people talk to met ai they do voice calls they you know you do video chats um so voice is going to be a big thing but i think the reality is that a lot of the time we're around other people and for the use case that we really wanted to nail which which i actually think is the most frequent and most important use case that we do on our phones is messaging so if you want to nail that what you need is the ability for a message to come in to not be distracting not be like center in your field of view but just be there um and then you need a way in whatever situation you're in to be able to quickly respond in like five or ten seconds um in a way that is not interruptive to your current interaction and is socially acceptable and doesn't feel rude.

35:48And so then you get to, okay, hand gestures. I mean, yeah, I think that there are going to be useful things for, I mean, in Minority Report, he's doing a fair amount of that. But for gaming and things like that, I think you'll do that. But like you said, you're not going to walk down the street like that, right? I mean, that's kind of - Pretty goofy. Yeah, it looks weird. Your arms get tired. It's much more the former than the latter in terms of reason why it doesn't work, but the latter is also true um so we needed something that was basically silent and and subtle so the questions are there are a few options for that one that people are working on is basically whispering right so you can like sub audibly even um pick up on the sound or you You can have some camera that can like look at your mouth and do like lip reading.

36:40That's still pretty weird in a meeting. I agree. Yeah, I agree. So it didn't pass my bar for kind of subtlety, even though it is silent in theory. So I think like, so I think you need to go for the neural interface. And the other thing that's nice about the neural interface is you can get really high bandwidth input to, so it's not like, you know, with like, You know, smart watches today, you basically like, you know, you can move your arm and it can pick up like a gesture or two, but it's a very low bandwidth. There aren't that many things that it can do. You need something that can basically be reading the muscle signal.

37:20So that way you can just control it very subtly. And this can do that. I mean, I can, I mean, already I haven't like, you know, we're not that optimized. I think we're going to get the autocorrect a lot better, but I'm already, I think, around like 30 words a minute type. Really? Yeah. Yeah. Man. I mean, how advanced do you think the band gets in its current form factor in terms of what it can do? I think quite a bit. I mean, basically, today, well, I guess you have the sensors which can pick up the signals from your muscles. But then on top of that, it's basically just an AI machine learning problem to be able to pick up what you mean by the thing.

38:04and right now um it's not particularly personalized so um so you get it out of the box and it needs to work with certain gestures and you've never used it before so it works with these like like you're you're doing kind of it's like i mean this isn't a big gesture but i mean this is much bigger than what it needs to be in the future and then in the and then for for um kind of the neural text entry and you're basically you can kind of think about it as if you have a mini pencil and you're just like writing out the letter but over time what should happen is that the ai learns your pattern for how you you know write each letter and you should be able to write make like increasingly subtle and invisible motions that it basically learns are your way of doing that letter or that input or whatever and i think the future version of this is that the motions just get really subtle and you're effectively just firing muscles in opposition to each other and making no visible movement at all um and it picks it up so personalized autocomplete via your wrist basically yeah so super fast because if you're not moving there's no latency from actually having to actually physically move and then retract after after making a motion um so i think the upper bound of that is very high the the other thing is um and that's just for typing which is one kind of modality but i think you can kind of there are all these different dimensions where you can use it to input into things and you could control a a hand in space that is like operating a UI, right?

39:49And it's like, there's all kinds of different things that it can do that I think will just be really interesting to get into over time. And, you know, we basically, we invented the neural band to work with the glasses, but I actually think that the neural band could end up being a platform on its own. I agree. To basically just interact with all of your electronics and devices and do all kinds of different things once we kind of get it to be a little bit more mature. So you could have an API for it that could theoretically plug into a smart home or something like that. That'd be wild. Yeah, yeah, yeah.

40:26The price point for these is also lower than I expected. It's only like 800 bucks. Who are these for from a, like, is this an early adopter thing? Like you care about the cutting edge, like this is for you. You're not making a ton of these, right? I assume this is not like a massive, going to be a massive thing for you. It's more to see how people use this technology. I mean, I think that this is going to be a big part of the future. I mean, my view is that, you know, there's between a billion and two billion people who wear glasses on a daily basis today for a vision correction. Like, is there a world where in five or seven years, the vast majority of those glasses aren't AI glasses in some capacity?

41:05like I think that that's it's kind of like when the iPhone came out and everyone had flip phones and it's like just a matter of time before they all become smartphones I think that these are all going to become AI glasses and the question is all right well there are eight billion people in the world not you know one to two so are there going to be a bunch of other people who also wear glasses I would guess yes I mean there are a lot more people who wear sunglasses some of the time um so yeah I mean I think it's a big category um and there's a lot going on here I mean there's i think what you see when you're building products is that v1 um you build what you think is going to be great and then you get a lot of feedback but you also didn't get everything exactly perfect in v1 so you know v2 and v3 just end up a lot better right it's not a it's not a coincidence i think that you know the first version of the ray-bans ray-ban stories um um we thought it was good then when we did the the second version of ray-ban meta i think it sold five times more um and it was just refined right it was it was like so i think that there's going to be some dynamic here where it's like you have the first version you learn from it the second version is just a lot more kind of polished and taken and it's in the software gets polished too not just the hardware and that just kind of compounds and gets better and better um and full that fills your vision that's still coming to.

42:33Yeah, I mean, we're working on all this stuff and we want to get it all to be as affordable as possible. The reality is that the more technology you want to cram into the glasses, the more expensive that is because you're putting more components in. We also want the glasses to be as thin as possible. And that's a process of miniaturization that happens. And similarly, the more technology that you cram in, the harder it is to make smaller. So as much as we can miniaturize this technology, it will always be true that if you put half the technology in, you'll be able to make even thinner glasses.

43:04And some people have different aesthetic preferences where they'll want, I mean, like fortunately, you know, thick glasses are kind of in style. But, you know, some people want thinner ones like yours. You can't fit many electronics. Well, now I don't feel cool. Yeah, you may have to rethink your aesthetic choices in the future. But see, on pricing, we... I mean, we did, we work on, on getting it to be as affordable as possible. And, um, you know, I mean, our view is that our profit margin isn't going to come from a large device profit margin. It's going to come from people using AI and the other services over time.

43:44Um, cause you'll pay a subscription for the AI or something or yeah, or use it and do commerce through it or whatever, whatever the different things are that, that people do. So we're not like a company like Apple whose primary or a large part of their margin comes from having a large margin on the hardware. But in general, yeah, we try to make it as affordable as possible. And my hope is that if we build another one of these, hopefully it's even more affordable. Or the other choice that we can make is put even more technology in it and keep the price point there. But I think you're going to have a few different price points.

44:18There's going to be the kind of standard AI glasses that don't have a display. And I think those will sort of range between$300 to$500 or$600, depending on the aesthetic and kind of how high fashion it is. Maybe even more than$600 if you get something that's really high fashion. But that's kind of the range that we've seen so far from the early Ray-Bans to some of the Oakley Meta Vanguards with all the kind of custom stuff in it, like optical lenses with a prescription. Then there's a category like this with the kind of a display that isn't kind of a full field of view AR display. and i think that that's going to be yeah on the order of a thousand dollars like maybe you get it a little better maybe it's a little more but i but you can call it in that area and then i think when you get to the full ar glasses um that'll be somewhat more to start and and i think like people will just want to have the whole portfolio and then the goal over time will be to get as much of that technology into as affordable and as thin of a form factor so you can just like have as many styles is possible.

45:39The holidays have arrived at the Home Depot, and we're here to help bring the excitement with decor for every part of your home. Check out our wide assortment of easy-to-assemble pre-lit trees so you can spend less time setting up and more time celebrating. And bring your holiday spirit outdoors with unique decor like one of our Santa inflatables. Whatever your style, find the right pieces at the right prices this holiday season at The Home Depot.

46:10Running a business comes with a lot of what-ifs. But luckily, there's a simple answer to them. Shopify. It's the commerce platform behind millions of businesses, including Thrive Cosmetics and Momofuku. And it'll help you with everything you need. From website design and marketing to boosting sales and expanding operations, Shopify can get the job done and make your dream a reality. Turn those what-ifs into... Sign up for your$1 per month trial at shopify.com slash special offer. Imagine fast hydration. Combined with balanced energy. Perfectly flavored with zero artificial sweeteners. Introducing Liquid IV's new energy multiplier sugar-free.

46:54Unlike other energy drinks, you know, the ones that make you feel like you're glitching, it's made with natural caffeine and electrolytes. So you get the boost without the burnout. Liquid IV's new energy multiplier sugar-free hydrating energy. Tap the banner to learn more. And so you've got these new Oakleys and your deal with El Sol Exotica means you can do other smart glasses with all their other brands, right? So I think that implies, right, that there will be future brands that have Meta Tech. We'd love to do it. Yeah. So you see it as like building out a kind of constellation of all these different form factors and price points.

47:28That's the goal. Yeah. What about all these other AI wearables that are glasses that are happening? Like there's the friend pendant I'm sure you've seen. There's all these like displayless non-glasses devices. Sam and Johnny Iver apparently working on something. There's a lot of interest in this. And I'm curious, is this something that since AI has really taken off in the last few years, you see opportunity there in addition to glasses? or are you still just the main focus is glasses? Well, our main focus is glasses because I think glasses are the best form factor for this for all the reasons that we talked about before.

48:04I think that anything else that you have to kind of fiddle with takes your attention away from the physical world around you. I don't think that there's any other form factor that can see what you see, hear what you hear, talk to you throughout the day and generate a UI in your vision. and then there's the whole augmented reality part about blending the physical and digital world but I don't know I mean people use different electronics so I mean I'm not gonna I mean I certainly don't think that it's like in the future all eight billion people in the world are doing the exact same thing I mean some people use their phone more some people use a computer more some people use an iPad instead of a computer right some people you know primarily you know watch videos on a TV.

48:49Some people watch videos primarily on a, on phone. So I do think that they're going to be different things. Um, my guess is that glasses will be the most important. Um, I think something like earbuds is kind of interesting too. I mean, Apple clearly is like by far the leader on that with AirPods. I think partially because they did a good job and partially because I think they gave themselves some, um, kind of unfair advantages with how they bundle it and couple it and have technology that works with the um with the phone that i guess they're now just starting to open up which is great but um but for a while i think it just made it impossible for anyone else to build anything like the airpods um watches i think are interesting um in some ways too and you're not a fan of the like the pendant thing the pendant trend that's kind of starting right now um i mean is it a trend is i don't know it's early i mean there's a lot of startups doing this stuff.

49:47I think it's an interesting idea. I don't want to be too dismissive. My point is that I actually think that they're... My guess is that different people are going to like different things, but that glasses are going to be the most popular. There was no new quest this year at Kinect. And I'm curious how you're feeling about the quests these days in VR, mixed reality generally as a category. It seems like glasses are really taken off. There's obviously a lot more quests out in the world. They've sold a lot more. but I'm curious, you know, not, they're not being a new one this year and also just how you're feeling about it these days, like the category.

50:21Yeah, no, I mean, I think we're, we're making progress on it. I mean, this year, what we focused on was the MetaHorizon creation tool. So we announced MetaHorizon Studio and MetaHorizon Engine, which are, are these basically foundational tools for creating worlds and content using AI. And that's going to go towards making it so that people can create a lot more content in VR. But I think that that's also going to be the case. And all that stuff, I think, should translate over to AR too. I think a lot of this content you'll be able to have there. I mean, glasses that are see-through may not be quite as immersive as VR, but you can deliver a lot of the same kind of holographic experience.

51:06And then I think a lot of these things will also end up showing up on phones, right? I think there's this huge opportunity with AI to, you know, you're browsing your feed on Instagram and Facebook and like it's like each story should be its own world that you can jump into. But like the, and you're starting to see some of this with some of the AI models, some of the stuff that Google has put out recently, for example, are like interesting glimpses of where that could go. But I think that there's this real sense that the whole stack for how you create those kinds of immersive experiences needs to get rethought.

51:55It's not just going to be people doing things in the same way that they've created 3D video games historically. I mean, that's a very, very intensive process where the tools are like very - Yeah, hybrid entry. Yeah. I mean, so my kids are kind of into programming and into making things. And we try to build different 3D worlds. And I think some of the stuff is just like intractable for them, right? It's like, I mean, they're still kids, so it's fine. But if you attract it to like a prompt, I mean, then it becomes - I think by the - Yeah. But with the MetaHorizon Studio, which I've been playing with, with them, and obviously this isn't primarily designed for an eight-year-old to be able to use, but my bar is if it's enough that I can kind of make something good with my eight-year-old, then that's pretty cool.

52:48You really can create all these interesting things, right? It's like you can define what the world dynamic is, like what kind of you want the world to be. If you want to put stuff in the world, you can do it. If you want to texture things differently, if you want to change the skybox, you can do that. You know, if you like, so I think it's, it's, I think it'd be a very different way of creating things that's fundamentally a lot more accessible, which will then unlock a lot more creativity. And there will just be a lot more interesting worlds and things to do. And that I think is going to be important, not just for VR and AR, but I think it's going to unlock all these experiences that billions of people will probably first see on their phones at some point.

53:28So that's what we're doing with the MetaHorizon studio work. It's this kind of agentic AI flow where people at different levels of sophistication can go in and create really interesting worlds and immersive environments. And I think that's neat. Then we paired that with MetaHorizon engine, which is basically this custom graphical rendering engine that we've been creating for two years now um it's this project that we've had for to build it from the ground up because you know previously we were using unity which is great but it's not really built for this use case i mean most games you know you you load a game it takes you know 20 seconds to um to kind of get into the game which kind of makes sense because you're loading this whole 3d world that you now need to be able to interact with but we want the worlds that you can interact with to feel more like jumping between two web pages or jumping between like two screens within a native app that's like really fast.

54:37So the whole like, okay, it has to take 20 seconds to like page this whole new world into memory was not going to cut it. So we rebuilt, we basically built MetaHorizon engine from scratch to be this graphics engine that can support rendering these kinds of worlds with high concurrency and the avatar system, the photorealistic avatars and all this with just a few seconds of load time. So it's more like a website or just a transition and an app. And that's the kind of thing that'll make it so that when you're in VR, you can jump between worlds easily. It's not like some big commitment or some big decision.

55:14You can just feel free to explore because it's not like you're gonna have to wait 20 or 30 seconds for the next thing to load. You walk through the portal. You don't like it. You walk back through the portal the other direction. And similarly, for things like within Facebook or Instagram, having the ability to kind of see a post and jump into a world, that's something that needs to have very low friction to do. So the MetaHorizon engine is this kind of core piece of technology. So yeah, I'd say on the Metaverse side, this year's announcements at Connect were more about kind of the software foundations than hardware.

55:51But you're still committed to the hardware. Yeah, I mean, the way that we do the hardware is we don't have it plans that there's a new device every single year. Sure. We have multiple device lines. There's sort of the higher end one where we introduce some new technology and then we try to get it to be as affordable as possible. So we did Quest 3, then we did Quest 3S. But it's not like every year there's one. it's basically, you know, most years there will be a new one of one of those two. And then sometimes there's like an off year where we're pretty much just tuning the software to get ready for the new paradigm.

56:27Got it. Okay. So Quest is still going. Yeah. We're first. Okay. Yeah. I've been saving this. I've got to ask you about this. You know, there is a tremendous amount of interest in your AI strategy right now. Unlike anything I've seen in the tech industry, honestly. Well, AI overall. AI overall, but I think like what you've done over the summer with the hiring and the super intelligence, you know, mission that you put out and all of that. Yeah. And we've been talking about AI as it relates to the hardware throughout this conversation. And, you know, AI was a part of my demo. It wasn't, I would say, like front and center.

57:02And it seems like a lot of the work you're doing now is to get ready for when it will be. And I'm curious, you know, when I was here actually for the demo, I got to see the pod of the new lab and see them at work. and they're in there cranking. Like you can tell. And I would love to know, like maybe we can start here. Like when you decided I need to change things and why you decided to go about it the way that you did. Because I think that's the thing that people were like, whoa, like this is crazy. I think if you're on the inside, it doesn't feel as crazy because the talent market is very small.

57:38It's kind of rational if you look at the numbers. But just the strategy, like walk me through like when you were like, okay, I want to make a change. This is what I want to do. Walk me through that. Yeah. I mean, this is an area that I just think it, I think AI and super intelligence are going to be the most important technologies in our lifetime. I think it's so important that it sort of demands its own hardware platform, which is a big part of why I'm so excited about glasses, because I think glasses are going to be the best kind of hardware device category to provide personal super intelligence to people.

58:16But I think AI is just this incredibly profound thing. It's going to change how we run the company. It's going to change how all companies run. It's going to change how we build products. It's going to change what products are possible. Change how creators do their work, right? So change the content that's possible, the mix of content, all these different things. So I think being on the frontier there is really critical if you want to continue just doing interesting work and pushing the world forward. I think just like with mobile, if you didn't invent the mobile phone, you could still do interesting work building apps.

58:55But I do think at some level you can do even more interesting work if you can both pair the software with the hardware experience. So we are definitely committed to being at the frontier and building superintelligence. I think it's going to be the most important technology, like I just said. And because of that, I just think we're very focused on making sure we build a leading effort. So over the last few years, we stood up an effort that was improving very quickly. So Lama was a good initial academic project. Lama 2 was a kind of good initial version of that as an open source release. Lama 3 was a big improvement over Lama 2.

59:47And then Lama 4 introduced some important improvements over Lama 3 too. but um i didn't feel like we were on the trajectory that we needed to be on to to basically like be at the frontier and be pushing the field forward and that was so i you know it's like i think every company at some point um goes through periods where you're not on the trajectory that you want to be on something and these are decisions that you get to make in your life or in building a company where the real question is not like, it's not like, is there going to be a moment where you feel like you're not kind of on the track that you want to be on?

1:00:25It's what do you do in that moment? And so I just decided that, you know, we should take a step back and build a new lab. And, you know, I think part of that was informed by the shape that I thought the effort should be. we have this real focus on talent density, right? And the idea is that you really want to have, this is like a group science project, right? So you want to have the smallest group of people who can fit the whole thing in their heads at once. And there's not many people who can do that. No, but you also want the group to be as small as possible, right? So there are some problems that we're working on around the company where like you can just have more people work on them.

1:01:12And even if the marginal productivity per person declines, you can just keep on scaling the net productivity of the effort. You know, our feed and ads recommendation is an interesting example of this, where we have a lot of people are just testing different improvements to the systems. And if one guy's sitting next to you, if that guy's experiments don't work that well, it doesn't necessarily slow you down that much. But I think building these language models is not that way, right? It's a small group effort. You want the smallest group of people that can keep the whole thing in their head and do the best work that they can.

1:01:49So each seat on that boat is incredibly precious and in high demand. You also don't want a lot of layers of hierarchy because when someone gets into management, their technical skills kind of start decaying pretty quickly. Even if they were an IC researcher a few months ago, now if they're spending all their time kind of helping to manage, then okay, after six months, a year, they might be less in the technical details than they were before. So I think that there's this huge premium on just having a relatively small, extremely talent-dense effort that is organized to be quite flat. And you're very hands-on with this team.

1:02:32well yeah i mean in the sense that i mean i'm not like an ai scientist yeah but the thing that i'm sitting near you i mean it's clear that this is like the priority so yeah the thing that that i'm focused on is one getting the very best people in the world to join the team so i've spent a lot of time just meeting all of the top researchers and folks around the field and getting a sense for who i think would be good here and um who might be at a point in their career where we can give them a better opportunity. That's one piece. Another thing that I'm very focused on is making sure that we have significantly higher compute per researcher than any other lab, which I think we are just way higher on compute per researcher than any other lab today.

1:03:16And as the founder and CEO, and because we have a strong business model that can support this, I mean, we make a lot of profit. A decent amount, yeah. Yeah, it's a reasonable amount. You can just call up Jensen and be like, more GPUs, please. It's not that simple. It's not that simple. And I normally text him with my glasses. But no, but it's, but there's a whole supply chain that goes into it. And the GPUs are one part of it. But then you also need to build data centers and get energy and get the other pieces and get the networking. And so, yeah, but the bottom line is we're very committed to that.

1:03:54and doing what we need to do to make sure that we have leading levels of compute. So we talked about recently how we're building this Prometheus cluster, which I think is going to be the first kind of single contiguous gigawatt plus cluster for training that I think has been built. In the world, we're building this Hyperion data center in Louisiana that I think is going to scale to five gigawatts over the coming years. and several other of these, we call them Titan data centers. They all have different Titan codenames that are each going to be one to multiple gigawatts. And that's a significant investment.

1:04:38I think it took a fair amount of conviction. So I think a bunch of conditions need to be met. Like basically you need to have a business model that can support it. You need to have a CEO who like believes in this very deeply, right? It's that they're just willing to like make that kind of investment for that. And then you need to have the technical ability to actually go build the things and bring them online. And I think we're one of, if not the only company in the world that meets all of those criteria. So, so yeah, I mean, other people will do other interesting things too. But no, I think that this is going to be very interesting.

1:05:21The other principle, though, that we have for the lab is, you know, it's split into different efforts, right? There's the lab that we call TBD, which it... That's what I saw. Yeah, that's the research lab. TBD was the placeholder name, but then it stuck because it's kind of a good vibe, right? It's like, all right, it's like a work in progress type vibe. um then we also have applied research and product um that in nat friedman's group and that team is working on a lot of research that goes directly into the products um so things that may not necessarily directly be on the path to super intelligence like um you know speech that passes the turing test and things like that but but are important for the products nonetheless.

1:06:05So we're working on all those things. And the research effort, the TBD effort, is truly a long-term research effort. So one of the principles that we have for the lab is just no deadlines. So people are asking, okay, when are we going to ship the model? This is very kind of, it's a strategy and it's also the values that we're trying to put in it is, I mean, all these researchers are very competitive. They all want to be at the leading edge. They know the industry is moving quickly. They're going to put a ton of pressure on themselves. Me telling them that something should get done in nine months or six months or whatever isn't going to help them do their job.

1:06:49It's only going to put another artificial constraint on it that makes them sub-optimize the problem. And I want them to go for kind of the full thing. I mean, It's like we're going for trying to build AI that can improve itself responsibly and that can – and we're basically building these models that bring all these modalities together to deliver the kinds of experiences that we're talking about. And, yeah, I mean, I think me putting a deadline on that is not going to be helpful. So, yeah, so I'm very focused on – and that's the nature of research, right? It's not engineering. It's not like engineering is when you know how to do something and you go and you need to put together a complex process to build it.

1:07:27Research is when there's several unknown problems. In an AI, I don't even think we have a sense of how many unknown problems there are. For something like glasses, we have a sense. It's like, okay, there's like 10 areas of unknown problems that we need to go solve. Like how do we get the right waveguides? How do we get the right laser display? No one's ever done this, but like we can try 10 different things in each one of those and kind of run it forward. In AI, I don't think anyone can definitively tell you how deep the problem space is. So it's very much research. That's fascinating. So, yeah.

1:08:01So what do you do to do that as well as possible? You get the very best team, talent density, make sure that people have the resources that they need, and clear all the other stuff that comes from running a big company out of the way. And that's kind of my job for them. Yeah. You were talking about the CapEx and the data centers. you obviously see something on the other side of that that will warrant that being worth it um but i'm wondering do you ascribe to these these bubble fears at all that people are talking about that we're in this massive uh overspending getting ahead of the skis bubble and maybe a company like meta will be okay because you guys do have a core business that makes a lot of money but yeah how do you think about this bubble talk that has been going on for the last few months especially um I mean, I think it's quite possible.

1:08:48I mean, I think basic, if you look at most other major infrastructure buildups in history, you know, whether it's railroads or fiber for the internet, you know, in the dot-com bubble, these things were all chasing something that ended up being fundamentally very valuable. In most cases, it ended up being even more valuable than the people who were kind of pushing the bubble thought it was going to be. But in at least all of these past cases, the infrastructure gets built out. People take on too much debt. And then you hit some blip, whether it's some macroeconomic thing or maybe you just have like a couple of years where the demand for the product doesn't quite materialize.

1:09:32And then a lot of the companies end up going out of business. and then the assets get distressed and then it's a great opportunity to go buy more. So I think that that, it's not, it's obviously impossible to predict what will happen here. There are compelling arguments for why AI could be an outlier and basically just, if the models keep on growing in capability year over year and demand keeps growing, then maybe there is no collapse or something. Um, but I do think that there's definitely a possibility, at least empirically based on past large infrastructure build outs and how they led to bubbles that, um, that something like that would happen here.

1:10:20Um, from Meta's perspective, you know, I, I think the strategy is actually pretty simple. it's um you know it's at least in terms of building out the infrastructure you know it's no one knows when super intelligence is going to be possible is it going to be three years five years eight years whatever is it never going to happen but i don't think it's never going to happen i i'm more ambitious or optimistic i think it's going to be on the sooner side but let's say let's say that you weren't sure if it was going to be three or five years like in a conservative business situation maybe you'd like hedge building out your infrastructure because you're worried that if you build it out, assuming it's going to be three years and it takes five, then you've lost maybe a couple hundred billion dollars or something.

1:11:09I mean, my view is that - That's a lot of money. Well, no. Well, I was going to say in the grand scheme, it is objectively a huge amount of money. Yeah. Right? I mean, didn't you just tell Trump you were going to spend like 600 billion? I mean, that's - I did. Yeah, through 2028, which is - That's a lot of money. It is. And if we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But what I'd say is I actually think the risk is higher on the other side. If you build too slowly and then super intelligence is possible in three years, but you built it out assuming it would be there in five years, then you're just out of position on what I think is going to be the most important technology that enables the most new products and innovation and value creation and history.

1:11:58So, I don't know. I mean, I don't want to be kind of cavalier about it. I mean, obviously, these are very large amounts of money and we're trying to get it right. But I think the risk, at least for a company like Meta, is probably in not being aggressive enough rather than being somewhat too aggressive. But part of that is we're not at risk of going out of business or something like that, right? If you're one of these, you know, companies that like an open AI or an anthropic or something like that, where, you know, they're raising money as the way that they're funding their build out. And, you know, there's obviously this open question of to what extent are they going to be able to keep on raising money?

1:12:46And that's dependent both to some degree on their performance and how AI does, but also all these macroeconomic factors that are out of their control. I mean, the market could get bearish for reasons that have nothing to do with AI, maybe something bad happens internationally. And then it could just be impossible to fulfill the compute funding, like the compute obligations. So I think that's, it might be a different situation if you're in one of their shoes. But I think for us, I think the clear strategy is it's just, I think it creates more value for the world if we kind of assume pretty aggressive assumptions about when this is going to be possible and take some risk that maybe it takes a little bit longer.

1:13:33Do you feel like the US is in a better place now to help with this and to help American companies succeed? I think you've done a lot of work with this new administration. And when we were here last year, you were saying you wanted to kind of stay out of it. But it seems like you realized, Was the realization that this is just so important that I have to play ball with this? Oh, well, I mean, the thing that I want to stay out of is partisan politics. Okay. But, I mean, we will always want to work with and have a good partnership and collaboration with governments, right? And that's going to be especially true in our home country, but it's also true in other countries around the world where we serve large amounts of people.

1:14:23So, yeah, but I'd say yes. I mean, I think this administration, for a number of reasons, is definitely more forward-leaning on wanting to help build out infrastructure. And that has, I think, been positive. And I think that that aspect of this is, I think the next few years are going to be very important for the AI build out and the AI infrastructure build out. And I think having a government, which it's important both at the federal level and in the states where you work, that want that build out to happen is fundamentally a helpful thing. Yeah. There was a line in your super intelligence memo that you wrote where you said, over the last few months, we have begun to see glimpses of our AI systems improving themselves.

1:15:13Yeah. I was really interested in that line. What specifically did you see that made you write that? Well, I mean, one of the early examples that we saw was a team that was working on Facebook that took a version of Llama 4 and made this autonomous agent that could start to improve parts of the Facebook algorithm. and it basically checked in a number of changes that are the type of thing that like a mid-level engineer would get promoted for really so yeah so i think that's like very neat it's like you basically have built an ai that is building ai that makes the product better um that improves the quality that people observe um to be clear this is still a uh a low percentage of the overall improvements that we're making to facebook and instagram um but i think it'll grow over time so that's what that's that's um one of the things that i was talking about when i said glimpses right it's i mean this isn't like okay the ai is improving itself at like an exponentially fast rate or something like that.

1:16:31I think that what we're seeing are early examples of AIs autonomously improving AI in ways that are having a positive impact on the experience that people get to. And that's how to think about superintelligence broadly is that when you're there, it's AI that is rapidly improving itself. That's what it means? Or is that too simple? Yeah, I think AI that can improve itself. And that's beyond human level. I think that there's there's this dynamic today where all the AIs are trained on data and knowledge that people have produced. So a lot of the systems seem to be kind of very broad and have all the knowledge of humanity.

1:17:12So maybe that's in some dimensions, it might feel like it is smarter than any one given person and kind of the breadth of what it knows. But I still think today, the systems are basically gated on human knowledge. And there is a world beyond that, right? Where I think you're starting to get into that with some of the thinking models where they can go out and solve problems in the future that no person can solve and then can learn from having solved that problem. And the pace of that improvement to me is somewhat less important than just the process of it. I'm not like a super fast takeoff believer because, um, in the sense that I don't think it's like going to be, okay, one day it can improve itself.

1:18:03And then the next day it's going to take over everything. Um, I mean, I, I think that there's like way more physical constraints. Like it takes time to build data centers. Like if you, a lot of frontier human knowledge comes from empirical experimentation, right? So you, um, you know, if you're, you You develop some new drug. You want to see A, if it works and B, if it's safe for people. How do you do that? You run a test where you give it to a handful of people, maybe more than a handful, but some statistically significant group of people. And you observe how that goes for a while to see both whether the positive effects are kind of long-lived and whether there's any negative effects.

1:18:49Okay. Well, if you're trying to run a six month or a 12 month trial, you can't do that in less than six or 12 months. I mean, maybe you can get a negative result sooner, but you're not going to be able to validate that you can get the positive result that you're looking for without having done that test. So I think that's also going to be true with AI, right? There are going to be some things that maybe something that's like a super intelligent system can just intuit or reason from first principles using the knowledge that we already have. But I think a lot of learning is going to be experimental.

1:19:24And I do just think these things take time, right? You have to run long-term experiments if you're trying to make long-term changes in the world. And that I think is going to be true for the AI too. Now, maybe it'll on average run smarter experiments. So per experiment, maybe it'll learn more.

1:19:41I think it will probably be able to figure out some things from first principle. It will definitely be able to figure out a bunch of things from first principles. But I don't know. I'm not like, I'm not on the camp of people who think it's going to be like overnight, this changes. I think it's going to be this very steady progression. We're just making our lives better. All right. Well, it's going to be a wild few years. Yeah. Mark, I appreciate you doing this. Happy to. Yeah. Congrats on the new show. Thank you. Appreciate it. Alex, I enjoyed that interview, but I have to ask, how does it feel to have Mark Zuckerberg in your neuro band?

1:20:15I hope he's not actually in the band, but I guess we don't know for sure. But seriously, thanks to Zuck for taking the time to be the first guest on Access. You can read more about what we talked about in my newsletter, sources.news. And Ellis, what are you plugging? Yeah, you could find me on Twitter, which I will continue to call Twitter, at at hamburger and at meaning.company. Access is produced in partnership with Vox Media. Please follow us. We're a new show. We need your support. We need your follows. Hitting that notification button to get new episodes. You can find us on Spotify, Apple Podcasts, all the other podcast apps that I don't know about.

1:20:58We're also on YouTube and video. Please check us out there at AccessPod. You can also find us on all of the socials at AccessPod. Smash that like button. Smash it. All right. Can't wait to tell my grandkids. Alice, that's our first episode. We'll see everyone next week.