推出 4D Creation 开放测试：NPC、4D 世界与 Roblox CEO Dave Baszucki 对游戏未来的展望

What proportion of you think your children will have at least one serious relationship with an AI throughout their lifetime?

I think zero right now.

Yeah.

I don't think we're close to crossing the human/AI barrier, but it is worth thinking about. Things are changing every day. We have to incorporate that. And then there's also some weird universal truth of things that stick around for 40 years, and you kind of have to blend them together. We feel if we build Roblox right, it might be the kind of thing that's kicking around in 40 years.

Hi, listeners. Welcome back to No Priors. 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. We talk with Dave about the future of AI in gaming, Roblox's investments in AI, how NPCs are going to change and really transform games, Roblox Studio—one of the largest coding platforms out there—and their 20-year mission to build the Holodeck. Welcome, Dave. Thanks so much for doing this with us.

It is so great to be here, and thank you for doing this at Roblox headquarters.

Yeah, exciting to be here.

Great to have you guys. Thanks.

I think we wanted to start with a big-picture question. 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 and what's going to happen in the future there. That's across things like world models, how you think about NPCs and their evolution, how you think about assets within games, and how you think about world creation. We'd just love to hear your big-picture views of where we're going 10 years from now. What's coming?

1. Roblox Builds the Holodeck

Two steps on this one. First, I know you both dabble in investing a little.

Yeah.

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. It portends a new category. One could call it the category of human co-experience. It's got the sizes of a bunch of companies from 20 years ago. You can see social-networking companies, you can see YouTube video, you can see toys, and you can see all of that. What it imagined is really a new category. 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.

It feels like Ready Player One, the movie.

Exactly.

Yeah.

We had always imagined that if this high-fidelity space was backed by physics simulation and reality simulation, you'd be able to do stuff. You'd be able to build a car, put wheels on it, and drive it around. You'd be able to go to a birthday party and blow out the candles. You'd be able to chop trees down and make a house out of it. That was literally the genesis of Roblox, and here we are, 20 years later. There's so much more to do.

One vision of AI is: How do you superpower the creation of that Holodeck, and how do you get it photorealistic as soon as possible? How do you get 10,000 people in it instead of 100? How do you do acoustic simulation that sounds realistic with 10,000 people? In that sense, AI is super interesting for trying to achieve a vision that's been around for 20 years and has arguably been in science fiction—Snow Crash, the Holodeck—for even longer.

2. Real Time Dreaming Emerges

There is another interesting vision, and I think it's useful comparing these product ideas. There's the future of just a single person by themselves dreaming what they want to dream. We sometimes call that the real-time-dreaming category. It's a category that we can see a bit of the trappings of in short-form video. There's a little bit of that in short-form video: someone's at 2:00 a.m. in their bed doom-scrolling.

Uh-huh.

It's reacting to dwell time. It's reacting to what you favorite, and you're getting a little bit of a prebuilt dream. I think that could go all the way to the famous Tom Cruise movie, Vanilla Sky, where he literally dreamed in real time 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.

So that's more world-building as content in a continuing way.

That is exactly right. If you put those at two extremes, one is a communication platform, high fidelity, with people you know. The other is real-time dreaming, where everyone's an NPC. Everything in between is possible, and I think we see everything in between ultimately as being possible. We may see weird product categories we never imagined.

People are probably thinking, does the text prompt turn into a voice prompt, and does it automatically turn into a video prompt? For example, “I want to learn French. Video how you're in a French café and start talking to an NPC.” I think there's a wide range of products that start to get into this space we're interested in, which is that 3D reality space.

Do you think that same space will exist for both consumer and business applications? People also talk about this in the context of, instead of doing a Zoom, why don't you go into a shared space with someone and meet with them? What degree of your life becomes virtual with that kind of technology?

I think on the communication side, people just want more and more fidelity. We can go back to paintings on the wall of a cave. People want more fidelity. We had smoke signals. We had the mail system. We had the Pony Express. We had the telegraph system.

Uh-huh.

We have the phone system. We have, arguably, the COVID-initiated video experience. We also had the COVID-initiated 3D experience on Roblox, actually, and we found that side-by-side video was used by a lot of young people as an early test case of how to co-experience.

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. If we were having this 4D immersive communication, we'd say, “Go to legacy analog mode and make it look like Zoom,” basically.

Uh-huh.

We could always do that, but we'd be able to do things we could never do on Zoom. You'd say, “Hey, let's pop up and go walk around my office. I want to show you something. Come and follow me,” kind of thing. I do think 4D simulation, as it gets photorealistic, will be a superset of video.

The other thing for business is that there are a few things that 4D—and I use 4D because it's not just 3D shape, but function in a way—will provide that video does not. One is acoustics. There are some really tricky physics problems when you're having a company meeting of 1,000 people. If we try to do it on Zoom, we have a bunch of squares. Who's live? Who's not live? How do we mix the sound? That's very tricky.

If we do it in a three-dimensional simulation, I can hear everyone with attenuation based on how far away they are, and it can actually be a much more natural experience. I walk closer to you, and I hear you. That may actually be a better human-interface paradigm.

There's one final technical thing that's going to be very, very difficult: when people try to sing “Happy Birthday” to you together. No matter whether you're on Zoom or in a 3D simulation, you're hearing them about 30 milliseconds later.

Late. Mm-hmm.

What I think we're going to find is that we're actually allowing each one of us to sing “Happy Birthday” with a time-forward extrapolation of each other and then remixing it. For the business side, I do think ultimately we're going to see some interesting conferencing solutions.

Roblox has this amazing, wild story with physics simulation, or the physics engine, as part of the birth of Roblox.

Yeah.

But today, or for the last decade, I think of this as an immersive world where people have their first coding experiences. They build amazing games. They interact with each other. I don't think of that world as being super focused on realism.

That's right.

Is that something that's changing for you guys? Is it enabled now? How should we think about that?

The way I would think about it is that we're pushing really hard on the physics-simulation capability, and at the same time, we have this incredible free market that we can't control on content generation.

I would say we just started talking about full acoustic simulation. I just tweeted about it. Proximity attenuation and all of these kinds of things, as well as echoes, are going to start happening. On the physics simulator, it's gotten better and better and better behind the scenes. It'll keep getting better.

But what we have found, I would say, is that the diversity of developers is beyond our expectations. Some of the things that have gone viral—Dress to Impress, for example, and Grow a Garden—are less physical. But they have leveraged cloud capabilities that we've built side by side. 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.

Dress to Impress was able to do some very interesting things with how they run contests and layer clothing on. The specification for our product could be: We just want to simulate the real world with 10,000 people. The more we go in that direction, the more we'll see creators using it in different ways.

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 or simulations?

I think we're really watching it. The biggest thing we've 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 to 100 or 10 or 1? We feel it's super robust to build a communication platform, and that's part of our goal of getting to 10% of gaming.

I think we're going to start unraveling the technology for this over the next 5 to 10 years. What's the most efficient way to synchronize the state of 10,000 people?

I see.

And when we—

So it's an infrastructure problem versus necessarily a physics simulation problem.

How do we—

Yeah.

—synchronize the state and the memory over the last 5 hours of 10,000 people? Where have they been? Where have they been in 3D space? What have they been doing? The world-model thing is very exciting because they're storing memory, literally video latency—a minute of frames. It's showing super-early promise.

Mm-hmm.

3. World Models Store Memory

I think the technical revolution is going to be: What's the ultimate synchronized state for 10,000 people? Is it in video latency? Is it in a much more native 3D format? Is it a hybrid, a new discovery of some 3D video latent space? To be determined.

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, and to store it not rasterized like video, but as vectors. That means the ability to play back anything that's ever happened on Roblox. That's 13 billion hours a month, and so the ability—

What do you think that does for you, the history of your world?

There's a sci-fi book by Arthur C. Clarke called The Light of Other Days, which talks about the fantasy of what would happen if we had infinite playback in our world. 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 that everything could be played back, but you could look at any playback—

Uh-huh.

—and so everything's completely transparent. There are no private conversations. It actually delves into that.

I think for us, we would use it very thoughtfully, in a privacy-compliant and judicious way. But for very simple things, you can imagine: There was a safety incident. Let's go back, put 5 cameras in place, listen to the audio, see what happened—

Mm-hmm.

—and make a good judgment call. That's kind of interesting.

There's a really special thing that maybe I did with someone in my family that's really special. Given that I have access to my history, I could play that back and possibly, because it's 3D vector data, not raster data, reshoot it and make a cool video—

Mm-hmm.

—of the best moment.

Yeah.

I think where it ultimately ends up for us is that there's a huge push now for people to find hybrid video ASDW human interaction data. People are starting to train on video with keyboard input, and the question is: Where do I find 200 million hours of this data?

The data we have—13 billion hours a month—can be reproduced from any camera angle and can interact with the 3D space. So it's very powerful data. It leads to a really interesting computer science program idea around how to create great NPCs that are more than just an LLM.

Okay, tell us more about that. Actually, that sounds really interesting.

Go check out on X or XFeed. 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. What we are starting to train is NPCs based on a combination of all the data we have in a privacy-compliant way—

Uh-huh.

—that are starting to navigate and starting to play games. The challenge, I think, would be—challenge level 1—everyone has access to an NPC that can get pretty good at playing any Roblox game.

Mm-hmm.

Uh-huh.

Super cool.

Challenge level 2 is: By watching, if I opt in in a privacy-compliant way, my behaviors—the way I act, literally my gestures, the way I look at things, the way I talk—

Uh-huh.

—would we allow you to have your own virtual doppelganger if you wanted?

Oh, that's fascinating. Yeah.

Number 3 would be: If I have a good virtual doppelganger, just as agentic AI is all the buzz in other areas, is there a simple user interface for agentic virtual doppelgangers? And that starts to get interesting—

Yeah.

—because one could imagine sending their virtual self out to go do something. It could be, you know, my son or my daughter wants to play with me. I'm having to do some work. Could my virtual doppelganger fit in for 15 minutes?

Uh-huh.

Sure, that's great.

Yeah.

That could be really interesting. There are a lot of other things, just like we're going to see agentic space in 2D, in the workspace, and in the productivity area. I think it's going to be interesting to imagine.

So you're basically using these 18 billion hours a month to train—

13 billion

13 billion, excuse me—

Yeah

to train—

Native—

A native NPC—

NPC model.

Yeah.

A Roblox privacy-compliant model that will never share the data. Then, over time—no ship date—we could make that customizable as your own virtual doppelganger, or for creators with the addition of a prompt. Upload Benjamin Franklin's history, or, if it's already embedded in an LLM, just say, “Hey, I want Ben Franklin.”

Yeah.

“And I want a happy Ben Franklin.”

Are there experiences that you imagine creators making with these NPCs that are really different from the games we have today?

I think there are a lot of standard game NPCs that have done such a good job. I won't claim when these types of NPCs would be good enough, but Grand Theft Auto, right? They've been working on this forever. That is a very rich world with a lot of very high-performance NPC behavior.

Imagining someday that many people can create them, I think, to make that happen, these platforms have to be very cloud-connected platforms rather than running locally, to spin up inference on all of these things.

But you could imagine that. You could imagine a very young creator saying, “I'd like to build an American history thing where you go and talk to Benjamin Franklin.” So I think we would want to enable both of those.

It's really interesting because the transition is—this is going to be a poor analogy, but there's a transition that happened in self-driving. We went from mapping out heuristics and sub-behaviors and fine detail to capture edge cases to just doing end-to-end models for deep learning—

That's right.

—and it sounds like you're basically considering doing the same thing for NPCs. In other words—

I think that's the exciting computer science capability. Rather than all of these wonderful techniques people have used—decision trees and—

Yeah, all the hand mapping.

—this and this—getting to a purely more general modeling solution would be really fascinating.

Yeah, super interesting.

What do you think the role of a Roblox game designer, or any game designer, is 5 years out from now? Do you think it's different?

I think I'm optimistic that, typically, with an industrial revolution or whatever revolution—we've chatted about stuff like this before—there's a thought that there won't be any jobs after this. I think I'm a little bit more optimistic that a new class of things will show up.

I think there's a huge opportunity for more creators. I'm not to the point where I believe that what we call AI slop today is going to take over. I think the human touch has got a long runway on it. You could imagine more diversity or quality of experiences supplemented by AI.

But I do think the role will be more leveraged, and the quality coming out of 5 people will be astounding.

Mm-hmm.

We'll come to expect that quality, so quality will go up. I think iteration on testing will go up. I think where we see platforms like Roblox going is that, just as people are starting to spin up agents to develop code, hopefully people on the Roblox platform, against the Roblox cloud, will be able to go away for 24 hours, and their agents will be tweaking and testing, spinning up a Roblox experience, and sending 20 NPCs into it on various client emulators.

That's fascinating.

There's an NPC on a phone, and there's an NPC here tuning the game. Hopefully, you'll feel more power in what you're building.

I'm also curious: There's a set of research around world models, and I'm sure you've seen that, that's just trying to directly generate game experiences—

That's right.

—with no physics engine underneath. You don't control the mechanisms. It's just video. What's your view on that?

I think it's a super exciting, interesting thing to think about. I think, like I said before, one question is: As the fidelity gets higher, is memory stored in video latent space?

Mm-hmm.

Or does there need to be another research breakthrough where memory is ultimately stored in some new latent space that's 3D or 4D?

Mm-hmm.

The fidelity is amazing. I think that's really beautiful. I think that the way we would see the ultimate architectures of these platforms is probably, for a while, not an all-in-one thing. We would probably see some hyper-efficient 1,000-person synchronization-state engine.

Mm-hmm.

We might also see some of the photorealism not being done server-side, but being done in intermediate spaces or client-side as well. We may see multistep pipelines where there's the raw framework, there's 3D upsampling, and then there's local 2D upsampling.

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." You could imagine a world model being used even in its early state today, rather than in a gameplay state, where rather than making a video, you talk about the video while you walk around.

You kind of say, "Put me in a Western world. I want to go left, I want to go right, add a few horses." And that video actually goes to a multiplayer experience. You could imagine that. I think what a lot of other people are wondering is: Does someone turn it, just like they did with short-form video, into a highly retentive, like, local dreaming product that could just react to what you're doing?

I think it's still early for both of those, but it's a big possibility. I would say that we are doubling down on a hybrid multi-AI tech stack that supports multiplayer natively.

Yeah. I think one of the ideas that's internationally very interesting has been the idea that you could generate serial micro-dramas as a format that's not yet that popular in the West, but seems like a good format for the technology.

Absolutely.

Yeah.

Micro-dramas—I've watched a few. It's really interesting. It's fun to see that some entrepreneurs tried to do that—I won't mention who—5 or 10 years ago—

Yeah, right.

—and then they naturally emerged, kind of thing. It's interesting to think: If those micro-dramas are easier to produce, they don't need live actors. But what is the plotline that really resonates? We'll see with that.

And you can mass-test that at scale if you're using AI, right? You can actually have—

That's right.

—a very large number of iterations. This is kind of one of the arguments people make around the Fermi paradox: Why haven't we seen alien life? The idea is there's alien life; they're just stuck in these virtual worlds now, where it's more fun than the real world.

I haven't heard that.

Yeah.

What's the famous quote? "Where is everyone?"

Yeah, exactly.

Where is it?

So one option is they're in some virtual landscape that's more compelling than the real world, and so therefore there's no reason to travel the universe.

Yeah, it could be—

Yeah.

—or it could be—

Yeah.

—we could really start conjecturing about the future evolution. Is it human or machine life, and all of that?

Yeah.

And once it gets to maybe human-machine life, what happens? So—

Yeah, the subsumption, yeah.

Yeah, that's interesting.

Yeah.

That's the Vanilla Sky theory.

Yeah. Exactly.

Or the optimistic thing is—

Yeah.

—more the holodeck thing.

Yeah, yeah.

Like, everyone's hanging out with other people rather than a simulation—

Mm-hmm.

—of themselves.

Yeah.

Are there experiences that you imagine from a communication or social perspective that you think are going to be just different or new, or that you're going to have real insights into from these really great NPCs? Because I'm sure you've seen what has happened with the rise of AI companions and how engaging they are, given they're just text interfaces.

There's a Black Mirror dating episode. I don't know if you've seen it, where they have a lot of virtual-doppelgänger power.

Yeah.

Obviously, Roblox is not doing dating. We're just talking about this right now—18-plus, ID-verified, all of that.

Yeah.

So let's not go crazy, but let's think about a future. In the Black Mirror episode, you literally can spawn the lives of 1,000 virtual doppelgängers. Those 1,000 virtual doppelgängers—you're on your new dating app, and in 3 minutes, someone else's virtual doppelgängers live 1,000 virtual lives with your virtual doppelgänger. Then they say, "What's the success rate?"

When the 1,000 virtual doppelgängers say, "98% of the time we had a good life," then you say, "Oh, we should go on a date—

Uh-huh.

—or something."

Yeah.

So I think—

Like social simulation. That's going to be useful.

There's a lot of things that I think are going to be really hard to project. I think one of the things we would focus on is: The more we can just build raw, high-performance infrastructure, have AI as a service, have all of these things capable in a cloud, have the same thing run on a phone as a computer, and have it auto-translate, hopefully we would just cover some of these things.

Yeah. I'm really optimistic about that, given the engagement that people see with these companions as just a text interface with the beginning of memory, right?

Absolutely.

Versus if you're Roblox, you have embodiment. You have, as you said, cloud persistence.

That's right.

And I think it's just going to make the characters much more interesting very quickly. Yeah.

I'm thinking through how we would embody memory for NPCs, which is more than just a prompt. One could imagine any NPC on Roblox, because we have a recording of the whole thing, actually having access to everything they've ever said or done in a very lean format, so they can go back through that and retrain on that.

Hopefully, that would be a native part of the platform. But I do know there's a lot of interesting engagement with these types of companions, and imagining fully 3D versions will be pretty fascinating.

Yeah. A friend of mine basically asks at dinner parties—he runs a well-known AI company—and his controversial question for the group is often, "What proportion of you think your children will have at least 1 serious relationship with an AI throughout their lifetime?"

I think zero right now.

Mm-hmm.

I do. I don't think—

Yeah.

—we're close to crossing the human-AI barrier. But it is worth thinking—

He meant it even in the context of, even if it's superficial and there isn't true intelligence behind it, just given the behavior with text, once you start having video, virtual worlds, 3D environments—

I can imagine a real-time coach or therapist in your earbud—

Mm.

—just hanging out all the time.

Yeah.

So I think for that kind of less evasive, more over-the-shoulder coaching, I think that seems like—

And that's already a major use case, it seems, for ChatGPT and other platforms, where mental-health advice or coaching or other things like that are actually one of the major use cases today. So, yeah.

Yeah. I think that will be arguably very valuable. Mm-hmm.

Can I ask an industry-perspective question, just given you've worked in gaming for a very long time now? Why do you think that there's a lot of discussion about how assets of different types are cheaper to create with generative tools? I don't think that's really changed the dynamics of the gaming industry yet. I don't know if you would agree with that claim or explain it.

4. AI Rebuilds Game Creation

One could say that no matter how cheap it is to create assets, the expectation of quality from consumers goes up at exactly the same velocity.

Okay.

And so what was not possible a while ago will become possible, and the expectations of consumers will go up. I think there's a general notion that the technology for the gaming industry has to move to the cloud, vertically integrated, and asset management can't just ship on a DVD or be a giant download.

Uh-huh.

I think we're really seeing this. We've launched dynamic LOD for textures. Meshes are on the way. You could imagine every single asset in the game—sound, 3D object, image—existing in various formats. The traditional format is a texture, a mesh, or an audio file.

Mm-hmm.

The future format could be an AI prompt, and it's generated procedurally on demand. There are these 2 things. They're both very AI-enabled.

One can then imagine that you have to be cloud-connected, because for all of these assets, if there's a full range of LOD, just like with video streaming and upsampling, low-end devices need low LOD, and high-end devices need 4K LOD. Ultimately, there's on-demand AI generation.

That then gets into, once again, one of the demos is probably on 3D upsampling. The notion is that if someone makes a very primitive experience on Roblox, they would augment it with a prompt and say, “I know it's kind of primitive, but make it kind of look medieval and more realistic,” and have all of those assets auto-upsample in 3D in the cloud for free, on demand. That's the kind of just-in-time thing we're thinking about: how AI can accelerate if you have a cloud platform.

Roblox is also, from the studio perspective, one of the biggest development environments in the world.

That's right.

What are you seeing from the perspective of AI coding there?

This is great. Once again, industry-standard ways of using a different model with Studio—so if you're in cloud code, you can control Studio—are super exciting. A lot of our users have started gluing these things together, with the hope that ultimately they can build stuff from there.

Roblox Studio itself is running a native assistant as well, and a native code engine. I think the pattern for that is that Roblox Studio will, more and more, follow many of those industry best practices for code, but in parallel needs to have an environment-generation component that's very unique to that.

I think the future of environment generation is, once again, any prompt you can imagine: text, image, video, possibly walking around in a world model to define a world, having it iterate on that, iterate on that to a primitive 3D skeleton. Do I like it? Then iterating on that to a fully functional game.

I think that's our vision: to run those side by side, so a first-time Studio user gets it out of the box. At the same time, we plug into all of your AI workflows. The final thing is that, because it's so cloud-connected, you have all of that cloud capability we talked about earlier: spinning up jobs, spinning up agents, running test plans, and keeping grinding away, trying something different.

Just out of curiosity, what is the cycle of development for a successful Roblox game today?

What we have seen—

Yeah.

Over the years, we've seen bigger teams. The top Roblox creators now are well into the tens and tens and thirties plus million a year. It's really serious stuff.

We've seen in the Roblox creator community, I think, a healthy sign that the average yearly revenue of creator number 1,000 is going up faster than that of creator number 1, so that's a healthy long tail. All the way out to creator 1,000, there's such a huge community of people making a living.

What I think our creators have found is that it's gotten both healthier and more competitive. I think we've seen the pattern now of much more live ops in that, because this is cloud, you can update your experience just like good websites, every day or every week. That's just a constant practice. I'd say update frequency is much more rapid.

Mm-hmm. Which is great for users because they get updated content at a faster rate and so on. Yeah.

That's right. Users love updates. They love new things. It keeps games fresh.

The other thing we've seen is, on the discovery side, we've tried to take the approach more and more that we want discovery to be completely transparent. These are all the things that are going into the discovery algorithm. We check it. There's a benefit to that transparency in discovery: it actually keeps a lot of pressure on us. “Oh, my gosh, someone's going to game discovery.”

Well, then we'd better make it really good, right, and transparent. I actually think transparent discovery and recommendations is an interesting trend for the industry worth talking about.

With that, we've also seen the content distribution get richer and a little less spiky. There are now, like, the top 20 experiences on Roblox are all really good and all vying for the number 1 slot, which maybe wasn't so true 3 or 4 years ago.

Zooming all the way out for a second, I think it was true 5 or 10 years ago when we met. It was true when I was talking to one of the members of your leadership team today.

They will say you're an extremely high-conviction, very long-term-oriented leader, right? So, the 20-year plan for Roblox is to do things in line with the vision we have. You're obviously also all in on AI, right? Very excited about the technologies and investing in them.

How do you square these things, where you have a very long-term view when so much is happening in the environment all the time?

5. Roblox Plays the Long Game

I think we always want to pair “take the long view,” which is one of our 4 big values, with “get stuff done.” It's almost like in those old Gartner charts.

Yeah.

“Take the long view,” not “take the long view.” “Get stuff done,” not “take the stuff.”

Yes.

The magic quadrant is “take the long view, get stuff done.” By “get stuff done,” we typically mean rapid iteration inside the company.

If someone has a view of some exciting new product and we're going to ship it in 6 months, that can be a very scary thing unless it can be broken down to, “We're going to ship this thing every week and iterate toward this,” but we have a good target of where we're going.

Inside the company, I'd say our AI team, our facial age estimation team, and our safety team—these are literally things working on a weekly basis, which allows a fairly fast reaction time. AI, I think, is arguably, as far as speed right now, just every day.

But you mean, you make it sound pretty simple in that all the things that you're trying to get done on a weekly or 6-month timescale are all aligned toward a vision that isn't really affected by it.

I think if you have your 6- to 12-month vision—

Yeah.

Hopefully that's moving not too fast.

Mm-hmm.

At the same time, the weekly iterations are moving very quickly in that direction. You could say the Roblox vision is ultimately to build the Holodeck. It's 10,000 people. It's multiplayer, real-time modification of the environment. It's NPCs.

Mm-hmm.

Photorealistic. That's a pretty stable spec for 20 years.

Yeah.

And so then—

I actually feel like that's the biggest takeaway for me here. I spend a lot of time with software CEOs, right? I think there's a lot of actual concern about whether the vision we had for the company, in terms of its durability or its value, is still as valuable as we thought 5 years ago.

For Roblox, perhaps because you were pointing at something so far away to begin with, it's pretty clear: AI only moves you toward it.

I think we—

You have crisped that up over the years.

We've crisped that up over the years. I would say 2 years ago, we were a little less crisp. We would say, “We're going for 1 billion DAUs.” That's a good spec with the spec I just mentioned. It's a little harder to operationalize around product management teams and that.

So we took a step—a stepping stone—and said we want to get to 10% of global gaming content on the way to that. It's actually a remarkably good step. It's about 300 million DAUs. It's about 20 billion. It's about 3X where we are.

It can be forensically torn apart. We can see what part is the U.S. market. We actually have a higher target for the U.S. than 10%.

That has been very clarifying. I think for a software CEO, always knowing both 3X and 10X is very helpful, because if you don't know 10X, then it's hard to sleep at night. If you don't know 3X, it's hard to plan forensically how we're going to operationalize that and how quickly we're going to get there.

One other thing that I think Roblox has done uniquely well is how you've thought about hiring and how you've rethought certain aspects of that. I think that's really led to people who have more bias to action, who are moving quickly, who are iterating rapidly.

Can you talk a little about how you've rethought some aspects of that, if you're willing to share that as part of this?

One is, we've been very public that we really want values-aligned people who are good problem solvers, who are very creative, and it's hard to do signal-to-noise in that kind of a hiring thing. So what...

We actually bought a company about 5 years ago that was trying its best scientifically to build 3D assessment tooling for some big companies to really find people who are problem solvers and creative thinkers.

That's part of what I was referring to.

Yeah.

I thought it was fascinating that you made that acquisition.

So we bought Imbellus. We got a great team of scientists, people that had been involved with the SAT test and all of that, and we've tuned that to our new college grad intern program so far. We've essentially, I believe, in a very fair meritocracy kind of way, taken the—whatever it is—10, 50, 60,000 people and run them all through what we've, I think, shown to be a very fair assessment, not weighted by any other social factor.

The people that come out of there give us incredible confidence that they're problem solvers. They do a bunch of other things really well. So we have this building wave inside the company from 3 years ago, 2 years ago, and 1 year ago, of these 100, 200, 300, 400 people that hopefully we're mentoring—some of the best future leaders in the industry.

Yeah, super interesting, because I have a number of friends who always talk about how you discover untapped sources of talent—

That's right.

—or how you find the best people in the world, and are there ways to test for that or sort of screen for that?

We've done a lot of testing on this—fairness testing. Looking backward on it, I'm very confident in where it's going to go for us.

Super interesting. Yeah.

It's a very systems-oriented approach. The one thing we have found, which is a little edgy, might be the correlation between traditional universities and our own testing.

Sorry, what's edgy about that?

Well, I think people typically think, "I went to Stanford. That's where all the top talent is."

Oh.

We have found that not necessarily to be true. And we have found talent at community colleges and small Midwestern engineering schools. Because we're assessing in our own way, we basically ignore the signal of where you went to university.

I've heard that anecdotally from other CEOs now, particularly for certain cohorts, like the cohort that came during COVID. For a lot of the top universities, it's perceived as a weaker cohort. I don't know the reason or the interpretation behind it, but it's interesting to see this shift in terms of how people even think about where some of the best talent is.

Yeah.

So it's interesting that you're saying that as well.

So we get to have self-determination. We get to say, "Colleges, you do what you want"—all of that. We're just going to do it in a fair way that we think is our way, and it's worked out.

Can you give, because it's so interesting, any intuition for some way you test for problem-solving capability?

I think we make public some of the tests that we use.

Okay.

And the cool thing is they're all built on Roblox. They're all tested by lots of people, and they're arguably really interesting three-dimensional problems, like programming a factory or using some new geometric language to program a robot.

Okay.

It's almost like turtle graphics.

Cool. Yeah.

And you can play with that and see how people—

We're going to try to get jobs tomorrow.

Yeah.

Yeah, yeah.

That'd be awesome.

I was thinking I'd play with that. Sounds fun.

I might actually be a little afraid to take the test myself, given how good some of our candidates are.

As a long-term predictor of the future, is there something happening in AI that you believe or hold with conviction, like 5 years out, that you think not everybody believes yet?

I think typically it's the other way. More people believe more stuff than I believe.

You think of yourself as a skeptic?

He's building the holodeck, but he's a skeptic.

I think I'm more about the power of time and the power of compounding rather than an amazing invention and just rapid iteration. One thing I am constantly amazed by is either the graphical user interface or Microsoft Excel. I think I was playing with Microsoft Excel in the '80s. And here we are 40 years later. It's not that different.

Yeah.

Still very useful.

It's an interesting complement. In the AI world, things are changing every day. We have to incorporate and do that, and then there's also some weird universal truths of things that—

Yeah.

…stick around for 40 years, and you have to blend them together. I think the good news is that we feel if we build Roblox right, it might be the kind of thing that's kicking around in 40 years.

Awesome. Thanks so much, Dave.

Thank you for joining us.

Thank you. Okay, cool.

Yeah. Good to see you.

That was really good—

Yeah.

…good questions.

Yeah.

Find us on Twitter at nopriorspod. Subscribe to our YouTube channel if you wanna see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-priors.com.