How AI Will Transform Roblox Games into Photorealistic Worlds | CEO David Baszucki

00:00:00What proportion of you think yourchildren will have at least one seriousrelationship with an AI throughout theirlifetime?>> I think zero right now. I don't thinkwe're close to crossing the human AIbarrier. But it is worth thinking thingsare changing every day. We have toincorporate and do that. And thenthere's also some weird universal truthsof things that stick around for 40 yearsand you kind of have to blend themtogether. We feel if we build Robloxright, it might be the kind of thingthat's kicking around in 40 years.

00:00:36>> Hi listeners, welcome back to No Priors.Today Allad and I are here with DaveBazooki, the founder and CEO of Roblox,[music] the 3D immersive world wheremore than 150 million users come everyday to hang out. We talked with Daveabout the future of AI and gaming, theirinvestments in AI, how NPCs are going tochange and really transform gameRoblox's studio, one of the largestcoding platforms out there, and their20-year mission to build the holiday.Welcome, Dave. Dave, thanks so much fordoing this with us.>> It is so great to be here, and thank youfor doing this in Roblox headquarters.>> Well, exciting to be here.>> Yeah, great to have you guys. Thanks.>> So, I think we want to start with a bigpicture question. I think um you allhave been really visionary in terms ofwhere all this is heading and AI isobviously having a big wave and effecton gaming and what's going to happen inthe future there and that's acrossthings like world models how do youthink about NPCs and their evolution uhhow you think about assets within gameshow you think about world creation so wejust love to hear your views of like bigpicture 10 years from now where are wegoing like what's coming>> yeah so two-step on this one I would sayfirst I know you both dabble ininvesting a little>> so there is there is a rope Robloxbusiness plan PowerPoint deck from

00:01:51almost 20 years ago that we pull outsometimes and we're amazed at thefidelity of it. It uh pretends a newcategory. One could call it the categoryof human coexperience. It's got thesizes of a bunch of companies from 20years ago. So you can see socialnetworking companies, you can seeYouTube video, you can see toys, and youcan see all of that. And what this whatit imagined is really a new category.Some have called it the metaverse. Somehave called it the the hollow deck whichis really the ultimate highfidelitysimulation where people can cometogether and do stuff.>> Feels like a Ready Player One the movie.>> Exactly. And so we do have that businessplan slide. And and what we had alwaysimagined is if this hype fidelity spacewas backed by physics simulation andreality simulation, you'd be able to dostuff. You'd be able to build a car, putwheels on it, drive it around. You'd beable to go to a birthday party and blowout the candles. You'd be able to choptrees down and make a a house of it. Sothat that was literally the genesis ofRoblox. And here we are 20 years later.There's so much more to do. But onevision

00:03:06of AI is how do you superpowerthe creation of that hollow deck? Andhow do you get it photo realistic assoon as possible? How do you get 10,000people in it instead of a hundred? Howdo you do acoustic simulation thatsounds realistic with 10,000 people? Soin that sense, AI is super interestingabout just trying to get a vision that'sbeen around for 20 years and hasarguably been in sci-fi once again, snowcrash, holiday forever. Um there isanother interesting vision and and Ithink it's useful comparing theseproduct um things that's much morewhat's the future of just a singleperson by themselves dreaming what theywant to dream and um we sometimes callthat real time dreaming category. It's acategory that we can see a bit oftrappings of with short form video.There's a little bit of that in shortform video. You know someone's 2 a.m. intheir bed doom>> scrolling. It's reacting to dwell time.It's reacting to what you favorite. Andyou're kind of getting a little bit of apre-built dream that I think wesometimes imagine could go all the wayto the famous Tom Cruz movie to VanillaSky where he literally real time dreamed

00:04:22for several years in an imaginaryuniverse where everyone around him wasprimarily an NPC and um he didn't evenknow he was doing it. So that>> so that's more world building as content>> that that is exactly right and so if ifyou put those at two extremes one iscommunication platform highfidelity withpeople you know the other is real timedreaming where everyone's an NPCeverything in between is possible and Ithink we see everything in betweenultimately as being possible and we maysee weird product categories we neverimagined You know, I I think people areprobably thinking, does the text promptturn into a voice prompt and does itautomatically turn into a video promptif there's like I want to learn Frenchvideo? How you're in a French cafe,start talking to an NPC? So, I think Ithink I think there's a wide range ofproducts that start to get into thisspace we're interested in, which is kindof that 3D reality space. Do you thinkthat same space will exist for bothconsumer and business applications?Because people also talk about this inthe context of you're instead of doing azoom why don't you go into a sharedspace

00:05:37with someone and meet with thatyou know what degree of your lifebecomes virtual.>> Yeah, we have seen um I think on thecommunication side people just want moreand more fidelity. We can go back topaintings on the wall of a cave. Peoplewant more fidelity. We have smokesignals. We have the mail system. Wehave the Pony Express. We have thetelegraph system.>> We have the phone system. We have thearguably the COVID initiated video.>> Uh we also did have the COVID initiated3D experience on Roblox actually and andwe found sidebyside video. A lot ofyoung people used Roblox as an earlytest case of how to co-experience. Ithink as the technology for multiplayer4D simulation gets better and morephotorealistic,it's almost going to be like video isthe downsampling. And it's like if if wewere having this um 4D immersivecommunication, we'd say go to legacyanalog mode and make it look like Zoombasically. And and we could always dothat,>> but we'd be able to do things we couldnever do on Zoom. you know, you'd say,"Hey, let's pop up and go walk around myoffice and I I want to show yousomething. Come and follow me." kind ofthing. So, I

00:06:52do think the 4D simulationas it gets photo realistic will be asupererset of video. The the other thingfor business is um there's a few thingsthat 4D and I use 4D because it's notjust 3D shape, but its function in a waywill provide that video does not provideis acoustical things. And there there'ssome real tricky physics problems whenthere you're having a company meeting ofa thousand people. If we try to do it onZoom, we have like a bunch of squares.What's who's live? Who's not live? Howdo we mix the sound? Which is verytricky. If we do it in a athree-dimensional simulation, I can heareveryone with attenuation based on howfar they are. And it can actually be amuch more natural experience. It's likeI I walk closer to you and I I hear you.And that that may actually be a betterhuman interface paradigm. There's onefinal technical thing that's going to bevery very difficult is when people tryto sing happy birthday to you togetherand because no matter whether you're onZoom or whether you're in a 3Dsimulation, you're hearing them like 30milliseconds later. So what I thinkwe're going to find is people areactually

00:08:08allowing each one of us to singhappy birthday with a timeforwardextrapolation of each other and thenremixing it. So I do um to you know forthe business side I do think ultimatelywe're going to see some interestingconferencing solutions.>> Roblox has this amazing wild story withphysics simulation and or physics engineas like part of the the birth of Roblox.But then today or or for the last decadeI have um I think of this as a immersiveworld where people have their firstcoding experiences. They build amazinggames. They interact with each other. Idon't think of that world has been asbeing super focused on realism.>> That's right.>> Um so is that something that's changingfor you guys? Is it enabled now? Likehow how should we think about that? Ithink what you um the way I would thinkabout it is we're pushing really hard onthe physics simulation capability and atthe same time we have this incrediblefree market that we can't control oncontent generation. I I would say we arejust started talking about full acousticsimulation. I just tweeted about it. Soactually you know approximate you knowattenuation and all of this kind ofthings and echoes that's going to starthappening on the physics simulator. It'sbehind the scenes gotten better andbetter and better. It'll keep

00:09:23gettingbetter. But what we have found I I wouldsay is the diversity of developers arebeyond our expectation. And I would saysome of the things that have gone viral,Dress to Impress, for example, Grow aGarden are less physical. Um but theyhave leveraged cloud capabilities umthat we've built side by side. you know,grow a garden was essentially able tobuild an experience where things keepgrowing even when you're not there byleveraging a lot of the cloudpersistence we provided. Uh, dress toimpress was able to do some veryinteresting things with how they runcontests and layer clothing on. So, Ithink we're generally, you know, thespecification for our product could bewe just want to simulate the real worldwith 10,000 people. And the more we goin that direction, the more we'll justsee creators using it in different ways.>> How much attention have you been payingto world models and what's beenhappening in the AI world in terms ofhow people are shifting how they dophysics simulation? Right.>> I think we're really watching it. Ithink that the biggest thing we havefocused on and I think we'll continue tofocus on is how do we get that 10,000player multiplayer type experience

00:10:39andhow do we fall back to 100 or 10 or one?We feel it's super robust to build acommunication platform and you knowthat's part of our goal of getting to10% of gaming. The the technology forthis I think we're going to startunraveling it over the next 5 to 10years is what's the most efficient wayto synchronize the state of 10,000people and and when>> it's a deep infrastructure problemversus necessarily a physics. How do wesynchronize the state and the memoryover the last 5 hours of 10,000 people?Where have they been? Um where have theybeen in 3D space? What have they beendoing? The the world model thing is veryexciting because they're storing memoryliterally video latence. You know, aminute of frames. It's showing superearly promise. I think the technicalrevolution is going to be what's theultimate synchronized state for 10,000people. Is it in video latence? Is it inmuch more native 3D format? Is it ahybrid new discovery of some 3D videolatent space? To be determined. I I willshare that the format we're using, partof the

00:11:54vision for us is to ultimatelystore the history of everything onRoblox and to store it not raster likevideo, but store it vector, which meansthe ability um to play back anythingthat's ever happened on Roblox. That's13 billion hours a month. Um and so theability>> What do you think that does for you? Thehistory of your world. Um, so there's athere's a sci-fi book by Arthur C. Clarkcalled The Light of Other Days, which umtalks about the fantasy of what wouldhappen if we had infinite playback inour world. It's like a real socialproblem. Um, because if everything we'veever done can be played back, um,actually in that sci-fi book, it's notjust everything could be played back,but you could look at any playback.>> And so everything's completelytransparent. there's no privateconversations. It actually delves intothat. I think for us, we would then useit very thoughtfully and privacycompliant and judiciously. But for verysimple things, you can imagine ah therewas a safety incident. Let's go back andput five cameras and listen to the audioand see what happened>> and make a good judgment

00:13:09call. That'skind of interesting. Um there there's areally special thing that maybe I didwith someone in my family that's reallyspecial given I have access to myhistory. Go play that back and possiblybecause it's 3D vector data, not rasterdata, reshoot it from make a cool videoof the best moment. And I think where itultimately ends up for us is there's ahuge push now for people to find kind ofhybrid video ASDW human interactiondata. People are starting to train onvideo with keyboard stuff and like wheredo I find 200 million hours of thisdata? The data we have which is 13billion hours a month is can bereproduced from any camera angle and caninteract with the 3D space. So it's verypowerful data. It leads to a reallyinteresting computer science program umidea around how to create great NPCsthat are more than just an LLM.>> Okay, tell us more about that. Actually,that sounds really interesting.>> Yeah, go check out on XRX feed. I thinkthe last day or two we released a bunchof uh native video demos and maybe wecan even play them back here. But whatwhat we are starting

00:14:24to train is NPCsbased on a combination of all of thedata we have in a privacy compliant waythat are starting to navigate andstarting to play games. The challenge Ithink um would be um challenge levelone, everyone has access to an NPC thatcan get pretty good at playing anyRoblox game.>> Super cool. challenge level two is bywatching if I so opt in in a privacycompliant way, my behaviors, the way Iact, literally my gestures, the way Ilook at things, the way I talk.>> Um would we allow you to have your ownvirtual doppelganger if you wanted?>> That's fascinating. Um and then umnumber three would be is if I if I havea good virtual doppelgangerum just as agentic is all the buzz inother areas. Is there a simple userinterface for agentic virtualdoppelgangers? And that that starts toget interesting because one couldimagine sending their virtual self outto go do something. It could be you knowmy my son or my daughter wants to playwith me. I having to do some work. Couldmy virtual doppelganger fit in for 15

00:15:39minutes? Sure, that's great. You know,that could be really interesting. Um,and there's a lot of other um thingsjust like we're going to see agenticspace and the 2D and the work space andthe productivity area. I think it'sgoing to be interesting to imagine.>> So, you're basically using these 18billion hours a month to train>> 13 billion>> 13 billion excel

00:16:01that will be a Roblox privacy compliantmodel. will never share the data. Andthen over time, no ship date, make thatcustomizable as your own virtualdoppelganger or for creators with theaddition of a prompt, you know, umupload um Benjamin Franklin's history orit's already embedded in an LLM and justsay, "Hey, I want Ben Franklin." Yeah.>> And I want a happy Ben Franklin. Arethere experiences that you imagine umcreators making with these NPCs that arereally different from>> I think there's a lot of um standardgame NPCsum that that I think have done such agood job. I wouldn't claim when thesetypes of NPCs would be good enough, butlike Grand Theft Auto, right? They'vebeen working on this forever. That is avery rich world with a lot of very highperformance you know NPC behaviorimagining someday>> um many people can create them. Um Ithink to make that happen theseplatforms have to be very cloudconnected platforms rather than runninglocal um to spin up inference on all ofthese things. But you could imaginethat. You could imagine

00:17:16a very youngcreator saying, "I'd like to build anAmerican history thing where you go andtalk to Benjamin Franklin." So, I thinkwe would want to enable both of those.>> It's really interesting becausebasically the transition is this isgoing to be a poor analog, but there's atransition that happened inself-driving. where you went frommapping out heristics in subbehaviorsand sort of fine detail to capture edgecases to just doing end to end uh modelsfor deep learning and it sounds likeyou're basically considering doing thesame thing for NPCs.>> There's that that's the excitingcomputer science capability is ratherthan a you know all of these wonderfultechniques people have used you knowdecision trees and this and this gettingto a purely more general modelingsolution would be really fascinating.>> Yeah. Super interesting.>> What do you think the role of like aRoblox game designer or any gamedesigner is five five years out fromnow? Do you think it's different?>> I think I think I'm optimistic thattypically industrial revolution orwhatever revolution I think we'vechatted about stuff like this before.There's a thought that there won't beany jobs after this. I think I'm alittle bit more optimistic

00:18:32that a newclass of things will show up and I thinkthere's a huge opportunity for morecreators. I think,>> you know, I'm not to the point where Ibelieve, you know, what we call AI sloptoday is going to take over. Like Ithink the human touch is got a longrunway on it. So you could imagine ummore diversity or quality of experiencesum supplemented by AI. So but I do thinkum the role will be more leveraged andthe quality coming out of five peoplewill be astounding>> you know but we'll come to expect thatquality so quality will go up I think uhiteration on testing will go up I thinkwhere we see platforms like Roblox goingis just as um people are starting tospin up agents to develop code I thinkpeople hopefully on Roblox platformagainst the Roblox cloud. You'll be ableto go away for 24 hours and your agentsare going to be like tweaking andtesting, spinning up a Robloxexperience, sending 20 NPCs into them onvarious client emulators.

00:19:47Here's an NPCon a phone. There's an NPC here tuningthe game. So hopefully you'll feel morepower in what you're building. I'm alsocurious. There's um a set of researcharound world models that I'm I'm sureyou've seen that is just trying todirectly generate game experiences withlike there's no physics engineunderneath like you don't control themechanisms it's just video. What's yourview on that?>> I think it's like a super excitinginteresting thing to think about. Yeah,>> I think like I said before, I think onething is going to be the higher thefidelity is memory stored in videolatence>> or does there need to be anotherresearch breakthrough where memory isultimately stored in some new latentspace that's 3D or four dimension. Butthe fidelity is amazing. I think that'sreally beautiful and I I think that thethe way we would see the ultimatearchitectures of these platforms isprobably for a while not an all-in-onething. We would probably see some hyperefficient thousand personsynchronization state engine.>> We might also see some of the photorealism not being done server side butbeing done in intermediate spaces orclient side ex as

00:21:02well. And we may seemulti-step pipelines where there's theraw framework, there's 3D upsampling,then there's local 2D upsampling. We'reactually going, oh my gosh, like peopleare putting together a bunch of pieces,you know, dedicated NPC capability,dedicated synchronization capability, 3Dupsampling, 2D upsampling, andpotentially world model stuff. you know,world model. Um, you know, you canimagine that being used even in itsearly state today rather than maybe agameplay state where rather thanmaking a video, you talk about the videowhile you walk around and you kind ofsay, "Put me in a western world. I wantto go left. I want to go right. Add afew horses." And that video actuallygoes to a multiplayer experience. Youone could imagine that. I think what alot of other people are wondering isdoes short form video does someone hitit just like they did with short formvideo into a highly retentive like localdreaming product that could just reactto what you're doing. I think it's stillearly for both of those but it's a bigpossibility. I I would say though we aredoubled down on a a hybrid multi-A techstack that supports multiplayernatively.

00:22:17Yeah, I I think one of theideas that's like internationally veryinteresting has been the idea that youcould generate um like serial microdramas as a form that is not yet thatpopular in the west but seems like agood format for the technology>> micro dramas. I've watched a few I meanit's really interesting you know if youit's fun to see like some entrepreneurstried to do that I won't mention wholike five or 10 years ago then theynaturally>> emerged kind of thing. It's interestingto think if those micro dramas areeasier to produce. They don't need liveactors, but you know what is theplotline that really resonates. Youknow, we'll we'll see with that.>> And you can mass test that at scale ifyou're using AI, right? You can actuallyyou know a very large number ofiterations. This kind of one of thearguments people make around the firmyparadox of like why haven't we seenalien life? And the idea is it's justthere's alien life. are just stuck inthese virtual worlds now where you knowit's fun the real world>> like what's the famous quote where iseveryone>> yeah exactly>> so one option is they're in some virtuallandscape that's more compelling thanthe real world and so therefore there'sno reason to travel the universe>> or it could be you know we we couldreally startconjecturing what's the future evolutionis it human or

00:23:32machine life and all ofthat>> and once it gets to maybe human machinelife what happens Yeah. The subsumption.Yeah.>> Yeah. That's a>> interesting.>> Yeah.>> That's the vanilla sky thing.>> Exactly.>> Or or the the optimistic thing is morethe holiday thing like everyone'shanging out with other people ratherthan a simulation of themselves.>> Yeah. Are there experiences that youimagine from a like a communication or asocial perspective that um you think aregoing to be uh just different or new orthat you're going to have real insightsinto from these like really great NPCs,right? Because I'm sure you've seen whathas happened with the rise of AIcompanions and how engaging they aregiven they're just text interfaces.There's some fun things that likethere's a Black Mirror dating episode. Idon't know if you've seen it>> where they have they have a lot ofvirtual doppelganger power. Andobviously I'm not So Roblox is not doingdating. We're just talking about thisright now, you know, blah blah blah 18plus ID verified, all of that. So let'snot go crazy, but let's think about afuture um in the Black Mirror episodewhere you literally can spawn the lifeof a thousand virtual doppelgangers.um that thousand virtual doppelgangers,

00:24:48you're on your new dating app in threeminutes. Um you know, someone else'svirtual doppelgangers live a thousandvirtual lives with your virtualdoppelganger. And then they say, "What'sthe success rate?" And when the like thethousand virtual doppelgangers say 98%of the time we had a good life, then yousay, "Oh, we should go on a date orsomething." So there's um>> social simulation.There's a lot of things we I think aregoing to be really hard to project. Ithink one of the things>> we would focus on is the more we canjust build raw high performanceinfrastructure, have AI as a service,have all of these things capable in acloud, have the same thing run on aphone as a computer, have it autotranslate. Hopefully, we would discoversome of these things. Yeah, I'm I'mreally optimistic about that given theengagement that people see with thesecompanions as just a text interface withlike the beginning of memory, right?Versus if you're Roblox, you haveembodiment, you have, as you said, cloudpersistence. And I think it's just goingto make the characters much moreinteresting very quickly.>> Thinking through how we would embodymemory for NPCs, which is more thanjust, you know, a prompt. Um, one could

00:26:03imagine any NPC on Roblox because wehave recording of the whole thing, theyactually have access to everythingthey've ever said or done in a very leanformat. So, they can go back throughthat, retrain on that. Um, so thathopefully would be a native part of theplatform. But I do know, yeah, there's alot of interesting engagement with thesetypes of companions and imagining fully3D versions will be like prettyfascinating. Yeah, a friend of minebasically asks um at dinner parties, hehe runs a well-known AI company and hissort of controversial question for thegroup is often, "What proportion of youthink your children will have at leastone serious relationship with an AI likethroughout their lifetime?">> I think zero right now.>> I do I don't I don't think we're closeto crossing the human AI barrier. Um butit is worth thinking. He meant it evenin the context of like even if it'ssuperficial and there isn't trueintelligence behind it just given thebehavior with text once you start havingvideo virtual worlds>> I can imagine a real time coachtherapist like in your earbud justhanging out all the time.>> Yeah. So I I think for that kind of lessevasive more over your shoulder coachingI think that seems like>> that's already a major use case it seemsfor uh some of the like chat GPT

00:27:18andother platforms where uh mental healthadvice or coaching or other things likethat are actually one of the major usecases.>> Yeah, I think that will be arguably veryvaluable.>> Can I ask like an industry perspectivequestion just given you've worked ingaming for a very long time now. Um, whydo you think uh like there's a lot ofdiscussion about how assets of differenttypes are cheaper to create withgenerative tools? Um, I don't thinkthat's really change the dynamics of thegaming industry yet. I don't know if youwould agree with that claim or explainit.>> One could say no matter how cheap it isto create assets, the expectation ofquality from consumers goes up atexactly the same velocity.>> Okay?>> And so, you know, what was not possiblea while ago the expectation of consumerswill go up.>> I think there's a there's a generalnotion that the technology for thegaming industryhas to move tocloud vertically integrated and assetmanagement can't just ship on a DVD orbe a giant download. I think we're we'rereally seeing we've launched um dynamicLOD for textures, meshes is on the way.

00:28:33So you could imagine every single assetin the game um sound uh 3D object imagecan exist in various formats.Traditional format is it's like atexture or a mesh or an audio file. Mhm.>> Future format could be it's an AI promptand it's generated procedurally ondemand. So there's these two things.They're both very AI enabled. One canthen imagine that you have to be cloudconnected because for all of theseassets you know full range of LOD justlike with video uh streaming upsamplingon low-end devices you need low LODhigh-end devices you need 4K LOD andultimately um on demand AI generationthat then gets into um once again one ofthe demos is probably on 3D upsamplingum the notion that if someone makes avery primitive experience on Roblox,they would augment it with a prompt andthey'd just say, "I know it's kind ofprimitive, but make it kind of lookmedieval and more realistic and have allof those assets auto upsample to in 3Din the cloud for free on demand." Thatthat kind of just in time thing is

00:29:48someof the stuff we're thinking about how AIcan accelerate if you have a cloudplatform. Roblox is also from the studioperspective one of the biggestdevelopment environments in the world.>> That's right.>> Uh what are you seeing from theperspective of AI coding there?>> Yeah. So this is great. So once againindustry standard ways of using adifferent model with studio. So ifyou're in cloud code you can controlstudio is super exciting. And so a lotof our users have started gluing thesethings together wi with the hopeultimately you know they can build stufffrom there. Roblox Studio itself isrunning a native assistant as well and anative code engine. And I think the thepattern for that is Roblox Studio willmore and more for code follow many ofthose industry best practices but inparallel needs to have kind of anenvironmental generationkind of component that's very unique tothat and I think I think the future ofenvironmental generation is once againany prompt you can imagine text imagevideo possibly walking around in a worldmodel to define a world having iterateon that, iterate on that to a primitive3D skeleton.

00:31:04Do I like it? And theniterating on that to a fully functionalgame. So as that I think that's ourvision is to run those side by side sofirst time studio user gets it out ofthe box at the same time we like pluginto all of your um AI workflow. I thinkthe final thing is because it's so cloudconnected that thing we talked aboutearlier spinning up jobs spinning upagents running test plans kind of keepgrinding away trying something umdifferent is like you have all of thatcloud capability.>> Just out of curiosity what is the likecycle of development for a successfulRoblox game today? the um what we haveseen>> is over the years bigger teams so thatyou know the topRoblox creators now are well into thetens and tens and 30s plus million ayear like it's really serious stuff wehave seen in the Roblox creatorcommunity I think a healthy sign thatcreator number 1 and what is theiraverage yearly revenue is going upfaster than creator number one so that'sa bit of a healthy long tale. And Iwould say all the way out to creatorthousand, there's such a huge community

00:32:19of people making a living. What I thinkour creators have found is it's gottenboth more healthy but more competitive.And I think we've seen the pattern nowof much more live ops in that becausethis is cloud, you can update yourexperience just like good websites everyday or every week. And that that is justkind of a constant practice. So I sayupdate frequency much more rapid>> which is great for users because theyget uh updated content at a faster rate.>> That's right. Users love updates. Theylove new things. It keeps games fresh.Um so I think the um the other thing wewe have seen we've um on the discoveryside we've tried to and more and moretaken the approach we want discovery tobe completely transparent.um like these are all the things thatare going into the discovery algorithm.We we check it. There's a benefit tothat transparency and discovery and thatit actually keeps a lot of pressure onus. You know, oh my gosh, someone'sgoing to gain discovery.>> Uh well, then we better make it reallygood, right? And and transparent. Iactually think transparent discovery andrecommendations

00:33:34is an interesting trendfor the industry worth talking about. Umbut so so with that we have seen alsothe kind of the content distribution getricher and a little bit less spikythere. There's now like the top 20experiences on Roblox are all reallygood and all kind of vying for thenumber one slot which maybe wasn't sotrue three or four years ago.>> Zooming all the way out for a second.Like I I I think it was true five or 10years ago when we met. Um it was truewhen I was talking to one of the membersof your leadership team today. They willsay you're a extremely high conviction,very long-term oriented leader, right?So they, you know, 20-year plan forRoblox and we're going to do things theway we have a vision toward. You'reobviously also all in on AI, right? Veryexcited about the technologies andinvesting in them. How do you likesquare these things where you have avery long-term view when so much ishappening in the environment?>> I think we um we always want to pairtake the long view, which is one of ourfour big values, with get stuff done.And so, um, the it's almost like inthose old Gartner charts, you know, takethe long view, not take the long view.Yeah. Get stuff done, not take the likethe magic quadrant is take the longview,

00:34:49get stuff done, we we typicallymean inside the company, rapiditeration. And so if someone has a viewof some exciting new product and we'regoing to ship it in 6 months, that canbe a very um scary thing unless it canbe broken down like we're going to shipthis thing every week and iteratetowards this. But we have a good targetof where we're going. And so I thinkinside the company we I'd say our AIteam, our facial age estimation team,our safety team, these are literallythings working on a weekly basis whichallows a fairly fast reaction time. AI Ithink is arguably as far as speed rightnow just every day. But you mean youmake it sound pretty simple in that likeall the things that you're trying to getdone on a week or a six month timescale, they're all aligned toward avision that like>> if you have your 6 to 12 month vision,>> yeah,>> hopefully that's moving not too fast.Um, but at the same time, the weeklyiterations are moving very quickly inthat direction and and you know that youcould say the Roblox vision isultimately to build the holiday and

00:36:04it's10,000 people. It's multi-playerrealtime modification of theenvironment. It's NPCs>> photorealistic. That's a pretty stablespec for 20 years. And so then>> I actually feel like that's the biggesttakeaway for me here. like I spent a lotof time with um software CEOs, right?And I think there's a lot of actualconcern about like is the vision we hadfor the company in terms of itsdurability or its value>> uh still as valuable as we thought 5years ago. And for Roblox perhapsbecause you were, you know, pointing atsomething so far away to begin with,it's like pretty clear, you know, Ithink AI only moves you.>> We have crisped that up over the years.I would say two years ago we were alittle less crisp. We would say we'regoing for a billion DAUs. That's a goodspec. With this spec I just mentioned,it's a little harder to operationalizearound product management teams andthat. So we we took a step a steppingstone and said we want to get to 10% ofglobal gaming content um on the way tothat. It's actually a remarkably goodstep. It's about 300 million DAUs. It'sabout 20 billion. It's about 3x where weare. It can be forensically torn apart.We can see

00:37:19what part is the USA market.We actually have a higher target for usthan 10%.>> Um, and so that has been veryclarifying. I think for a software CEO,always knowing both 3x and 10x is veryhelpful because if you don't know 10x,then it's hard to sleep at night. If youdon't know 3x, it's hard to planforensicallyhow we're going to operationalize that,how quickly we're going to get there.>> One other thing that I think that umRoblox has done uniquely well is uh howyou've thought about hiring and how youyou've rethought certain aspects ofthat. Yeah.>> And I think that's really led to peoplehave more bias to action who are movingquickly, who are iterating rapidly. Umcan you talk a little about how you'verethought some aspects of that if you'rewilling to share that as part of>> Yeah. I mean, one is we've been verypublic that we we really want valuesaligned people that are good problemsolvers that are very creativeand it's hard to do signal to noise andthat kind of a hiring thing. So what weactually bought a company about fiveyears ago that was doing their bestscientifically to build 3D

00:38:35assessmenttooling for some big companies to reallyfind people who are problem solvers andcreative.>> That's part of what I was referring to.That was fascinating that you did thatacquisition.>> So we we we bought Embellis. We got agreat team of scientists, people thathad been involved with the SAT test andall of that, and we've tuned that to ournew college grad intern program. So farwe've essentially I believe in a veryfair meritocracy kind of way taken thewhatever it is 10 50 60,000 people runthem all through what we've I thinkshown to be somewhat of a very fairnon-weighted by any other social factorassessment and the people that come outof there we have an incredibleconfidence that they're problem solversum they do a bunch of other thingsreally well so We we kind of have thisbuilding wave inside the company ofthree years ago, two years ago, one yearago of these 100, 200, 300, 400 peoplethat hopefully were like mentoring someof the best future leaders in theindustry.>> Yeah, it's super interesting because Ihave a number of friends who always talkabout how you discover untapped sourcesof talent or how do you find the bestpeople in the world and are there waysto test for that

00:39:50or sort of screen forthat?>> We we've done a lot of testing on this,fairness testing, um looking backwardson it. I'm I'm very confident in whereit's going to go for us.>> Super interesting.>> Yeah, it's it's like a verysystemsoriented approach. The one thingI I we have found which is a little edgymight be the correlation betweentraditional universities and our owntesting and and we we hire>> Sorry, what's edgy about that?>> Well, I think people typically thinklike I went to Stanford, that's whereall the top talent is. We have foundthat not necessarily to be true and andwe have found community college, youknow, the small Midwestern engineeringschool, like because we're assessing ourown way, we we kind of we basicallyignore the signal of where you went touniversity.>> I've heard that anecdotally from otherCEOs now, particularly for certaincohorts, like the cohort that cameduring COVID. Yeah.>> For a lot of the top universities isperceived as like a weaker cohort. Um,and I don't know the reason or theinterpretation behind it, but it'sinteresting to see the shift um, interms of how people even think aboutwhere is some of the best talent. So,it's interesting that you're kind of>> saying, so we get we get to be haveself-determination. We get to saycolleges, you do what you want, all ofthat. We're just going to

00:41:05do it in afair way that we think is our way andit's worked out. Can you give just uhbecause it's so interesting like anyintuition for something you some way youtest for problem solving capability?>> I think we make public some of the teststhat we use and they the cool thing isthey're all built on roadblocks. They'reall tested by lots of people and they'rearguably really interestingthree-dimensional problems likeprogramming a factory or like you'reusing some new geometric language toprogram a robot, you know, it's almostlike turtle graphics and you can playwith that um and see how>> we're gonna try to get jobs.I would be I'm I'm actually might be alittle afraid to take the test myselfgiven how good some of our candidatesare.>> You know, as a long-term predictor ofthe future, is there something happeningin AI that you believe or hold withconviction like five years out? Do youthink not everybody believes yet?>> That I believe that not I I thinktypically it's the other way. Morepeople believe more stuff than Ibelieve.>> You think of yourself as a skeptic? Uhwell, I think of

00:42:20myself as um>> he's building the holiday, but he's askeptic. I>> I I think I'm more like a the power oftime and the power of compounding ratherthan like an amazing invention and justrapid iteration. One one thing Iconstantly am amazed by is either thegraphical user interface or MicrosoftExcel. I think I was playing withMicrosoft Excel in like the 80s>> and here we are 40 years later it's notthat different.>> Yeah.>> So very useful.>> So it it's kind of like this interestingcompliment in AI world. Things arechanging every day. We have toincorporate and do that. And thenthere's also some weird universal truthsof things that stick around for 40 yearsand you kind of have to blend themtogether. So I think that's how we I Ithink the good news is we feel if webuild roadblocks right it might be thekind of thing that's kicking around in40 years.>> Awesome. Thanks so much Dave.>> Thank you for joining us.>> Thank you. Okay cool.>> That was really good questions.>> Find us on Twitter at No Priors Pod.Subscribe to our YouTube channel if youwant to see our faces. Follow the showon Apple Podcasts, Spotify or whereveryou listen. [music] That way

00:43:35you get anew episode every week. And sign up foremails or find transcripts for everyepisode at no-briers.com.