The Exchanges

Q Maybe, uh, take us, um, a little bit behind the scenes in terms of the core infrastructures. We have this global audience, uh, that interacts in real time and creates three D worlds in, in real time, um, from a, uh, yeah, core infrastructure perspective to scale up and down. How does that work?

A So it's an incredibly complicated in-house technical stack in order to present a really robust and simple interface to the user. So we have our core data centers. Um, we have about a 100,000 servers that we maintain with our own infrastructure, um, running our own software stack on those. And that that's the heavy lifting backend. That's how we persist all of the data. Um, you have an avatar in the virtual world. It's how we keep track of what you're wearing, what the social graph is, um, all of the items in the world, all the three D worlds. We have edge data centers around the world. I think we're up to 17 of those, and so wherever you are, there's geographically located near you an edge data center so that we can have low latency, so that the experiences can be really responsive, because there's not a lot of telecommunications delay. And then we, of course, have the clients, which run on every device, so you could have a game console, you could have a VR headset, a lot of our users are on mobile devices, especially phones and tablets, Laptop, a desktop. And so we present whatever device you're on, wherever you are, we present a window into this three D world where you can interact with real people.

AI assessment note: “we have our core data centers. Um, we have about a 100,000 servers”

Q Full production. Okay. Wonderful. So, um, so what, what does it do? You alluded to this a second ago, but what does it do?

A So AI code assist, um, It's primarily targeted at helping people who are new to programming on the platform to quickly become better programmers. So for this specific feature, sort of setting aside the long-term vision, this specific feature is not enabling no-code creation of sophisticated new algorithmic development. That's not our intent. It's to say a human being is going to be the software developer, but we want to remove a lot of The sort of the learning curve for them, and we want to automate a lot of the tasks that might be more boilerplate. So with AI Code Assist, you're sitting in your text editor for writing software code inside of our studio tool. You start to write what you want a piece of code to do, or you start to write the actual code itself, and it will recognize and sort of like an email when it auto suggests, you know, do you want to say yes, absolutely, or no, I can't make it tonight. It auto suggests, but instead of auto suggesting a sentence, it auto suggests 10 or 15 lines of code that perfectly fit into your program. So it might be the equivalent of you're writing a leaderboard for a game. Maybe you're making, you know, Virtual field hockey. And it will, if you're starting to look through the players and assign scores, numbers, and it will recognize what you're doing and say, hey, here's an example of how do you iterate throughout the player. Here's how…

AI assessment note: “instead of auto suggesting a sentence, it auto suggests 10 or 15 lines of code”

Q So still following the chronology, um, AI image generator, uh, was, was the, the next one. What does that do?

A Yeah. So AI image generator, um, so you've probably seen tools, um, You know, for making sort of clip art stock photos where you, you type in a description, like, you know, cathedral with Socrates or something, and it, and it produces some image, right? Uh, and they're a lot of fun to, to play with, and there's a lot of services like this. Um, so what we were targeting instead was, because our environments are three D, um, one of the building blocks of making a three D scene is that you separately define the shape of objects, the geometry, usually by a polygon map. And at this point, I think everybody has seen the sort of like, you know, white outline, wireframe, kind of behind the scenes stuff. So you just, you define the geometry. Then you essentially paint the geometry with materials, and that a lot of the detail is not actually in the shape. It's just sort of painted onto the surface as a, you know, texture, so to speak. And there's a couple levels of that. One is, it's, it's literally the color that each pixel will be when the object appears on the screen. The more sophisticated level, which is what most of Roblox's technology is based on, is called, uh, physically based materials. And the idea is instead of painting the colors of objects, you're painting the chemical and physical properties of the material, how it's going to reflect light. Is it smooth and shiny? Is it, i…

AI assessment note: “painting the chemical and physical properties of the material, how it's going to reflect light”

Q And maybe gives a, give us a sense for the size of the engineering team to maintain this. Is that a huge organization?

A So the, the company as a whole is about, uh, 3000 people at this point. We're, we're based in Northern California. And it's heavily engineering dominated. So almost all of that is engineers or engineer associated professions. Um, a subset of those sort of associated professions are say like the research scientists, the machine learning scientists, the data analysts, uh, as well. And, um, we work on the, essentially everybody's working on the entire stack. Obviously there, there are specializations for different areas. Um, But I specifically want to call it the machine learning and AI, because I think that that's been the fastest growing area of our investment in our people, has been in taking what has been a core of adoption of AI tech for about six years now at the company, originally starting in safety, and now going into all aspects of the business, especially our generative AI, which is where most of our recent product releases have been focused.

AI assessment note: “the company as a whole is about, uh, 3000 people at this point.”

Q kind of work, um, in the context where, as we know, Generative AI, um, you know, doesn't typically get you a hundred percent of the weight there. Is the idea that it's indeed part of a pipeline, therefore the Generative AI is going to produce a, a draft, and then the creator is going to adapt it for the last mile, or does it work, um, in a different way?

A So everything we've done for generative AI in the past, and certainly our current vision for where we're going in the future, is that we want to take the tasks where the creativity is separable from the execution of the creativity, and give you automation of the execution without taking away the creator's intent and agency. And we love cases which I think both the texture creator and the material Generator do, where the output is something that it, if you're satisfied with it, you can ship that. And that's good for, especially if you're new to creation, where you might not have the skill set to go beyond that, and you're limited by your skill set. But as with the code editor, everything we do has the vision of, we want you to always be able to, if you have additional skills, to then take that output and edit it and customize it. And so it's really about Getting you off of the blank page and sort of springboarding creation, but at no point do we ever want to take away your ability to customize it down to the individual pixel if you want.

AI assessment note: “take that output and edit it and customize it. And so it's really about Getting you off of the blank page”

Q And StarCoder, you mentioned it, um, a couple of times. That's another research project, right?

A Yeah, so StarCoder was a, a multi-site collaboration between, uh, academia, multiple companies. Um, one of the, the leads is, uh, Arjun Guha, who works with us in Northeastern University. And there was this problem that there was this, um, arms race essentially between, you were hearing about these large language models. So, so things, um, Um, like GPT and all of these companies were producing them and claiming there's about, they were all closed source. So you can't really get in there and use it. You don't know what it is. You're restricted by a company's, um, terms of use, and you can't really deeply evaluate. You can only go through their interface. You can't sort of get into the actual technology. Um, and so we felt like we have a need for A language-based AI model that we can have control of. But a lot of people have this need, and it would benefit the world and the advancement of science and technology if instead of each company making their own, we got together with several groups and said, let's just make one and then share it publicly and share what we've learned. And we focused initially on the source code generation aspect, Because it was something where it was easy to get, it was a key problem for us, of course, because we wanted to do code assist, but it's also something where it allowed us to, it's easy to get licensable, legally, ethically reusable source code t…

AI assessment note: “Yeah, so StarCoder was a, a multi-site collaboration between, uh, academia, multiple companies.”

Q And those were technical and engineering kind of, kind of learnings?

A Yeah, absolutely. In a lot of, um, As I mentioned right at the beginning, a lot of the challenge of AI is we have decades of experience in software engineering and how to do traditional software development in a way that's maintainable, in a way that's efficient, that avoids bugs, that has security best practices. For AI, because you're essentially programming by feeding data into the process, um, We're not starting from scratch. There is work on this, but it's not the level of sophistication that we have for software engineering as a field as a whole. And so a lot of the learnings in AI are not on the technology side. It's not about creating a new AI model from scratch. It's about learning how to prepare data, how to augment the data so that you don't need quite as much of it, learning how to normalize or regularize the data so it all sort of fits into a standardized format. Learning how to prevent bias so that if there's some bias in the underlying training set, you know, maybe it preferred there are more red images than blue images or something. Um, you don't want that to, that artifact to come out in the net result and how to prevent things for generative AI, like hallucination, where it's, it's sort of extrapolating meaninglessly beyond the training data. So most of our, our AI papers are about either applying a model in a new way, like applying control net To animation. A…

AI assessment note: “Yeah, absolutely. In a lot of, um, As I mentioned right at the beginning”

Q a big Part of the infrastructure and games or experiences, but a big part of the, um, infrastructure that you provide is, is a, is a series of creative tools. So before generative AI, and we're going to spend the bulk of this conversation talking about generative AI, how did it work? What was the sort of spectrum of tools that people had at their disposal to create those experiences?

A So, um, because as, as we discussed Everything, all the content on the platform is created by the users. We have a suite of creation tools. So on desktop and laptop platforms, we've released our free tool called Roblox Studio, which is available to everyone. And I think one of the most important and unique things about Roblox, which really differentiates it from, from all other three D creation, um, throughout the entertainment industry, Is that on Roblox, it's not just the, the tool suite is free, is that publishing is instantaneous and free. And so it's the only platform where you can do for three D what today you can do for micro blogging, like text or images, which is one person without some sort of app store approval, without becoming a publisher, or that's, you can go and create something that you want to share with the world. Push a button and make it live in minutes around the world on all these different platforms. Um, and you can even make money on it. So you can monetize your content directly, control who sees it, and then you can price objects within that. And there's no other tool chain in the world that allows you to do that. And so that's sort of the core of the creation engine. The technology is super important, but the business and the community are really where the magic is of, That process, which is that they have instant access to the world. So they get feed…

AI assessment note: “Studio, which is sort of a best-in-breed, easy-to-learn, the three-d creation tool”

Q So continuing our tour, uh, control net. What is it?

A So this, so this is in the category of, um, I would break down our, our generative AI work across, um, Four categories. There are things for communication, like we do natural language translation. You can type something in Spanish, and I can see it in English in real time. Um, there's safety, which is things like the voice moderation or the image moderation. Um, creation tools, which are, I would say, Avatar Auto Setup is probably the most powerful, but a great example is the material generator we talked about. And then there is research, where it's, it's not product, It's, um, fundamental work we're doing that'll probably affect multiple product lines inside the company, and usually where we're doing it in collaboration with universities, sometimes with other companies, and we're sharing the results with the world. So we're trying to advance the state of the art aggressively, and it's not about a specific product, it's about changing the way we think about AI and how powerful it is. And, um, I would say our, you know, three of our recent Contributions in that space have been Starcoder and Starcoder II, which are large language models, and that there's a large collaboration with a lot of participants. ControlNet and AdaptNet. And AdaptNet is a version of ControlNet, so let me talk about ControlNet specifically. The idea of ControlNet was when we had something like, and this is …

AI assessment note: “let me talk about ControlNet specifically. The idea of ControlNet was when we had something”

Q And to the extent that you can talk about it, is that, uh, something that's, uh, fully homegrown? Are you using, uh, third-party technologies or open source? How does it work?

A So, yeah, so every, everything we do, um, uses whatever the best tool for the job is to, to present features to our community. So, um, and, and a lot of the time we are then, you know, if we make something new, such as, uh, the optimizing Luau compiler and type inference, we release it back to the community as open source. Or we've published, um, and we can get into this a little more later, um, uh, Starcoder, for example, with some of our research on this, how do you make software that programs? And so we shared our LLM with the world that we had built, um, as fully open source that anybody could build on it. And we also, more importantly, shared the methodology for how we trained it, how we taught it based on not just our own data of Lua code, but we trained on other code. And so it was able to learn from looking at Python or Java code, how to write Lua code for Roblox. And that was called a domain transfer. You learn one thing that's nearby, and then you transfer that knowledge to a different domain. And we developed specific techniques for how do you do domain transfer for programming languages. And we've also done a lot of analysis with seeing how programmers use the tool. And we've published all of this as, as open, uh, peer reviewed scientific research. So the, um, the actual tool today is our own custom front end, and then we're using a mixture of third party and propri…

AI assessment note: “actual tool today is our own custom front end, and then we're using a mixture”

Q leadership perspective, um, how did the whole generative AI, uh, sort of, uh, efforts start? Was it, uh, was it like a big wake up call when, I don't know, ChatGPD came out, or was it something that you guys had been Thinking about for a while and, and then, uh, was there a decision to go all in into generative AI or was it more of an iterative process?

A Yeah, that's a great question. And I think, let me define some terms because I think, um, being knee deep in the technology every day, um, you have a different perspective. And, and the first thing is that it is, it is very early days of AI. This is, um, I think not just for Roblox as a company and our general industry, but I think for the, the world as a whole, um, this is going to be something that affects everybody is a really powerful technology, but it's, it's not about, You know, next year the AI boom is over. It's, it's, this is, you know, the, the markets and other stuff will do whatever it does, but I think the core technology has decades before the, the full impact is felt. It's, it's going to be a long-term transformative, uh, effect on our industry and one that everybody has to really be thoughtful of at every level. So in that context, I think a lot of this hit the public consciousness with some of the products that you talked about. Um, But the reality is this has been a very long, I mean, some of this technology stretches back into the, you know, seventies and eighties, and it was just waiting for the right time and the right scale. And we needed all of the pieces. We needed GPUs for some of this. We needed internet scale data. We like, there was just all these pieces that needed to be in place. Um, but it wasn't something that came out of the blue in terms of th…

AI assessment note: “it wasn't something that came out of the blue in terms of the core technology.”

Q I'd love to talk about the people, uh, making all of this happen on the generative AI front, um, starting with yourself, maybe a quick word on, on your background and your journey to, to this role, and then how does the Gen AI team, the R&D team, um, you know, who are they, how many, and, uh, how are they organized?

A So everyone, um, you know, we have radically different Backgrounds for people and sort of ways that they, they found their way to Roblox. Um, part of that is that Roblox is really a, a unique business in terms of the space we've carved out. Um, and we've been doing this, uh, since about 2004. I think we might have incorporated in 2005. So, um, We've been doing it for a long time, but it was sort of, you know, we were the, the lone UGC, three D social platform. And so, um, so there aren't people who come to Roblox having worked at other social three D UGC platforms where everybody has to have something adjacent they did. And so we have people who, um, came from academia and many of our researchers, for example, were professors at prestigious universities before they, You know, working on these kinds of problems, working on AI or three D or distributed systems or networks before they came to Roblox. We have people who come from video streaming search, uh, companies, um, from, you know, traditional text or image based social media, people who come from video game companies from film companies. And so we get this, this real mix of people and, um, I feel like I spent, you know, and also we've grown so much, especially in the last five years as a company. I mean, we went from a few hundred to a few thousand, um, very quickly as our user base grew. And in terms of, um, one of the pers…

AI assessment note: “we have radically different Backgrounds for people”