This transcript is generated with the help of AI and is lightly edited for clarity.
//
CADI:
We really haven’t seen robots operate in the real world yet—getting their own data and then training on that. It’s a very small subset of robots that have been able to successfully go out in spaces, because no one really wants a hundred-pound robot in their house that, like you said, falls on their cat. That would just be catastrophic, right?
I was building out PokéTax—just my spin on filing a tax form, like Pokémon games. And it requires some level of thought and guidance. Like, what if you never filed any of your taxes in Pokémon because you’re always earning money as you battle trainers? So you really have to provide some level of guidance.
//
PARTH:
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.
REID:
Welcome to another episode of Reid Riffs. We’ve got our real Parth, not just Parth AI—and it’s also the real Reid, not just Reid AI. And we have an excellent guest. Cadi, thank you for joining us.
CADI:
Yeah, I’m super excited to be on. Thank you guys so much for having me.
REID:
It’s our pleasure and honor. And why don’t we start with—okay, so Parth, how did you select Cadi?
PARTH:
Yeah, I’m super excited to have Cadi on. I think we’ve known each other now for maybe three or four years. When I met Cadi, she was working on a hackathon, trying to build a kind of copilot for game design and game development. And me being obsessed with video games, I was like, oh my God, I want this. This would be awesome. I saw what Cursor did for me for programming—but imagine if we had something similar for making games. So that’s when I met Cadi. And over the last couple 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, and now she works in robotics as a product operations specialist. She has this incredible exploratory career arc that’s been really exciting to follow—seeing how she connects the dots across all these different problem spaces, using the AIs to fill in our gaps. So I figured it would be awesome to accelerate your own creative projects and see what you can do, and bring you onto the show to see what you’ve been learning over the last few years. So welcome, Cadi.
Tell me more about how you got into AI, what your initial goals were, and where you ended up—talk about your journey to this point over the last few years.
CADI:
So overall I had a pretty unconventional path into AI. I was working on creating games in Unity—I was just super obsessed, really wanted to create something. I was working on 2D platformers and such, and I was really lucky at the time with robotics just being there, and jumped on, because we were basically programming VR interfaces with robotic dogs. It was kind of my foray into actually starting to program and learn. And at the time we were just getting introduced to GPT, having that as the magical thing where you’re literally copying and pasting code snippets in. So it’s hilarious to see where we are now.
But that really spiraled into me doing haptic suits, coding more VR stuff in Unity, to now working in AI accounting for a bit, and then moving over to more of the data-collection, robotics portion of: how can we really gather the data, and what is the data that’s needed for the next generation of embodied intelligence, or robotics in general?
REID:
Very cool. So go a little bit into 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?
CADI:
They have a lot more crossover than I initially expected—I was quite surprised. The easiest way to think about it is teleoperation, for example. You’re basically in these VR headsets in a game world, operating these robots. And on a software level, you get to see the advent of these world models coming into play—being able to simulate and recreate these environments and have robots reasoning through them. I think the main bridge being crossed is: how can we simulate 3D worlds, physical realities, and have physics within them?
And that’s the bridge that’s being crossed. World models and robotics are taking this to the next level of ambition—really going end to end in many ways.
REID:
By the way, quick follow-up. What are the biggest gaps between them so far? Because it’s the classic 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?
CADI:
I think this was really able to be represented when Genie came out, and it was just a bloodbath of the stocks of Unity, Roblox, all these gaming companies. They provide a really great take on what the gaps currently are in these world models—where physics and world models are not yet deterministic. There’s also no object permanence in some cases; it’s only short horizons where you have the ability to see and interact with the world.
So with these in mind, it’s like: how do you create these longer-form interactive experiences in games, other than just a quick first-person shooter or a quick tycoon game? On top of that, if you don’t have determinism, how can you create multiplayer games? That’s something you’re not yet able to achieve. So I find that’s the gap, in some sense. But I also find there’s an opportunity to bridge it—for example, a lot of people are talking about, why don’t you use world models as an input into these more deterministic systems to create these games? So there’s that path, versus the more implicit path of pure straight AI creating these hyper-realistic games.
PARTH:
Yeah. What clued me into this paradigm was a demonstration of Minecraft—I think it was running on Oasis, maybe on the Etched chip. So it was more of a demonstration of the chip, but it was a Minecraft generator trained on Minecraft. And it had the memory of a fish. Anything that went outside of the frame would be forgotten. So if you walked up to something and then turned around, everything would change, because it wasn’t able to remember and hold on to the details of the world beyond what it could see in front of it. What is Genie—could you clarify that? How would you describe it to someone that maybe didn’t get to play with it or see it yet?
CADI:
I think it’s basically: you take a prompt and you’re able to see and create all these different worlds and interactive environments. There was Genie 1, which was the very beginning, where you’re able to see the introduction into an interactive world; Genie 2; and Genie 3 now, which is the next level of interactivity—the possibility of having the world at your fingertips with just a prompt.
PARTH:
Right. Like, the world unfolds as you explore it on the fly, instead of being predetermined.
CADI:
Right, yes.
PARTH:
What’s a recent project where AI let you move from an idea to a working prototype faster than you could have before?
CADI:
I’ve been playing a lot with how AI and games overlap. Prior to this, I was struggling a lot to figure 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 this horrible HTML/CSS file. It just broke; I could not get it to work. And now I feel extremely supercharged, being able to just create these games. One of the projects I was working on is basically a spin on the imposter game, where everyone’s a duck and there’s one killer duck in the mix.
Your goal is to act like an NPC and complete a set of tasks. So that duck is trying to monitor which character is actually the human. And it’s been amazing to have AI help me iterate on so many core gameplay loops. It’s amazing to go to bed, be in the shower, and have AI spin up five gameplay loops, send them to friends, take that feedback, and just rip at it again. The core thing is: how can we make the game fun? That requires a lot of time, effort, and research—and being able to minimize that and get it into the hands of people is incredibly powerful.
REID:
I’m curious what you’ve learned so far about how much AI understands the concept of fun. I’m expecting you to say uneven—but, having just made my token prediction of uneven, what’s the actual answer?
CADI:
I would definitely agree. I think AI can approximate—based on the games it’s probably seen, played, and been trained on. If you give it examples of the games you enjoy and want, it’ll approximate exactly what the core gameplay mechanics are and give you suggestions. I’ve noticed that if it’s something fairly new, it struggles; it requires a lot of context building still. So I wouldn’t say AI has a good grasp on fun yet. It still requires a lot of time and effort, really honing what the gameplay mechanics are. It’s just faster in the way that it provides you the code to see it in real time.
PARTH:
Yeah, I see the same thing. Most of what I have to do is play the game and then point out things. My job is to put the game in the hands of other people, then play the game and be like, okay, what is anti-fun? And then I go to the AI and I’m like, this experience was very frustrating, we should fix it. And the AI comes up with six alternatives to patch that experience, and I’m like, oh, these are pretty good ideas, we can try them out. But definitely it doesn’t understand fun like we understand fun.
And that’s the very human level of being in the game and playing and trying to have your own experience. Which is exciting, because I hope we’re not automating the part of experiencing the game—although it is cool to have AI play the game before you do, highlighting and finding errors. You kind of have it bug-fixing and mapping things out.
CADI:
Yeah, I agree. There’s this level of taste in there, especially on top of story creation—if you’re doing role-playing games. I was building out PokéTax, just my spin on filing a tax form, like Pokémon games. Even the dialogue: the first pass of the iteration, I was like, well, it’s the stereotypical, let me battle, you don’t go to the next level. But I’m like, wait—what if? It requires some level of thought and guidance to the AI. Like, what if you never filed any of your taxes in Pokémon because you’re always earning money as you battle trainers? There’s that whole system. So you really have to provide some level of guidance in some cases.
PARTH:
Speaking more about guidance—can you talk about a prototype and point to the moment where your input made most of the difference, where that guidance comes into play?
CADI:
I spend a lot of time on—and I love—the visuals of a game. It makes a huge difference: what is production-grade quality. I spent a lot of time figuring out how to make the assets what I want. I threw AI at Blender—just try it, just go—and it did not get what I want. It was quite an approximate version of the trees, the setting, just this serene space that I wanted. And I find that I enjoy drawing out particularly the assets that I would want. So I just used paper and pencil. I’d draw it out, give it to GPT, have it run an image, and be like, yep, that’s it—and iterate on the prompt. I’d then take that and create a character sheet. Having that available made it so much easier to go from 2D to 3D—whether that was throwing it into Blender or, I think, image to three.js was one of the things I used a lot. It was just so much faster to iterate and get to where I wanted to be.
REID:
Two follow-ups. A tiny one, but I’m very curious: why hand-draw before putting it in? Why not voice-prompt an image generator, go cycles on that, and then put it in? I’m curious what the learning is for the creativity workflow. So that’s the small one first, and then a second one in a moment.
CADI:
For me, I have a particular vision of what I want the assets to be and how they look. AI in general—this is very hard for me, and maybe this is a skill issue—but I struggle to get to where I want. There are always these small little details that it misses, even if I speak to it. So I find that providing reference images, or any kind of approximate of what you need, just gets you closer to one-shotting it, instead of having to be like, this bow tie is off, it’s not this color, I’d want it in this style. So there’s so much you can minimize if you come in with a vision. But if you didn’t, I would typically do the same thing you mentioned—just continuously iterate with AI.
REID:
Well, given that some of us—like me—don’t have any good drawing skills, that’s my only path forward. But I was curious what the alternative is. Now, the other thing: in games, there’s the theory of fun, which you talked about and gestured at, and there’s also the richness of the world, which parallels physical intelligence. So what have you found in the world-building part—where both amazing capabilities of AI have happened, and also still-TBD challenge elements? Because that’ll be part of the lens in terms of getting into the physical world.
CADI:
That’s a really interesting question. For me, the challenge is the level of granularity—how much can you drill down to each individual asset. It’s incredibly easy to throw AI at Blender, or at these worlds, and be like, just generate me something. But the asset barely moves. How granular can you get? If you have a packet of tea, there’s the packet itself, there’s the tea bags, and then there’s the tea inside. So how much can you drill down? It’s fascinating to see AI being thrown at these problems—just go at it continuously on a loop, iterating again and again at the level of granularity, to reach where everything can be represented fully within these worlds. That’s the gap I see more.
PARTH:
More detail, make it more hyper-realistic, more detail—three hours later, I’m like, oh, this is not bad. And then I’m trying to put it in the game, and it’s like, remove the detail so it renders properly. Speak a little bit about your work and the tools you’re using—Blender especially, the latest models. It’s been so much fun seeing AI use your computer, use the tools on our computer, use Blender, building 3D assets. What are some tools that people may not know about that you think could deeply impact their work? Maybe top three, top five.
CADI:
For me, what I’ve been playing around with a lot these days—I love pixel art games; these are kind of childhood, like Octopath Traveler and stuff like that. So I’ve been playing around with PixelLab, plugging that into Aseprite, and seeing what characters I can generate, and the varying quality. This has been a struggle for me previously, using things like GPT just to generate pixel art assets. Now it’s gotten a lot better across the board on all fronts—not just GPT, but PixelLab is just fantastic to use, and definitely worth a shot. It’s also really fun to see different spaces and try them. And I always love to use the world models in general, like World Labs, and see what I’m able to create, using those as reference images for the 3D worlds and getting to those particular assets I’d like.
REID:
Let’s also dive into one of the things I think is a particularly important part of what we’re doing in all these creativities—there’s an amazing amount of capability that AI adds, some amazing creativity, all forms of amplification intelligence. But one of the things we’re learning on this journey is where the elements of human contribution are for amazing results. This gets to questions of taste, questions of judgment. What have you been finding is the most important add from you, and from people, in the taste and judgment? We got a little bit of it—Parth, like, I gotta play the game to get the theory of fun—but how would you elaborate on that?
CADI:
For me it’s definitely related to style. A lot of the game is figuring out the visual effects—what’s the art that you want to bring into people’s hands. Art style is a huge part of what makes things fun. On top of that, I still think AI really struggles to hit exactly where you need to be. I’ve tried having it rig a system and have that as an asset within my game, and it still hasn’t been able to do that properly. Maybe there are other products out there doing it better than just having Fable or Opus run at it. But it’s a lot about the fine-tuning. You don’t really need to do the things like reskinning an asset—things that take up so much time—but rather focus on the highest-leverage points to get things to be fun. And to me, that’s the small individual details of those assets: the style consistency, is this accurately represented in the game, and does this make sense for a character to be enjoyable and fun for people?
PARTH:
Very cool. 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 self-driving cars driving around the parking lot, and my dad was like, oh, this is going to be the future. And then it took like 15 years—a long time. And now I take Waymo everywhere. I kind of put off buying a car until the car could drive itself. Do you feel that with robotics? It seems like a lot of people work in robotics, you work in robotics—and I have this skepticism of, like, some people say, oh, it’ll be here next year, and I’m like, I don’t know.
I have robots in my house—very minor, very specialized robots for taking care of the cats. There are minor levels of automation. But I’m not sure how comfortable I’d feel with a robot walking around my house. Like, just don’t hurt the cats—that’s a simple expectation I have. How do you think about how long the robotics wave will take to unfold, and how do we get the best version of this robotics future? How do we make the best possible future come to fruition?
CADI:
Yeah, this is a really good question. It seems like there’s still quite a bit of iteration that needs to be done. The US in general seems to be much more focused on a research-and-development phase rather than a deployment phase. We really haven’t seen robots operate in the real world yet, and we haven’t gotten the flywheel of robots getting their own data and then training on that. It’s been a very small subset of robots that have successfully gone out in spaces—because no one really wants a hundred-pound robot in their house that, like you said, falls on their cat. That would be catastrophic, right? Or trips and goes out the window and harms someone.
So we haven’t really seen robotics in general—the data. It’s like LLMs had their moment with the text available on the internet. So where do we see that with robotics? There’s this question of diversity and what it means to really be in these 3D spaces and train AI to understand physics and 3D worlds. That has yet to be solved, so a lot of it is trying to discover what that means. There’s YouTube data that’s possible; there are world models, and also egocentric data sets. But is that enough? That still remains the question.
PARTH:
On egocentric—is that like the first-person—
CADI:
Yeah. iPhone data. Yeah.
PARTH:
So it’s first-person POV, doing a task or exploring—that kind of data set?
CADI:
Exactly. Looking at your hands—basically those people who used to film themselves making Subway sandwiches, very similar to that. It’s very interesting to see. And there’s this level of granularity that hasn’t yet been reached—what’s sensor- and tactile-related. Do we need these in these world models and data sets? We’re not able to take a YouTube video and be like, yeah, this guy applied X amount of force to grab this edge of a paper and fold it. So that’s the limitation. It’s very interesting to see where we’ll be. I know it’s a roundabout answer, but robotics still needs some time to really truly take off. We’ll see a lot more robotics in commercial environments in the next three to five years. Residential is a much scarier place to deploy a robot. But we’ll see. And I know everyone loves humanoids—I think humanoids are not the best form factor.
PARTH:
I was going to ask this question. What is it about the humanoid form? And also—there are plenty of things we need robots to do that would require shapes or abilities that I’m not sure humanoid-sized robots can have. What are your thoughts on that?
CADI:
Yeah. I think about it like: do you really need someone with hands to grab boxes? Or could you have this Roomba go under a pallet, pick it up, and move hundreds of boxes at once? Why would you need human hands? It’s extremely cool to see humanoid robots exist, and they could possibly even solve the issue of cross-embodiment of data—a lot of the data sets coming out are egocentric, your human hands. But how does that translate to robots using grippers? It’s not quite one-to-one. That’s why there are UMI grippers and all these other things.
So there’s a lot yet to be discovered. Maybe it’s a play on humanoids being the flashy thing, on top of it being way easier to use the existing data set that’s on the internet, on top of the fact that it’s cool to have a human robot in your own home. I think Sunday Robotics has one on a Roomba—a very cute little robot. I don’t think it’s quite hands; I think it’s more of a gripper type. But it’s definitely worth playing around with various form factors. I think humanoids are just one of many robots that should be explored.
PARTH:
I recently got a tour—my friend works at a robotics startup called GrayMatter Robotics, and I visited their factory floor. They’re basically building factory robots. And I noticed, looking around, that if you’re trying to build a ship or a fire truck, all the robots they have are different-sized arms on a treadmill. 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 materials science, the chemistry of the environment. And then I look around, and it’s one guy sitting behind a desk, and like 15 robots on a factory floor assembling a fire truck or a giant naval ship. And it made me think—huh, maybe a lot of the most interesting, most of the value of robotics might come from shapes that are not two legs, two hands, and a face. It made me think about it.
CADI:
Yeah. The most laborious tasks shouldn’t be in a human form factor—it should be something that can exponentially help produce stuff. So it’s a question people would need to answer, especially in more commercial environments: what really makes sense on the floor.
PARTH:
Awesome. Well, this has been super exciting—I feel like I’m learning so much every time I talk to you. Thanks for joining us today, Cadi. And before we leave, do you have a question for either Reid or me to answer?
CADI:
Yes. I’d love to open the floor. I’ve been hinting at this—language models have been learning a staggering amount of knowledge from just the internet, from text. But there’s also—I think it’s called Moravec’s paradox—where machines can write poetry but can’t really understand these physical spaces. So I’m very curious: what’s the most important thing about the world that you think can’t be learned from the internet?
REID:
Well, it’s a great question. A little bit of quick framing: one part of how I look at these models is that they’re alien intelligences that are highly learned and trained to be human intelligences. That’s part of the reason you get a jagged edge—even with the trove of knowledge in large language models. Some things superhuman, some things equivalent to human, and some things contextually weak. That’s similar to how they lack context awareness.
Now, when you get to the 3D world, you can break it into three categories. One is very structured environments—factory floors, et cetera. Those are generally relatively easy to align that jagged edge well. Still a lot of work to do, but it’s fairly straightforward. But as you get out—and this is part of what Parth was saying with autonomous vehicles and driving—you get to a general environment, and you’ve got a lot of unknowns and uncertainties. We move to a hyper-alert state when something breaks our normal parameters. Like, you see a little bouncing ball come out as you’re driving, and you go: wait, is a kid going to follow the ball? Is a pet going to follow it? You go to a hyper-alert state. That’s the kind of thing most of these models get to, on a lot of specific learning and training—the YouTube data you were mentioning. And that’s a well-understood gap. Well understood that it is a gap, not necessarily a well-understood gap.
So that’s one. But here’s one most people don’t get to. You get to the structured environment, the general environment—and now, in the 3D world, you also get to a competitive environment with other agents. One of the ways I’ve illustrated this in the autonomous world: we get all these agents driving cars—well, that’s fine when they’re in a world without hostile counter-agents. People can figure out how to mess that up pretty easily. So what happens when you start having competitive agents in this environment too? That’s yet another thing. You see it a little bit with cybersecurity online. It’s the question of what happens when you have competitive agents in the 3D world—things that don’t necessarily align with the outcome your design intends. Those are all areas of information that are still very much TBD, and not from pure large language models or data on the internet.
CADI:
Yeah, that’s very interesting. It reminds me of when people let AI agents go rogue in Minecraft—just spawning hundreds of them and seeing what would happen. I’ve seen not as much progress after that, on how we can really let agents run loose within these environments.
REID:
Exactly. Well, Cadi, a pleasure and an honor. And Parth—always good to see the real you versus the AI image.
PARTH:
Oh my God, I’m so glad we got to have you on, Cadi. Thank you so much. This has been incredible.
CADI:
Yeah, I appreciate it. This token grant has been really exciting—to be able to create things, figure out what I want to build, and really build these things and bring them to life. So thank you guys so much.
REID:
Possible is produced by Palette Media. It’s hosted by Aria Finger and me, Reid Hoffman. Our showrunner is Shaun Young. Possible is produced by Thanasi Dilos, Katie Sanders, Spencer Strasmore, Yimu Xiu, Aman Suri, Danny Garrison, Trent Barboza, and Tafadzwa Nemarundwe.
ARIA:
Special thanks to Surya Yalamanchili, Saida Sapieva, Ian Alas, Greg Beato, Parth Patil, and Ben Relles.