Transcription
You know, in a video model, you might predict the next likely sequence or the next most entertaining frame. What world models do is they actually have to understand the full range of possibilities and outcomes from the current state and based on the action that you take, generates the next state, right? So the next frame. And so it is a much more complex problem than traditional video models. So to me, it is a world that is accurately generated based on the actions that you take as a result of what's already been generated.
Hi listeners, as you may know, I recently wrapped up the AIE code conference in New York. And while I'm traveling, I do like to visit top AI startups in person to bring you interviews that you don't find on any other podcast that just does a Zoom call.
General Intuition, or GI for short, is a spinout of a 10-year-old game clipping company called Metal, which has 12 million users, but in comparison, Twitch only has 7 million monthly active streamers. Metal collects this data by building the best retroactive clipping software in the world. In other words, you don't need to be consciously recording. You actually just have Metal on in the background while you're playing, and you hit a button to clip the last 30 seconds after something interesting happens. It's very similar to how Tesla and self-driving does bug reporting. If you've ever done a self-driving bug report in Teslas, the result is that Metal has accumulated 3.8 billion clips of the best moments and actions in games, resulting in one of the most unique and diverse datasets of peak human behavior, actively mining for the interesting moments. They were also very prescient in navigating privacy and data collection concerns by mapping actions to these visual inputs and game outcomes.
As you saw on our FE Lee and Justin Johnson episode with World Labs and with the recent departure of Yan Lun from Meta, there's a lot of interest in world models as the next frontier after LLMs to improve on spatial intelligence and to work on embodied robotics use cases. Deepmind has been working on this with Genie 1, 2, and 3, and SEMA 1 and 2. And this year, OpenAI seemed to finally agree because they've been leaning on LLMs a lot, and they made the news by offering $500 million for Metal's video game clip data.
Our guest today, Py, turned down that money and instead chose to build an independent world model lab. Instead, K.A. Ventures led the $134 million seed round, which is Venode Kosa's largest single seed bet since OpenAI. We're able to get an exclusive preview of GI's models, which unfortunately we cannot show you directly, but I can confirm they were incredibly humanlike, and we chose to include the first 11 minutes of the demo discussion. Even though I couldn't show it to you, it may be hard to follow, but I tried to call out what was noteworthy for you to know as your likely reaction if you were watching along with us.
Now, enjoy the world's first look at my first look at General Intuition.
So, what I'm about to show you is a completely vision-based agent that's just seeing pixels and predicting actions the exact same way a human would. Um, and so, yeah, what I'll show you here is what this looks like four months ago. So this was, uh, so again, this is just an agent that's seeing, that's receiving frames, um, and it's just predicting action. So you can see it has like a decent sense of, um, uh, of being able to, you know, navigate, um, it tabs scoreboard, just like gamers always tap the scoreboard. So these are purely, these are pure imitation learning.
>> I see. So I see slicing the knife.
>> Yeah, exactly. So it's doing everything that like humans would. In this case, here's, here was the first interesting part that we saw, like it gets stuck, and then it has, memory as well. So you see it can get unstuck. Um,
>> How long is the memory?
>> Uh, four seconds. Yeah. Uh, four seconds. Okay. So, this was four months ago. This was maybe a few weeks after that. So, you can, you can see there's like, it's still doing the scoreboard thing, but there's still, uh, quite like [laughter] And these are bots, too. So, you can see that
>> It's very human. Let's just say that.
>> Yeah. Uh, and then, um, right. So, this was really like the early days of research. You can see, right, it does one thing and then goes for another. Um, and then we've been scaling, right, um, uh, on data and compute, and also we've just been making the models better. Um, and this is where we are now. So what you're seeing is, um, pure, like I said, pure imitation learning. This is just a base model. There's no RL, no fine-tuning. Uh, this model sees no game states. It is purely capable, not sequence, etc.
>> It's purely predicting the actions, um, from the frames. That's it. Um, and it, this is playing against real humans, uh, just like, um, uh, like a human would play, and it's also, it's running completely in real time. So there's absolutely everything here plays exactly like a human.
>> Do you give it a goal?
>> Yep.
>> It just figures out its goal because obviously it's trained on a scene.
>> Yes.
>> Um, and I, I, I picked, right, I picked a sequence where also it doesn't do well initially. So you can see like this is, this is just like a, a sequence, a random sequence.
>> But this is the, I mean, it looks like it's doing well.
>> So, um,
>> Oh. Okay. Yeah. Watch.
>> Yeah, that's pretty good. Maybe too good. Um, this is what my favorite part. So you can see it does something that like, um, here, like a human would never do this. Then gets unstuck, then has four, realizes which, and then in the distance.
>> So you're saying one, it makes a mistake that a human will never make, but it unsticks itself.
>> And two, what we just saw is it is doing superhuman things.
>> Yeah.
>> Okay. Yeah. Um, I mean, there are things that that that humans do, obviously. Um, but because it is trained on, on the highlights of things, all the exceptional things, it's inheriting those.
>> Yeah. So it's not like move 37 where we RL'd our way into something.
>> Yeah. Replicating superhuman or like
>> Superhuman the baseline of our dataset is peak human performance.
>> Yes. Yeah.
>> Um, okay. So that, that's the agent. Uh, so now what I'm going to show you is we then are able to [snorts] take those action predictions and we're able to label any video on the internet using those actions. So, um, um, and so this is, this is just frames in, actions out. Yellow is the, uh, model prediction, or sorry, yellow is ground truth, purple is the model prediction, and then bottom left is compound error over the entire sequence, and then this is reset per prediction.
>> Reset meaning you, every now and then you reset.
>> Yeah. So this just means it resets the baseline. Um, and so this basically, a single error in the entire sequence compounds here, but it doesn't compound here, if that makes sense.
>> Yeah. Um, so, and again, this is just seeing frames, right? It's not, it's not seeing any of the, any of the actions. Um, and so, you know, so what we did, right, is we, we, we trained it on less realistic games and we transferred it over to a more realistic game. And then, and this is where it gets really exciting, we transferred it over to a real-world video, which means that you can use any video on the internet as pre-training.
>> What is it predicting?
>> Um, it's predicting it as if you were controlling it using keyboard and mouse. So if you were, if you're basically playing the sequence as a human is.
>> Is there some sense of error or
>> Uh, so that's why you, you transfer to more realistic games first.
>> Yeah.
>> And then you transfer to real-world video because you can't get a sense from ground truth from, from the real-world video yet.
>> Um, let's see. And then, um, so what they don't, so I'll show you here. Uh, this one is also, um, this is the same, uh, uh, agent that I just showed you. This is playing against other AIs.
>> This one's playing against bots.
>> Yeah. Um, the previous one was against players. Uh, but with the sniper, it doesn't really matter that much, as you'll see. [laughter] It's like, uh, so one, one thing that's really interesting is you notice that it behaves differently as it has like different items, right?
>> That makes sense. Yeah. Intuitively.
>> Yeah.
>> I think there's also a question about egocentricity versus the third person. Does it matter?
>> Um, the third person, I think, will be very, very helpful if you're, for instance, trying to control multiple objects in an environment later on.
>> Uh, right now, I think having fully, imperception, first person is quite helpful.
>> Um, this one's also, this is the policy itself.
>> What do you mean, this is the policy?
>> The agent.
>> Yeah. Same constraints that I just told you about.
>> Yeah.
[music]
>> Like this, where, right, the, where it hides, that to me was just incredible, like just from, from knowing, being able to to
>> Also hide when you see it.
>> Exactly. Yeah. Yeah.
>> Um,
>> And it needs the spatial intuition to go, well, this is hiding, and that's not hiding.
>> Exactly. And, and, and right, while it was reloading. Yeah. Um, okay. So that, so those are that, that's the policy, and this is a completely general recipe, meaning we can scale this to any environment. Uh,
>> Is this work closest?
>> Okay. No, let's keep going on demos until
>> Um, I was going to go into research.
>> Yeah. Yeah. Um, okay. So, and then this is, this is these. So what I'm about to show you are the world models. Um, there's a few really, really interesting parts about our world models. So the first is, uh, we actually made the decision to, uh, transfer, um, sorry, we made the decision to, um, pre-train world models from scratch, but also we've actually been able to fine-tune open-source video models to get a better sense of physical transfer. Um, and so one of the things that you'll notice here is like our world models have mouse sensitivity, which is something that like gamers absolutely want, right? So you can have these like very rapid movements, which you couldn't do in any other world model. Um, and so this is a hold-out set. So this clip was never seen before at training time. Um, and so you can see it has, it has spatial memory. This is, this is about a 20-secondish generation. And here's what's fascinating. This is an explosion that occurs. And you can see that in the, um, in the physical world, right, the camera would shake, and in the game, that would never happen. So you see, you see the world model inherits the, the physical world camera shake, but the, the actual, um, uh, game never does that. Uh, which, which is, which is sort of that, that to us was quite fascinating, right? Also, did the models that I just showed you that we used to transfer over from video, the two of those combined will allow us to like push way beyond games in terms of training.
>> This is another interesting. So, this is the world model. This is rapid camera motion. So, like again, this is stuff that we're literally just taking one second from here in the context and the actions and replaying it here, right? Um, and so you, you'll never essentially have, um, uh, like what we're saying is the skill that you see in the clips, that like the speed and the movement, that also pays off at training time when you're doing world models.
>> This is my favorite example. Uh, so this shows that the world model is capable, um, of performing, uh, with partial observability. So what you're going to see is, um, again, you're replaying the actions from here in here, just using one second of video context. Everything after that is completely generated. Um, so what you're going to see is the model is going to encounter, in this case, smoke. Normally now models break down. What you actually see is comes out of the same place. Um, and so it's capable of, um, of even with partial observability, still maintaining, um, its position in the world. Um, and then here it is also interesting. So this is, uh, this is sniping. So this is, gives this gives you like a, um,
>> Reaction time.
>> Uh, like the fact that it can do depth and like sequences in completely different views, right? So this is a completely different view than if you were to be outside of that view, right? And so it's, it's able to maintain consistency, um,
>> While zooming in.
>> Yeah, exactly. Um, uh, and so, um, yeah, so you can see, uh, so even while this goes out of scope, right? Watch. And then it can, and it comes back, and you'll see it's still, still there.
>> Yeah. Um, and so, uh, yeah, this is the work that that Anthony, who has been working on.
>> I'm just wondering how much game footage you have to watch in order to find these things.
>> [laughter] We can ask Anthony. You know, it's, it's, I'm, I'm sure he's not going to be too excited to play these games, uh, [laughter] afterwards.
>> Um,
>> You're not playing. You're just watching.
>> Yeah. Yeah. Yeah. Um, great. Okay. So, those were the models. Um, see, these are interesting. So, we also were able to distill into like really, really tiny models. Um, so this is, for instance, a, um, a long sequence on a very, very tiny one. You can see it makes like a bit more stupid mistakes. Uh, like it, it does things that are not as optimal. Um, but
>> I haven't seen anything yet.
>> Uh, at the beginning, it was running into a wall for free. Yeah. Exactly.
>> Um, uh, I mean, I do that too.
>> Yeah. [laughter] Yeah.
>> Um,
>> It's looked, I mean, it's doing pretty well.
>> Yeah. And, and again, all these models are running completely in real time there. So there's no, uh,
>> Okay. So I was thinking your main model does real time anyway. What's the goal of distilling? Is it cost or
>> Uh, yeah. Parameters?
>> Yeah.
>> Yeah.
>> This is the interesting one. Peaks the corner. That's what we mean by like the fish and the poor reasoning aspect is humans actually, they sort of simulate the optical dynamics of their eyes and how they actually spat. Right? You've seen all this.
>> Yep. Um, exactly. And so, uh, like even in like real, this is kind of interesting, like even in like the real world, um, with, uh, for instance, YouTube data, right, you have to first solve for pose estimation. Then once you have pose estimation, maybe you do something like inverse dynamics, right? Where you basically are able to like somehow label some of the actions that you're seeing. And then you still have to account for optical dynamics of like where your eyes are actually looking before the decision, cuz like there's just three levels of information loss where when you're playing video games, you're actually simulating the optical dynamics with your hand, right? And I think that like that's, I think why games are a better representation of sport reasoning initially than, um, uh, than YouTube videos, for instance.
[music]
Okay, we're in the GI offices with the CEO. Welcome.
>> Thank you.
>> Thanks for having us in your office.
>> Yeah, excited to be here.
>> If I'm in New York and you're one of the hardest raises of the year, I have to come and visit and, uh, thanks for taking some time on the weekends.
>> Yeah.
>> Yeah.
>> So, you've raised 133 million seed, uh, for General Intuition. Most people didn't care about you, I, I guess, cuz GI is new, but, uh, more gamers would have a middle.
>> Mhm. And before that, you ran probably Statemer, like the largest depth stage. Um, what's your reflection on just that then journey of like, now you're an AI founder?
>> Yeah, you started off root stage.
>> Yeah. I think, um, I grew up with Tourette's. Uh, I spent most of my time as a teenager coding and playing video games. Uh, so in that sense, it doesn't feel that much different. Um, but I think for, uh, so yeah, so I started the largest private server RuneScape, worked at Dr. Suborders for three years for Ebola, and then on like satellite, satellite-based map generation for disaster response, um, which was already like very AI-related adjacent. I built some models back then and then started Metal, which became one of the largest social networks and video games. I've always been kind of like AI-like adjacent, you know? I, I'm a self-taught engineer. Uh, so for me, the modeling itself always felt a little foreign. I actually had to, uh, take a ton of tons of classes over the summer and early this year to get better at it. Uh, because I, I, it still felt like, like I was really, really good at at the infrastructure side, and I had written like our, our transcoders for Metal myself. So I was very, very familiar with CUDA and like the GPU side and all the video infrastructure that we were using, uh, for this stuff. But the modeling side itself was, was still quite foreign. Um, luckily, obviously, we have, I have really, really good co-founders, but they, they essentially put a bunch of coursework together for me to, to go complete to get really, really good at understanding the fundamentals better. I think for me, I had seen inside of the labs that had really good, uh, leadership with fundamentals at top, and also the ones that didn't, and I think the ones that did were just like much better. Um, and so for me, um, yeah, I wanted to be more like that. So in that sense, it was a bit, it was first very foreign, and then now I feel pretty comfortable with everything. And but yeah, like I think for, um, there's a lot to be explored starting in video games and also reverse engineering. Like I think the interesting thing about reverse engineering is it kind of teaches you to look at problems very differently. It's like the ultimate form of deductive reasoning in a way. Uh, and so, um, uh, so this is just how I think, how I operate, and so for me, it's, it's been a really, really interesting journey. Uh, you know, I don't claim to have any of the credentials or or skills that some of the other guests have had on, but hopefully it will make for a good time.
>> Yeah. Well, your co-founders, uh, definitely bring a lot of that different ability, and you bring a lot of the, I guess, gaming expertise with cheese. We'll see what I bring to the table.
>> Yeah. Just, just a little bit of history of Metal. Let's establish Metal for those who don't know. Uh, ability, clicks, yeah, the year, um, that's you have more active users, concurrent users than Twitch, something like that.
>> Yeah, on the creator side, I think. And the reason is because Metal is a lot more like Instagram than it is like Twitch. So people, um, so the way you think about Metal is it's, it's a native video recorder. Like unlike something like Twitch, where you actually have to use other software to record and stream to Twitch. Um, it's not a streaming software, it's actually a video recording software. And a lot of gamers love to put things like overlays on top of their. Um, and as a result of that, we have sort of the largest dataset of ground truth action-labeled video footage on the internet by maybe one or two orders of magnitude.
>> Yeah.
>> What, what's an example of an overlay? Like, Naomi overlay? I usually think of as CAD.
>> Yeah. Yeah. Also, um, controller overlays, for instance, if you're playing, um, like let's say you're playing, uh, console.
>> Yeah. Like flight simulator. You get like, you know, the joystick and all, all the things. So you get the actual actions that people take inside the games, as well as the frames of the games themselves, which is a loop, right? Because it's essentially, you perceive, then you act, and there's a state update, and then you perceive again, you act, state update, which is like roughly precisely what you use in order to trace to train these agents.
>> Yeah, it's, it's almost perfect training data. We were showing you, showing me in the demo when we show some B-roll here on, uh, how you don't log key, it's very important for you to log action. Yeah. When did you figure this out?
>> Oh, um, maybe starting a year and a half ago. Yeah. And, and we realized that like figuring out this side of the research for us was, we very much never wanted to be in a position where we eroded privacy or something like that. So we never wanted to actually log like a W or A or S and a D, which for researchers, the fact that we don't do that, like often sounds strange, like why wouldn't you do that? But I think for us, the privacy
>> We get the data.
>> Yeah. I, I think, you know, a lot, a lot of the, the, um, the researchers, they did, they hadn't quite understood yet that you can actually just get away with just doing the actions. Um, and the reason is, like, at training time, having the actual keys is noise anyways. Like, if there is text in the screen and you would want to, in theory, make that, um, part of the training, then like reading text from a frame is like really easy. And so for us, if we actually can, so we convert, basically, hit, you hit the, uh, the input, we convert it to the actual action. So we had thousands of humans label every single action you can take in every single video game over the past year and a half. Uh, which is an enormous amount of action labels. Um, yeah. So when you act, you, we, we get the actual, um, action itself, and then it being at training time, you, you can, for like, a general set of that, of that game, convert back into computer inputs if you want to, but you can never do it for any individual person. And so that, for us, from, from like a design perspective, was was important. So we, we figured all that stuff out. Then we actually started pushing, um, like we already had features as well with this. So for instance, like gamers already love to be able to navigate their clips by like things that happened. So we have an events capture system, and then we also have the overlays where you actually just want to overlay and render the actions on top of your clip. We developed kind of in tandem with the feature set itself. And then obviously, when, when world models became a thing, and it's very, [clears throat] very clear that all the all the data for this was precisely like that sequence, yet we were able to sort of be first to market, recruit the best researchers, and start a lab.
>> Yeah, that's, uh, that's incredible. Uh, one more question on Metal before we move forward. It's been 10 years.
>> Yeah.
>> What is the, I don't even know how you roll something like this. I'm just kind of curious and like the opportunity to ask you what really worked.
>> Yeah. Uh, that you became so, so huge because I'm, you're not the only one. Yeah. But, uh, I'm sure it's performance and everything, but
>> A few things had really worked. I think the first was a lot of our competitors were focused on solving the social network and a recorder at the same time. And that never, like our bet was really that we could get so many people to record with us that we could bootstrap the network on top of that, and that worked so well. Everyone was sort of distracted trying to bootstrap a social network. We were just focused on building a really, really good capture tool, and then we got tens of millions of people to use that, which then we were able to bootstrap a network on top of the share behaviors. We already had like the profile behaviors and the share behaviors, obviously, but the actual content consumption piece and and the sharing piece really only came after we hit critical mass. It was actually early days during COVID when like the network really accelerated. Fortnite happened, which was really important, and I think also the fact that Discord existed, um, uh, made it quite a different time than, uh, when other types of networks of these types had launched because Discord essentially was like the connective tissue already between gamers that like never really existed before. And so I think those combination of things really, really made it. I think we also built a product that, for instance, with with most video recorders, you have to remember to start and stop the recorder. So, you have to go into the application, then hit start, then start your game, and then, um, you know, maybe you'll play games for three hours, then you'll close the game, then you have to close your video application.
>> Then you, well, then you have to process like a multi-gigabyte file. Uh, then you have to upload those somewhere. And so like this was a pain for people. And so what we did is we just ran this kind of recorder. When you hit that button, it does a retroactive video record. So all the recording initially is in memory. And then when you hit that button, it exports only that sequence to disk and syncs it to your phone. And so that that became super popular. It also was interesting about it also means that you're not sort of behaving or acting differently because it's always there and you can just export whatever happens, which is also very, very helpful for for training, obviously. Um, the thing you were, the first to do that.
>> Yeah.
>> The thing you were explaining just before this was is similar to how Tesla does their bug reports, right? You're driving from the disengage autopilot, you're like, they're like, "Well, tell us what happened."
>> Exactly. Exactly. See, see, you're driving. Tesla doesn't want to train on the like 10 hours of you driving through a desert where nothing interesting happens. You have the clip button on the steering wheel. Something interesting happens, either while FSD is engaged, and I'm not sure if you can use it without FSD as well, but you hit the clip button, it basically uses that precise sequence to mark, which is then more helpful for training because it's more unique as a training time.
>> Yeah. Yeah. I mean, so one thing, we're going to get to this on the agent side. One thing I, I, that does pop up is, well, a lot of life is boring. A lot of life is going for me. A lot of, a lot of playing games is doing the boring stuff that is not capable.
>> Yeah. Somehow using the generalized fight.
>> Yeah. [laughter]
>> Yeah. Yeah. It makes you think, right?
>> It makes you think.
>> Yeah. Yeah. It's also quite interesting, like I showed you the models, like what happens when you increase the size of the context window, um, and how behaviors actually are largely shaped by the size of the context window. Yeah. That, that to me was like one of the most interesting, uh, parts about the research, um, made me think about our own behaviors in a way.
>> Yeah.
>> Let's talk about also the like forming the gene. Uh, on your website, you're 12 for the three co-founders.
>> Yeah.
>> And just let's talk about how this team comes together because you may not yourself don't have that academic network. Yeah. You manage the people.
>> Yeah. I started reading all the research papers by that time. I was already pretty deep into like having a decent understanding of of not world models in particular, in particular LLMs and and transformer-based models. And so, um, there was Genie, there was SEMA. Those two were really, really interesting. In SEMA, in particular, was interesting because what they do is they basically take 10 games and then they, they have a graphic in SEMA, uh, where you can see kind of the precise actions that are inside of those games that they mapped. And I believe they found something like a 100, um, which are actually actions that also exist in the real world. And, what they did was they then, I believe it was specifically for navigation, they did a nine-one hold-out set. So they, they, they trained, um, an agent on the nine games and then, um, had it play the 10th game, the hold-out game, but then they also trained a specialized agent just on the 10th game, and they compared how good they did. And my, and if I recall correctly, it did roughly as well, um, playing the 10th game on navigation specifically on the hold-out on the nine-game agent than it did on the one-game agent. And that to me was really interesting because that's precisely the type of data that we had, right? And so for us, the thinking was, okay, what if we did exactly what LLM did? What if we use, right, this, um, uh, right, so LLMs were trained on predicting like text tokens on words on the internet. What if we predict action tokens on essentially what is the equivalent of the Common Crawl dataset, but for interactivity?
>> Vision input?
>> Yeah. Action output.
>> Correct. That's it.
>> Well, I, I think actually, I'm going to double back a little bit to like a question I had, which is one of the, one of the reasons why I thought you would prefer keyboard and mouse over actions is the action is potentially unbounded. Right? You can jump, walk left, walk right, but then also look up, look left, bench. It's, it's unbounded. So, it's huge, isn't it?
>> Yeah, I think problem.
>> Yeah. There, there's benefits to the action space being small to start with. So, I think we're, we're going to start with anything that you can control using a game controller, but yeah, long term, we want to actually predict maybe like action embeddings and have models sit inside a general action space to be able to transfer out to other inputs as well.
>> Got it.
>> Yeah.
>> Okay. And then let's keep going on on the research side. So, uh, Genie. Yeah.
>> And then do the co-founders.
>> Yeah. So there was the Diamond paper, there was Genie, and then there was SEMA. The Diamond Paper for me was really interesting because they had actually managed to get this world model called Diamond running on a consumer GPU. I believe it was 4090 at 10 FPS, and you could play it. And they did that on like 90 hours of data, like 95 hours. I think it was 87 hours, and I think 80 or something like that. That was just incredible, right? That they had something playable on that little data. So I actually cold-emailed the entire group of students, and I was, and I, I told them, hey, I think we have this thing. And then it was pretty interesting. So like right when that happened, a lot of the labs, uh, also started started understanding what we had. And so we started very aggressively, multiple labs tried to bring us in in various ways, and and they were part of that. Like they basically were seeing that happen. And I think for them, that also kind of like solidified how real it was. And then when we chose to do our own thing, you know, initially we thought that we were going to have to just work on world models, right? So, so we thought, okay, the main benefit of this dataset is is is like Genie is is world models. What we didn't realize at the time is that we have so much of this data is that we can essentially do these world models in parallel and take the equivalent of like the LLM, mostly on imitation learning, and then use the world models after that to get into like our off-stage, right? And so for us,
>> And eventually getting rid of robotics, something like
>> I mean, ideally, you got rid of the imitation. Yeah, the imitation learning, but yeah, we essentially realized that that we could get so far on just imitation learning. The way to look at it is we essentially, like, let's, let's take the LLM analogy. We essentially have sort of the internet or like Common Crawl, if you will, and every single lab is trying to simulate that, right, in order to get similar data in order to train their agents. And so for us, the reason why we stayed independent and we just did our own thing was we think we essentially leap every single company that's forced to either be consumers of world models or or build world models and and and take this, take this foundation model bet for switch the board, spatial temporal agents, and be in a place where, you know, we have a lot of customers years before any of the labs even get there. And maybe the most similar, um, comparison is like what Anthropic did with code. Right? Anthropic just focused really, really hard on nailing the code use case. Their models are incredible for it. A lot of their customers use it for it. So we just want to become incredible at this spatial temporal agent use case, and likely that starts in like game simulation, and then using world models, we can then start expanding out to to other, um, areas. So would you show me a little bit of how it does generalize up?
>> Yeah. Demos. Um, but although games is kind of common error.
>> Yeah. Games and simulation. Um, I, I, I would, I would specify as game engines and particular. So even if you're, for instance, uh, simulating human behavior in Omniverse because they're trying to create better training data for factory floors, um, you can use it.
>> Yeah. Maybe Meta has a similar dataset because of the Quest.
>> I never really asked them. I never really looked into the Meta Quest specifically. So you need a few things. You can't just like, there's lots of companies that have like maybe recorders, but you also need the public graph, otherwise you can't train on the data, right? You can't train on people's like private videos that they have saved somewhere, right? And so I think you, you need the social network graph components, um, because these videos need to be on the internet to rank.
>> No, to train on them. Yeah. I, I mean, I think, I think generally people, like people don't want to train on, like, like because these things, they live on your device usually, right? Um, and you can't train on anything that lives on your device. Like you actually need to go and upload and do your thing, right? For Meta specifically, I think also VR, the scale of VR is still pretty small. The amount of, um, environments in VR that are that that have like consumption at scale is probably in like the hundreds. Um, whereas on PC, it's probably in the tens of thousands, right? And so you get a lot less diversity. Um, the three-dimensional input space of VR is pretty interesting. We see some of this too, obviously. And so yeah, I do suspect, you know, Meta, Meta starts using these types of things, but it's unclear to me whether they can get to like a similar scale of data or diversity on the environments as we can.
>> Yeah, there's a lot of challenges there.
>> Yeah. Um, okay. I want to take this in, in like a few different ways, but I guess let's, let's fill out the papers. Uh, maybe one more to mention is Gaia. Yeah. Which, uh, I actually, I interviewed the guy authors, but that too seems like the particular, uh, insight that brought it over seed.
>> Yeah. So, so Anthony, too, who led the, um, research on Gaia 2, is also one of the engineers that joined our team. So it's all the Diamond, uh, the core contributors for Diamond, and then Anthony, um, and we just had three more researchers join this week. It's been a good week. And yes, I, I think a lot of the approaches in Gaia 2 were heavily inspired by Diamond, and then Vansa, who was one of the, um, authors of Diamond, also already was at Wave by the time that I emailed them, and he also realized what this was and realized that, you know, you could scale world models to a much larger, like scale, and decided just to make the leap as well. So I think everybody that sees a dataset makes a leap because it's, but it takes a while to wrap, wrap your head around it because it's like, oh, it's video games, right? Like intuitively, it doesn't make sense. And then when you actually understand and you see, right, how we've been able to transfer it to physical world video and things like that, then it makes sense, and then everybody tends to jump. I would call it video games and call it RL. So then, yeah.
>> If I lived in San Francisco, maybe I would. Yeah. [laughter] Uh, uh, just a quick note, cuz we actually cover all these papers in in latest club. Uh, SEMA 2 did not seem to have as much in Tekken SEMA 1, and I don't really know why. They, they did a lot more work. G3 had a ton of impact, but I also felt like because you could play with the model or people, it just seems an extension of all those things. I guess, like any quick takes on SEMA 2, G3, which both this year is what?
>> Yeah, I'll, I'll talk about SEMA 2. The steerability of SEMA 2 was to me the most impressive part because lighting up the action sequences and the text conditioning is is quite hard to do, right? And so that, and the fact that they were like, it's also quite interesting that it means that they can sort of use Gemini as as part of the flywheel, right? Where, um, where you can sort of scale, scale this orchestrator as like an independent, almost like a puppet master, if you will, and then like, in theory, Gemini could orchestrate many instances of of SEMA, right? That to me is the most, most interesting part is where I, I tend to agree with this, where like I think our models will initially be used as like, like you'll have like an orchestrator VLM of sorts that's kind of like managing instances and instructing them, um, and I think SEMA showing that you can do this was was fascinating. Also, the fact that you could, um, they didn't just have text conditioning, but they also were able to do like drawings and markings, uh, of where to go. They really took an interesting end-to-end approach to me, that I, I look forward to seeing a lot more of. Um,
>> But you talking to them, like you saw is that the one collaborative room?
>> Yeah, I, I think the, um, yeah, we're very friendly with DeepMind. We like them a lot. I just saw the team not too long ago, and I think, you know, big fans of their work.
>> The, the headline that kind of shade from Alice Heath's coverage review. Yeah. Is you're the biggest bet that Invido Cross has made since OpenAI.
>> Yeah. How did that conversation start?
>> Okay. So what if I know it's style and maybe I'll get slapped in the fingers for revealing this or whatever, but
>> Forgive me if I were bad.
>> Um, is he asked you to like draw a 2030 picture of your company? And I think he just picked N plus five years, whatever. I don't know.
>> I did the same to you. Yeah. Um, he asks you to like walk that back from first principles all the way from today, and and and he asks you to do that flawlessly where he can challenge any assumption, any part of the vision, that that, and he asks you questions, right? He has a very technical background. He also has a bunch of technical people on his team, and he truly backs people that have these like very large visions on that vision and the ability to defend it alone. Um, and that's what he did for us. Um, and I think that's why he made that. So I think also, through this, uh, through through this question, he gets to know a lot of things about how technical you are. He gets to know how well you think from first principles, because if that, if that vision is not connected to something real, it's very easy to suss it out by asking good questions. Um, and then, and then he just backs fully. I think like he, he really gets in your corner, um, if it's the right fit. And yeah, they've been incredible partners. They, they've opened so many doors for us.
>> I had to ask the question. Any, like, it's a, it's a very notable story. Uh, obviously a lot of work went into it, but it's also worth it when come out of side.
>> For sure. One of the things also wanted to, I, I guess I kind of asked this question out of sequence, but, um, one of the things that excites me about talking to you is there are a lot of people like you who are founders of business and businesses that along the way have a ton of data, and yours happens to be highly valuable. You pursued before deciding to do an independent journey and also talk to other companies about potential licensing or acquisition. And something is your learnings from those periods. But also like, one, one version of this is very simply, how do you value data?
>> Yeah, I don't think you can value it unless you actually model it yourself and see what the capabilities are. That's my, that's my real outcome.
>> You say model, but train a model.
>> Yeah.
>> But that's obviously like not doable for, for everyone. Um, and also, I think my general advice would be, as model capabilities increase, you, and models are also like, you know, these, uh, fuel ML, they're very, very good at labeling as well, generally, right? What I was afraid of when I was having some of these conversations was, okay, like, you know, as, as the cap, the capabilities increase, you're just going to need less, uh, ground data, and like you can do more model-based data generation or synthetic data generation. I would recommend if you're going to do large data deals, like just try to get like a large chunk of equity in the company that you're doing it with. Um, if you can. Now, a lot of them won't do this, but I think that to me would, or just go do the research, figure out what's actually possible. In our case, we were quite lucky in the sense that this is actually the foundation data,
>> Right? And I think, right, like that's not true for, for every dataset. I think, you know, we just happened to to hit a particular gold mine.
>> But you, you also did, you read Kab, you did the action thing like one or five years ago. Yeah. So you were.
>> Yeah. That's the thing, like you, you have to be grounded, right? And I think a lot of the, um, and, and I think that's the hard part. And I think a lot of what's interesting is you can also kind of look for if like scaling laws already exist on your data type, which like for video, there were some, but for these like input action labeled, uh, sets, there really wasn't any. The other question is like, does it go into LLMs? Does it go into, uh, world models? Does it go into like, what type of model is it going to be used for? And I think that's an important thing to know. And so I just want to, you know, if, if you're having these conversations with labs about data, just like make sure that you actually understand like what it's going to be used for, cuz that's a very, very good way for you to like make the decision yourself about whether you want to pursue that. Now, a lot of them won't tell you that, and I think,
>> You know, I think in, in that case, you don't generally just don't want to do it, because like, I think, I think for our case, like we really cared that, like, for instance, there weren't going to be competing products with game developers built, right? Because we didn't want to like bite the hand that feeds us, and I think we are part of the games industry. So those questions, I think are normal. And then we eventually decided, you know, he just has the data, we're just going to go do it ourselves, and that's when the rest happened.
>> Yeah. And
>> He assembled the team that didn't, uh, take advantage of that. I, I feel like that's, you've aligned a lot of stars in order to make GI burn.
>> Yeah.
>> That other data founders, they, at the beginning of the journey.
>> Yes. Or I'm a data founder. Founders.
Who happen to have beta, but they have a main business, right? I I don't know if you there's two sides to this, right? There it's really easy to be super naive about it and like I had a lot of people tell me initially, oh, it's not that valuable. You're just like making this up and and and so for me like doing the work and actually understanding it myself was a really, really big part of of building that confidence and go start the company.
But a lot of times it is true that like model capabilities increase so quickly that like the certain data you just don't need anymore. Um and so I think it is it's really important to like get people to do the work such that you can make these types of distinctions and and and so so my recommendation would be go build models with your data, see if you can create any sort of capabilities that that aren't clearly already there um or on path to being there, and then figure out um where you go.
"Yeah. I didn't want to ask this earlier, but you gave me the opportunity to uh when you say do the work thing, do coursework and all that, and your co-founders gave you some homework."
"Yeah."
"Uh is this like some books? I mean, courses?"
"No, this was um Francois Flores. Flores. So he has a little book of deep learning and then he also has a full course that he's published um uh on his website. I went through the entire course uh over the summer. I believe it's like something like 30 or 40 lectures, which also take-home projects and things like that. Um and I would recommend anybody uh uh does this. It it goes through right history of deep learning, like the the topology. It takes you through um the linear algebra, the calculus, eventually end up with like chain rule and by this time you've you've done like all the the the more important concepts. It takes you through how do you create neural networks using uh using these concepts that you've learned."
"Wow. This is super first principles. This guy and I've I've I've had the the the uh opportunity to spend some time with him as well. He is one of the most first principles people I've met in my entire life. I'm convinced like I actually asked him why did you curse? He said, oh, because I thought all the other curses weren't right. And because because he is so first principles and he can only explain things from like everything you see and how he explains this thing. It's everything is from first principles, including like the history of deep learning itself was part of of the course. And um yes, he goes uh um so all so he goes through everything and then uh and by the end of it, I think you like I now have like a pretty good intuitive understanding of how everything works, but obviously still right like I I like to describe it as um I'm like the the guy who just got his driver's license. I can drive the car and like my co-founders are like the F1 drivers that like have done this for years. They know where all the um uh where all the the gaps are. And and so I I enjoy getting to learn from them. The cool thing is also that world models is just like a very, very new space. And so, you know, I I get to bring ideas to the table that like no one thought of. And not because I'm great at this, just because it's such a new space that like people just haven't tried it yet."
"Um so"
"To get a hit on definition."
"Yeah. What are world models to you?"
"You know, in a video model, you might predict the next likely sequence or the next most entertaining, the next most entertaining frame. Um what world models do is they actually have to understand the the full range of possibilities and outcomes um from the current state and based on the action that you take, generates the next state, right? So the ne the next frame. And so it is it is a much more sort of complex problem than than traditional video models. So to me, it is it is a world that is accurately generated based on the actions that you take as a result of what's already been generated."
"And just to fact check, uh that needs to understand physics. It needs to understand if I'm building a type of material, you need how it interacts with some type of material."
"Yeah, I think the interactions is the most important part. I think the reasons why world models are so fascinating. One of the things that I did when I was studying over the summer was I tried to actually build a super rudimentary um PyTorch based physics engine, which I would not recommend writing a physics engine in PyTorch for obvious reasons, but I wanted to be able to um because it's differential, so you can uh you can generate the model."
"But yeah, exactly. You can. And then you can um uh uh train. And so I wanted to, you know, I got so many people ask me about, you know, why aren't you just using uh why aren't you just simulating or generating this data. Um and I really wanted to understand from first principles why. And I think the most important thing that I figured out was the compute complexity of simulation goes up really, really rapidly with three variables. First, the number of agents in an environment. Second, uh their doth, so their individual freedom."
"Yeah. And then third, the information that each action reveals. Um so, like, um for instance, if you if you have a text action or a speech action, the environment can change so much based on whether you say right, water or fire, that the outcomes are going to be completely different of like how a human would behave in that type of situation. And so it goes up so quickly with those three variables that at some point you just hit a point where you just want to maximally bet on either video transfer or generation of these environments using world models, because that type of stochasticity is just incredibly difficult. But it's already very, very present in a lot of the video pre-training uh that goes into into these world models, right? And so I think for for us, it is more so about making a maximal bet on video transfer and interacting with things that are difficult to simulate. And the steerability is also really interesting with text uh than it is on betting against simulation or something like that. And so I think there's still a large market for for traditional simulation engines, specifically in areas where video is really hard to get."
"Is this exactly what the big lads are also saying when they're talking to that?"
"I honestly haven't talked about the bigs to the bigs like since we started working on them ourselves. I think people are more reserved with what they share with us. Yeah."
"Of course. With him said, from your question. How would you contrast your version of world models with Fei Yamun?"
"Yeah. So I don't know exactly what Yangun is doing today. My understanding it's based on the Fija like L Japa approach, which is so I'll start with Fei. I think what's really interesting about Fei's approach is that you in some way are able to reuse the the um the splats, right? In game engines and in things that let you stay in verifiable domain. Um which I think is a really interesting approach. Um however, my understanding is they're currently not interactive, which in my opinion is like the whole point of of world models, right? It's it's environments. They're great environments. And I think from a business perspective, I think they they picked a really important part of the tool chain, but to me, that's not really uh a world model. But my my guess is they'll get there, right? They'll they'll start generating."
"Yeah. They just really use it."
"Yeah. Exactly. Exactly. And I think right, Fei is one of the like founders of the entire space. Um so I think it's going to be really interesting to me on on on what maybe that interactive piece looks like for me to really judge their approach. I I think we interviewed this before uh we interviewed her with Justin Johnson uh her co-founder. He was he was more focused on the physics side of things and game trying to have good news. I I I I do think that basically that the splats if you just add more dimensions on I guess the forces acting on them then then you get to drag 2D out of the box because you basically these are virtual atoms that then has all the blow physics applied to them."
"Yeah, I'm uh I'm excited to see what that looks like when they actually release it. It's really hard, really hard for me to comment on anything. I really like the um uh the the frame-based approach um because all of our video or all of our training data is in this format."
"Yes. Yeah. So they we actually asked them about this and they were like, yeah, it's possible, but we they're choosing the splat."
"Yeah. Yeah. And you can also go from splat to frames, right? I'm sure you can write like at some, it's it wouldn't be easy like you'd have to actually render out the environment. So sure, it's not it's not going to be a simple problem, but like in theory, it has to be something that you can do if you really wanted to. So like because it's almost like having a more sort of ground truth three-dimensional representation of the underlying world, right? So I think it's an interesting approach. Um it might be overkill, right? Uh uh you're also dealing with like a much larger like degrees of freedom on the output space, right? So so who knows how well it scales. I like the fact that like I think these video models also use things like autoencoders, right? You can actually have the world models predict like much smaller um uh maybe like a"
"Resolution or size."
"Yeah, exactly. And then you can use like diffusion upscaling or methods like this to actually um uh enrich. And so I think that world models just allow a much more or world models in my sense for a much more like controlled space that that that we know really well. Um I'm not suggesting their approach is wrong. I'm just, you know, like this is I think what we really like about it. Honestly, Yan's podcast that he did, I don't remember which one it was, but a long time ago where he where he basically proclaimed LLMs to be a dead end, um was one of the things that inspired me to do this. I think this is very consensus around models people. Basically, everyone heard this is like stops with their LM and just filter to world models. I would say that the main perspective I asked this exact question to Nome Brown from Open AI and he was like, well, learning the civil models, right? So there's basically that the different things in our system uh"
"Yes. So"
"Yeah. I I I'm not one to proclaim LLMs are dead ends personally. I think um I think they're actually quite useful in particularly as orchestrators. Like the way I think about is as humans, right? We had sort of a three-dimensional world, then we invented text as like a in a way, a compression method, right? So you had we invented text in order to communicate with each other in a in a common way uh in a way that actually compresses all this information that we are perceiving in three-dimensional space into just like a single sequence. And I think that allowed sciences to emerge, it allowed so many literature, like so many parts of the world that we that we cherish. So I think it's a critical part of uh of the whole picture. I also agree that that uh it's very, very clear that they do build sort of the internal implicit world models inside LLMs um and so I think they'll be very helpful as things like orchestrators. Um the problem is when it comes to the generalization. I think text has a generalization backbone when most of the the um when most of the the pre-training is is is text, right? Or or or largely text sequences, then I think you want that backbone to be kind of more spatial and plural in nature, and then also just have text like as one of the as as part of that. And I think the actual argument of um of LLMs is also for instance, the auto-regressive nature of the prediction itself. So the um the fact that it's running the entire output right through the transformer and then in order to predict the next token, which doesn't like the environment in the real world is continuous, right? It's always it's always changing. And LLMs kind of just forget about that, right? I think a lot of the the the argument isn't first, right? So I think the the fact the fact that like text doesn't necessarily generalize well to sufficient coral um context, and then the auto-regressive nature of the prediction and using text for that, right? So I Those are those are the two main arguments. Um I think I think text prediction is just one of the actions that is going to come out of of these, you know, these these policies and world models. I think speech and text generation will just be >> one of the actions that that can that can be a part of that. I think that there will just be labs coming at this problem from both sides. Um and everyone ends up in roughly the same place. And the same place will be whatever people think is cool. Uh right? Like whatever the consumer gets >> whatever is closest to the GI >> yeah. And so I don't think there's like a clear answer. I think it's really interesting to come to come at it from the world modeling side, but it's also because we have to, right? Because like text is largely commoditized, we can import all the text. I think it's interesting and tempting. Yeah, it makes sense that you can probably recover it's sort of like you're taking a step back, you're starting your branch of the ML research sheet, but you might guess there just end up recovering all the other tech stuff emergingly."
"Yeah. Yeah. We can import a lot of that research, right? A lot of that is um"
"That's really cool on the on the research side. Let's talk about the stuff that GI is producing more like the the biggest of research and products output. You mentioned the word customers. What are your current customers?"
"Yeah. So, we're already working with some of the largest game developers in the world."
"Yeah."
"Uh we're also working with game engines directly. And so really what we're doing at the moment is replacing essentially the player controller inside of a game engine. So anything that you're currently that maybe like behavior trees or things that you're deterministically coding, we hope to replace with a single API, which is just you stream us frames and we predict actions. And that can be inside an engine or it can be um eventually even inside the real world. Hopefully those are then also steerable. So the models that you saw weren't steerable yet, but I think we want to get to a point where they're fully steerable."
"Well, steerable means like, well, I want you to share figure anything else out in the pre."
"Yeah, I think it's it's text conditioning on the generation. So yeah, the ability to to you're right, we want to get to a point where you can generally and that's why it's called general intuition, where we can sort of can mimic the intuition of all these gamers into humanlike behaviors in any situation. Um as I mentioned also the lab is named after this quote from AlphaFold, which is wouldn't it be amazing if we could mimic the intuition of these gamers who are by the way only amateur biologists um on his path to um he tried to get an AI to train fold it to generate a lot of data for for AlphaFold. And so for us really the the north star, right? What we hope to get to one day is being able to represent scientific problems in three-dimensional space and then have a space-temporal agent capable of perceiving that space and using hopefully also the the right the the text reasoning capabilities that LLMs have today, in addition to the space-temporal capabilities, to be able to work on the other side of that problem. So for that, for us is is sort of the north star. That's why like, you know, we're we're sort of trying to be hyper-focused based on core workloads, the same way that Entropic was hyper-focused code and use that to then get into organizations and expand from there."
"Yeah."
"Just as a side note, since you mentioned Entropic, any idea what they did on this to to solve."
"No, out of any lab, I probably know Entropic at least, to be honest. Yeah. I admired him though."
"Yeah. Well, the the the current working theory is that they had a super lucky um roll of the ducks, [laughter] but we'll and and then he compounds from there. That sounds like a nice story. I'm sure it's not that."
"Yeah. Okay. So, um why did game developers want this?"
"So, if you're a game developer, how well you're actually retaining players is like um if you have a game that's already at skill is like decently dependent on how good your bots are. So, if you're logging in at an obscure time, let's say 3:00 a.m. in America, and your player liquidity is low, then you need really, really good bots to keep those players engaged."
"Is this known? Is this a thing?"
"Yeah, for sure. Like Fortnite and whatever."
"A lot of human. Yeah. Um and so if if you're like as a human, do I want to play against bots?"
"Usually, it's not just bots. It's like players mix in with bots because you don't want to play just against bots, but it's better to have a full game than to have like an empty game."
"Yeah. Um and so I think as long as it's part of the environment, I think it's okay."
"That means you also have to sort of grade that skill level."
"Yeah. Yeah. Which we can do. Um because we have we know exactly how good people are at these games."
"Yeah. Yeah. I think for us um bots is kind of like step one. Uh right. So what what I was showing you is we're building a general agent that can sort of play any game um in real time. But really that extends into all of simulation, right? Like in GTA V for instance, people are genuinely role playing real life, >> right? And so they're actually behaving in quite aligned ways with with the goals they set for themselves. So you have all these examples represented in video games, right? You have Truck Simulator, Power Wash Simulator, >> Power Wash, >> Power Wash Simulator where like actually the behaviors that you'd want uh in nature to be able to perceive. They're all there."
"Most Yeah. It's it's really like how seriously some gamers take Truck Simulator. Um if you haven't seen this, you should watch it. Yeah. They buy the whole like truck driving set and they're doing the job of a truck driver."
"Yeah. What I mentioned to you, we have more people at any given time on Metal playing with steering wheels in like Truck Simulator and these types of games than Whimo has cars on the road. Um it's a ridiculous stat, but it's true."
"Yeah. Yeah. I mean, so you know, I I used to think that quality to self-driving, you kind of just the interplay a lot of GTA 5. Um he Yeah. I mean, so I'm bad at it."
"Yeah. Our bet is not that we can zero-shot any of these things. It's just that like the next self-driving company can maybe have collect 1% of the data because right also for instance clips already self-select into negative events and and adversity, right? And so like a lot of our data set because already the highlights >> is is really um precisely what a lot of these companies spend like their last 20% doing >> right? And I think that's the main argument. If you're if you're another company that's looking at what we're doing, I think the thing that people are not that people won't understand is that anything that that you're currently doing in pre-training, as long as your robot can be controlled using a game controller, we hope that we can move that to post-training for you. So our bet is not that we can create the next self-driving car company. It's just that the next self-driving car company hopefully only needs 1% of the data or maybe 10% of the data, I don't know, right? To be able to deliver a really good product."
"Yeah. Yeah. It's also the the term that comes to mind a lot is active learning. I don't know if you've uh used to identify with that. Start it got less [clears throat] cool for a bit in order to see the uptrend uh bit which which obviously you have the best data set for the sort of high intensity or you say negative but feeling for negative it could be negative all part of it."
"Yeah, for sure. I think negative events is just because it's the most common term that people use for like if you're if you're Tesla, you want the crashes, you want like >> right, right, but it's only gaming, it's yeah."
"So, you know, the model that you saw obviously had really, really incredible moments and and that was Yeah. That um that it had a large representation of people at their best."
"Yeah."
"And worst. Yeah."
"Yeah. Yeah. Amazing. Okay. Cool. Uh any anything else on the customer development side that you want to sort of fetch off?"
"Yeah. Um uh we're also already working with robotics companies, but again, the and manufacturing, but the key is that the robot has to have gaming inputs. So our bet is not that we can transfer over to like higher do robots than the keyboard and mouse. It's really just that we can move the hard work of of pre-training hopefully to post-training."
"Yeah. It's like kind of like the foundation model that is a very good basis to start."
"Yeah. You're going to you're going to give us frames and and likely some text >> or you'll license the model to because they've been wondering."
"Yeah. Our our business model is initially going to be an API, like the Anthropic API. Um but you also saw for instance, some of the video labeling models that we've been able to develop. So um the goal is for any company to be able to take in their uh their video data as well, and we can create first obviously custom versions of the policy for you, the agent. If that doesn't work, then um we we've already working with a customer that that is doing we distill a model and and they uh turn that into a product for themselves."
"So people can engage with you on the agent level, API level, people can engage with you on the sort of model level. Can you also buy data?"
"No, we don't sell data."
"Okay, cool. So that's the that's the business. Um and is there a world in which I I mean I I think this is on your landing page if you are, you know, Frontier Labs for for world models. Is there a world in which there is a more sort of application layer thing that you that comes out, like a chat GPT for whatever?"
"Yeah. You're going to see us launch a few things on on Metal itself that are going to blow your mind uh as a result of this this um this agent. I'll I'll leave it to the imagination for now."
"If people typically grate out, email."
"Yeah, on the world modeling side, like I think one people underestimate is that Metal is already one of the largest, you know, video consumption platforms as well. People watch millions and millions of videos a day. Um so um world auto-based entertainment and things like that. While it's not like a focus for us right now, I think we'll be like on the consumer side, we have the ability to move very, very quickly here, um and and get it integrated in a way that I don't I don't think anyone else can. Yeah, you could theoretically do video gen like Sora, like what is what is that one? What's the Meta one? Meta m not not real. Yeah, you could theoretically generate clips that nobody play, but you know, it's a device."
"Yeah, I I I think for us the games being so human-centric is like a really big part of what makes us special. Like I I actually actually just don't think that would work. Like one thing that we are really excited about though, I'll give you one sneak peek of what we're thinking about is what if you could literally replay any of the clips that you have inside a world model or your friends can play them. Like I showed you a model that already took part of your clip as a context."
"Instant replay, enter that world."
"But it's also how we go from imitation learning to RL, right? Cuz like it's part of our research roadmap anyways to make every single every single clip on Metal playable. Um so uh yeah, who's who is to say that that doesn't apply to just the actual clips that you take?"
"Yeah. Yeah. Interesting. Can you say more about the RL potential?"
"We describe Metal as as the episodic memory of humanity in simulation. So when you take a clip, really the way to think about it is you get the highlight of what is maybe three hours of playtime, right? You maybe get like two to three minutes of the things that were the most out of distribution, right? It is genuinely your episodic memory um of that playtime in simulation, the things that you most want to remember and share. We want to be able to load uh and this is the work that Anthony who is doing. The reason why we built world models is every crash that you run into in Euro Truck Simulator or American Truck Simulator or a driving game. We want to be able right and again, these are ground truth labels, so we know precisely the actions that lead up to the negative events. They're also title labeled when people upload it onto the platform, they say okay, it's a crash, right? And so we can select all these events and if we can put them inside a world model, we can go into right, we can um uh we can train reward models to then reward based on how you perform in clips that actually contain negative events [snorts] for example. And so for us, it's it's very much about um uh right, we can we can create this this this like LLM moment on imitation learning, but actually making every single clip on the platform playable um at billions of clips scale is how we go from imitation learning to RL."
"Cool. Uh we covered a lot of it. Uh is there anything else that you want to do before we to grapple with the the long-term vision stuff?"
"Yeah. Yeah. I think I think for us um this is a very, very ambitious long-term bet. We need the best researchers in the world that that that that want to work on this stuff. It's really exciting not being extremely data constrained. Um like we really get to like we get so many learnings every week that we didn't think were possible and it makes it for for a joy working here. Also the other thing is because we have such a large data moat, we don't have to be as concerned as the LLM companies about publishing because"
"We don't want to be able to"
"Exactly. No one can replicate the models, right? And so for us um we really want to bring back like the original culture of of open research, which is why we did the partnership with Qoutai in France. Can you say I I actually didn't."
"Yeah, we just did a um we just announced our partnership with Qoutai in French, which is an an open science lab in Paris, one of the best research labs in the world. Um Eric Schmidt, I believe, funded in addition to some some French people. They are essentially acting as the partner that's currently doing a lot of open research on the data. We also want to partner with universities who um because like we do believe this is the frontier, but it's so data constrained that really no everyone has their hands tied behind their back right now. And so we want to help fix that. So for instance um we want to work with the universities to build like negative event prediction models for maybe like trucks in India on all the truck data where all these crashes occur. We have all these things that we know we can do that we just haven't at the time to do. Um and so if if you're listening to this and and you're uh maybe an academic institution or something and you want access to some of this data and a research um in educational research fashion, I think we're we're quite open to doing that because we want to educate people and uh Yeah. And other than that, we just want to work with the best infrastructure and uh research engineers on the planet as we're going into scaling, you know, runs that have thousands, tens of thousands, eventually hundreds of thousands of GPUs."
"Yeah. Yeah. Amazing. Uh I primed you this as like the closing question on flight. It's a little bit that no cost 330 there was I didn't know."
"Yeah."
"So what does GI become in?"
"Yeah. In 2030 we want to be the gold standard um of intelligence. Uh and any sequence uh long enough is fundamentally coral, right? Which I think is um so by nailing coral reasoning, you go after the root problem of intelligence itself. What the world looks like is we want to have AI. So I sort of group um the sequences of AI in three stages, and I credit Andrej Karpathy for for teaching this: bits to bits, atoms to bits, and bits atoms, and then atoms to atoms. In the atoms to atoms stage, I want like I want GI models to be responsible for 80% of all the atoms to atoms interactions driven by AI models. Uh uh and and and the reason the reason for that is because we were able to unblock intelligence so quickly and robotics, like intelligence is the bottleneck that supply chains actually converged on gaming inputs as their as their primary input methods and they converged on essentially simpler systems that let us do a lot more a lot quicker. So we are essentially the the 80% market approach. And then you have lots of companies that have kind of like specialized maybe human robot OS stacks that are that are the other 20. And then so I want to be responsible for 80% of all the atoms atoms interactions driven by uh by these models and be the goal center for intelligence and maybe 100x more in simulation because I think simulation will actually be the larger market initially. So I think in simulation um because you have very little constraints uh also from a safety perspective, simulation is much easier. So I think a lot of the takeoff initially is in simulation. So a lot of the simulation use cases like what I mentioned, scientific use cases, I'm really, really excited about. And so um yeah, 80% of atoms atoms interactions uh coming downstream from these types of phases of world foundation models, and then 100x more in simulation. Yeah. Yeah. It reminds me a lot of that uh what Mark and Villa from the Chaz Institute are doing with virtual biology because you can do a lot simulation and you can do."
"Yeah. Oh, you can do it a lot faster uh with interest. Uh amazing. Thank you for inviting us to your office. Yeah. And thank you for sharing a little bit about your training. Thank you. Yeah."