Making taxes fun with Pikachu and AI | Cadi Zhang

Cadi Zhang joins Reid Hoffman and Parth Patil to explore how generative AI is changing game development, world models, and robotics. Drawing on her work across Unity games, VR teleoperation, AI accounting, and robotics product operations, Cadi explains what virtual worlds can teach embodied intelligence—and why physical robots still lack the tactile, real-world data that can’t be scraped from the internet. She shares how she uses GPT, Blender, PixelLab, and Aseprite to prototype games faster, including a duck-themed imposter game and PokéTax, her Pokémon-inspired tax-filing game. AI can rapidly generate code, gameplay mechanics, and 3D assets, she says, but it still can’t judge whether a game feels fun, maintain a consistent art style, or model the precise force needed to fold a sheet of paper. Reid, Parth, and Cadi discuss simulation-to-reality gaps, the limits of current world models, robot safety, humanoid versus task-specific form factors, and why human taste remains the decisive creative skill.

• 0:00Speaker 00

  We really haven't seen robots operate in the real world yet, getting their own data and then training on that. I think a very small subset of robots that have been able to successfully go out in spaces. Because no one really wants, you know, 100 pound robot in their house and like you said, fall on their cat. That would just be catastrophic, right? I was building out Poki Tax was just like my spin on filing a tax form and like Pokemon games. And it requires, I feel like, some level of thought and guidance to AI. Like, what if you never filed any of your taxes in Pokemon? Because you're always earning like money as you battle trainers. And so it's like, you really have to provide some level of guidance.

• 0:36Speaker 01

  The Token Grantee Program gives $1,000 a week in tokens to high potential creators already deep in AI across film, gaming, comics, print, and digital art. There are no tool restrictions. The freedom to choose is the point. Grantees also get access to my own custom fleet of agents and an ongoing collaboration with me. The goal: to close the gap between an idea and the world.

• 1:03Speaker 02

  Welcome to another episode of Read Riffs. We've got our real Parth, not just Parth AI. It's also the real Read, not just Read AI. And we have an excellent guest, Katie. Thank you for joining us.

• 1:18Speaker 00

  Yeah, I'm super excited to be on. Thank you, guys, so much for having me.

• 1:21Speaker 02

  It's our pleasure and honor. And why don't we start with, okay, so Parth, how did you select Katie?

• 1:27Speaker 01

  Yeah, I'm super excited to have Katie on. Katie, I think we've known each other now for maybe like three or four years. When I met Katie, she was working on a hackathon, trying to build a kind of co-pilot for game design and game development. And me being obsessed with video games, I was like, oh, my God, I want this. Like, this would be awesome. Like, I saw what Cursor did for me for programming, but imagine if we had something similar for making games. And so that's when I met Katie. But over the last couple of years, I've gotten to know her as she's exploring AI, but also getting to know her wide-ranging career arc. So she has experience in AI for accounting, some crypto transaction analysis, as well as like now she works in robotics as a product operations specialist.

• 2:09Speaker 01

  So she has this incredible exploratory career arc, which has been really exciting to follow. And seeing how she connects the dots across all of these different problem spaces and using the AIs to fill in her gaps has been super inspiring. So I figured, I mean, it would be awesome to accelerate your own creative projects and see what you can do and bring you onto the show and see what you've been learning over the last few years. So, welcome, Katie. Tell me more about how you got into AI and what your initial goals were and where you kind of ended up. And talk about your journey to this point over the last few years.

• 2:45Speaker 00

  So I think overall, I had a pretty unconventional path into AI in general. I was working on basically creating games in Unity. I was just super obsessed, really just wanted to create something and was working on 2D platformers and such. And kind of like entered in and was really lucky at the time with robotics just being there and jumped on because we were basically programming VR interfaces with robotic dogs. And it was kind of my foray into like actually starting to program and learn. And I think at the time, it was, we just started getting introduced to like GPT

• 3:25Speaker 00

  and just having that as like the magical thing where you're copying and pasting codes, snippets, literally in at the time. And so it's like hilarious to see kind of where we are at now. But that kind of really spiraled into me doing like haptic suits, coding more VR stuff in Unity to now, you know, working in AI for a bit, for AI accounting, and then now moving over to more of like the data collection robotics portion of, you know, how can we really gather the data? And what is the data that is needed for the next generation of like embodied intelligence or robotics in general?

• 3:59Speaker 02

  Very cool.

• 4:01Speaker 02

  So go a little bit into kind of how building game worlds and intelligence for the physical world, what do they have in common? What do the fields have to learn from each other? What's the bridge, as it were, from bits to atoms? As you're moving from game worlds to intelligence for the physical world?

• 4:21Speaker 00

  They have a lot of crossover than I initially expected. I was quite surprised. And I think the easiest way to even think about it is tell the operations, for example. You are basically in these VR headsets in a game world, operating these robots. And I think like on a software level too, you kind of get to see the advent of these world models coming into play. Just really being able to simulate and just recreate these environments and have robotics or robots reasoning through them. I think the main bridge that is being crossed, or I guess the similarity here is

• 5:01Speaker 00

  like, how can we basically simulate 3D worlds, physical realities on top of having physics within them? And I think that's like the bridge that's being crossed. But world models or robotics is kind of taking this to like the next level of ambition, really just going end to end in many ways.

• 5:19Speaker 02

  By the way, quick follow up. What are the things that you've seen so far that are the biggest gaps between them? Because it is a classic kind of Silicon Valley AI point of view. Oh, you build it all in simulation, then it moves much faster and it learns the physical world. But there's unevenness in some parts of the ground. What are the biggest gaps you've found so far?

• 5:37Speaker 00

  Yeah, and I think this is really able to be represented. I think when Project Genie came out and it was just like bloodbath of like stocks of like Unity, Roblox, like all these like gaming companies. And I think they provide like, I think, a really great take on like what is the gaps currently in these like world models? Where essentially like physics and world models is not yet deterministic. Also, there's like no object permanence in some cases. It's only like short horizons where you're kind of like having the ability to see and interact with the world in a way. So with like these in mind, it's like, how do you create these longer form, interactive experiences in games?

• 6:18Speaker 00

  Other than just like a quick first-person shooter, a quick tycoon game? On top of the fact that if you don't have determinism, how can you create multiplayer games? It's just something that is not yet you're able to achieve. So I find that as like the gap. In some sense, like I find that there's also an opportunity to even bridge this where, for example, I think a lot of people are talking about, why don't you use world models as an input? Into these like more deterministic systems to create these games? So there's like kind of that path versus the more implicit path of just pure, like, straight AI and creating these hyper-realistic games.

• 6:58Speaker 01

  Yeah, I think I showed an early, what clued me into this paradigm was I saw a demonstration of Minecraft. I think it was running on Oasis, maybe on the etched chip. So it was like a more of a demonstration of the chip, but it was like a Minecraft generator, trained on Minecraft. And, but it had no, like it had memory, like the memory of a fish. So anything it would see that was outside, anything that went outside of the frame would be forgotten. And then, so if you walked up to something and then you turned around, everything would change because it wasn't able to like remember and hold on to the details of the world beyond what it could see in front of it. What is Project Genie? Could you clarify that? Like, how would you describe that to someone that maybe didn't get to play with it or see it yet?

• 7:40Speaker 00

  I think definitely it's like, basically you're taking a prompt and then basically able to see and create all these different worlds and interactive environments. I would say, you know, there was Genie 1, which was definitely kind of the very beginning where you're able just to see the introduction into just, you know, interactive worlds. Genie 2 and Genie 3 now, which is kind of the next level of interactivity. You know, the possibilities of even having the world at your fingertips with just a prompt.

• 8:11Speaker 01

  Right, like the world unfolds as you explore it on the fly instead of being predetermined, right?

• 8:16Speaker 00

  Yes.

• 8:17Speaker 01

  What's a recent project where AI lets you move, let you move from an idea to a working prototype faster than you could have before?

• 8:24Speaker 00

  So I've been playing a lot with, you know, how AI and games kind of like overlap. I think prior to this, I was struggling a lot figuring out how to use AI and bridge that over to games. I remember the first time I was like, just let it vibe code, just let it go. And it wrote me 10,000 lines of like this horrible HTML CSS file. Just, it broke, could not get it to work. And I feel like now, I feel extremely supercharged in a way of having AI and being able to just create these games. I think the one of the projects I was working on was essentially, it's kind of a spin on the imposter game

• 9:05Speaker 00

  where everyone's a duck and there's like one killer duck within the mix. And your goal is to act like an NPC and complete a set of tasks. And so that duck is trying to monitor to see which character is actually the human.

• 9:23Speaker 00

  It's really been amazing to have AI help me iterate on so many core gameplay loops. It's just amazing to just go to bed, you know, be in the shower and then have AI just spin up five gameplay loops, send them to friends, take that feedback and just rip at it again. Um, I think the core thing is like, how can we make the game fun? And that requires a lot of time, a lot of effort and research. And it's being able to minimize that and really just kind of get it into the hands of people is just incredibly powerful.

• 9:53Speaker 02

  I'm curious what you've learned so far in this is how much does AI understand, you know, kind of well, the concept of fun? I'm expecting you to say uneven, but but like, given my just made my token prediction of uneven. What's the actual answer?

• 10:12Speaker 00

  I would definitely agree. I think it's, I think AI can approximate. I think you can approximate based on, you know, the games that it probably has seen, it has played and it has like acknowledged what has been trained on. And so like, for example, if you give it, uh, just like examples of the games that you enjoy and want, it definitely approximate exactly like what are the core gameplay mechanics and give you and provide you suggestions. I've noticed like if it's something like fairly new, um, it it struggles. It is, it's not, I think it requires a lot of context building still. So I wouldn't say AI has a good grasp on fun yet. Um, it still requires a lot and a lot of time and effort being put into it of just like

• 10:58Speaker 00

  really honing, what are the gameplay mechanics? It's just faster in the way that it provides you like the code to like see it in real time.

• 11:05Speaker 01

  Yeah, I I've seen, I see the same thing. I'll, I'll be working on a game and actually most of what I have to do is play the game. And then just point out things I'll just, I, I, it's like, my job is put the game in the hands of other people. And then see and just play the game and be like, okay, what is anti fun? And then I go to the AI, I'm like, this experience was very frustrating. We should fix it. And then the AI comes up with like six alternatives to patch that experience. And I'm like, oh, these are pretty good ideas and we can like try them out. But definitely it doesn't understand fun like we understand fun. And I think that's like the very human kind of level of like being in the game and playing and trying to have your own like experience, uh, which is exciting because I think I hope that like we're not automating the part of experiencing the game.

• 11:45Speaker 01

  Although it is cool to have AI playing the game before you do and highlighting and finding errors. Like you kind of have it like bug fixing and like mapping things out.

• 11:55Speaker 00

  Yeah, I agree. I think there's like this level of taste that is in there, especially on top of like story creation. If you're having these like role playing games, I think I was building out like Poki tax was just like my spin on filing a tax form and like Pokemon games and, um, just like even the dialogue, the first pass of the iteration, I was like, well, it's just like, you know, the stereotypical like, let me battle you, don't go, you know, to the next level. Uh, but I'm like, wait, what if, you know, and it requires, I feel like some level of thought and, um, guidance to AI. Like what if, you know, you never filed any of your taxes in Pokemon? Because you're always earning like money as you battle like trainers and stuff like there's that whole system.

• 12:38Speaker 00

  Um, and so it's like, you really have to provide some level of guidance in some cases.

• 12:43Speaker 01

  Speaking more about guidance, like, can you point, uh, you know, and talk about a prototype and point to the moment where your input made most of the difference, like where you feel that that guidance comes into play?

• 12:52Speaker 00

  I think for B, I spend a lot of time and I I love just kind of the visuals of a game. I think it just makes a huge difference of what is like production grade quality. I spent a lot of time figuring out how to like make the assets what I want. Um, I threw AI at Blender, just like, just try it, you know, just go. It did not get what I want. Um, it was quite like an approximate version of like the trees, the setting of just, just like this kind of like serene space that I wanted. And I find that for me was like, I enjoy just like drawing out like particularly the assets that I would want. And so I just use paper and pencil.

• 13:32Speaker 00

  I would just draw it out, give it to GPT, have it run an image and be like, yep, that's it. And iterate on the prompt with, you know, playwright. I would then basically take that, create a character sheet. Um, and so having that available just made it so much easier to go from like 2D to 3D. So either if that was like throwing it into Blender or if it was throwing it, I think to it's like image to 3JS was one of the things I used a lot. Um, it was just so much faster to iterate and get to like where I want it to be.

• 14:03Speaker 02

  Two, um, follow ups, a tiny one, but I'm very curious, which is why hand draw before putting in? Why not like voice prompting an image generator and then going cycles on that? And then putting it in? I'm curious like what the learning is for the, the, the kind of the creativity workflow. So that's the small one first and then a, a second one in a moment.

• 14:29Speaker 00

  I think for me, it's like, I really, I have a particular vision of what I want the assets to be and like how they look. I think AI in general, it is just very hard for me and maybe this is a skill issue, but it's like, I, I struggle to like get to where I want. There's always these small little things and little details that AI misses, even if I speak to it. And so I find that, you know, in general, with just providing like reference images or any kind of like approximate of like what you need, just gets you to just one shot and like closer to the step. Um, instead of having to be like this, like bow tie is off, it's not this color or, you know, in this manner. I would want it in this style.

• 15:13Speaker 00

  So it's just like, there's so many things you can probably minimize if you had like coming in with a vision. But if you probably didn't, I would say I typically would do the same thing that you would mention, which is just continuously iterate with AI.

• 15:26Speaker 02

  Yeah, well, given that some of us like me don't have any, you know, good drawing skills. That's my only path forward, but I was curious what they, what they, what the alternative is. Now, the other thing is a little bit like there's, there's, there's kind of in games is kind of the theory of fun, which you talked about before and kind of gestured at. There's also the kind of the richness of the world, which then parallels the physical intelligence. So what have you found in kind of, as it were, the, the world building part of it? Where both amazing capabilities of AI have happened and also still yet TBD challenge elements? Cause that'll be part of the lens in terms of getting into the physical world.

• 16:10Speaker 00

  Yeah, and I think that's a really interesting question. I think for me, what I see is like the challenge is like the level of granularity. Like how much can you drill down to like each individual asset? It's just like, I think it's like incredibly easy just to throw AI at like Blender or, you know, at these worlds and be like, just generate me something. But the asset barely moves, you know, like how granular can you get? And so it's like, if you have like, uh, like a packet of tea, there, you know, there's the packet itself. Um, there's like the tea bags and then inside as well as just like the tea. So it's like, how much can you drill down? And I think it's very fascinating to see now AI being thrown at these problems be like, just go at it

• 16:53Speaker 00

  continuously on a loop, iterating again and again at the level of granularity to reach where everything can be represented fully within these worlds. Um, I think like that's like the gap that I see.

• 17:05Speaker 01

  More, more detail, making more hyper realistic, more detail. Three hours later, I'm like, oh, this is not bad. And then I'm trying to put in the game, it's like, remove the detail so it renders properly. You speak a little bit about your work and the tools you're using and Blender, especially the latest models. It's been so much fun seeing AI like use your computer, use the tools on our computer, uh, use Blender, building 3D assets. What are some tools that people may not know about that you think could deeply impact their work? Maybe some of your like, maybe like top, top three, top five, maybe.

• 17:38Speaker 00

  I think for me, what I've been playing around with a lot these days is, I mean, I love pixel art games. I think these are kind of like childhood, like Octopath Traveler and stuff like that. Um, and so I've been playing around with like Pixel Lab and just like plugging that into Aspiride and seeing what kind of, you know, characters I can generate and just like the varying quality. I think this has been a struggle for me and previously using things like, you know, GPT just to generate pixel art assets. Um, I think now it's gotten like overall across the board, a lot better, um, on all fronts, like not just GPT, but Pixel, uh, Pixel Labs is just fantastic to use, um, and definitely worth a shot.

• 18:22Speaker 00

  I think it's also really fun to just kind of like see different spaces and try them. And so I typically always love to just use the world models in general, like world labs. And just seeing what I'm possibly able to create and using those as like even reference images for the 3D worlds and getting to those like particular assets that I would like.

• 18:43Speaker 02

  Let's also dive into one of the things I think is particularly important part of what we're doing in all these, these kind of creativities, which is there's an amazing amount of capability that AI adds. There's actually some, uh, amazing creativity that AI adds, you know, all forms of amplification intelligence. But like also one of the things that we're learning on this journey, and I think it's important is kind of see all the, the, the AI cast forward. But also where are the elements of kind of human?

• 19:13Speaker 01

  Contribution for amazing, you know, kind of like results. So, and this kind of gets to kind of questions of taste, questions of judgment. What have you been finding that has been the most important ads from you? And from people in the kind of taste and judgment? We got a little bit of like Parth, like I gotta play the game to kind of get the theory of fun. But what, like, how would you elaborate more of that?

• 19:40Speaker 00

  I think for me, it's definitely related to like, style. And I think a lot of the game actually is figuring out just kind of the visual effects. Like, what is like the art that you want to bring into the game, into people's hands? I think art style is just a huge part of like, what makes things fun. Like, of course, there's that on top of the fact that I still think like AI for me really struggles to hit exactly where you need to be. Like, I've tried having it rig a system and kind of like have that as like an asset within my game. And it still hasn't been able to do that properly.

• 20:18Speaker 00

  And maybe there's possibly some other products out there that might be doing it way better than what just having like Fable or like Opus run at it. But I definitely think it's a lot about it, it's like the fine tuning. It's like, you don't really need to do the things of like, you know, reskining an asset, just like kind of things that takes up so much time. But rather focus on like, what is like the highest leverage points to get things to be fun? And to me, that is just like the small individual details and of those assets. Just like, you know, the style consistency, is this like accurately represented in the game? And does this make sense for a character to be kind of like enjoyable and fun for people?

• 20:58Speaker 01

  Very cool, yeah. I kind of like, kind of, I kind of want to take it back to your theory on robotics. When I was in high school, we saw the first prototypes in Silicon Valley of like self-driving cars, kind of driving around the parking lot. And my dad was like, oh, this is gonna be the future, this is gonna be the future. And then it took like, it took like 15 years, like it took a long time for them to, and now I now I take Waymos, I take Waymo everywhere. So it's finally here. I kind of put off buying a car until the car could drive itself. Do you feel that with, I mean, especially in robotics, it seems like a lot of people work in robotics, you work in robotics. I kind of have this like skepticism of like, some people like, oh, it'll be here next year. I'm like, I don't know, I don't know if I trust a robot to, like, I have robots in my house, very minor, like, very like specialized robots for like handling, you know, taking care of the cats.

• 21:43Speaker 01

  But this is like minor levels of automation. I'm not sure I would, like, how comfortable I feel with like a robot walking around my house, you know, like, just, just don't hurt the cats, you know? Like, that's a simple, like, expectation I have. How do you think about how long this will take to, this like the robotics wave will take to unfold? And how do you think we get the best version of this robotics future? What, like, how do you think about like where this goes? And how do we make the best possible future come to fruition?

• 22:11Speaker 00

  Yeah, this is a really good question. I think definitely it seems like there is still quite a bit of, you know, iteration that needs to be done. I think the US in general for me seems to be much more focused on kind of a research and development phase rather than a deployment phase. We really haven't seen robots kind of like operate in the real world yet. And we haven't really gotten the flywheel of like robots getting their own data and then training on that. It's been, I think, a very small subset of robots that have been able to like successfully go out in spaces. Because no one really wants like this, like, you know, 100-pound robot in their house and, like you said, fall on their cat. That would just be catastrophic, right?

• 22:55Speaker 00

  Or like trip and like go out of the window and just, you know, harm someone. So I think definitely, we haven't really seen robotics kind of like, in general, the data. It's like LLMs had their moment with kind of just the text available on the internet. And so like, where do we really see that being with robotics? And there's this like question of like diversity and what it means to, you know, really be in these 3D spaces? And like basically train AI to understand physics, 3D worlds, that has yet to kind of been solved. And so I think a lot of it is trying to discover like what that means.

• 23:34Speaker 00

  You know, there's YouTube data that is possible, you know, there's also world models and also like egocentric data sets. But yet, like, is that enough? It's like still remains to be the question.

• 23:45Speaker 01

  Ego-centric, is that like the first person?

• 23:48Speaker 00

  Like human, like, yeah, iPhone data.

• 23:50Speaker 01

  So it's like first-person POV, like, doing a task or like exploring like that, that kind of data set?

• 23:55Speaker 00

  Exactly, it's just like, you know, looking at your hands, just basically those people who used to, you know, like film themselves, like making Subway sandwiches. Like similar, very similar to that. Yeah, it's very interesting to see. And I think there's this level of granularity that has not yet been reached, which is, you know, what is sensor and tactile like related? Do we need these in kind of like these world models and data sets? Like, we're not able to take a YouTube video and be like, yeah, this guy applied X amount of force to grab, you know, this edge of a paper and fold it. So I think like that's the limitation and it's just very interesting to see kind of like where we'll be.

• 24:36Speaker 00

  I know it's a very roundabout answer to say that I think robotics still need some time to like really truly take off. I would say definitely we'll see a lot more robotics in commercial environments in like the next, like, you know, three to five years. I would say that residential is much more of a, like a scarier place to deploy a robot. But we'll see if, I know everyone loves humanoids. I think humanoids are not the best form factor.

• 25:03Speaker 01

  I was gonna ask this question, actually. What, yes, like, like, what is, what is it about the humanoid form? And then also like, there are plenty of things we need robots to do that would require a shape or abilities that I'm not sure humanoid-sized robots can't have. What are your thoughts on that?

• 25:18Speaker 00

  Yeah, because I think about it in the way it's like, do you really need someone with hands to, you know, grab boxes? Or could you just really, you know, have like this Roomba grow under a pallet, pick it up? And then just move like hundreds of boxes, you know, at once? It's just really like, why would you need human hands? I think humans in general, it's extremely cool to see, you know, humanoid robots exist. And possibly even solve like the issue of like cross embodiment of like data where a lot of the sets or data sets that are coming out are like egocentric. It's your human hands. But how does that translate over to, you know, robots that are using grippers?

• 26:01Speaker 00

  So it's not like quite like one-to-one. And so that's why, you know, there's Umi grippers, there's all these other things. And so I think there's a lot yet to be discovered. And maybe it's just a play on, you know, humanoids are kind of like the flashy thing. On top of it being way easier to kind of use this existing data set that's on the internet. On top of the fact that I think, you know, it's cool to have like a human robot, just like in your own home. I think like Sunday Robotics has like one on like a Roomba and just like a very cute little like robot. I don't think it's quite hands, I think it's more of a gripper type. But it's definitely like worth playing around with various form factors.

• 26:46Speaker 00

  I think humanoids are just kind of one of many robots that should be explored.

• 26:52Speaker 01

  Yeah, I, I recently got a tour of, my friend works at a robotics startup called Gray Matter Robotics. And I visited their, their factory floor. They're basically building like factory robots. And I noticed, and I was looking around and I noticed that like, if you're trying to build a ship or if you're trying to build a fire truck, all the robots that they have are like different sized arms, but on a treadmill. And they're like, and in their case, they have to focus on getting the robot to understand the right amount of pressure, the materials it's working with, the like material science of that, the chemistry of the environment. And, and then I look around, it's like one guy sitting behind a desk and then like 15 robots on a factory floor that are like assembling a fire truck or like a, or a giant naval ship.

• 27:35Speaker 01

  And, yeah, it made me think, I was like, huh, like, like, maybe, maybe a lot of the most like interesting, like most of the value of robotics might come from shapes that are not like two legs, one, you know, two hands and a, and a face. But it made me think about it, yeah.

• 27:51Speaker 00

  Yeah, because I think that maybe the most laborious tasks like shouldn't be just like a human form factor. If, you know, it should be something that can, you know, exponentially just basically help produce stuff. And so it's, I think it's a question that I think people would need to answer, especially in, I think, like more commercial environments of like what really makes sense on the floor.

• 28:12Speaker 01

  Awesome, well, this has been super exciting. I feel like I'm learning so much every time I talk to you. But thanks for joining us today, Katie. And I guess before we leave, do you have a question for either Reid or me to answer?

• 28:25Speaker 00

  Yes, I would just love to like open the floor in general. But I think for me, like I've been hinting at, you know, in general, that language models have been learning like just a staggering amount of, you know, knowledge from just the internet, from text. But there's also like, I think it's like called Moravix paradox where machines, you know, they can write poetry, but they can't really understand these physical spaces. And so I'm very curious, like, if so, do you find like? What's like the most important thing about the world that you think that can't be learned from the internet?

• 29:00Speaker 02

  Well, I think one of the things, it's a great question. There was a little bit of kind of very quick framing. So one, part of how I look at these models is that there are alien intelligences that are, that are highly learned and trained to be human intelligences is part of the reason why you get a jagged edge. And that's true even with the, the trove of knowledge in large language models. And it's part of the reason why I think, you know, part of the comparison and knowledge representation is the jagged edge. Some things superhuman, some things equivalent to human, and some things really kind of contextually, you know, weak. And that's similar to also how they lack context awareness. Now, when I think when you get to the 3D world, there's kind of, you can break it into

• 29:45Speaker 02

  three kind of categories. So one is like very structured environments, factory floors, et cetera. Those ones are generally speaking relatively easy to align that jagged edge well. And not easy, like still a lot of work to do, but it's like, get it straightforward. But then as you begin to get out, you go to, and this is part of what, you know, like Parth with the autonomous vehicles and driving, as you get to a general environment, and you got a lot of unknowns and uncertainties. And so the kind of the question around like, like we kind of move to a hyper alert thing when something breaks our, our normal parameters, like, like you see a little bouncing ball come out as you're driving, you go, wait, is a kid gonna follow the bouncing ball?

• 30:30Speaker 02

  Or is a pet gonna follow the bounce? And you go to a hyper alert thing. And that's like the kind of thing that most of these models on a lot of specific learning and training, specific data, you were mentioning YouTube, et cetera, get to. And that's, I think, a well-understood gap. But well understood, well understood that it is a gap, not necessarily a well-understood gap in terms of, so that's one. But here's one that I think most people don't really get to. So you get to the, okay, structured environment, general environment. And now, when you get to the 3D world, you also get to kind of call it a competitive environment with other agents. And one of the ways that I've illustrated this in the autonomous world circumstance

• 31:10Speaker 02

  is, hey, we get these, all of these, these agents driving cars well, that's fine when they're in a not like a world without hostile counteragents. Like, people can figure out how to fuck that up pretty easily, you know? And so what happens when you start having competitive agents in this environment too? And that's yet another, even another thing. And you see that a little bit with, you know, even cybersecurity online, whatnot, but it's kind of like the question of what happens when you have competitive agents also in the 3D world? You know, things that don't necessarily align with the outcome that your design intends

• 31:55Speaker 02

  to be. And I think those are all areas of information that is still very much TBD and not from pure large language models or data on the internet.

• 32:07Speaker 00

  Yeah, that's very interesting. It reminds me of when, you know, people let AI agents go rogue in Minecraft of just like spawning, you know, hundreds of them and seeing what would happen. I think like that's something yet to like, I have seen not as much like progress like after that of, you know, how can we really let agents run loose within these environments?

• 32:27Speaker 02

  Exactly. Well, Katie, a pleasure and an honor. And, you know, Parth, always good to see the real you versus the AI image.

• 32:37Speaker 01

  Oh, my God. I'm so glad we to have you on, Katie. Thank you so much. This has been, this has been incredible.

• 32:43Speaker 00

  Yeah, no, I appreciate it. This token grant has been really exciting, you know, just to be able to create things and just really figure out, you know, what I want to build and just really build these things and make it to come to life. So thank you guys so much.

• 32:57Speaker 02

  Possible is produced by Palette Media. It's hosted by Ari Finger and me, Reid Hoffman. Our showrunner is Sean Young. Possible is produced by Tanasi Delos, Katie Sanders, Spencer Strasmore, Imo Zoo, Aman Suri, Danny Garrison, Trent Barboza, and Tafadzwa Nima Rundwe.

• 33:14Speaker 03

  Special thanks to Surya, Yala Manchili, Sayida Sabieva, Ian Alice, Greg Beato, Parth Patel, and Ben Ralas.