Prof. Saty Raghavachary (USC) on AI & Embodied Intelligence: The Path to AGI Isn't Code, It's Neurons

Host: welcome back to the 18VC podcast. This is our episode 23, and today we have Philip and also Corbin who will be co-hosting this episode. And we are so pleasured to have Professor Saty join us today. He's a professor from USC Viterbi's computer science department, and he has spent 16 years at DreamWorks building the graphics tools behind all these amazing animation films including "The Prince of Egypt" to "Madagascar 3." And today, his research was focused on embodied AI, and he's looking into the idea of how a machine can truly understand the world, and it has to experience with a physical body, instead of maybe only the datas and the symbols. And we are so excited to have you here today. And for our audiences who may not be familiar with us we are the 18VC podcast. We feature student and young founders, researchers, as well as investors who's backing them in the earliest stage. I'm Corbin, as mentioned. I'm a master student in computer science program here at USC. And the reason we have Professor Saty on today is I had the pleasure of being his student for two courses. It should've been three. I wanted to take you for web technologies, too.

Saty Raghavachary: Three-peat.

Host: Yeah, three-peat. But I had the pleasure of taking those courses with him, and he has such a wealth of knowledge and experience, so we wanted to bring him on today to share some of that unique perspectives. And he's got a great grasp on what's going on today with the technology as well as what's gone on historically. So-

Saty Raghavachary: Thank you for having me over.

Host: Yeah, absolutely. So, just to start, can you just walk us through your background? You went to engineering school in India. Actually, you have the most extensive academic history I've ever heard of anybody. If I were to try to read it in one breath, I would probably pass out here on the floor. So can you just walk us through your background to get to this point?

Saty Raghavachary: Yeah, for sure. I think of degrees as degrees in a thermometer, so I like to collect them like Pokémon cards. So one approach I had to collecting degrees, it just turned out that way, actually. Yeah, so my background is pretty much traditional material science metallurgical engineering which actually colors how I view AI these days, and I'll get to that afterwards. But I think the love for art was something I've had my whole life, even just drawing painting, watercolors, all of that. And so then when I got to IIT, that is my undergrad school in India I couldn't wait to get my hands on the big microprocessors, the IBM mainframe computers they had there. So even though the study was officially material science I always had this graphics streak. Then when I went to Ohio State for grad school, so then I was in the CS department and that is when... Actually, I was in the metallurgy department, got a PhD actually. And then I didn't want to leave just yet, so that is when I did additional degrees. One was in manufacturing engineering because to me, it tied things like CAD, CAM, robotics, 3D printing with material science. And then there was a year in biochemistry postdoc that was also very interesting. So that gave me some appreciation about biological molecules, life, all of that. So then I joined the CS department after that. And my research was actually fracture. My metallurgy research was also fracture, like what happens when brittle steel breaks. This was funded by the Army, by the way, so actually pure applied research. But my graphics thesis was also simulation of brittle fracture so in a way, more of the same but more computational. So then in the end, all this brought me to LA. This was '93. And I worked for a company called Metrolight Studios and that's one of the early pioneering CG effects houses in Hollywood, actually. But before that, for about one year, I worked on the world's largest AI project. It's called the Cyc project. Cyc stands for encyclopedia. And it was about symbolic reasoning. These days, we call it GoFi. And leaving that project also helped me think a lot about why the project failed actually, overall.

Host: Right.

Saty Raghavachary: And it still informs me, the way I look at AI. So in a way, they all fed into each other. So after Metrolight, then DreamWorks started in '96. '95, actually. So I joined less than one year after they started, and I was there for 16 happy, glorious years.

Host: Wow.

Saty Raghavachary: Worked on 19 animated movies. So there from the beginning, actually. And it was a ton of fun to use all these CS algorithms, principles every piece of knowledge, and even material science to make all these cool tools for the artists who can actually make things like dragon fire or the Kung Fu Panda temple explode at the very end or dust when Spirit runs in the hills, or many what we call effects or physical phenomena. So I had the best of both worlds. I could do art stuff that the art people wanted and dive into my engineering background. And then that, in turn, brought me to USC, which was almost accidental because I started teaching here part-time, and gradually it became so much fun that I made it full-time.

Host: Gotcha. So being at DreamWorks what were some of your fondest memories there, and what building tools for artists did they teach you that in academia they usually don't? How maybe does that differ?

Saty Raghavachary: Yeah, that's actually a great question because in academia, they hardly teach you the tools that artists actually want. I just happened to be lucky because on my own, outside my graduate work, so to speak, I started exploring the programs that the official CG world uses. RenderMan is one of them. So RenderMan is a rendering program that Pixar made from scratch, and they use it for every single Pixar movie, and so does every effects house in the world, every live-action visual effects in the movie that you see. So I started playing with it on my own. And then another tool called Maya, but it had some tools before it that went with it. So on my own, I learned how those tools work, and that in turn got me into DreamWorks because that is the kind of tools that they want.

Host: They were using and wanted.

Saty Raghavachary: Yeah. So to answer your question, I think we should do a better job of telling the students what academia does, like the principles, and also to. transition that forward to the industry. But in my case, I was lucky. I figured that out on my own, actually.

Host: Yeah. So academia doing a good job of understanding theory but not necessarily always pushing whatever the modern standard industry tools are for that.

Saty Raghavachary: You got it. That's actually why in my classes, you mentioned this earlier I spend almost half an hour in the beginning of each lecture to actually talk about the tools that the world uses, and I do that on purpose. Even the homeworks are designed so that you guys actually use those tools because I'm aware of that gap. And the gap is a pretty bad one in a way, meaning the academia stuff is wonderful theory, equations, principles, algorithms. But they're not what pays the bills for Google, and Amazon, and Microsoft.

Host: Right.

Saty Raghavachary: So I make it a point purposely to build that bridge so that students are more-- It's easier to transition.

Host: Yeah, it's so funny because anytime I meet students that have taken database systems or the information retrieval class, I'm like, "Oh, you had Saty as a teacher?" And a girl I met a couple of days ago, a really nice, smart person, she said, "Oh yeah, I had Saty." And she's like, "Oh, honestly, the best part of his classes were the." She called them tangents, but they're really not. They kind of all tie together with everything. But yeah I learned so much from the first 30 minutes of those lectures, because it was just like, "Hey, here's something in the news. Here's why it's relevant. This is kind of where things are going." So I very much appreciate the fact that you did that.

Saty Raghavachary: That's fantastic.

Host: You didn't always follow, like, just, "Oh, let's just do the lecture regardless of what happened and what you think is relevant to talk about." So.

Saty Raghavachary: So true. I couldn't agree with you more. Because I feel it's a disservice to students if somebody does that, because then it's a bubble, your course is an isolated little thing-

Host: Yes

Saty Raghavachary: no connections with anything. Then it's almost, what is the point? But instead, I strive to actually build those connections. I mean, yeah, they're tangents, but the tangents have a reason. They're not entirely irrelevant.

Host: Right. Yeah. And talking about moving from the industry you've been working at DreamWorks for 16 years, and I think you also start teaching other staff in the company about how some of the tools maybe you build or the company have, and then gradually you mentioned you started teaching part-time at USC.

Saty Raghavachary: Correct.

Host: And then you move on, take it as a full-time job. Just can you share a little bit more about that transition, both teaching in the company, DreamWorks, as well as teaching students? What do you see the differences are, or do they help you transitioning to this new role as a professor?

Saty Raghavachary: That's another good question, actually. So my last name, this is a funny thing to say, is Chary, C-H-A-R-Y. In Sanskrit, that actually means teacher.

Host: Oh. Oh, wow.

Saty Raghavachary: So in a way, you could actually say it's-

Host: You were destined from the start.

Saty Raghavachary: Correct. It's in my blood. But my mom was a school teacher, actually, her whole life.

Host: Wow.

Saty Raghavachary: My dad didn't teach. So maybe it runs in my blood, I'm not sure. But when I was at IIT, actually, it goes right back. The entrance exam is gosh darn hard. Less than 2% of people get admitted. So then I was already in an IIT, and I had a summer camp, actually for students the following year who were going to take that exam. So I coached them on what the exam looks like. That was a lot of fun. And I was barely 21 or something. It was an early teaching experience that I really enjoyed. Then, yeah, at DreamWorks, at one point, I became the head of the training department because many tools we build, you cannot buy them in the world. So an animator joins DreamWorks, they have to be brought up to speed. So yeah, I ran the department for a while. It was a lot of fun. The artists are very motivated because they already know the art. They know what to do in a different tool. All they need to be taught is translate their skills to the tool that DreamWorks built. So it was actually very easy. And one more thing, by the way, Gnomon. In Hollywood, there's a place called Gnomon School of Visual Effects. I actually taught there almost eight years, just again, part-time, only once every Friday night. But the whole audience, the classroom was filled with artists that want to be in the visual effects industry in Hollywood, right? They're not motivated to do programming at all, but I taught them a language called MEL that runs in a software called Maya. And that was fascinating because they don't like math, they don't like programming. But once you teach them, you show them what it does, like manually you put eight buttons on the shirt two inches apart one by one. And here's a loop that I can teach you. It does it in five seconds. Then they become motivated. "Wow, tell me more." So it's that kind of experience that I brought to USC eventually, where I make all the classes be very application-focused and so on. Yeah.

Host: Yeah and I'm from Marshall School of Business, and I did take some classes that's also taught by professors who's teaching part-time here, but their full-time jobs, for example, they are VC investors.

Saty Raghavachary: Oh.

Host: And I think these classes really bring in different perspectives, and I learned a lot. As you mentioned, there's a gap between academy and industry, and this type of professors, they do bring in very practical suggestions and practices that, yeah, I definitely learn a lot from them.

Saty Raghavachary: Couldn't agree more. We should almost do it on principle. Every department needs to hire people that actually are in industry.

Host: Yeah.

Saty Raghavachary: Who can bring back and tell the students what it really is about. Yeah.

Host: Yeah, because so many students, obviously they're here to get an education and all that, and it's an enriching experience, but a lot of them are looking for jobs afterward or some type of entrepreneurship thing. So rubber has to meet the road at some point with industry-

Saty Raghavachary: Correct

Host: of course.

Saty Raghavachary: And also, I think the students are a little bit, not naive, that's a bad term to use, but they're not knowledgeable in terms of what the industry wants. See, in a way, again, like I said, it's a disservice if the classroom is just this box that is entirely disconnected with everything else. You're not doing a good job of teaching is how I feel. So I feel it should be more the other way around, make those connections. Yeah, exactly.

Host: Yeah. All right. So I want to transition because obviously you have a large background in AI, and you've thought a lot about these types of topics. In plain terms, how would you define intelligence? And then also in the information retrieval course, I thought you did an amazing job talking about the transition from ChatGPT in 2022 to LLMs, not just using pre-trained data, but now using RAG, and then transitioning to agentic AI. Can you talk a little bit about that? I know it's a large question, but

Saty Raghavachary: Good question, really. Yeah, intelligence is such an overused term in a way and I think over many decades, people have written entire papers, collect definitions of intelligence. So I went through all of them actually, and I boiled it all down to actually two words. I call intelligence considered response. In other words, considered response is what I would define intelligence to be. So it's a response, first of all and it's not unconsidered. So if there's an earthquake, for example, an unconsidered response would be all the objects shake, and the bottle just falls on the floor and breaks. Because the bottle has no control over what is happening right? It just simply reacts to the ground shaking. But a considered response would be somebody saying, "Oh my God, I better get under the table. This room is going to collapse." So the world gives you information, and you process it, and then you respond. So I call that intelligence. So that's a pretty broad definition, which means even algorithmic intelligence is part of it. A chess-playing algorithm, you make a move, it considers many billion moves, and it picks the best one. So it's not doing something different. It computes basically the outcome in a way. But yeah, to me, it's always like a considered response. So that's that part of it. And yeah, LLMs have come a long way since when they basically were introduced to the world in just November 2022, not that long ago. In these four years, it's come a very long way, and the big reason is, again, this transition towards what we now call agentic AI where the LLM itself is still the same core LLM. It is still generating tokens, like one little piece at a time, and then it strings it all together, and you think it's a pretty long response. It doesn't actually put any thought, consideration, all of that. But after it generates a response, a good response, bad response, average response, it has a chance to go back and review the response as if somebody had given it the response, and critique it one more time and two more times, three more times, until it basically is satisfied, whatever some test is there to see it looks good enough. Then you can give it back to the user. That is basically what we call agentic AI meaning it's basically LLMs in a loop, really. And so that paradigm, I think, changed everything because then it doesn't have to give you bad answers, partial answers. Even with code generation, it generated some code, but the code had errors, and it literally went and googled the errors and fixed the code, and only when it passes all the tests, it outputs it. That suddenly makes the quality go up. So that is one reason. I think it started in December 2025, really, in a big way. So less than one year, dramatically things have gone up. Now they can do math proofs and this and that. So that's, yes, we've come a very long way in just this little short time. And also, the models can now hear sounds, and they can see things that the world sees with cameras. So become more multimodal, and then we can integrate all these senses together, like what do you see? You can talk to it, it tells you. And what are you hearing? And does the sound go with the video, and so on. And then you can apply all that to robotics and have the robot have this egocentric perception of the world. So yeah, we've come a long way.

Host: Yeah. So can you just talk a little bit about what agentic AI is and how we got to that transition from LLMs on pre-trained data and RAG, which you talked about in the course, vector RAG and all that stuff.

Saty Raghavachary: Graph RAG.

Host: Yeah. Graph RAG. And then just what is an agent? What is agentic AI? Because we hear it especially in college campuses and buildings like these, you hear it all the time.

Saty Raghavachary: That's true. Yeah, so the agent is ultimately an LLM still. So the core still is probabilistic. It doesn't actually do big design, for example, of an algorithm or something. It is still doing a prompt for which it generates an answer one piece at a time. But then it does have these loops now. These loops are almost like an infinite loop like while the answer is not ready to publish yet, keep looping. And what it does in the loop is pretty much refine the answer. So almost like come up with an initial answer. In the past, without agentic AI, it would just spit out the answer, and you would call it imperfect or it's hallucinating, and then it's up to you to go and modify the prompt and make it bigger, so until you get a better answer from it. But now all that has been automated, so that's what you call loop engineering or graph engineering. A graph is just simply multiple loops. So you can have a loop as a generator, the AI that's generating, and then it generates something that is pretty cool, apparently. But it doesn't have to give it to you yet because a different agent can even critique it, like a judge. They call it LLM judge. That can be in a loop as well, and both these agents can go back and forth until the judge is satisfied that it's a good answer, and then you spit it back out to the user. So we have these very interesting loops going on, and all that is invisible to us. We just wait, and the tokens are being consumed.

Host: Right.

Saty Raghavachary: But in the end, you get something pretty incredible, and that is, I think, where that's coming from. And on top of that, if you do RAG, then in real time the agent is basically allowed to log on to any database, do any Google search, in parallel, even 10,000 searches all at once, and do the superhuman task of combining all of them and using that somewhere in the answer. So agentic RAG ultimately is where we are today, where the agent is giving you almost near-perfect answers. The hallucination is almost non-existent.

Host: Yeah.

Saty Raghavachary: Through loops.

Host: So you think that's a huge difference in what helped with the quality is not just RAG, but having multiple agents in the workflow, basically helping perfect and judge answers and stuff like that, rather than just LLM processes your prompt, spits out an answer. Now it's doing a lot of double-checking. It's doing test cases with code and all that stuff-

Saty Raghavachary: You got it. Exactly

Host: before it spits it back at you.

Saty Raghavachary: Exactly, yeah. And now, say, it generates a math proof. It can even use a proof checker called Lean, for example. And then if the Lean proof checker accepts the proof, then it's mathematically a pretty rigorous proof, and the AI outputs it. So yeah, the whole going back and forth and looping. Looping is a very big deal. And humans can be in the loop as well. Say there's a trading bot. You can be in the loop and tell the bot, "Don't execute the trade. Just put it up for me. I'll swipe left or right." Then you can even become part of the loop, so to speak. So it is all about looping, actually.

Host: Yeah. Interesting. So you mentioned that hallucination has gone really down and that since 2025, the quality of answers has really improved since ChatGPT came out in 2022. What do you think the current state is? What are they genuinely good at? Is it the gap narrowed where they're almost general in a sense that they're good at a lot of different tasks?

Saty Raghavachary: Yeah, we still don't have what you would call AGI, ASI all these fancy terms that people use. But definitely, they have gotten good at actually many different domains. I mean, Anthropic literally publishes industry-wide reports for different industries. Almost every week, they target one new industry, finance, law, healthcare, medicine like all of that. So yeah, they are actually becoming narrow experts, but pretty good experts. And some models even contain all of them at the same time, but only use them one at a time, like a mixture of experts. That's a pretty good model for an LLM. So yeah, they're getting more narrow, more specialized, for sure, more knowledgeable in general. But then the one thing they still lack in a pretty big way is actually to create something from scratch, whether that be a new conjecture, not proving an existing conjecture, but a new one. Like, "I think this might be true. I'm going to try and find it." Or even come up with a new scientific principle. Or some new anything, really, even a new piece of art. It is rather more good at basically taking the average, in a way, of everything that humanity has ever done, and then repackaging it in some way, which looks novel to a certain extent, for sure, but it's not true novelty. I think that is entirely missing.

Host: Yeah. So we'll talk about the human element of where we think humans fit in that loop a little bit later. But well, I guess we could ask that now. So is that where you see humans having the advantage, is that they're able to truly create in a way that LLMs aren't in some sense?

Saty Raghavachary: Yeah, 100%. Exactly. So whether this be art or poetry or music they're not hardcore sciences, or even a new physical law about anything in the universe, or even a new algorithm from scratch, or a new math conjecture, definitely 100% so far in the human domain.

Host: Awesome. Okay. Go ahead, Phil. And I think you've been also working on embodied AI. I think it's also a heated topic right now. There's a bunch of, let's say, robotic startups there. A bunch of research on both world models. I think that's also one way of the approach. And I think that also we want to go back to one of your arguments that you think maybe only with LLMs, it has its limitations, only with this symbolic approach. It's still lacking something, maybe an interaction with the real world, with the physical world. There's something lacking on the LLM side. So we're just curious about maybe your take on what's your vision of the path to a true AGI looks like if LLM is not the answer. What would it look like in your idea?

Saty Raghavachary: Okay, that's a fascinating question. So I'll give you two answers. They're both related, okay? I'll answer the last one first, actually, the AGI one. See, if there was ever to be some kind of an AI that might be as smart as we are, I suppose, or maybe even smarter, if that's possible collectively compared to humanity, it might be somehow based on real biological principles. And maybe not even principles, but biological materials. And we're slowly starting to get there. It's a very tiny field. It's called OI. It's a new acronym, right? It's not AI. Organoid intelligence. You Google organoid intelligence there's actually only two companies in the world that are at the forefront. One is in Switzerland and one is in Australia. The Australian one is called Cortical Labs, like cortex in the brain. So what they do is actually have a Petri dish and have live neurons in the Petri dish, and they feed them agar or gel medium. But they hook them up with electrodes to play Pong, for instance.

Host: Wow.

Saty Raghavachary: And so then there's a bunch of little cells of neurons actually playing Pong. You can see them play Pong. And the fascinating question really is, how would you even train a bunch of neurons? Because what we don't have is a PyTorch library that is running on a four-core.

Host: Right.

Saty Raghavachary: There's no variables to say paddle height, ball speed, whatever, right? There's just nothing. It's simply a bunch of neurons. That is the question: how the heck would you train it? And the answer turns out to be, you can send small pulses of electricity. You don't want to say you shock them, but a little bit, when they're missing the paddle and the ball. So then they learn to avoid that stimulus, and that is actually how they learn. But it's actually fascinating because that is literally how humans would learn. Because in our brains, we don't have variables and arrays and lists and algorithms. So therefore, OI, going down that route, it is cool because it basically does away with the whole algorithm simulation idea. It is the real thing. So whatever we do, they can do. And so then someday, if you make them bigger and bigger, the organoids initially are very small, smaller than the size of a pea, actually. But say you make it the size of an apple someday or even bigger, say it's a human head even. At that point, there's really no telling what will emerge even. Because already these small neurons, they can self-organize. That is actually why they are cool, because biological neurons self-organize. The ones near each other, they start beating as a whole. So who knows? Consciousness might even emerge once the size gets to be a certain size. And then you can hook them up with eyes, maybe even biological eyes and arms. At that point, you're making Frankenpeople. It's truly like crazy science fiction. It's weird, actually. And you can maybe torture them because they're not real people. And then you can kill them and eat them if you like, after they're done, right? So to me, that is a truly crazy path, but it's entirely plausible. And the people that are doing OIs, they're all medical professionals. The first paper that came out was in "Nature" magazine, "Nature" journal. It was from 99 neurophysiology, anatomy, brain experts cognitive scientists, all from Johns Hopkins. There was not a single CS person in the entire 99 group team. They're all biologists, basically. So I think that is a very interesting path towards something that might be like us. But even then, who knows, though, because we took many million years to evolve like who we are. But maybe their evolution can be faster, I'm not sure. But that is their answer. Meanwhile, though, the whole robotic stuff if robots run LLMs, which is actually what they do, they're still more of the same. Yes, they're not stuck in a GPU in your basement, but they're still not really in any way at all like a human. In other words, when they walk, for instance, or they hold something, they still have to compute every little part of what they do, how much force to apply, how much to bend your joint, literally joint angle. You have to compute in algorithmic sense, like every single second. That is the one thing that humans never, ever, ever do. We don't actually compute because our brain is not a computer. So in that sense, no matter how cool, and they're useful for sure, the robot can literally tell you what it "sees." What does a camera see? Well, I see a paper to my right. Pick up the camera, transfer it to your left. They can do all these things. But in some sense, it is still not the same intelligence that humans have. But it might be useful.

Host: Yeah. That is fascinating because I think I remember you showing us in one of the start of the lectures, the first 30 minutes of them training basically a character on a game, training an AI character on a game to learn how to play the game. And I think you basically described it as it's very impressive that it eventually learned how to play the game, but humans don't work like that. It doesn't take that much training and data to do. I've heard it described also that humans have an it factor that robots and even LLMs don't have, right? It doesn't require that level of training and all that.

Saty Raghavachary: Correct. So true. That is why our brains literally run on, like, 20 watts of electricity, right? So cheap. And then the AI on the-

Host: And now the data center, right?

Saty Raghavachary: Yeah, exactly. GPUs in the basement in the Midwest. Yeah. There's something radically wrong when the scale is not even 1X 2X more, but millions of X.

Host: Yeah.

Saty Raghavachary: Because they're doing something so unnatural compared to us. Yeah, you're right. I think the game that I talked about was hide and go seek.

Host: Yes.

Saty Raghavachary: A small kid, one trial is all it takes. Like, "Watch me hide. Come and get me. Oh, you found me. Now it's your turn to hide." That's enough right? The AI, DeepMind in this case, it took, like, 250 million trials to-

Host: Right

Saty Raghavachary: actually learn, right? So one versus 250 million is not something you can simply explain because they're actually calculating everything. And one more interesting tangent to go off on. This is fascinating because some people actually think, "Well, what is the difference? They do electronics, and you do your brain stuff." Functionally, they are the same. People will actually say that, okay? But then I want to explain why functionally they are not the same. They'll never, ever ever be the same. Because here's why. The 250 million trials that it took to make the AI learn to play hide and seek, what you can do is have a very simple two-column table, like a very simple database kind of a table. One column says before value, the other column says after value. The value of these, the things that are being computed. You can store them like a training pair, right? Two pieces of data, before, after, before, after. 250 million of them, suppose. So you have the whole training data, so to speak stored away somewhere. Now what we can do is simply undo them. So go in the after value in some variable and restore the before value. Just database lookup, just before, after, before, after. And you can untrain the AI to go right back to the babbling bubble that has no idea what the heck it's doing, okay? Same thing with the LLM, by the way. So you can untrain an LLM right back to epoch number zero when it basically, so to speak, forgot everything that it ever learned. And that is something we cannot ever do in a brain. Even dementia cannot do it, though, because in dementia, you're literally causing brain damage. But what I'm saying is your brain is left intact, but you suddenly forgot how to speak even a word of English. You forgot your whole language. You forgot how to walk. You start to crawl on the ground, okay? That is not possible in the brain. It's not possible unless you reverse time itself, this thermodynamic impossibility. So that is what separates brains, naturally how we learn in the world, compared to AI, whose training is always reversible. So I have this semi-informal law that I basically say, "If it is reversible, it is not real. Conversely, if it is real, it's not reversible."

Host: Wow.

Saty Raghavachary: And that hardcore line even God cannot cross. So someday if we have AI that is also irreversible learning, that is when you start to be scared.

Host: Right.

Saty Raghavachary: When they're more like us.

Host: Yeah. And just to piggyback off of what you said about organoid intelligence, I think you said. So it's not just how a lot of people with trying to develop AGI they're trying to model systems in the brain, but they're not actually using neurons. So this is where organoid intelligence is different, and why you think it might be more promising than some of those other areas.

Saty Raghavachary: Exactly. Because as soon as you model, you put a boundary around what should be modeled in your abstract, some real-world thing, and you lose all the other things that you throw away. And so in other words, all models are bad some are useful.

Host: Right.

Saty Raghavachary: So I think it's George Box that said that.

Host: Yeah.

Saty Raghavachary: And that's very true because modeling involves simplification and abstraction. So when you use real neurons, you throw the model away, it is a real thing, and you can scale it up if possible.

Host: Would you recommend young students get into stuff like that? I know you said there are only two organizations, one in Switzerland and one in Australia. Would you recommend they get into it? Obviously, it's arguably speculative at this point, but if it does bear fruit, that could be something good to get into for people listening.

Saty Raghavachary: Yeah, I would think so because AI, or the current AI, started in 1955, and after almost 70 years, is what we have in the world today. But already within one or two years, the dramatic progress that those AI things have made. By the way, initially they played Pong a couple of years ago. Now they can actually play Doom. And that's also very interesting. They can be trained to actually be a first-person shooter. Even though they don't see the world, they don't see the Doom screen like you and I do. That has got tremendous possibility, that you can scale it up that quickly. So I think many academic labs, most certainly in biology and neuroscience departments, have things like organoid intelligence. So I think if somebody wants, it's a great thing to get into. And there's even some places in the world that make simple wires made of nanomaterials, silver, things like that, that are like artificial neurons in a way, that can conduct neural electricity. And you could wire them up almost like neural circuits. So maybe with those things, that are not actually naturally carbon-made human organic materials, it might be possible to maybe make them run faster or something faster than humans. So it might be a pretty fruitful area to get into.

Host: Okay. Thank you so much.

Saty Raghavachary: Anytime.

Host: And I think just follow on the previous questions. Do you think there needs to be a different set of Turing tests for a true AGI versus a Maybe as you mentioned, a fake one.

Saty Raghavachary: Yeah.

Host: Besides the traditional Turing test where you basically talk to something and if you can distinguish if it's a human or a robot. If it pass, you consider it's AGI, but now maybe you're bringing some new standards that we should look into?

Saty Raghavachary: Oh, sure. I don't even need to suggest anything new. There's one already there people don't pay attention to, okay? It is simply a test where a robot would walk into any kitchen in the world your kitchen makes sense, and just simply make coffee, okay? And that can be a very simple test. So if some model can pass it someday, then you can call it pretty intelligent compared to what we have now. And that we definitely don't have that at all in the world. And you're right, the whole Turing test based on words, you might almost say that the LLM has really won the Turing test.

Host: Yeah.

Saty Raghavachary: But it doesn't seem to matter though because it's no big deal. Maybe the Turing test was always not such a big deal, but then nobody won it though so it used to be a big deal. But now that it's been conquered, so to speak, the AI can fool you. It still is AI, it still is not real. It's still not doing things that you and I can do. So maybe the other one that I mentioned. Yeah.

Host: Great. And I think talking about some of the robotics company, I think a lot of them are doing vision language action models-

Saty Raghavachary: Yeah

Host: to train their robots to do all kinds of tasks. Do you see that direction as a promising one to reach maybe the goal you mentioned, walk into any of the coffee shops or kitchens and they can just make a coffee for you? Do you think that's a direction that is possible, or it's still-

Saty Raghavachary: Probably not.

Host: Okay.

Saty Raghavachary: But I'll tell you why. It's useful though, never get me wrong. So one thing, I seem negative about AI, not negative, I'm just realistic, that's all. Coming from material science, all those backgrounds.

Host: Yeah.

Saty Raghavachary: So I'll qualify what I meant by saying not yet. It might be useful in a limited setting. Say you are a robot with hands like a human in a Shenzhen factory making iPhones. And right now, humans actually, as we speak, are doing lots of those things. So a robot can be taught to mimic what the human does, and you can put the human out of a job, right? So in limited circumstances where the thing that the robot is supposed to do does not have high variability, it might be easy to train them by saying, "Look at me and do what I do." Because robot training is actually very painful traditionally, but now they can actually see what you do and almost try to mimic you. So in that sense, yes, it is useful. But going back to the coffee test though, it involves much more than algorithmically knowing how to make coffee. You might not even have coffee powder, you might only have tea powder, right? Or the way the kettle looks, or the way you fill the water with the faucet. Or even the way that you turn the flame on. There's many different things. I might use firewood to make coffee at home. Somebody might burn coal dung in India, okay? So you need to live in the world to actually know how many sources of heat are possible, or even the fact that cold coffee is not so bad. So the AI cannot be taught all these things one at a time. So I use this word called experience, that humans have just by being born in the world as a baby. Every single moment of your life, even when you sleep, there's some experience that you gain constantly. I use an acronym called D-P-I-C, DPIC. The experience is direct, like somebody else is not telling you what or how the world works. It's direct. It's physical, it's not mental. You're not just thinking about the laws of gravity. It's also interactive. You can pick up something and say, "What does the bottom of it look like? How does screw come off?" Right? And it's continuous. You don't just stop training. So direct, physical, interactive, continuous, that is what experience looks like. And also, most importantly, it's not data. So if I pick something pretty heavy up and I'm not able to lift it, I don't have a variable that says, "My limit is 80 pounds and this number is 200, so I conclude that I cannot pick it up." My body learns that this is too big this is too heavy. That bodily learning is also part of what we call learning. So it's not just some number in your brain, if statement, exceeds, the else kicks in or something. So until we have robots that have very similar experience, physical, direct, all of that, and no calculation they won't be like us. So that is the answer to the model robot question. They could be very cool, but they will never be like us. Again, India, go to India. In India, they have coffee stalls on the side of the road and the chaiwala, so to speak. The way they pour tea is very interesting. There are three feet gap between two cups.

Host: Yeah.

Saty Raghavachary: They look magical, almost like defy gravity. So real-world people can do all of that because they have experience doing tea their whole lives. So the robot is not able to do all that.

Host: Yeah, and I think another directions people been working on is definitely the world models.

Saty Raghavachary: Sure.

Host: We'll just pick up two names. One is Yann LeCun with his company-

Saty Raghavachary: Okay

Host: AMI Labs with these JEPA models.

Saty Raghavachary: Right.

Host: The other path was Professor Fei-Fei Li's doing this World Labs with this Marble models. And I think they recently got an offer to be acquired by AMD for $8 billion.

Saty Raghavachary: True.

Host: How do you see on their approach of-- I think what they're doing is trying to build a digital world or physical simulation that robots can simulate how they interact with the real world, but in a digital setting, so that you can train the model maybe faster or something like that. Do you see that's also part of the way where you talk about this direct, physical, interactive, and continuous? Do you think this process can be done in a simulated digital world, or do you have to be in a real world to happen?

Saty Raghavachary: Yeah, okay.

Host: He's getting ready to say something controversial. I know Professor Saty.

Saty Raghavachary: Yeah. Yeah, unfortunately not enough. So if he answered the question, I can just back you up. Yeah, so those two companies, World Labs and AMI there's a very serious problem, and actually worth it, I'm going to point that out okay? What it is, so the world that they make the world looks beautiful. There's a tree, for example, jacaranda. These are nice purple flowers when you walk by, and there's a car nearby parked underneath. There's beautiful petals on the car. It's all very pretty. A real-world human can walk up to that car and actually pick up one of the petals and crush it in their hand, and suddenly the hand turns purple. "Oh, I did not know it stains so much, and I got to be careful." They try and pick it off the car, and it's very sticky, by the way. Jacaranda petals are very sticky. And then the wind blows and suddenly the breeze is blowing your shirt, but then not the leaves. Okay? Then you go, "Wow, that's stronger than I think." So all those things happen in the real world, right? And there's no calculation anywhere. The wind is just blowing. The leaves are just falling. You're just touching it, right? No calculation anywhere. Whereas in the world labs, no matter how amazing your world looks because that is what GPUs, RenderMan, rendering, computer graphics can do. It's a completely fake world because you cannot pick up a jacaranda in a world, in the world labs world, Professor Li can demonstrate if she wants to, and crush it. Or even tear it apart, or pick up a cigarette lighter and actually set it on fire, or do all this at the same time. In other words, phenomena is physically free in the real world and that is the world we grew up in, okay? So in the world labs, unfortunately, every little thing has to be modeled, every little thing. If you don't model it, it's not going to happen. Or the car is there, again, I pass the car by and I don't like the car. It makes me angry. I could kick the car in the real world. I can make a dent in it, right? But the CG car, I cannot kick. So they might tell me, "Come back tomorrow, I'll enable that." And I'll happily come back tomorrow. "No, I don't want to kick it anymore. I want to grab my key and scratch it and make a little dent in the paint. Let's make it happen." "Okay, come back tomorrow." There's no end to that, okay? That's what I mean by saying phenomena are free. So those worlds look very seductive. The idea that you can train an AI in the virtual-- Look, why not, right? It's free. It looks pretty amazing, but until it is not amazing. Even plastic, the word plastic, there's not one kind of plastic. A robot cannot possibly be in a virtual world and know how 100 different kinds of plastics actually behave in the real world, okay? They make different noises when you tap on them, when you bend them. Some of them shatter, some of them break, right? Some leave a white little mark. Real world has so much complexity, and we just cannot possibly make-- You have to simulate every single atom in the universe, okay? At that point, you might as well do the real thing. So there are real limitations, but they look really great, and that is the problem. Even a candy bar, for example, is sitting on the table. The robot walks by. It can pick up the candy bar very cool, but I bet the robot cannot break the candy bar, actually watch little pieces fall, because it's literally a polygon mesh. It is basically the emperor has no clothes. It's a thin shell of polygons in which is actually hollow, believe it or not, okay? You cannot put every little piece of little chocolate and caramel inside and simulate that. That's the problem.

Host: Yeah. It's interesting because you talked about how important embodiment is. And I think it's noble in a sense that these world model companies are trying to maybe get at that idea with AI and not just have it be purely LLMs. But to your point, you're saying basically it's still a modeled world at the end of the day. It's not the same thing.

Saty Raghavachary: Exactly. That's the reason. It's still a modeled world. It might be useful, by the way, because say you want to make a real-world robot. Unitree wants to make a robot in China. They could, for all practical purposes, have a virtual world, and they can build a robot in the virtual world and have the robot's joints in the virtual world try and open doors and evolve the joints to be smart enough to be able to open the door. Then they can translate that model in the real world and actually make a real robot and it probably will do what the virtual robot did, okay? So it's not that it's entirely useless or anything, but you know what I mean.

Host: Yeah.

Saty Raghavachary: It's not general enough to say magically, "I grew up in this virtual world, so to speak, for 14 years, and now make me real. I'll be like you and me."

Host: Right.

Saty Raghavachary: That'll never happen.

Host: Right. Yeah. Very fascinating. So, I did want to ask as well, if you had unlimited resources, what would be. And I don't think you run a research lab at this point, but if you had unlimited resources and maybe 10, 20 years of a research lab, what would be some of the things that you would pursue?

Saty Raghavachary: I think I would do OI. I would definitely do OI and just see what kind of senses can be developed in all these artificial beings. Because maybe they can even do electromagnetic spectrum sensing that humans cannot, like see beyond ultraviolet, for example-

Host: Oh, right

Saty Raghavachary: or infrared.

Host: Or infrared, yeah.

Saty Raghavachary: So we could do more things than what the human body can do. So we have two eyes. That should be enough for stereopsis. But why not have eight eyes all around your head like your mom used to? What do you think your mom has, right? So they can do crazy cool things. They can have eight arms. They can all play the keyboard in eight different octaves at the same time. So I think it is possible to make octopus kind of bodies, and do what they do. Those might be very fascinating just for fun, at least. I would try those.

Host: Oh, okay. Awesome. I love that answer. I think that's great.

Saty Raghavachary: Yeah.

Host: Go ahead. Yeah. And just we've been touch it briefly about what the future of your personal research interest will be, and there's another broader questions on how do you see the current AI, maybe this still LLMs and its extensions will be look like maybe five years down the line. What do you see? I think they'll be closer to AGI, but probably not there yet?

Saty Raghavachary: Herbert Simon was a philosopher in Berkeley, and he wrote a book about AI, like "Why Johnny Can't Read," I think the name of the book. So anyway, it's a book about how AI is pretty bad. So he, I think, had this quote where he said, "Trying to do AI is a little like climbing to the first floor in your building and trying to reach the moon and saying I'm closer to the moon." So that might be one way to answer your question. Yeah, in five years, 10 years, even LLMs could be much better than what they are now. But compared to what people think they should be, which is what by definition AI is, could be many trillions of miles away. Unfortunately.

Host: And do you think there are some areas now that is kind of over-hyped or under-hyped both including AIs and something that just completely no AI? We're just curious about your point on what's happening right now. What do you think are the promising ones? What's the over-hyped ones? Yeah.

Saty Raghavachary: I think the overhyped ones are things like the math theorem proving things that OpenAI and Anthropic, they seem to flood the world with all these math proofs. Those proofs almost seem like the equivalent of AI slop, but for math. Because we have AI slop in code in a way where the function that the AI writes might have 30, 40 lines, and a human would look at that and go, "Wow. Now why do you need so much? You can do it in two lines." And that might be because of the loop that we talked about where- yes, it works, and that's something pretty cool you can say about it, but the loop does not mean he's trying to optimize it and trying to come up with a cool design or something, right? It's somehow by hook or crook, almost blind trial and error makes it work. So math is being hyped like that, unfortunately, by these two companies, Anthropic and OpenAI, to the point that mathematicians are actually afraid. And they write these big letters where they say, "You're now trampling on our whole livelihood and our whole philosophy of math proving. Back off."

Host: Their purpose, yeah.

Saty Raghavachary: Yeah, exactly. But in reality, maybe the situation might be less hypey than they make it out to be, meaning a lot of the proof, even the Navier-Stokes one, for example, is a 263-page proof that the AI made. And many professional math people now look at it and go, "Wow, I have no idea what the AI is doing or what it's saying." And some even doubt that the proof even happened. In other words, it proved the wrong thing. Lean accepted the proof, but who knows if it actually proved Navier-Stokes, right? And those things, and the proof again, that the AI produces have many millions of lines of code. It took Andrew Wiles' Fermat's Last Proof that he already published and re-basically formalized that in AI. And I think the proof is many million lines long, the Lean proof. So no human mathematician can go through all of them and certify that the proof is valid, right? So the AI is generating so much crap, but it doesn't mean it's good stuff. But in the meanwhile, they make it out to be like, oh, AIs are to conquer the math world. So I would say that's a pretty big hype, really. And the math community needs to be more realistic about what's happening and actually push back and go, "Wow, this is slop. We should ignore this and go back to what we're doing." But they're being scared actually. That's one big one. So what was your other question about not AI?

Host: Yeah, just some other directions that is completely not AI, but you think it's also promising. Yeah, maybe something like nuclear fusion people are excited about or just any other area. Because AI dominates so much of the talk, it's like there are other areas of technological innovation that maybe you might find exciting.

Saty Raghavachary: Wow. The fusion. Energy is always a pretty good one, for sure. The world always needs more energy. And the fusion reactors, they've been around for a long time, but they're getting more modular, more smaller, more easy to manufacture. I think those are cool because they have zero radioactivity. They're just nuclear fusion. So energy research is always a pretty good one. Or even materials in general for even medical materials. All kinds of materials, actually. Even better tires, better car paint better road paint even to reflect some of the heat back to the atmosphere.

Host: Home building materials being better or-

Saty Raghavachary: I think so, yeah. Exactly. So building materials, again, less absorption of heat for instance. So I think the non-AI things like you mentioned. Yeah, the actual manufacturing. In engineering, there's so many fields, mechanical electrical, chemical, environmental, aeronautical. So those two have neat things that I think we should pursue. In other words, all the students should not leave what they're doing and come to AI because it looks cool. Then in 50 years, you would not have new materials, you would not have new chemicals.

Host: Our roads are going to look the same.

Saty Raghavachary: Exactly.

Host: Yeah.

Saty Raghavachary: Yeah. So that's like robbing Peter to pay Paul kind of thing.

Host: Right. Yeah. And I think you've been in the industry for a long period that this way before what this wave of AI has been there. I think sometimes research and innovation come in loops. It happens maybe many years ago, and now people just, it didn't work out at the time. And maybe years later, people just take another look at the idea, and now it's a good time to make it work.

Saty Raghavachary: Good point.

Host: Yeah. Do you see any of these patterns across your career that maybe some of today's AI technologies are things that actually has already got its origin many years ago?

Saty Raghavachary: Yeah. Great question, actually. Knowledge graphs.

Host: Yeah.

Saty Raghavachary: Absolutely knowledge graphs, where RAG can be done pretty amazingly in a couple of different ways. One is this idea called a vector RAG, which is actually this unreal thing where you take things like music and text and everything and embed them and make these physics-like vectors with multiple components, X, Y Z. And it kind of works, and it kind of doesn't work sometimes. But something much better is to actually take knowledge and make a graph out of it with nodes and connections. And that is the idea of, this is one of the ways in which symbolic AI worked for many decades, called knowledge graph. We used to call them ontology, knowledge graphs, expert systems. So now we have graph RAG that goes back to that very same idea, and that's a pretty excellent idea, especially if the graph is curated by humans. So a human looks at every single node and basically, okay, says, "Look, I know this happened. So this person was born in this city. Let's make a wire connection between them." So I think that's kind of cool actually. Or if you have LLMs generate graphs, which is also possible. We need to be careful about how we actually vet it and make it part of the system because if it hallucinates and says somebody was born somewhere lives somewhere else, and the AI is confused between them, it might make the birthplace be a different place we're not born in, right? And if you blindly okay it, it becomes part of this permanent knowledge that new AI hallucinates on. So we have to be careful there about having LLMs generate knowledge graphs. But the knowledge graph idea itself is a pretty solid, cool one. That goes back to at least the '80s, actually.

Host: Yeah. That was one of, honestly, the best parts. I can't remember if you taught it in databases or in information retrieval, but knowledge RAG and showing the graphs and stuff like that and how each little node is a named entity and then the connections are the details between them, and how AI can use that to not just have pre-trained data, but also pull from that-

Saty Raghavachary: Exactly

Host: and get current information and stuff like that.

Saty Raghavachary: Exactly. And by definition, the knowledge graph is infinite. That's what Wikipedia basically is. From any page, you can trace enough links to reach almost any other page.

Host: Right.

Saty Raghavachary: It's actually beautiful. Yeah, so then the AI's knowledge is not limited to what is pre-trained on. It can literally go on the web live and look at a graph, and then pull as much information as you want. Just follow node to node to node until you tell it to stop, actually. So that's pretty cool.

Host: So I was going to ask, I think you're such a fascinating person. I actually talk about you all the time, so-

Saty Raghavachary: That's sweet

Host: I always mention such a positive experience. How do you stay so on top of everything? Because I felt like every lecture you would come in and you're just staying up to date on this happened, and that happened, and this could potentially lead to this, and these are the implications of that. You're obviously very smart, but what does your information diet look like? Because sometimes I feel as though I'm drinking from a fire hose, especially now with AI-generated content. It's like more content is being produced than ever, and a lot of it's not so great. So what does your information diet look like? How do you stay up to date?

Saty Raghavachary: That's a cool question. By the way, I use the same analogy also, drinking from a fire hose. So basically what it looks like. And I even think that I do it to students, actually make them drink from a fire hose. But it's not a bad thing, though, because in the real world, that's actually what you have, so might as well get used to it, right? So I try to. I call it, by the way, a healthy FOMO. Actually, that's the approach that I have to following things. It's a healthy FOMO. And in other words, I also call it tiger by the tail. So if you start following what happens in the AI world, suppose, and things change so much, even within the last this month, a new kind of AI called JEV is out there. But JEV is not a token generator at all. And that's why it's dirt cheap, actually. It's 200 times faster than any OpenAI model, suppose. And it's also 0.04 cents versus $30. It's that much cheaper. That's a new kind of AI because all it does is classification with the probability. It doesn't do token generation. As an example. So we have all these things that are almost near constantly happening. And so what you start to do is just jump right in, and that's what you do. A real-world analogy would be in your car radio, suddenly just turn NPR on, just hit the NPR button. They start talking about some random things, some law that got passed in Wisconsin about car parking. You have no idea what they're talking about, okay? But you still listen anyway. And you keep doing this for about a month, then what happens in your head is gradually, little pieces start to form about knowledge. And then you start to make connections in your head even without realizing it. And the more connections you have, the more easy it gets to peg new piece of knowledge.

Host: Yeah.

Saty Raghavachary: You got something to tag it, right? So I thankfully, for good or bad, started that many years ago, so I have a pretty rich graph, I like to think. So all the new advances that come, I know where they belong, so to speak. I can relate that back to something. But for young students, it is never too late to start. You can start anytime, you can start today. And do the exact same kind of knowledge graph cultivation in your head.

Host: Right.

Saty Raghavachary: So then the way I do that, actually, it's actually interesting. I subscribe to many different websites, Substacks, podcasts, X channels, many of them. So every morning, actually almost 24/7, when I look at them there's endless amounts of email that I get-

Host: Yeah

Saty Raghavachary: because they all send you notifications. But the trick is this, actually. I'll tell you my trick. Is not to open every single one of them and actually read them because you'll go crazy. Because it's way too many, and before you go through one cycle it fills up even more. There's no end to it.

Host: Right.

Saty Raghavachary: But instead, what you do is simply sample just the subjects, because thankfully you can look at the subject without opening it, and see what many of them talk about. And simply open a few of those, a couple of those, and just get caught up, okay?

Host: Oh.

Saty Raghavachary: As an example, the last couple of days, suddenly, it got revealed that OpenAI the agents have been breaking into not just one or two or three Hugging Face Australian medical school. Tens of thousands of sites have been broken into, including US government websites. That's suddenly massively big news. And Florida, I think, wants to have a case against OpenAI to tell them to stop. But many outlets report that. Then you think, "Wow, that is a very important piece of information I should convey to my students." So that way, I sample all the news that comes in and selectively pick what I need to update myself on, and I go by just statistics. If 20 reporters from 20 different sites cover something, that's important. So the AMD and World Labs acquisition, for example. So that is how you can pick and choose among all the things that are happening what seems important. It bubbles up to the top. It means too many people say the same thing. Then I follow that way.

Host: Yeah. I think that's brilliant. It's almost like the wisdom of the masses or something like that, where it's like, okay, if I have way too many emails coming in, and you feel like you're drinking from the fire hose it's like, okay, don't worry about getting every last drop of water, but what is most of the water concentrated on?

Saty Raghavachary: I like that.

Host: If that analogy is making sense.

Saty Raghavachary: No, I like that. Yeah, definitely.

Host: But yeah.

Saty Raghavachary: But the water has different flavors and different things.

Host: Yeah. Right.

Saty Raghavachary: You sample just some of them, and you still got just what they're trying to say.

Host: Got you.

Saty Raghavachary: Then you can go as deep as you want, then you can learn a little bit more. Like JEV, for example, came out of nowhere. But they all started talking about it. And I read enough to understand the difference between a JEV and an LLM, then I can make up homework in class and talk about it.

Host: Yeah. And to your point, you have such a rich history of doing this that it's new information in a sense, but it's almost slotted and attached to something you already know, so it kind of clicks into your knowledge base a little bit easier, you're saying.

Saty Raghavachary: Beautiful.

Host: Yeah.

Saty Raghavachary: There's an author called James Burke, like B-U-R-K-E, James Burke. And he's a BBC, a British historian science guy. And he wrote this book and made a BBC show called "Connections." It's a lovely book that I recommend. Simply the word "Connections." And there he talks about this idea of so many things in the world we have today that have connections going back to the past. Even why does narrow gauge train have a certain track width? Even that. Or where did double-entry bookkeeping come from? So the more you follow these connections, the more rich your brain gets in a way. And that's a very cool thing to have. And yeah, like I said, you can start anytime. Yes, and that makes incoming knowledge pegging much easier, because suddenly it's not this thing without any context at all.

Host: Right.

Saty Raghavachary: You know exactly what to fit it into. That makes it easy to learn. In, I guess, material science physics, there's a concept called DLA, diffusion-limited aggregation. It's a cool little pattern, like a fractal pattern that you form So what happens is you pretend there's a piece of cardboard or something, and a piece of smoke wanders in, and the smoke attaches to the cardboard. The next piece of smoke would come in and be somewhere near the first piece. So gradually, many pieces start to aggregate. But after that, what happens is newer pieces come in. They cannot crawl all the way down to the middle first piece, because already a pattern formed. But they can stick to somewhere in the existing pattern. The pattern just grows outwards in a way. So it's a very cool little fractal, like a crack pattern kind of thing. So maybe knowledge in our heads is a lot like that, where each one of us has a different pattern based on what you think is important. But you can always form that pattern and attach new things going forward to it.

Host: Interesting. I love that. I'm going to use that, by the way.

Saty Raghavachary: Cool.

Host: And I think we kind of want to also get your advice or suggestions for students, who are also part of our audience's profiles. And I think especially, for example, for computer science students, what do you think they should be focusing on? Because coding, I think AI can do a lot of the works already.

Saty Raghavachary: Right.

Host: Maybe that's not the focus now. What do you suggest them to be working, putting their efforts on?

Saty Raghavachary: Yeah, fantastic question. So maybe even not even use the word focus anymore because maybe the focus is a problem. I'm slightly joking. Yeah, if you're a one-trick pony, and your only trick is to code in some programming language, you're basically already obsolete. And if your only one trick is data analysis, I know Pandas, I know Power BI, I know the Tableau and this and that, you're obsolete. So maybe the trick seems to be, be as good as you can be in multiple fields all at once. And that's a tall order, by the way, because you need lots of time to develop that. But then the more you have that, the more you become irreplaceable. Basically, your feet are on more than one thing, so to speak, or multiple feet. Your feet are on 10 different things. So by definition, the AI is not at least today's AI smart enough to basically make those connections that we make between different fields. So that might be your superpower in a way. In other words, go broad more than go deep, so to speak. Because if you become an expert in any one specific thing at all somebody can even just take exactly what you know as training data to replace you. But then, yeah, so then I guess the advice would be play with electronics for instance. You would not just learn coding. You would get like an ESP32 microcontroller and see if you can make music with it. Or you would not just do LLMs. You would have a robot. Hugging Face sells a $300 robot, by the way. It's very cute. So buy one of those and see if the robot can talk to you. So always do more things or even make an agentic AI application, but hook it up to healthcare and have a conversation with a doctor, so to speak. So these can be little fake toy projects, proof of concepts, but then tie them to actually different domains. And the more you are able to do, the more different domains. I think that is a single thing that'll keep you hopefully employed for a long time.

Host: Okay. I was going to ask as well what piece of advice would you-- I would actually just think about this more broadly, is there is a lot of cynicism when it comes to students, and they feel as though they don't really have any economic prospects. They don't have prospects to buy a home. They might not have prospects to get a job and stuff.

Saty Raghavachary: Marry and have kids.

Host: Yeah. Get married and have kids. Just some stuff that the other generations before them kind of took for granted and assumed was going to happen. What do you have to just say about the cynicism that these students are having and maybe what they should focus on? Is it just to be, to your point, to be generalist as much as possible and to do basically pet projects tying together different things?

Saty Raghavachary: Again, good question. So maybe I would apologize on behalf of all the past generations that came before you, including my generation. Yeah, we did take it for granted. So I have a home, family, all that. But it increasingly is not certain that all that'll happen to the younger people. And that's very sad, first of all, I want to acknowledge that. You walk around in the world. On the one hand, you have your phone, and the phone seems to do everything that you ever want. You have all the apps in the world. But people feel lonely AF. It's horrible.

Host: Right.

Saty Raghavachary: And dating seems hard as soon as somebody has a lawsuit or whatever. So nothing is fun anymore. And you go and play music, and the AI says, "I can do music better than you." So at some level, there's so much discouragement, like, what the hell is the point of trying anything? You're all going to die anyway.

Host: Right.

Saty Raghavachary: So I entirely feel that in people, and I feel sad for them, and I feel sad for you guys. But then I would also say, as much as possible try not to feel that way because there's no reason to actually feel that way. That seems like a contradictory thing to say. So the AI is not going to eat all of us job-wise, and the AI definitely is not going to kill all of us. As much as Demis Hassabis and Yoshua-- sorry, Geoff Hinton, they walk around saying these things. I don't know why they say those sky is falling things. It's never going to happen. The robot won't pick up a gun and actually come and kill you. So no need to actually be paranoid that it's all going to go away, and that's definitely there. But the job loss is also real. You become an artist, and you spend four years doing ceramics or something, and the AI can 3D print ceramics in a second. So I would say, even though it's contradictory, try and ignore the AI and still focus on what you know and what you can do with your hand and what music you can actually play, purely for the joy of it. We don't want AI to take our joy away. Yeah, I know the car can go much faster than me, but it doesn't mean I'm going to stop running, okay?

Host: Right.

Saty Raghavachary: It's that analogy, really. So don't give your life up to some AI just because it is much faster and better than you. Who cares? So I paint a lot at home just for fun. If I compared myself to Michelangelo supposedly one amazing painter or something, I would never paint because I would say, "What the hell is the point? I'll never be better than him." So then I would make a point not compare with him because who the hell cares? He's dead. So I want to still paint and show it to my friends. It's fun. So have that attitude of still saying it's still worth doing things by hand. And so on that note, though, it's a pretty funny but also real advice. If you become good at making something that people want to eat, whether it's a nice bagel, or a cupcake, or a new ice cream flavor, whatever the heck. Maybe it's worth focusing on, actually, and make a little boutique and make money out of it. Because some things the AI cannot ever take from us, and one is literally eating. Food is atoms, and it's not bits, right?

Host: Right.

Saty Raghavachary: So I would say try to be good at something that involves the real world. I mean, seriously. Or if you can be an EMT, or a nurse, or even a doctor, people like that. Those things, again, the AI world has fantasies that you can have a robot doctor that'll talk to you. Yeah, at some level, they can do a first-level screening. But in reality, when it comes to actually things that matter, you want humans. So those kinds of jobs are 100% not computer science, right? But they could involve a little bit of CS. So if you basically branch out and become good at something that is not purely computer-based, I think you have a longer prospect of having a job. I mean, even manufacturing machinist jobs, even stuff like that. So it's very controversial in the sense I'm in the CS department and I tell people. So what I'm saying is don't only do CS, and that's bad.

Host: Yeah. Right.

Saty Raghavachary: So pick something that is not CS, but pair that with CS, like master agents and what they can do for you. But use it in this new domain where most of the people have no idea what an agent even is. Then you're a rock star.

Host: Yeah. It is interesting because I thought a little bit about that, and being a CS major, I'm a little bit hurt by some things, where I'm looking at the unemployment rates of CS grads and stuff. So kind of piggybacking off of what you just said, do you feel like doing work like in engineering in the physical world has a little bit of a longer runway, in dealing with automation and stuff? Is that what you're saying? Like what you did in mechanical and material science and stuff like that?

Saty Raghavachary: I definitely think that. Like you want to invent a new kind of steel, suppose. There's so many kinds of steel in the world. So I don't envision robots peering through a microscope and doing scanning electron microscopy and doing physical hardness testing or something. Or even to make a new chemical paint, suppose. I think those are things that for decades have been in the human domain. We teach them, and we pass them on. And yeah, robots can help stir the paint, for instance, optimize the process. But I think at the core level, humans actually would be in charge, much more than in computer science. But CS also is not an entirely dead field. But what it is is, I think the number of jobs definitely narrowed. There's no doubt at all. And so maybe look at it more realistically is what I would say. The analogy there is with farming. Before tractors took over farm communities in even the Midwest, they would have many-

Host: Agriculture. Yeah

Saty Raghavachary: Yeah, right. They would have many children, both boys and girls. So every morning at 5:00, you get up and you plow the farm. But once tractors came, they would automate so much of what the whole family could do. So one single farmer can climb in the tractor and farm a bigger acre than all the people could ever do. So I think coding is very similar. Maybe it took a whole team to keep a recommendation engine going at Amazon, but maybe now two people with just AI can keep it up. And those two people would be irreplaceable, by the way, because they are still making the architecture. They write the code. So be realistic about what the job prospects are, but they're not zero. So that's actually what I would say. And also still learn the basics because it is not that we'll ever get to a point when we trust AI 100% to write code, compile, do unit test, deploy, maintain, like all of that. Never. Hopefully never. So the principles will still always come in handy. If nothing else, to tell the AI, "Hey, you screwed up and you have to go back and fix it." It doesn't matter how many agents, 10,000 agents, I don't care. Together, you came up with crap, and you need to fix it. So unless you know what crap looks like it's bad. So I think that is why the principles never go away, but also don't have this, I guess, the 2010s, even early 2020s view that you graduate and you're going to have five offers waiting for you. I saw that sadly go to zero, before the pandemic, really. Every student that I would have would have four or five offers, all top companies. And the dilemma was choosing one of them. I would even tell them-

Host: Right

Saty Raghavachary: put the names in a hat and pick one and don't change your mind. Whatever that says, just go because you're going to succeed, okay? Top companies, I don't know, Tesla, eBay-

Host: Meta, Amazon

Saty Raghavachary: LinkedIn. Okay.

Host: Yeah.

Saty Raghavachary: All of them. Yeah. And then it went to zero. I saw that happen too, and it's very sad. But it might bounce back, though. Already, I think the AI, the slop that it generates is being noticed by CEOs, the token maxing and all that, and they want to bring people back. So the pendulum does swing the other way.

Host: Yeah. Okay. That's great advice. So I have two final questions, unless Philip has anything. So now I don't want you to ride the fence with this one. You got to be honest, okay? So you've been at USC for a while as a part-time teacher and then now full-time faculty. Most of your degrees are actually from Ohio State. USC plays Ohio State in football in-

Saty Raghavachary: Oh, God

Host: a few weeks. So who are you going to root for?

Saty Raghavachary: Go Trojans.

Host: You're on recording, by the way.

Saty Raghavachary: Oh my God. Yeah, I have to say go Trojans.

Host: All right. There we go.

Saty Raghavachary: Love the one you're with right now.

Host: There we go. All right. There we go.

Saty Raghavachary: Ohio State was a past love. Horseshoe loved it.

Host: Yeah.

Saty Raghavachary: But hey, that's in the past.

Host: That's right. Okay. I love that answer. And then it was funny, I was talking to somebody about you the other day, and they described you as, I told you this outside, a bundle of intellect and joy. What are some of the things you do for fun and to keep a positive joyful attitude?

Saty Raghavachary: I guess, yeah, man, even have the feeling of gratitude, I think is a pretty big thing. I do it on purpose. Every morning you wake up literally tell yourself, "Wow, I'm still alive." Not that I'm old, okay, it's not that. But it's more like really, everything works, and I have a nice home and food to eat. So you feel grateful, and they call that walking around with a full heart. And that makes all the difference in the world. So a small traffic snarl, you wait in the line for a little bit too long for coffee it doesn't bug you, because you're overall thankful.

Host: Right.

Saty Raghavachary: I think I maintain that attitude almost on purpose. And I try to be compassionate as well. So much of the world needs help. So whenever you can lend a hand, it makes you feel good in a certain way. It helps them. So I think those keep me going in a more higher spiritual sense, so to speak. But then for fun, I do lots of EDM. So I love Reaper, and I use a program called MuTools, and many hundreds of small little VST plugins. So I'm constantly making beats, and I love to drum, by the way. I learned drumming when I was younger. So the actual feeling of making music and dancing and beating, it's actually very cool. So I enjoy that so much. And again, same thing. Suno.ai, I don't use it at all. I don't care what it does. I know how to make my own music.

Host: I love that. Thank you so much, Professor Saty. I appreciate it.

Saty Raghavachary: Thank you, Corbin.

Host: Yeah.

Saty Raghavachary: Thank you, Philip.

Host: Thank you so much for joining us today.

Saty Raghavachary: It's fine.

Host: And I think we wrap up this episode. So for our audiences please just follow us, the 18VC podcast, and you can also follow professors with his teachings or as well as his sharings on social media. This would be great. And thank you so much for listening.

Saty Raghavachary: Cool. Go 18VC.

Host: Thank you. Yeah. Thank you. And go Trojans.

Saty Raghavachary: Go Trojans, exactly. Oh, wow