Google DeepMind Lead Researchers on Genie 3 & the Future of World-Building

**[Podcast Script Format]**

**Host (Erik Torenberg):** All of the applications basically stem from the ability to generate a world that just from a few words, you look at it and like there's a world that's generated in front of your eyes and it's amazing that it's happening. I was very excited about how far can we push that. And it's at the point where like a human who is not an expert will watch it and think it looks real. And I think that's pretty incredible. GD3 from Google DeepMind can create fully interactive, persistent worlds in real time from just a few words. Today, we're joined by the team behind it, Shlomi Fuchter and Jack Parker Holder from Google DeepMind, plus Anjane Midha, Marco Mascoro, and Justine Moore from A16Z. We'll talk about how it works, the special memory that keeps worlds consistent, the surprising behaviors we've learned, and where world models are headed next. Let's get into it. Jack, Shlomi, GD3 has taken over the internet. We're honored to have you on the podcast today. Has the response surprised you? Reflect a little bit about the reaction.

**Jack Parker-Holder:** We weren't sure how big it's going to be, but today I felt definitely that we have something that was for a long time coming, basically being able to generate environments in real time. I think a lot of work that was done in Google's deep mind and outside pointed to that direction, but we really wanted to make it happen, and I hope we have, yeah.

**Host (Erik Torenberg):** Team, why don't we reflect internally a little bit about what we found so game-changing about G3 and why we're so excited to have this conversation. Mark?

**Marco Mascorro:** Yeah, for sure. I mean, first of all, it's an amazing model. I think there's a lot of excitement around the special memory, the consistency across all the frames. I think this is the first time I can see like you can have some sort of interactive way of doing this stuff with videos because it used to be like you would do one problem and you would have 15 seconds of a video, but now you can actually have some sort of interactive kind of element to it, which I think is very exciting. So can you elaborate a little bit more on your insights on this? How was, for example, figuring out what data you should collect, how you make it very interactive, and keeping the flow of the whole video, which I thought was phenomenal.

**Jack Parker-Holder:** Sure, yeah. So I think you kind of highlighted a few capabilities, sort of the length of the generation, the consistency of the world, maybe diversity as well of the kind of things you can generate. I think the main thing is that obviously we made progress in quite a few different fronts, right? In separate efforts, right? So we had this Genie 2 project that was much more sort of like 3D environments that it could generate. And it wasn't super high quality. It felt like it coming from Genie 1, but it wasn't the same quality as things like VO2, which the State of the Art video model at the time came out in December, roughly exactly the same time it came out a week later than Genie 2. And obviously internally, there was a lot of discussion between the two projects about the different directions we were pursuing. and then Shlomi had also worked on Game & Gen which is the Doom paper as people know it which I think you guys also wrote a nice piece on straight after that came out so I think that also attracted a lot of attention and so we felt that across these different projects we had quite a lot of interesting things that would naturally kind of combine and we could basically take the most ambitious version of the combined project and see if it was possible and fortunately it was and I think the timeline is probably the bit that surprised many of us because obviously we set ourselves these goals and we tried very hard to achieve them but you can never be totally sure how it's going to actually feel when you've got to that point. I think it ended up being something that resonated with people a lot more than maybe we expected but we were always believers.

**Shlomi Fuchter:** Yeah, I'll just add to this that I think the real-time component is really important and not many people experience it firsthand but we really tried in the release to at least have a few trusted testers interact with it and also get the feel of it by adding these overlays that show what happens, how people can use the keyboard to control it. And I think there is something magical about the real-time aspect. I felt it for the first time when our model, like game engine model started working fast enough and we were just like, oh my God, it's actually, I can actually walk around. And it was a bit of a wow moment. And yeah, I think there is something when it responds immediately, that is really magical. I think that's kind of sparked the imagination of many people when the Doom kind of simulation came out and here we really wanted to push it to somewhere. We weren't sure it was going to work. So it was definitely at the edge of what's possible, I think. That's how we felt. So we just said, yeah, let's try and see if we can make it happen.

**Host (Erik Torenberg):** I think you guys, I don't know if this was on purpose or not, but you perfectly timed it when everyone on X and Reddit and everywhere was making those videos of characters walking through, games, but they obviously weren't interactive, they weren't real time. And then you guys came out with this release that was like, now this is an actual product and it blew folks away. I'm curious, because you can imagine so many different applications for this, right? Like more controllable video generation or making it much easier to create games, even personal gaming where someone's just kind of creating their own world. They walk through like RL environments for agents, robotics. Are there any particular use cases that you're most excited about?

**Shlomi Fuchter:** I think all of the applications basically stem from the ability to generate a world just from a few words. And I think for me, kind of like this potential, when I started looking at video models, I think it was pretty early when I think it was one of the models were like Imagine Video, which was modeled by Google Research. But there are a lot of models that they were very basic compared to what we have today. But the ability to simulate something, like you look at it and like there's a world that's generated in front of your eyes and it's amazing that it's happening. and I think at this point I was very excited about how far can we push that, right? So I think there was one way to do it and Genie is definitely another way to make it a bit more interactive. So I think all of the applications basically stem from this core capability. So it can be entertainment, of course, as you said. It can be training agents. It can be helping agents to reason about the world, education. So I don't think any particular application is more important than others. I think it's really up to how developers in the future will be on top of that.

**Jack Parker-Holder:** Yeah, I would give basically the same answer in the end with a different journey to get there, right? Which is, I personally myself worked in reinforcement learning for a few years before starting the GE project in 2022. And the motivation originally was that in RL at the time, we had this problem where we'd say, which environment should we try and solve, right? Because once you've already done Go, which people thought was years or decades away, and then that was solved in 2016. Well, it's solved, but we reached superhuman level in 2016. And then StarCraft, three years later, which is not particularly a long time for something incrementally significant. So it was 2021 time, it was a big question of what should we try and do with RL. We know that the algorithms can learn superhuman capabilities if they have the right environment, but we don't know what the environment would be. And so we were working on designing our own ones with colors. But then instead, it seemed like the more promising path when you had the first text image models coming out. Whereas what if we just think long-term, what's the way to really unlock unlimited environments? That being said, over the course of the project, and originally we started it, I guess, in 2022, it was very focused on that one application, but it seems quite clear now that this could have a big impact on all those other areas you mentioned. So I think it's like language models in 2021, maybe. You probably wouldn't have guessed an IMO gold medal a few years later, but it would come that fast. as a direct application of that technology, right? It was probably, oh, it can help me with my emails or whatever it was. And I think it's really cool to build these kind of new class of foundation models and then see what people can imagine doing with it. And that's one of the really exciting things about sharing the research preview, right? So you've got this kind of feedback. So we're hoping a lot of these things can happen.

**Host (Erik Torenberg):** One of the things in the research preview post, Jack, that blew me away was this, and it wasn't even your first GIF, I think, in the blog post, it was either second or third. You had this visual of somebody painting the wall with the paintbrush and then the character moves out of, right? Like out of, to a different part of the wall, paints and then moves back. And the original paint is still there. And I didn't believe it. I was like, there's no way. And then I read, and you're right, it was described as a special memory. So the persistence part for me, I'm not taking away from all the other stuff. The interactivity is amazing. But I think broadly speaking, folks expected that at some point, video generation, for example, would become real time. When I saw the Genie 3 post, it was like, okay, they actually went and did it. But the special memory, the persistence was when I kind of sat up in my chair and I was like, how did that happen? Could you talk a little bit about when did you discover that as an emergent property or was that a specific design goal? What's the backstory on that because that feels like a big unlock, Jack. Why don't we start with you?

**Jack Parker-Holder:** Yeah, so that's a great question. I'll say a few things. So the TLDR is, it was totally planned for, but still incredibly surprising when it worked that well. So that specific sample, when I saw it, it was hard to believe. I actually wasn't sure that the model generated it for a second. I was like, that took me to watch it a few times and really check and freeze the frames and look back and check that it was the same. But going back a few steps, so obviously Genie 2 had some memory, right? So this got kind of lost because Genie 2 came at a time when there were lots of announcements, very exciting announcements. I mean, VO2, only a few days later, It was a busy time of the year. And the main headline act was that we could generate new worlds at all, right? So that was the thing that we wanted to emphasize. But it did have a few seconds of memory. And we had a couple of examples, like I created a robot near a pyramid, looked away, looked back in the pyramids there. But it's like kind of blurry. It's not perfect. But some other models around the same time or more recently didn't have this feature, right? So people kind of indexed to that because they didn't notice the early signs of it in the Gini 2 work. And then for Genie 3, we basically went much more ambitious on the same sort of approach, right? And we made it like a headline goal for ourselves, is can we make the memory be what it is, right? We said we want minute plus memory and real time and this higher resolution all in the same model And those are kind of conflicting objectives right So we sell ourselves this kind of technical challenge And we said if we target this then it just about feasible and it'll be pretty incredible. And then you still don't know, obviously, it's going to pan out. So then when you get to the end of the research, seven months later, to see the samples, it still is quite mind-blowing, to be honest. So yeah, it's kind of planned for, but still pretty cool and exciting when you see it because it's like research projects aren't sure things are they? One thing that we didn't want to do and we didn't want to build an explicit representation right so there are definitely methods that are able to achieve consistency and they did that through an explicit some 3D you know nerves, oncocardial splatting and other methods that basically say okay if we know how the world looks like we use this kind of like prior assumptions on how the world remains static pretty much then we can build representation and know what you're looking at. So that's great, I think, for some applications, but we didn't want to go down this path because we felt it's somewhat limiting. And I think, so we can definitely say that the model doesn't do that. And that generates like frame by frame. And we think this is really key for the generalization to actually work. Every time someone interacts with it for the first time and they like test, they look away and then look back, I'm always like holding my breath. And then it looks back and it's the same. I'm like, whoa. it's still really it's very cool it's very cool and how long is this special memory I don't know if you can talk about it you mentioned a minute plus but is there some sort of like measure that you have is it like can you keep it for half an hour or what is the limit on that there is no like fundamental limitation but currently the current design we're limited to one minute of this type of memories yeah it's also a real time trade off for the guests as well we felt that because of the breadth and the other capabilities that like a minute was sufficient.

**Host (Erik Torenberg):** One more question that related on the, between Gini 1 to, like for example, NLMs, like you have DeepSeq R1, like they saw in this paper, like the longer they keep it running, they suddenly will see like these interesting behaviors, like the model will start like reasoning or like would give like a, oh, I'm wrong in this, I should self-correct. Do you see anything in kind of like these scaling from two to three, do you see any sort of like interesting behavior that you were not expecting that suddenly just appear by increasing the amount of data and the amount of compute?

**Jack Parker-Holder:** Yeah, I'll just say, I think there is a bit of like overall, definitely like many generative models, we see that improvements happen with scale. I think that's not secret. And I don't think it's not the same type of intelligence, I would say, like an LLMA has. I'm not sure if reasoning is the right term, but we do see that some definitely things like It can infer from if you approach like a door, and it makes sense for the agents to maybe open it. So you might see that it's starting to do that, for example. Or there's some better world understanding that happens over time, and it's just like things look better and more realistic. So I think these are the trends that we've still observed. Yeah, and from G2 to G3, I think the real world capabilities really increased, right? So on the physics side, some of the water simulations, you can see some of the lighting as well. Like a really breathtaking. I think we have this example of the storm on the blog. And that one I think is super cool. And it's at the point where like a human who is not an expert will watch it and think it looks real, right? And I think that's pretty incredible. Whereas Virginia 2, it was like, it kind of understands roughly what these things should do, but you know it's not real, right? You can look at it and you can clearly see that it's sort of not completely photorealistic. So I think that's quite a big leap on the quality in that side.

**Host (Erik Torenberg):** Yeah, one of the things that was really cool in all the examples was the water is sort of a great way to see, like, does it understand, like, what the world is and how objects interact? And that example someone posted of the feet going in the puddle was amazing. But then there was also that example of, like, a cartoon character. it was more of an animated style who was like running across this kind of green patch of land and then ran into this blue kind of wavy thing that looks like water and he started swimming which I thought was really interesting like were there particular things you had to do around that for the model to be able to understand like how characters should interact in different environments and different styles what you're basically describing is like the real breadth different kind of environment terrains and worlds and things like that, like water or walking on sand versus going downhill and snow and how the agent's sort of interactions should differ given the terrain that they're in. And I think that that really is a property of scale and breadth of training. So this is very much like an emergent thing. I don't think there's anything really specific we do for this. Again, you hope the model has learned this because it should have a general world knowledge. It doesn't always work perfectly, but in general it's pretty good. So for the skiing examples, you do go fast when you go downhill and then when you try and go back uphill, it's very slow, if not at all possible. When you go into water, obviously you hope, as you said, that the agent will start swimming and splashing. And this does typically happen. When you look down near a puddle, hopefully you're wearing Wellington boots. like this kind of stuff does just kind of make sense and I think it feels pretty magical because it very much aligns with what you were thinking about the world and the models just generated it all so yeah that's that's also one of the really exciting things for sure yeah and on top of that one kind of trade off that typically we have is that we want the models to do two things we want the model to create the world in a way that looks consistent so like Jack said like if you walk in rain or in pulse than probably wearing boots. But if we provide it with a different description or like the prompt is saying something else, we want it to still follow the prompt. And there is some tension here because some things are very unlikely, right? You might say I want to wear flip-flops and jump in the rain or whatever. Then the model still has to try and create something that is very unlikely. And that's where typically, you know, video models maybe find it more challenging and that's where our models might find it more challenging but still successful to a surprising degree to go into this kind of low probability area and I think that's really, in a way that's what we want many people they want to just look at the video that looks like their own, maybe this room but something a bit more exciting and that's where I think this is the magic of the models that they can take you to places that maybe are not so likely to be in reality. The text following is really amazing in this model. And that does feel really magical. I think this is something that the VO does really well as well, right? Like pretty much what you ask for. It's really well aligned with text. And we have that with Genie 3. So you could describe very specific worlds and really kind of like arbitrary, silly things. And it pretty much works. we actually had this discussion because people were very disappointed to find out that the video I made of my dog actually was not my dog's photograph I just described her in text and yeah I don't know if that's a big secret but it looks exactly like her and the model just kind of knows I think that's pretty amazing so I think that that's actually a really important capability that we didn't have with Genie 2 as well because we relied on image prompting. And so there was some transfer issue where you rely on imagine to generate the image. And that often does look really good, but it's not necessarily a good image for starting the world. Whereas going directly from text, you get the controllability for anything you want. Plus it just kind of naturally works because it's in the correct space for the model to do its thing. And that's something really powerful. And why is that, Jack? What do you think led to such a massive instruction following or text adherence gain? Because it's a pretty hard thing to do.

**Jack Parker-Holder:** Well, I mean, our team had never really worked on this. And so Genie 1 and 2 both worked with image prompting. And so obviously, for this next phase, we leveraged a lot of the research done internally on other projects. And personnel-wise, I mean, Shomi's obviously been co-leading the VO project. And so we were able to kind of build on a lot of other work and ideas internally and that basically may allow us to turbocharge progress. So if we'd done this by incrementally building ourselves on an island, it would have taken a lot longer than being part of Google DeepMind where we have these teams that have a lot of knowledge in different areas and lean and build on. Which I think is super exciting about our big industry right now is that we have so many experts in different areas that we can seek out advice and help from. And Shlomi, a question for you on that is, having let the VO3 work, which is kind of mind-blowing, is there a reason why this is Genie 3 and not like VO3 real-time? So I think it's definitely a bit different, right? Like Genie allows you to navigate environment and then maybe take actions, right? And that's not something that Veo at this point can do. but there are other aspects that are different that Genie doesn't have, right? Genie doesn't have audio, for example. So we just think it's while definitely there are potential similarities it's sufficiently different. Also another thing is that at this point Genie 3 is not available as a product and we do think about it as a product that is mainstream and became very popular and what the future holds, I don't know But I mean at this point we just felt it sufficiently different in terms of what capabilities and how we think about this So Gini3 is pretty much a research preview right It not something we are releasing at this point Something we think about a lot is what are the edges of a modality? We're talking about this all the time, which is the lines start blurring pretty quickly between real-time image and video and then real-time video and interactive, whatever, world generation, world model. I don't think we have a good word for what Genie3 is yet, but you guys called that world model, which is, I think, a great term. But in your mind, where does the video generation modalities stop and real-time worlds start? And do you think in the future, are these converging into basically one modality? Or if you had to predict over the next few years, do you guys think actually, yeah, these will diverge into completely different disciplines? It seems like they share kind of one parent today, which is video generation, but where is the world going, do you think? Are these two completely different fields?

**Shlomi Fuchter:** From my perspective, they're different. So I would say modality is one thing, right? We have text, we have audio. Even within audio, there are different types of submodalities. Speech is not the same as music. We have different products for music generation, and we have other models for speech generation, speech understanding. So even within one modality, you can have different flavors. and then of course you have video and other things so I think basically I would say the modality is one dimension and another is how fast or how quickly we can create new samples and completely orthogonal maybe direction or dimension is how much control we have so I think we kind of picked specific direction or a specific vector in the space for Gen3. I think different products, different models can try and go in a different direction. I think the space is pretty big and there are a lot of trade-offs to be made. So, yeah, I don't know. I think it really depends. Some people believe there is, you know, one model that's with everything or I think there is still open-ended what's the best way. Like we're in a place where engineering is a big part of our research, right? And actually making those, like it's not a paper, right? where we want to build something that people can actually use. So I think this really makes it like abstract ideas go to get you to some point, but to actually build things, you have to make some concrete decisions. And I think it kind of forces you to decide what you want to do and what you're going to do. Yeah, I think this is a really interesting point, Mike, and ultimately it has to be driven by technical decisions and also the goals, right? So if you look at the models right now, we obviously made a choice that we want VO3 and Gini 3 to be separate projects this year, right? And if you look at them both as they are right now, they have very different capabilities that the other models not have. And technically to combine all of that already into one model would be, I think, very challenging. I mean, VO3 is clearly a higher quality threshold than Gini 3, right? And it has very different priorities, right? So then the natural things you could say, oh, well, you know, what if we just took these together and combined them? But that may not be the best next step for either of those two models, right? So it may not be the case that the thing that the other one has is actually the most compelling thing for a completely different experience. And I think that given the breadth of interest in both models, right, there's actually... quite a small set of people that are really actively using both. And they tend to be more folks like yourselves who are just more broadly interested in AI, right? Rather than really downstream use cases. So you mentioned agent training, for one, which is very high-action frequency, requires more ecocentric, sort of, I guess, more worlds where tasks can be achieved. It doesn't require high-quality cinema, videos you could generate with a bio model, right? It's quite different. And then on the filmmaking element, I mean, I'm not so sure that Genie 3 is really there at this point. And that would be necessarily the goal. I don't know. On filmmaking, Justine can do some pretty incredible things with the filmmaking tools today. You'd be surprised. Give me access and I will make amazing films with Genie 3. I guess that did kind of get to one of my questions though, which is the work you guys are doing is incredible and you clearly probably have so much going on in your brains just to coordinate training these models and managing these teams, how much do you also have to think about, like, what are the downstream use cases of the model when you're training it? Because you could imagine a world in which you're just like, we don't really know or care what people are going to do with it yet. We're just going to go in the research direction we think we should go and see what happens. But based on how you guys are talking about it, it sounds like you've also been pretty thoughtful around what are the different capabilities or features needed for different potential use cases, at least, of different models.

**Jack Parker-Holder:** Yeah, I'll just say that basically we have some applications in mind, but that's not what's driving the research. It's more about how far can we push in this particular direction? Can we make all of that work really great quality, really fast generation, real-time, very controllable? I think that's what drives us. I think they're to have, to develop Genie Free. And the applications kind of like follow, and I don't think, you know, to be honest, I don't know what would be the applications for. Like, I think we're very surprised. You know, I'd like to mention like, they're free, people find new ways in how it can be useful. And to prompt it, we have like visual stuff, you know, people just discovered it, right? We didn't even think about it initially. So I expect kind of the same thing, and I think that's why I am excited for more people to be able to access in the future. And in general, our approach is to make sure that over time there is more access to the models we build. And I think that's the only way to discover what's the real potential. I guess one somewhat related to that, how do you think going forward, like Genie 4, 5, or any other models, what is top of mind right now? If you wanted, for example, to focus on, I don't know, like seems like gaming could be one of the applications, having multiplayer type of games, we have two special memories or two different completely views, but at some point they merge. How are you thinking on like going forward? Like what's next? Is it like scaling these models just on more data, more compute? Is it creating this sort of like multi-universe type of things where you have multiple players, multiple people looking at the same model, but in different views? What's top of mind for you guys?

**Jack Parker-Holder:** Top of mind, I think for the next few days might be a vacation. after that maybe walking my dog in the real world and then I think you mentioned a bunch of really interesting things to be honest and like I think we are we're still collecting a lot of feedback on this current model right and I think that in general we are most interested in building the most capable models right and so we would hope to have even broader impact in future and really enable other teams to do cool things with it, both internally and externally. And for me, it's like, I just started this with a very, very focused vision about AGI. And I still think, honestly, what I'm excited about for AGI, which is more embodied agents, I really believe this is the fastest path to getting these agents in the real world. And I think we made a big step towards that. but I'm still sometimes even more excited about applications I never thought of that come up from other people seeing the model. So I think it's kind of this trade-off of, obviously you want to focus on some applications but then you want to be open-minded about others and I think that's the real joy of building models like this is you get to see all of these people can be way more creative than me with it. So I think that there's always really cool things that we can do and I honestly can't really tell you in one year what the biggest application will be. But we'll definitely be trying to build better models.

**Shlomi Fuchter:** Yeah, I'm really excited. I think we're only as impressive, you know, maybe the model is. I think they're very far from actually simulating the world accurately and being able to do, kind of put a person in there and then do whatever they want. And I mean, when I say far, it doesn't mean it's far in terms of, you know, calendar time because we are really in an accelerated timeline. but it feels like there is more work to do to get there. And I think I just imagine once we can actually, whatever the form factor would be, but stepping to this world and just kind of like maybe tell it how you want to, what you want to experience. There's so many applications. Imagine, for example, someone is afraid of talking to people on a stage or in a podcast, right? They can simulate that, right? or you can have someone who is like afraid of spiders, they can maybe actually see themselves getting over that. So that's like, you know, just one example of something that actually my wife thought about it. It's not my idea. So I think it's really like there's so many things, right? So I think this is just, it all hinges on the ability to simulate the world and maybe put ourselves in it, maybe seeing yourself from the side and potentially having agents interacting with things. and yeah, the realism and really making it work in the way that is similar to our world, I think it's really key.

**Host (Erik Torenberg):** I am actually personally petrified of skiing and the model is already quite good at that. So I might, when things quieten down, spend some time because I promised my wife that our children would grow up knowing how to ski and we getting close to the age where I have to live up to my promise And I not sure if I want to do it yet We have to improve the model for you Jack so you can actually get that in distribution I hope so. We were just talking about before we started that we might see applications like in robotics. I mean, Jack, you were talking about embodied AI and like now like limitation in robotics is the data, right? Like how much data you can collect. And now probably you can just generate a lot of different scenes that you were not able to do before, purely from just recording videos or so. So I think that's another thing that is pretty exciting. And congrats on the model. It's phenomenal. On the robotics application, there was a conversation that I was listening to from Demis yesterday where he was talking about your guys' work on Genie 3. And he mentioned that there's an agent, I think you guys call it Sima, right? which can then interact with the genie agent. And as I was hearing him describe it, which was kind of breaking my mind, which is that you had one simulation agent asking the world, asking the genie agent to essentially create a real-time environment for it to interact in, right? Which was when I realized, oh, the way you guys have built it, it's composable with other agents. Can you talk a little bit about why that's so important for robotics, like Marco was saying, and what are the major limitations today that you think we'd have to overcome as a space to make the robotics sort of progress, the rate of progress in robotics much faster than it is now?

**Jack Parker-Holder:** So we designed it to be an environment rather than an agent, right? So Genie 3 is very much like an environment model. We don't see it as like an agent itself that can think and act in the world. It's more just a general purpose sort of simulator in a sense, right? that can actually simulate experiences for agents. And we know that learning from experience is a really important paradigm for agents. That's how we got AlphaGo, because AlphaGo learned by playing Go by itself, trying