The ultimate vision for Roblox, held for 20 years, is a high-fidelity physics-simulated space where people can co-experience — build a car, drive it, go to a birthday party, chop trees — and AI's role is to superpower the creation of that holodeck and make it photorealistic, scalable to 10,000 people, and acoustically simulated.

(00:00:00) What proportion of you think your children will have at least one serious relationship with an AI throughout their lifetime? (00:00:05) I think 0 right now. (00:00:06) I don't think we're close to crossing the human AI barrier, but it is worth thinking things are changing every day. (00:00:14) We have to incorporate and do that. (00:00:16) And then there's also some weird universal truths of things that stick around for four years and you kind of have to blend them together. (00:00:24) We feel if we build Roblox right, it might be the kind of thing that's kicking around in four years. (00:00:36) Hi, listeners. (00:00:37) Welcome back to No Priors. (00:00:39) Today, Elad and I are here with Dave Baszucki, the founder and CEO of Roblox, the 3D immersive world where more than 150 million users come every day to hang out. (00:00:49) We talk with Dave about the future of AI in gaming, their investments in AI, how NPCs are going to change and really transform game. (00:00:59) Roblox has (00:01:00) Studio, one of the largest coding platforms out there, and their 20-year mission to build the holodeck. (00:01:06) Welcome, Dave. (00:01:08) Dave, thanks so much for doing this with us. (00:01:09) It is so great to be here, and thank you for doing this in Roblox headquarters. (00:01:13) Well, exciting to be here. (00:01:14) Yeah, great to have you guys. (00:01:15) Thanks. (00:01:15) So I think we wanted to start with a big picture question. (00:01:18) I think you all have been really visionary in terms of where all this is heading, and AI is obviously having a big wave and effect on gaming. (00:01:27) and what's going to happen in the future there. (00:01:29) And that's across things like world models. (00:01:31) How do you think about NPCs and their evolution? (00:01:33) How do you think about assets within games? (00:01:35) How do you think about world creation? (00:01:37) So we'd just love to hear your views of like big picture. (00:01:39) 10 years from now, where are we going? (00:01:41) Like what's coming? (00:01:41) Yeah, so two step on this one. (00:01:43) I would say first, I know you both dabble in investing a little. (00:01:47) So there is a Roblox business plan PowerPoint deck from almost 20 years ago that we pull out sometimes and we're amazed at the fidelity of it. (00:01:59) pretends a new category. (00:02:01) One could call it the category of human co-experience. (00:02:04) It's got the sizes of a bunch of companies from 20 years ago. (00:02:07) So you can see social networking companies (00:02:09) You can see YouTube video, you can see toys, and you can see all of that. (00:02:14) And what it imagined is really a new category. (00:02:18) Some have called it the metaverse, some have called it the holodeck, which is really the ultimate high-fidelity simulation where people can come together and do stuff. (00:02:29) Feels like a Ready Player One in the movie. (00:02:31) Exactly. (00:02:31) And so we do have that business plan slide. (00:02:35) And what we had always imagined is if this high-fidelity space (00:02:40) was backed by physics simulation and reality simulation, you'd be able to do stuff. (00:02:45) You'd be able to build a car, put wheels on it, drive it around. (00:02:49) You'd be able to go to a birthday party and blow out the candles. (00:02:53) You'd be able to chop trees down and make a house of it. (00:02:57) So that was literally the genesis of Roblox. (00:03:01) And here we are 20 years later, there's so much more to do. (00:03:06) But one vision of AI is how do you superpower (00:03:09) the creation of that holodeck. (00:03:11) And how do you get it photorealistic as soon as possible? (00:03:15) How do you get 10,000 people in it instead of 100? (00:03:19) How do you do acoustic simulation? (00:03:21) That sounds realistic with 10,000 people. (00:03:24) So in that sense, AI is super interesting about just trying to get a vision that's been around for 20 years and has arguably been in sci-fi, once again, snow crash holodeck, forever. (00:03:36) There is another interesting vision, and (00:03:39) And I think it's useful comparing these product things that's much more, what's the future of just a single person by themselves dreaming what they want to dream? (00:03:50) And we sometimes call that real-time dreaming category. (00:03:54) It's a category that we can see a bit of the trappings of with short-form video. (00:04:00) There's a little bit of that in short-form video. (00:04:02) You know, someone's 2 A.m. (00:04:03) in their bed doom scrolling, it's reacting to dwell time, it's reacting to what you favorite, and you're (00:04:09) kind of getting a little bit of a pre-built dream. (00:04:12) That, I think, we sometimes imagine could go all the way to the famous Tom Cruise movie, Vanilla Sky, where he literally real-time dreamed for several years in an imaginary universe where everyone around him was primarily an NPC, and he didn't even know he was doing it. (00:04:33) So that's more world-building as content. (00:04:36) That is exactly right. (00:04:38) And so if you put those (00:04:39) at 2 extremes. (00:04:41) One is communication platform, high fidelity with people you know. (00:04:47) The other is real-time dreaming where everyone's an NPC. (00:04:50) Everything in between is possible. (00:04:52) And I think we see... (00:04:54) everything in between ultimately as being possible. (00:04:58) And we may see weird product categories we never imagined. (00:05:03) I think people are probably thinking, does the text prompt turn into a voice prompt? (00:05:09) And does it automatically turn into a video prompt? (00:05:12) If there's like, I want to learn French video, how you're in a French cafe, start talking to an NPC. (00:05:19) So I think there's a wide range of products that start to get into this space we're interested. (00:05:24) interested in, which is kind of that 3D reality space. (00:05:29) Do you think that seems useful exists for both consumer and business applications? (00:05:32) Because people also talk about this in the context of you're instead of doing a Zoom, why don't you go into a shared space with someone and meet with that? (00:05:39) I think we have your life becomes virtual. (00:05:41) Yeah, we have seen, I think on the communication side, people just want more and more fidelity. (00:05:47) We can go back to paintings on the wall of a cave. (00:05:51) People want more fidelity. (00:05:53) We have smoke signals. (00:05:54) We have the mail system. (00:05:55) We have the Pony Express. (00:05:57) We have the telegraph system. (00:05:59) We have the phone system. (00:06:01) We have the, arguably the COVID-initiated video. (00:06:05) We also did have the COVID-initiated 3D experience on Roblox, actually. (00:06:10) And we found side-by-side video, a lot of young people used Roblox as an early test case of how to co-experience. (00:06:19) I think as the technology for multiplayer 4D simulation gets better and more photorealistic, it's almost going to be like video is the downsampling. (00:06:30) And it's like if we were having this 4D immersive communication, we'd say go to legacy analog mode and make it look like Zoom, basically. (00:06:40) And we could always do that, but we'd be able to do things we could never do on Zoom. (00:06:46) You know, you'd say, hey, let's pop up and go walk around my office. (00:06:49) And (00:06:49) I want to show you something, come and follow me kind of thing. (00:06:52) So I do think the 4D simulation, as it gets photorealistic, will be a superset of video. (00:06:59) The other thing for business is there's a few things that 4D, and I use 4D because it's not just 3D shape, but it's function in a way, and will provide that video does not provide is acoustical things. (00:07:12) And there's some real tricky physics problems when you're having a company meeting of 1000 people. (00:07:19) If we try to do it on Zoom, we have like a bunch of squares, who's live, who's not live, how do we mix the sound, which is very tricky. (00:07:29) If we do it in a three-dimensional simulation, I can hear everyone with attenuation based on how far they are, and it can actually be a much more natural experience. (00:07:40) Like I walk closer to you and I hear you, and that may actually be a better human interface paradigm. (00:07:47) There's one final technical thing that's (00:07:49) going to be very, very difficult is when people try to sing Happy Birthday to you together. (00:07:56) And because no matter whether you're on Zoom or whether you're in a 3D simulation, you're hearing them like 30 milliseconds later. (00:08:05) So what I think we're going to find is people are actually allowing each one of us to sing Happy Birthday with a time forward extrapolation of each other and then remixing it. (00:08:15) So I do (00:08:17) for the business side, I do think ultimately we're going to see some interesting conferencing solutions. (00:08:23) Roblox has this amazing wild story with physics simulation or physics engine as like part of the birth of Roblox. (00:08:32) But then today, or for the last decade, I have, I think of this as a immersive world where people have their first coding experiences, they build amazing games, they interact with each other. (00:08:43) I don't think of that world as being super focused on realism. (00:08:47) That's right. (00:08:48) So is that something that's changing for you guys? (00:08:50) Is it enabled now? (00:08:51) Like how should we think about that? (00:08:53) I think what you, the way I would think about it is we're pushing really hard on the physics simulation capability. (00:09:00) And at the same time, we have this incredible free market that we can't control on content generation. (00:09:06) I would say we are just started talking about full acoustic simulation. (00:09:10) I just tweeted about it. (00:09:11) So actually, you know, approximate, you know, attenuation and all of this kind of (00:09:17) of things and echoes, that's going to start happening. (00:09:19) On the physics simulator, it's behind the scenes gotten better and better and better. (00:09:23) It'll keep getting better. (00:09:24) But what we have found, I would say, is the diversity of developers are beyond our expectation. (00:09:32) And I would say some of the things that have gone viral, Dress to Impress, for example, Grow a Garden, are less physical, but they have leveraged cloud capabilities that we've built side by side. (00:09:47) Grow a Garden was essentially able to build an experience where things keep growing, even when you're not there, by leveraging a lot of the cloud persistence we provided. (00:09:58) Dress to Impress was able to do some very interesting things with how they run contests and layer clothing on. (00:10:04) So I think we're generally, the specification for our product could be, we just want to (00:10:11) simulate the real world with 10,000 people. (00:10:14) And the more we go in that direction, the more we'll just see creators using it in different ways. (00:10:20) How much attention have you been paying to world models and what's been happening in the AI world in terms of how people are shifting how they do physics simulation? (00:10:28) I think we're really watching it. (00:10:30) I think that the biggest thing we have focused on and I think we'll continue to focus on is how do we get that 10,000 player multiplayer type experience and how do we fall back (00:10:41) to 100 or 10 or one. (00:10:44) We feel it's super robust to build a communication platform. (00:10:48) And that's part of our goal of getting to 10% of gaming. (00:10:53) The technology for this, I think we're going to start unraveling it over the next 5 to 10 years is what's the most efficient way to synchronize the state of 10,000 people. (00:11:06) And when we... (00:11:07) That's a deep infrastructure problem versus necessarily a physics simulation. (00:11:11) synchronize the state and the memory over the last five hours of 10,000 people. (00:11:17) Where have they been? (00:11:19) Where have they been in 3D space? (00:11:21) What have they been doing? (00:11:22) The world model thing is very exciting because they're storing memory in literally video latents. (00:11:29) A minute of frames, it's showing super early promise. (00:11:33) I think the technical revolution is going to be what's the ultimate synchronized state for 10,000 people? (00:11:39) Is it in video latents? (00:11:41) Is it a much more native 3D format? (00:11:43) Is it a hybrid new discovery of some 3D video latent space to be determined? (00:11:50) I will share that the format we're using, part of the vision for us is to ultimately store the history of everything on Roblox. (00:12:00) and to store it, not raster like video, but store it in vector, which means the ability to play back anything that's ever happened on Roblox, that's 13 billion hours a month. (00:12:14) And so the ability... (00:12:15) What do you think that does for you, the history of your world? (00:12:18) So there's a sci-fi book by Arthur C. (00:12:21) Clarke called The Light of Other Days. (00:12:24) which talks about the fantasy of what would happen if we had infinite playback in our world. (00:12:30) It's like a real social problem because if everything we've ever done can be played back, actually in that sci-fi book, it's not just everything could be played back, but you could look at any playback. (00:12:44) And so everything's completely transparent. (00:12:46) There's no private conversations. (00:12:48) It actually delves into that. (00:12:50) I think for us, we would then use it very thoughtfully and (00:12:54) privacy compliant and judiciously, but for very simple things, you can imagine, oh, there was a safety incident. (00:13:02) Let's go back and put five cameras and listen to the audio and see what happened and make a good judgment call. (00:13:10) That's kind of interesting. (00:13:13) There's a really special thing that maybe I did with someone in my family that's really special. (00:13:20) given I have access to my history, go play that back and possibly because it's 3D vector data, not raster data, reshoot it from a, make a cool video of the best moment. (00:13:33) And I think where it ultimately ends up for us is there's a huge push now for people to find kind of hybrid video, ASDW, human interaction data. (00:13:45) People are starting to train on video with keyboard stuff and like, where do I find (00:13:50) 100 million hours of this data. (00:13:53) The data we have, which is 13 billion hours a month, can be reproduced from any camera angle and can interact with the 3D space. (00:14:01) So it's very powerful data. (00:14:03) It leads to a really interesting computer science program idea around how to create great NPCs that are more than just an LLM. (00:14:13) Can you tell us more about that? (00:14:14) Actually, that sounds really interesting. (00:14:15) Yeah, go check out on XRX feed. (00:14:17) I think in the last day or two, we released a bunch of native video demos and maybe we can even play them back here. (00:14:23) But what we are starting to train is NPCs based on a combination of all of the data we have in a privacy compliant way that are starting to navigate and starting to play games. (00:14:37) The challenge, I think, (00:14:39) would be challenge level one. (00:14:42) Everyone has access to an NPC that can get pretty good at playing any Roblox game. (00:14:48) Super cool. (00:14:50) Challenge level 2 is by watching, if I so opt in a privacy compliant way, my behaviors, the way I act, literally my gestures, the way I look at things, the way I talk, would we allow you to have your own virtual doppelganger (00:15:06) if you wanted. (00:15:06) Wow, that's fascinating. (00:15:08) And then #3 would be is if I have a good virtual doppelganger, just as agentic is all the buzz in other areas, is there a simple user interface for agentic virtual doppelgangers? (00:15:22) And that starts to get interesting because one could imagine sending their virtual self out to go do something. (00:15:29) It could be, (00:15:31) my son or my daughter wants to play with me. (00:15:35) I'm having to do some work. (00:15:37) Could my virtual doppelganger fit in for 15 minutes? (00:15:40) Sure, that's great. (00:15:42) That could be really interesting. (00:15:44) And there's a lot of other (00:15:46) Some things just like we're going to see agentic space in the 2D and the workspace and the productivity area, I think it's going to be interesting to imagine. (00:15:54) So, you're basically using these 18 million hours a month to train 13 billion 13 billion, yeah, to train a native FPC model that will be a Roblox privacy compliant model, never share the data. (00:16:06) And then over time, no ship date, make that customizable as your own virtual doppelganger or for creators with the addition of a prompt. (00:16:17) Upload Benjamin Franklin's history or it's already embedded in an LLM and just say, hey, I want Ben Franklin. (00:16:25) And I want a happy Ben Franklin. (00:16:27) Are there experiences that you imagine creators making? (00:16:32) I think there's a lot of standard game NPCs that I think have done such a good job. (00:16:43) I wouldn't claim when these types of NPCs would be good enough. (00:16:46) But like Grand Theft Auto, right, they've been working on this forever. (00:16:50) That is a very rich world with a lot of very high performance, you know, NPC behavior. (00:16:57) Imagining someday many people can create them. (00:17:02) I think, to make that happen, these platforms have to be very cloud-connected platforms, rather than running local to spin up inference on all of these things, but you could imagine that you could imagine a... (00:17:18) very young creators saying, I'd like to build an American history thing where you go and talk to Benjamin Franklin. (00:17:26) So I think we would want to enable both of those. (00:17:28) It's really interesting because basically the transition, this is going to be a poor analog, but there's a transition that happened in self-driving where we went from mapping out heuristics. (00:17:37) in some behaviours and sort of fine detail to capture those cases, to just doing end-to-end models for deep learning. (00:17:44) And it sounds like you're basically considering doing the same thing for NPCs. (00:17:48) In other words, I think that's an exciting computer science capability is rather than a, you know, all of these wonderful techniques people have used, you know, decision trees and this and this, getting to a purely more general modeling solution would be really fascinating. (00:18:04) Yeah, super interesting. (00:18:05) What do you think the role of (00:18:07) like a Roblox game designer or any game designer is five years out from now. (00:18:13) Do you think it's different? (00:18:14) I think I'm optimistic that typically industrial revolution or whatever revolution, I think we've chatted about stuff like this before, there's a thought that there won't be any jobs after this. (00:18:30) I think I'm a little bit more optimistic that a new class of things will show up. (00:18:36) And I think there's (00:18:37) a huge opportunity for more creators. (00:18:40) I think I'm not to the point where I believe what we call AI slop today is going to take over. (00:18:49) I think the human touch has got a long runway on it. (00:18:53) So you could imagine more diversity or quality of experiences supplemented by AI. (00:19:01) But I do think (00:19:03) the role will be more leveraged and the quality coming out of five people will be astounding, but we'll come to expect that quality. (00:19:14) So quality will go up. (00:19:15) I think iteration on testing will go up. (00:19:19) I think where we see platforms like Roblox going is just as (00:19:25) People are starting to spin up agents to develop code. (00:19:28) I think people, hopefully on the Roblox platform against the Roblox cloud, you'll be able to go away for 24 hours and your agents are going to be like... (00:19:39) tweaking and testing, spinning up a Roblox experience, sending 20 NPCs into them on various client emulators. (00:19:48) There's an NPC on a phone, there's an NPC here tuning the game. (00:19:51) So hopefully you'll feel more power in what you're building. (00:19:54) I'm also curious, there's a set of research around world models, and I'm sure you've seen that it's just trying to directly generate game experiences with like, there's no physics engine underneath, like you don't control the mechanisms, it's just video. (00:20:09) What's your view on that? (00:20:10) I think it's like a super exciting, interesting thing to think about. (00:20:15) I think, like I said before, I think one thing is going to be the higher the fidelity, is memory stored in video latents? (00:20:24) Or does there need to be another research breakthrough where memory is ultimately stored in some new latent space that's 3D or 4 dimension? (00:20:33) But the fidelity is amazing. (00:20:35) I think that's really beautiful. (00:20:37) And I think that (00:20:38) The way we would see the ultimate architectures of these platforms is probably for a while not an all-in-one thing. (00:20:46) We would probably see some hyper-efficient thousand-person synchronization state engine. (00:20:53) We might also see some of the photorealism not being done server-side, but being done in intermediate spaces or client-side as well. (00:21:03) And we may see multi-step pipelines where there's the raw framework, there's 3D (00:21:08) upsampling, then there's local 2D upsampling. (00:21:11) We're actually going, oh my gosh, people are putting together a bunch of pieces, dedicated NPC capability, dedicated synchronization capability, 3D upsampling, 2D upsampling, and potentially world model stuff. (00:21:26) World model, you can imagine that being used even in its early state today, rather than maybe a gameplay state, where rather than (00:21:36) making a video, you talk about the video while you walk around. (00:21:40) And you kind of say, put me in a Western world, I want to go left, I want to go right, add a few horses. (00:21:45) And that video actually goes to a multiplayer experience. (00:21:49) One could imagine that. (00:21:51) I think what a lot of other people are wondering is the short form video, does someone hit it just like they did with short form video into a highly retentive, like local dreaming product that can just react to what you're doing? (00:22:05) I think it's still early for (00:22:06) both of those, but it's a big possibility. (00:22:09) I would say though, we are doubled down on a hybrid multi-AI tech stack that supports multiplayer natively. (00:22:18) Yeah, I think one of the ideas that's like internationally very interesting has been the idea that you could generate like serial micro dramas as a form that is not yet that popular in the West, but seems like a good format for the technology. (00:22:32) Micro dramas, I've watched a few. (00:22:34) I mean, it's really interesting. (00:22:35) You know, a few, it's fun to see (00:22:36) like some entrepreneurs tried to do that. (00:22:38) I won't mention who, like 5 or 10 years ago, then they naturally emerged kind of thing. (00:22:44) It's interesting to think if those micro dramas are easier to produce, they don't need live actors, but (00:22:52) what is the plot line that really resonates? (00:22:55) we'll see with that. (00:22:57) And you can mass test that at scale if you're using AI, right? (00:22:59) You can actually have, a very large number of iterations. (00:23:02) This is kind of one of the arguments people make around the Fermi paradox of, like, why haven't we seen... (00:23:07) alien life. (00:23:07) And the idea is it's just there's alien life. (00:23:09) They're just stuck in these virtual worlds now where, you know, it's better than the real world. (00:23:13) I heard that. (00:23:14) Like, what's the famous quote? (00:23:15) Where is everyone? (00:23:17) Yeah, exactly. (00:23:18) So one option is there's some virtual landscape that's more compelling than the real world. (00:23:23) And so therefore there's no reason to travel the universe. (00:23:26) Or it could be, you know, we could really start. (00:23:29) conjecturing what's the future evolution? (00:23:32) Is it human or machine life and all of that? (00:23:35) And once it gets to maybe human machine life, what happens? (00:23:39) Yeah, there's subsumption, yeah. (00:23:40) that's interesting. (00:23:42) Yeah. (00:23:42) That's the vanilla sky. (00:23:44) Yeah, exactly. (00:23:45) Or the optimistic thing is more the holodeck thing. (00:23:49) Like everyone's hanging out with other people rather than a simulation of themselves. (00:23:53) Yeah. (00:23:54) Are there experiences that you imagine from a like a communication