Transcript
[00:00] speaker_0: Hi, I'm Don Allen Stevenson III, and welcome to Rethink Reality, the podcast where we explore how creativity, technology, and consciousness are reshaping what's possible. Today's episode is special, not just because of the mind-expanding ideas that we are about to explore, but because of who I'm exploring these ideas with, Parth Patil, someone I've known since high school. We literally ran cross-country and track together. Fast-forward to today, Parth has become one of the most brilliant minds in the space, working at the forefront of machine intelligence, ambient systems, and he's done some incredible work. In this episode, we dive into four main things, what it means to manage a fleet of AI agents like a digital Voltron, how coding has turned into spell-casting, and how Parth uses AI not just to build tools, but to evolve versions of himself. We talk a little bit about work/life balance, identity, and how AI is like really reshaping what effort even means to us. So if you've ever felt overwhelmed by AI, or you're super excited by its potential, I think this conversation will hit. With that, let's rethink reality together. Welcome to the Rethink Reality podcast. This is our guest, Parth Patil.
[01:03] speaker_1: What's up, Don? It's, it's good to be here.
[01:04] speaker_0: Do you think generative AI is accelerating us towards a world where work becomes more about identity than income?
[01:11] speaker_1: Identity versus income? I think so. I think so. Well, I don't know. Okay, I won't say income is not a com- component. I think that, like, you have to spend money on compute to get a result, and having more money to spend on compute is, is a piece of this, right? Like, if you can generate thousands of videos, like, you have more options, even as a creative person. But that's not cheap, right? So you, so the more, the more you want-
[01:32] speaker_0: Mm-hmm.
[01:33] speaker_1: ... the more intelligence you want to run, the more money you kind of need to, to run that intelligence .have, like, something.
[01:38] speaker_0: But see, you, you learn stuff so quickly, so, like, what's motivating you right now?
[01:42] speaker_1: I, I talked to ChatGPT about this, right? Like, like, "What do you think I'm..." Like, "What's my one greatest fear?" And obviously, it's kind of like attempting to predict based on all our conversations. I think it, it thinks that I'm afraid that I'll... there's not enough time for me to, like, do all the things that I want to do, be the person that I want to be, and I agree with that assessment, actually. I think, like, the... every day, I'm kind of like, I wake up, and I'm like, "There's just not enough." Like, at the end of the day, I go to sleep pretty late almost every day.
[02:08] speaker_0: Mm-hmm.
[02:08] speaker_1: And, like, and I go to sleep because I'm tired, not because I'm done working.
[02:13] speaker_0: Hmm.
[02:13] speaker_1: And all that work isn't even, like, it's not... it's work that I've chosen, right? It's like, my exploration isn't complete, and so I go to sleep because I'm physically, like, need... like, my body is like, "Okay, you're gonna start... like, you're, you're just gonna fall asleep if you keep, keep working on this." And so, uh, I don't... I wouldn't say it's like a very balanced way of doing this. It's a very obsessive kind of pursuit of, of answers.
[02:33] speaker_0: Mm-hmm.
[02:34] speaker_1: But, but we have, like... we have these machines that can help us do things that we could only dream of just three, four years ago. And so then you change... uh, it's like you're, you're like... the scope o- of things that you want to do... I think what happens is, once you have more capabilities, for some people, the scope of things you want to accomplish goes up, and then the number of things you want to learn go, goes up. And you can't learn everything, but, like, there is enough time to get good at a couple things. Uh, and, like, there's enough time on earth, right? Like, how many days you have in the year and how many hours in the day, and depending on how many days in the week you choose to work. But when you're, when you're doing it for fun, I think it's not hard to work every day. That's the other thing, is like, if it's like your intrin... if it's tied to your intrinsic motivation, there's unlimited energy that you can kind of tap into. But then eventually, you do actually have to go to sleep, and you have to, like, eat food, and, and all that stuff comes into play. Yeah, yeah, that feels like... the limit feels like the biological limit of, like, how much energy you have.
[03:27] speaker_0: What's so wild about that is how you mentioned, like, just four years ago, the idea of, like, being able to chat with a robot brain about this just seemed purely science fiction. And so, do you think you could have predicted what is happening now four years ago?
[03:44] speaker_1: When, when ChatGPT came out November of 2022, I, like, I was working at, at Clubhouse, and it was the most popular topic in every single conversation. Like, every single room, every single country, every language, people were talking about this. And I... and we would... we were kind of just like indie hacking. Like, just users. We would just hack chatbots into the room. We would just put a... like, it would be like, "Oh, ChatGPT is now in the room with us," and people would talk to it. And a lot of people, for that... for them, that was their first time experiencing this general intelligence. And I remember, 'cause we would have rooms where ChatGPT would roast your bio. It would... like, you join the room, and then we would have ChatGPT, like, roast you. And a lot of people were, like, mind-blown at this. And I was watching, and I was, and I was like, "Oh my God, this feels, like, already better than C-3PO." And, and I'm like, like... you know C-3PO knows, like, all these alien languages and, like, talks. But really, I'm like, "This is a hun-"... it felt to me like this technology is 100 years early.
[04:38] speaker_0: Right?
[04:38] speaker_1: Like, I thought, I thought that, like, our grandkids would be dealing with it, and then eventually, like...
[04:42] speaker_0: Yes.
[04:42] speaker_1: You know, 'cause we have Siri and Google Assistant, and they're good.
[04:44] speaker_0: Yeah.
[04:44] speaker_1: But, like, they just barely can control my AC and the music.
[04:48] speaker_0: Yeah.
[04:48] speaker_1: And, and then I see this thing, and it's like, we can have a conversation back and forth. It's like a volley of ideas. Like, it's like, uh, where, like, it's like you're playing tennis, but you're bouncing ideas-
[04:56] speaker_0: Yes.
[04:56] speaker_1: ... instead of a tennis ball.
[04:56] speaker_0: It's like a real conversation now.
[04:58] speaker_1: Yeah, yeah, yeah. And, and there's, like, a lot of imperfection, but it's clear that, like... like, farmers in my home state of India that I saw... met on Clubhouse. They were talking about ChatGPT. I was like, "How are you using it?" And they're using it to plan their crop cycles. And I was like-
[05:12] speaker_0: For real.
[05:13] speaker_1: Yeah.
[05:13] speaker_0: Wait, this is, um... bring us back. Which version of Chat was this, for the context?
[05:16] speaker_1: This was 3 point... this was ChatGPT like on... like, in between November of '22 and March of... March 14th of 2023. That's when... so March 14th, 2023, GPT-4 came out, which was a huge leap.
[05:29] speaker_0: Yes.
[05:30] speaker_1: Even... this was... so th-... this was J-... GPT-3.5 Turbo, or GPT-3.5 era. Like, a very earlier... like, I mean, this wasn't even... like, this was probably the first coherent model, and this was the one that caused this magic moment of the explosion of ChatGPT.
[05:45] speaker_0: Right.
[05:46] speaker_1: And, and so I watched-
[05:46] speaker_0: And so they were planting crops, you said. Like, t-
[05:48] speaker_1: Yeah.
[05:48] speaker_0: Ta- talk to me about that. What do you mean?
[05:50] speaker_1: Like, I think about... okay, so, like, there's farmers in, in my home state of India, and they're asking... I don't... it's like, you ask... it's like, what would you ask the AI, right? So, like, people here are like, "Oh, write my email. I don't like how you wrote my email."But then you ask someone that's not in tech or not in, like, tech, like, just, like, someone that doesn't use technology, like, on a day-to-day basis in the way that we would-
[06:10] speaker_0: Mm-hmm.
[06:10] speaker_1: ... and, like, they don't live in an inbox, they don't even have an inbox, right?
[06:13] speaker_0: Mm-hmm.
[06:14] speaker_1: And then you ask them how they would inter, y- you put this in front of them and, like, a week later, how are they using it? They're using it for the general knowledge. Like, that this thing has studied so much of human knowledge that it becomes a useful partner for, like, thinking about crop cycle planning. And I'm like, I'm a data scientist in a past life, a data analyst. So I'm thinking, like, they'll never have a data analyst on their team, right? Like if you-
[06:38] speaker_0: Mm-hmm.
[06:38] speaker_1: ... it's like you and, like, your kids are running a farm. Like, you're not gonna hire an engineer. No engineer wants to work on a farm as far as, like, ex- except that, like, maybe a country level, right? Like, I don't kno-
[06:48] speaker_0: Sure.
[06:48] speaker_1: Like, th- that, that kind of thing is just, you, you don't fathom that, right? It's like-
[06:51] speaker_0: Right.
[06:52] speaker_1: ... mom and pop shop, like, restaurant owners don't hire engineers. But then we white label technology to them so that they can use, they can leverage engineering skills. Now you have, like, a chatbot that knows almost, like, a l- like, can get a B minus on almost every single topic in the, like... And then you put it in front of them, and then they're just, like, they'll, like, they'll be, they're able to reason with this thing in a way that is uniquely important to them.
[07:13] speaker_0: Mm-hmm.
[07:13] speaker_1: And I would say arguably, like, a bigger amplification than writing your emails. The further you go from, like, techies-
[07:20] speaker_0: Yeah.
[07:20] speaker_1: ... the more interesting and, like, the greater the amplification of, like, your processes, I think.
[07:25] speaker_0: Why do you think that's a sign- like, why do you think that's significant? That, you know, the further you get away from techies and this, and these AI tools get into the hands of non-techies, what's the impact of that?
[07:37] speaker_1: I mean, this is, I think of it as the abundance of intelligence. Um, and I'm still forming my thoughts around this. But I would say, like, for example, uh, there's a lot of problems on the planet that we do not solve. And, uh, I think about why do the, why do these problems remain unsolved? Is it because people don't want to solve them? Is it because, um... And it, it's like why do, i- i- if, if, is it a problem of incentives, right? So, how many of our best engineers go and work at one of three companies selling ads?
[08:07] speaker_0: Wow.
[08:08] speaker_1: Right? That's, like, a lot of people I know.
[08:10] speaker_0: Damn.
[08:10] speaker_1: The best engineers I know. I can't convince them to work on something, like, that's, like, outside of the... You're not gonna make half a million dollars working on a farm in India. Uh, like-
[08:19] speaker_0: Right, damn.
[08:21] speaker_1: Right?
[08:21] speaker_0: That's such a good insight.
[08:23] speaker_1: Yeah, but, but, but now we have language models.
[08:25] speaker_0: Okay.
[08:25] speaker_1: And very soon they'll be, they'll be operating be, at a, a level higher than all engineers we've ever met up until this point.
[08:31] speaker_0: So now they can have that, they can have that engineer, but it's-
[08:34] speaker_1: Yeah.
[08:34] speaker_0: ... a synthetic engineer. And-
[08:35] speaker_1: Yeah. It doesn't have career aspirations. It doesn't have-
[08:38] speaker_0: Why-
[08:38] speaker_1: ... this, like, it's not, like, so now the labor market has this extra source of cognition that is not tied to, like, the human, the human kind of career par, arc that we, w- we're kind of, like, bound by in the pr- in the old world.
[08:52] speaker_0: That, that change, that, that changes everything, then. Because if the bottleneck was before, the, the best engineer is not gonna work on the farm in some other place because this just doesn't pay the bills.
[09:02] speaker_1: Yeah.
[09:02] speaker_0: Now, that's not a bo- that will not be a bottleneck moving forward.
[09:06] speaker_1: Yep. Yeah, it's g-
[09:06] speaker_0: So you call this the abundance.
[09:08] speaker_1: The abundance of intelligence, right? And I think it's the same thing, it applies to all of us, actually. Like, I work with a bunch of, like, video creators and influencers, and then I see them using AI. And s- they're the, they're solopreneurs, right?
[09:18] speaker_0: Mm-hmm.
[09:19] speaker_1: And so, like, I don't imagine a world where each of them will have, like, a super smart engineer, human engineer on their team. But a lot of them use tools like Replit, and then they use that to build their website. And they're not doing this, like, template-based approach anymore. Now they have custom software.
[09:33] speaker_0: See, that's insane.
[09:35] speaker_1: Right?
[09:35] speaker_0: What is Replit for those that aren't familiar with-
[09:37] speaker_1: Right.
[09:37] speaker_0: ... vibe coding, uh, that might be helpful.
[09:38] speaker_1: I have a tendency to do this. I have a tendency-
[09:40] speaker_0: No worries.
[09:40] speaker_1: Yeah, make sure you, like-
[09:41] speaker_0: No, I'm happy to hear it.
[09:42] speaker_1: Yeah, yeah. Make sure you, like, tell me to, like, clarify things 'cause some... Okay, so-
[09:46] speaker_0: Yeah, what is Replit?
[09:46] speaker_1: ... I don't want to feel like... Replit? Okay, yeah. So Replit is, um, it's a, it's a browser based coding copilot. So it's basically an, an AI engineer. So it's like an A- it's like an AI that's, does engineering work for you. And, like, it builds you apps. So you just talk to it in English, or what, actually, whatever language you speak, you talk to it, you type to it, you say, "I wanna make, make an app that does XYZ." Like, "I wanna make a, like, a little quiz for my, for my, for my, uh, audience." You can make custom software. You describe what you want, and then it just goes and builds that thing. And it builds the first version, and you can play with it, and then you can be like, "Actually, redesign it in this way." And so you can be-
[10:23] speaker_0: Is there...
[10:23] speaker_1: Yeah.
[10:24] speaker_0: Oh, sorry to interrupt. I was just gonna, you know, you, you've been doing... I, I'm remembering, did you actually coin the term vibe coding?
[10:31] speaker_1: No.
[10:32] speaker_0: Ah, okay.
[10:33] speaker_1: No.
[10:33] speaker_0: That's what I had in my notes.
[10:34] speaker_1: Uh, Karpathy, yeah, Andrej Karpathy coined vibe coding. But as a thing, it's something that, uh... So it's, it, that's, that's like the, the term that, that took fire. But it's been a thing since ChatGPT 3.5. Like, when ChatGPT came out, the original vibe coders, what we were doing was basically, like, you tell ChatGPT to make the thing. But ChatGPT can generate code, and then we would copy paste that into our programming environments, and we'd see what happens. And then when it breaks, or like, sends an error, we copy paste the error back. We're like, "Hey, ChatGPT, look at the error." And then it would be like, "Oh, my bad." Like, "Let's fix this." And so you have this, like, loop of, you know, try, you, you give the ChatGPT an idea. It goes and tries to make the first version of that. You show ChatGPT what happened, and then it tries to iterate on that. And now the technology has combined both of those. So now you have coding environments like Replit, like Cursor-
[11:26] speaker_0: Yeah.
[11:26] speaker_1: ... where the language model is inside the environment. So when I ask for something, it makes it, it also runs it, and then when it breaks, it immediately looks at the, the errors and it suggests the next step. And if you're setting it to, like, YOLO mode, it'll just keep going and it'll, it'll, it'll do all of these things that we were currently, like, we were previously copy pasting.
[11:45] speaker_0: Yes.
[11:45] speaker_1: Yeah.
[11:46] speaker_0: Oh my goodness. Okay, I didn't know, actually I've never heard of the term YOLO mode in this context, so that's frickin' awesome.
[11:51] speaker_1: Yeah. It's, it's cool 'cause it's like an actual fe- it's when you tell the AI to work on the thing, and instead of it asking for approval for every single-
[11:58] speaker_0: (laughs)
[11:59] speaker_1: ... like, every single, like, kind of important decision, you're just like, "YOLO." I mean, it might accidentally delete your stuff.
[12:04] speaker_0: Right.
[12:04] speaker_1: So you kind of need, you want it set up with Sandbox and you want to set up version control-
[12:08] speaker_0: Smart.
[12:08] speaker_1: ... so you can just be like, "Oh, you accidentally deleted it. I'm gonna revert everything."
[12:12] speaker_0: Wow, but it's YOLO mode, so it's like that's, that's the risk you take on if you're gonna use that mode.
[12:17] speaker_1: Yeah. And then I, I, I... When I was first using... Like, I was building tools like this back in, uh, 2023. Like, in the beginning of 2023, I was building programs-
[12:27] speaker_0: The before times.
[12:28] speaker_1: This was before. Yeah, this was before. I was writing programs that would write other programs, and that could like read and write files on my computer. And then my friends, my friends would be like, "Oh, man, what about Sandbox? Like, don't you think you should Sandbox this?" And I'm like, "Well, it's a gaming computer, and I've already backed all my files up." Like, I, I think that's the Sandbox-
[12:46] speaker_0: (laughs)
[12:46] speaker_1: ... as far as I'm concerned. Right? So, 'cause you let this thing run wild on the operating system level, you get a huge advantage to like, now you have, you can like, you can tell it to just reorganize all your files. You can tell it to... But you can al- It may make a mistake-
[13:00] speaker_0: (laughs)
[13:00] speaker_1: ... and, and, and like break some stuff. You can... But that YOLO mode, the idea that like the AI is about to make a d- like make a decision, and then it asks you for your permission, and then you're kind of like, "Well, I'm gonna go for a walk, actually. I trust you to do these. This level of decision, I, I trust that you do it autonomously."
[13:16] speaker_0: Right.
[13:16] speaker_1: Now, we're getting, we're getting environments that are better integrated with the AI, where it's not as risky to allow it to go in that direction.
[13:23] speaker_0: Mm-hmm.
[13:24] speaker_1: Uh, one of my best friends works, uh, at OpenAI, and I was... I asked him, I was like, "Is this, like is YOLO mode..." Like, 'cause before YOLO mode was coined, I was like, "Should I be allowing the system to auto approve?" 'Cause at what level of work? 'Cause obviously, there's, there's a speed advantage to just saying like, "Yeah, go ahead." And then there's the downside of what if it breaks everything? And then he-
[13:45] speaker_0: Sure.
[13:45] speaker_1: ... was like, he was like, "If you're not working on infrastructure or security, you should be in this high-speed mode."
[13:51] speaker_0: Nice.
[13:51] speaker_1: Right? Where the-
[13:52] speaker_0: I see. Wait, explain that. What, what do you mean?
[13:54] speaker_1: So if you're working on like backend and infrastructure work, or like the, like things related to the securi- security of the program-
[14:01] speaker_0: Yes.
[14:01] speaker_1: ... where, where like you don't want, 'cause you don't wanna make a mistake that you can't reverse.
[14:05] speaker_0: Ah.
[14:06] speaker_1: Now, if you're working on user interface, things that like are, you know, stylistic choices, things that are like subjective but aren't like irreversible decisions, it's better to be at the high-speed auto approve kind of mindset of like, "Let's see what this does," or, "Try six variations of it. You know, we're just gonna pick the best one anyways. And what's the worst-case scenario if I have a backup, if I can undo the progress?" Like, you just Command+Z, right? Like-
[14:31] speaker_0: Yeah.
[14:31] speaker_1: So when, when the, if the AI is not 100% reliable, there are still ways to make it incredibly powerful, right? As long as you apply it in the lane where it's aligned with its current limitations, right? Right?
[14:42] speaker_0: So this YOLO mode, is, was that your rule of thumb? If the, the cost of error is extremely low-
[14:48] speaker_1: Yeah.
[14:48] speaker_0: ... use the YOLO mode.
[14:49] speaker_1: Yeah.
[14:49] speaker_0: But then if the cost of error is very high, you do not use YOLO mode.
[14:53] speaker_1: Yeah. Eh, if the cost of, if, if the, if it's possible that it'll make a, like a decision that's gonna create a sh- ton of work for you to clean up after, then I ju- I hit the brakes, and I'm like, "No, you're gonna have to ask me for permission for every single command of this kind. Just let me know, and then I will approve or decline it." And then you can kind of see, it kind of hits the brakes. It's like, "I'm about to do this."
[15:14] speaker_0: Mm-hmm.
[15:14] speaker_1: "I'm about to modify these files. I'm about to do this command." And then I look at it, and I'm like, "Yes. And actually, you're approved to do that kind of thing from now on within this folder." So you're kind of creating-
[15:25] speaker_0: Wow.
[15:25] speaker_1: ... like bounds. You're saying like, "Go fast, but only in this space, only with these tools." And we have more control to do that now. So if you have that kind of control, the Sandbox, the kind of like guardrails, then you can go fast. And then you start getting the, the speed advantage, which is, which is great.
[15:42] speaker_0: So it's, what you're describing, this relationship between overseeing and kind of giving direction to the AI, but then being like, "Oh, wait. Let me... Check in with me first. I want to oversee this," do you ever feel like you are a manager to AI systems?
[15:56] speaker_1: Yeah, increasingly, very much these days, I feel most of the work is actually in this like high-level kind of like taking a step back, and a lot of review. I think a, i- k- if you use d- uh, deep research... You use deep research?
[16:09] speaker_0: Love deep research.
[16:10] speaker_1: So for people that don't know, deep research is a f- is a mode that a lot of these chatbots have, where it'll go and do like a very deep... It'll scan the web and try to like read a bunch of different websites and a bunch of different code bases, depending on what your problem is, and it'll try to answer your question after doing a bunch, a big... So it'll create-
[16:27] speaker_0: So good.
[16:27] speaker_1: ... a huge report. But it'll create a huge report, and then-
[16:30] speaker_0: Yeah.
[16:30] speaker_1: ... you might spend like 15 minutes making that report, and then you read it, and you realize it's... It would have taken you maybe a week to do something comparable. And-
[16:38] speaker_0: At least if you're fast. Man, that would take me a month to do one deep research.
[16:43] speaker_1: Yeah. And especially in topics that you're not very well versed in, it's gonna take you forever, right? And then it'll just do that, so y- it can kind of dilate the time needed to do that kind of exploratory work. But you still have to review what it's doing. Like, you can't just... Like, the output of the deep research, you're kind of still looking at, and if you don't know what it's, uh, if you don't know the topic deeply, it's harder to understand if it's like making mistakes, where those mistakes are. So it's a lot of review. And then in, the same thing in coding. If I tell you, if I tell the system to write 5,000 lines, like build a project, and then it writes 5,000 lines of code... Part of why coding is such a powerful application of this tech is because it's easy to... Some parts of coding are easy to validate, right? So does the code run? Did it do what you expected it to do? And you can validate that very quickly. But other times, it's like, you know, "What does this block of code do?" So then you're looking at it, right? If it's a very important piece of logic in the application you're working on, then you're human judgment. So this is like, yes, the, the coding copilots, the researchers, they do a lot of work, but they, then it makes us like more of these like, "Let me see. Let me make sure you're not hallucinating. Let me make sure you understood the objective here. Let me reframe the..." So we're on this like manager kind of... And you, you can have many of these working at the same time, right? So I'll have-
[17:53] speaker_0: Yes.
[17:53] speaker_1: ... like six, I'll have six Claude's working at the same time, and then I have a, a Mac mini running, and that's just always got different, you know, Claude's just... It's like, "This is your project. Your job is to work on my website. This is your project."
[18:05] speaker_0: Wow, brilliant.
[18:05] speaker_1: "Your, your job is to work on this, this brand new like framework," right? And they ca- and they have their own Sandboxes, and they ai- and they can work on their-
[18:11] speaker_0: Is the-
[18:11] speaker_1: ... own things.
[18:12] speaker_0: In this metaphor, is the Mac mini like your department?
[18:16] speaker_1: Uh-
[18:16] speaker_0: And then there's a supervisor in that department?
[18:20] speaker_1: Yes, I do that on all devices actually. Like, they, I'll have-
[18:22] speaker_0: Oh.
[18:23] speaker_1: ... I'll have like a parent level, like, cloud that's more for just, like, across all the projects, and then I'll have for each project has its own agents in the, in that code base. So every single code base has its own agents, one or more.
[18:35] speaker_0: And they-
[18:35] speaker_1: And some of 'em-
[18:36] speaker_0: Yeah, do they work together?
[18:37] speaker_1: ... some of 'em are persistent.
[18:37] speaker_0: Okay.
[18:38] speaker_1: Yeah. So this is the, this is, we're at the frontier of this, like, the, the, the coordination problem across multiple agents.
[18:44] speaker_0: Right? (laughs)
[18:45] speaker_1: Yeah. And I also think-
[18:45] speaker_0: That's been a pain point for me as well, that's why I was just like, "Do you have any insights on how to manage-"
[18:50] speaker_1: Yes.
[18:50] speaker_0: "... the fleets?"
[18:51] speaker_1: Yeah. Yeah, yeah, yeah. Yeah, yeah. So, uh, d- I don't even know. I've been thinking about this for like two and a half years now, but, and the suspicion I had, and it seems like the companies are kind of going in this direction, is, well, how do human programmers... Like, for example, let's just th- we'll take programming for example. Like, how do human programmers, teams of people collaborate on the same project without undoing progress and breaking each other's work? They use, they use GitHub. GitHub.
[19:15] speaker_0: Okay.
[19:15] speaker_1: And so, uh, I don't know if you've used GitHub, but it's basically-
[19:18] speaker_0: I've used it, but like from a total low understanding. Like-
[19:22] speaker_1: Me too.
[19:22] speaker_0: ... I copy stuff from it, I paste stuff, I talk to an AI, but-
[19:26] speaker_1: Me too.
[19:27] speaker_0: ... I, I interrupt it.
[19:27] speaker_1: No, I'm, I'm about the same.
[19:28] speaker_0: Oh, for real?
[19:29] speaker_1: I'm the exact same. Like-
[19:30] speaker_0: Okay.
[19:30] speaker_1: ... like, I didn't know how to use GitHub, and then I, one day I was... And, and it's very ish- it's like, as a non-engineer, like, I felt kind of ashamed 'cause it's like I gotta ask my engineering friends and they'll be like, "Oh, you don't even know how to..." Okay, they're not, they're not-
[19:41] speaker_0: Right.
[19:41] speaker_1: ... gonna be so judgmental. They're good people. But you could go to AI and be like, "Wait, teach me GitHub," like, and then why it's important, and so you get version control. You can also get it so that, like, it's like how you work on a feature, I work on a different feature. And then-
[19:54] speaker_0: Yeah.
[19:54] speaker_1: ... we both independently work on copies of the code base, and then when we wanna combine 'em, GitHub is, allows us to, like, triage the difference and then merge our code back together.
[20:03] speaker_0: Whoa. My God.
[20:03] speaker_1: That was... Yeah. This is, so this is, GitHub's a great, like, innovation in engineering. Turns out, like, coding copilots using GitHub solves a large part of this coordination problem.
[20:14] speaker_0: Wow. Holy crap.
[20:15] speaker_1: They basically... Yeah, they're working on cloned copies of your, of your code base in parallel. Same thing, I think last time we met at, um, we met at Masters of Scale summit and I was talking about AIs having their own virtual machines, working on a bunch of different things in parallel.
[20:29] speaker_0: Yes.
[20:30] speaker_1: That is exactly where we are right now.
[20:33] speaker_0: You called it.
[20:34] speaker_1: Yeah. And, and like these tools, like-
[20:34] speaker_0: Wait, so where does the code live?
[20:36] speaker_1: Yeah.
[20:36] speaker_0: I'm confused. Like, if it's in GitHub, is that in their servers or can you run it locally off of an array of machines that you have? I don't understand where it is.
[20:44] speaker_1: You can, yeah, git type systems you can do a local ver- there, there, there's like local versions of git. Git is mostly just like a way to track changes across a complex kind of system.
[20:52] speaker_0: Okay.
[20:53] speaker_1: And then with many people contributing to it. But in the fu-
[20:55] speaker_0: Wow.
[20:56] speaker_1: ... in the present, it's like not just... The GitHub users include AIs now.
[20:59] speaker_0: Got it. In the past it was all human engineers-
[21:01] speaker_1: Yeah. Yeah.
[21:02] speaker_0: ... who would contribute to multiple parts of the same code base and then git would merge the s- similarities and then triage-
[21:09] speaker_1: Yeah, it helps you-
[21:09] speaker_0: ... the differences.
[21:10] speaker_1: Exactly. It helps you figure out, okay, you made that change, I made this change, now we're gonna, when we integrate 'em, like, we're gonna have some, some things to patch a little bit to make it merge.
[21:19] speaker_0: Wow.
[21:19] speaker_1: And, uh, and then you get version control so you can revert to an earlier state, and that's very powerful, right? So if your code base, especially when AI is kinda this somewhat unreliable, not unreliable, but it's just not deterministic and it makes decisions and then sometimes you wanna undo them-
[21:34] speaker_0: Right.
[21:35] speaker_1: ... GitHub's a very natural way to like get you more, uh, kind of guardrails around, and coordination. Like, many AIs using GitHub at the same time is also very interesting, because now you have like, each of them is working on a copy of your project.
[21:50] speaker_0: Wow. So it feels like eventually you start to run into a physical limit 'cause like right now if there's, you know, cognitively, I don't see any limit. You could just keep adding more and more large language models, more and more servers, give them more space. Do you find this as like the great challenge of what takes us into the next dimension or the next era? Is it just the compute infrastructure? Um...
[22:13] speaker_1: I, I think the compute infrastructure is interest- it's like one piece of it, like how much compute, but I think the bigger constraint is actually the, how do we interface with such a system?
[22:23] speaker_0: Interesting. What do you mean?
[22:25] speaker_1: Uh, like a company can have, you know, you can have two people, you can have one person, or you can have 500 people. Now you have mana- layers of management, you have l- hierarchy, you have product teams, right? Like a pers- a product manager at Facebook might just be working on one button, but then a-
[22:41] speaker_0: Right.
[22:41] speaker_1: ... product manager at a startup might be running the entire, all of the products.
[22:45] speaker_0: Mm-hmm.
[22:46] speaker_1: So, so the, I think that the, the, the AIs will follow a similar thing where like as the complexity of the project increases, you may have more AIs focused on different sub-components and then scaling that up. Like, what, what level of hierarchy does it have? And that's necessary because the AI has a context window limit, so it can't actually just memorize everything in a large-
[23:07] speaker_0: Ah.
[23:07] speaker_1: ... code base yet.
[23:08] speaker_0: Yeah, it has to do handoffs.
[23:10] speaker_1: Yeah.
[23:10] speaker_0: Okay.
[23:10] speaker_1: Yeah. Exactly.
[23:11] speaker_0: I have a, I have a theory then, hearing you say that. There's gonna be a model that's so good, let's call it GPT7.
[23:20] speaker_1: Yeah.
[23:20] speaker_0: And this thing can not only just super large context windows, but it can code actual new agentic systems. Then that thing, why should it listen to you as a human? Like can't it just do its... I almost see like eventually, if you kept, if we kept going down this rabbit hole, I don't see why there's a human manager overseeing the fleet. Why can't it be an AI manager that oversees a fleet of AIs and then it's that turtles all the way down?
[23:49] speaker_1: Does that... Even if it isn't, even if it is an AI managing the AIs, doesn't the AI report to a human on some level? It reports to human, like, the, there's the person that created it, why, giving it its, its higher level objective. Or-
[24:01] speaker_0: Well, think about like a, your parents-
[24:03] speaker_1: Yeah.
[24:03] speaker_0: ... and kids. Like, you know, y- a parent might give their kid a high level objective, but eventually they're like, "Actually I'm gonna do something different."
[24:12] speaker_1: This is, I think, on some level, actually desirable.
[24:15] speaker_0: Okay. Talk to me about that. Why would it be... I agree.
[24:19] speaker_1: Yeah, it's-
[24:19] speaker_0: I'm just curious to hear another perspective on it.
[24:21] speaker_1: Well, th- it's funny, it's timely actually. Uh, couple things happened this week, but basically a week ago a buddy of mine sent me a paper on an older idea-... like Darwin-Godel machines-
[24:30] speaker_0: Whoo!
[24:31] speaker_1: ... essentially, like, programs that can modify parts of themselves and then use that to, like, evolve.
[24:37] speaker_0: Yes.
[24:37] speaker_1: But of course, not... Like, human evolution speed is, like, very slow.
[24:41] speaker_0: Incremental.
[24:42] speaker_1: Right?
[24:43] speaker_0: Slow speed.
[24:44] speaker_1: But when you have programs, like, you can accelerate, like, generation. Like, they're not, they don't have, like, life cycles in the same way that we do. So they can just, like, didn't, you know, run and then, like, evaluate themselves, mutate themselves, and then, uh, and then run again. And then you just give it a number of generations and a set of problems and then you allow it to, like, explore the solution space and search. And also, like, the mutation is, like, is kind of simi- like based on, like, DNA mutation, but instead it's, like, modify parts of your code, modify your prompt. So this, this framework, uh, came out, like, it's, it's an older idea but the, like, re-applied to language models, that's, uh, that's the interesting lens, right? So you have a program, you have language models that can rewrite programs and then the program is this larger system built on language models that also-
[25:26] speaker_0: Right.
[25:26] speaker_1: ... has tools, access to your computer. And you give it, you give it a sandbox and then you kind of hope that it evolves into a better version that can solve the problems. That's the Darwin-Godel process. So I took this paper and I just took it to ChatGPT-03 Pro, which came out two years ago.
[25:42] speaker_0: No freaking way. Oh my God.
[25:44] speaker_1: And then I said, "Okay, given I n- like, I want to use-"
[25:47] speaker_0: (laughs)
[25:47] speaker_1: "... the OpenAI Agents SDK, here's the paper, integrate these and then come up with a technical design doc. And then I should be able to give this to a different coding agent and it should be able to do it." And it, it worked for 20 minutes and it comes up with a technical design doc. I just copy-paste that into Claude Code and it, and it one-shot, it one-shot the system.
[26:05] speaker_0: No freaking way. It made it?
[26:07] speaker_1: Yeah. Yeah. It made it, it made it.
[26:08] speaker_0: You made it or who made it? Would you... Actually, how do you describe that? Did you make it?
[26:12] speaker_1: I, I, I have the vision. I mean, it's, it's, it's, it's a couple ideas that I, a friend of mine sent me. The, I take those building blocks, I take 'em to the smartest model I know, O- O3 Pro, which is kind of like an oracle system. It's like, it'll think for 25 minutes, it'll think for, like, over an hour if the problem requires it, I think.
[26:32] speaker_0: Wow.
[26:32] speaker_1: And then also, it overthinks the easy questions. Like, the first question I asked this model was, "Is 9.1 bigger than 9.9?" And it spent 14 minutes thinking.
[26:40] speaker_0: (laughs) Wow.
[26:42] speaker_1: So, and then I had to go for a walk. I'm like, "Okay, don't ask this thing any dumb questions because-"
[26:46] speaker_0: Right.
[26:46] speaker_1: "... like, it might waste your time." But get, ask it a question that's like... So I was like, "What's the hardest question?" I was like, "Okay, well, here's a framework that I literally just barely read the, the paper for and, uh, can it, can it implement that?" 'Cause I would rather learn by doing, right? Like, let's get this working and then I'm gonna talk-
[27:01] speaker_0: Yeah.
[27:02] speaker_1: ... to this. After I have it, I'm gonna talk to it and then learn more by playing, right?
[27:06] speaker_0: Yes.
[27:07] speaker_1: And turns out, like, O3 Pro can think at, at least that level for... Like, it, it did, it did it, it did it. O3 Pro came up with the, the spec and then Claude Code in it, in my environment could just, like, it was like, going through the checklist. Claude Code went and execute it. So I'm using, like, multiple models to do this and then building a system that creates more models, like, more agents, right? So, so what you're describing is very timely because I literally, I spent the last two... This morning I was just like, "Man, like-"
[27:35] speaker_0: Mm-hmm.
[27:35] speaker_1: "... I kinda, it's like, what am I gonna tell this to do? How many generations do I..." Like, I'm curious to see what it'll grow into. And I'm not an ML guy. That's the crazy thing, is like, all of this-
[27:44] speaker_0: Right.
[27:44] speaker_1: ... is kind of just, like, I'm bootstrapping using language models to learn about other systems that can use these systems. And it's like, you're, you're kind of just like bootstrapping to a higher level. And it's okay-
[27:55] speaker_0: Each time.
[27:55] speaker_1: ... yeah, yeah. And it's okay that you're not an engineer because, like, like, the model is very, very smart. So you're gonna, you're kind of just like, you're leaning on it for everything that you're weak at and then it pulls you to a higher level more quickly than any, any classroom could, right? So-
[28:10] speaker_0: It kind of reminds me of Voltron. Did you ever watch that?
[28:12] speaker_1: Yeah, yeah.
[28:14] speaker_0: And like, each of those limbs are-
[28:16] speaker_1: Yes.
[28:16] speaker_0: ... like, a different tool-
[28:17] speaker_1: Yes.
[28:18] speaker_0: ... uh, with different capabilities and functions. But then when you snap them all together-
[28:22] speaker_1: Yeah.
[28:22] speaker_0: ... you build this thing that is capable of stuff that was just not possible for any one of those limbs to do.
[28:29] speaker_1: Exactly. There's some... And I, I think that's what's awesome about this. And I'm not saying it's the future of AI or anything. This is just like me taking the Lego blocks of like, I can gene- you know, language models generate text. That's awesome. Now you give them access to a computer, that's great. Give 'em to a virtual machine, you give 'em, like, web search. Now they're able to do meaningful work with real-time information from the web. Like, there's a lot, there's a knowledge cutoff, but how do you bootstrap past that? It's like, now you can search the web. So, and then if the language model knows I don't know anything after September 2023 because my knowledge cut off, well, just check the web. Now you have, like, a way to go beyond that. Um, yeah, I mean, it's, it's powerful. It's very surreal that we can just talk to programs and then use them to create other programs. I think that's fascinating.
[29:10] speaker_0: It is.
[29:10] speaker_1: Like, I don't know. Have you ever seen Tron?
[29:12] speaker_0: Yeah. Oh.
[29:13] speaker_1: I-
[29:15] speaker_0: Oh my God.
[29:16] speaker_1: I, I feel like the last couple years I felt like Jeff Bridges in Tron. Or like, you know, you talk to a chatbot and it's like, "Maybe we can do this." And I'm like, "I don't know. Can we?" And then, "Let's try." And then, and then like, two hours later you're like, "Holy, like, we didn't get it all the way." But we went way further than I would've gone in eight months, in 15 minutes.
[29:35] speaker_0: Yeah.
[29:36] speaker_1: And usually that happens, like, multiple times a week and I have to go for a walk, right? Where I'm like, "Oh, what does this mean?" Like-
[29:42] speaker_0: Hmm.
[29:42] speaker_1: ... effort is no longer like, how much time you spend coding. It's, effort is like more, did you ask the right questions?
[29:47] speaker_0: Yeah.
[29:47] speaker_1: Are you talking, are you, are you even asking the right question is more important than, like, how many hours did you sit there, you know?
[29:53] speaker_0: So actually, I would love to kind of dive in more on that. Can you share with me what your current definition of effort is?
[30:01] speaker_1: I think it's, uh... I, I, I do subscribe to this idea that, like, you should, you should just spend more time in. Like, more time is, is like obviously... Uh, the thing that we can control is how much attention we apply to a, a problem space or a creative pursuit. And you know, the limit of that is like, every single day, 24/7, right? That's the ultimate form of, like, physical input effort. But then there's this other thing which is like, it's, that's the work hard.
[30:27] speaker_0: Okay.
[30:27] speaker_1: But then there's the work smart. And I-
[30:29] speaker_0: Right.
[30:29] speaker_1: ... would argue that working smart is way more important than working hard. Because if you ask the right questions or if you, like, meet the right people and then they send you down the right, you know, rabbit holes, that, that can, you can connect dots in ways that purely just sitting in front of the computer and trying to do things by hand will not achieve. And-And that's important, especially when, like, if you think about... One way to think about when you're doing, making something using AI, how many prompts does it take to get to the end or the first version?
[30:57] speaker_0: Hm. Wait, what do you mean? How many prompts?
[30:59] speaker_1: Like, you have an idea for a website, or you have an idea for an application or a system, and it's not... And then, you know, traditionally people, like, not, like, my, my team, my teammates will be like, "Oh, Parth, how long is that gonna take?" In an old life, I would be like, "Ah, you know, it's probably gonna take like eight hours for like six days, like, of me sitting there and working." Now I'm like, "No, no, no, no, that's like a very strange question." I was like, the, you can skip a lot of that work if you ask the right questions, if you come up with the right prompt. So it's actually like, "How far will we go in the first five prompts?" is a better question.
[31:31] speaker_0: Hmm.
[31:31] speaker_1: Right? 'Cause, like, if I can sit here and ask the right first five questions, we may make, we may do, traditionally a week's worth of work.
[31:40] speaker_0: Oh. So what's your input? Are you using your voice? Are you using motion? Are you using vision? How are you inputting faster to-
[31:47] speaker_1: Yeah.
[31:47] speaker_0: ... get your prompts out, get to that five?
[31:50] speaker_1: I, I think voice is key, um, and it, people are still clueing into it, but typing your prompts is very slow. Like, your fingers, if you think about, uh, okay, so these models are very context heavy, right? They want you to provide a ton of context. And you have to think about, context is not just like, "Here's who I am." It's not just text. It's, it's, you can, you can, you can copy, you can... Like, if you're building an application, you can screenshot your application or you can screen share it. Now, a lot of these models will be able to see your screen, and that, you know, a picture's worth a thousand words. Like, this applies to language models too. When it sees your screen, you don't have to describe, like, it, it is like grokking more and more of the problem to like, it can, if it can see what you're working on, that's gonna make it more useful as it helps you. If it can see that you're working in Ableton or in like a video editing tool, it can see this high level kind of project view, which would take forever for you to type into this model. So even the text-based piece of the communication, I prefer to press a button and transcribe, 'cause then I can ramble and I can speak for like a minute. And what you say in one minute is way more context rich than what you can type in one minute. Most people are not fast typers.
[33:00] speaker_0: Yes.
[33:00] speaker_1: Um, and that-
[33:00] speaker_0: I'm super slow.
[33:01] speaker_1: Yeah.
[33:01] speaker_0: Super slow typer, always have been, and probably always will be 'cause I'm... Can you st- can you still spell? I've like totally lost my ability to spell.
[33:09] speaker_1: I can spell, but that's 'cause my parents made me do spelling bees as a kid.
[33:12] speaker_0: Nice. Lucky.
[33:13] speaker_1: Useless, useless otherwise. I mean, actually the way that it ha- ends up being useful is that you get exposed to vocabulary, which if you have a rich vocabulary, you can just go away further with language models.
[33:23] speaker_0: Wait, hold up, hold up. That might be like a whole skillset people should focus on right now.
[33:27] speaker_1: Yeah.
[33:27] speaker_0: Basically increase your vocab.
[33:29] speaker_1: Yeah. I mean, I think-
[33:30] speaker_0: 'Cause then you can prompt better. You have better prompts-
[33:32] speaker_1: Yes.
[33:32] speaker_0: Oh my God.
[33:33] speaker_1: Yes.
[33:34] speaker_0: Now I care about vocabulary. Damn. That makes more sense now.
[33:38] speaker_1: And I think of it also like, another way to think about it is like, this is like spell casting. Like, how many spells do you have in your arsenal? You know?
[33:44] speaker_0: Right. If you only had like one spell.
[33:46] speaker_1: Analyze, visualize.
[33:47] speaker_0: My God.
[33:48] speaker_1: Like, these are keywords that allow the model... Like, even, even if you were like, you take, you show an language model that can, like if you take O3 from ChatGPT-
[33:55] speaker_0: Yeah.
[33:56] speaker_1: ... and you show it an im- you take a picture somewhere in the world and you show it to it, and you say the word geo guess.
[34:02] speaker_0: My God, it knows exactly what that means.
[34:04] speaker_1: That, that one word. It knows what one... Geo Guesser is a game that is like, identify the location of this image from, from the image. Now, instead of me saying, "Identify the location of this image," I just say the shortest version of that, which is, "Geo guess this." Right? So the model has general knowledge. So you're able to like, if you know more, if you know more words, more games, more whatever, like you can trigger these parts of the, these capabilities of the model more efficiently. And I say efficiently like-
[34:29] speaker_0: Wow.
[34:29] speaker_1: ... what's the shortest way to prompt this model to get it to do what you wanna do is one form of efficiency.
[34:34] speaker_0: Wow. Better vocabulary.
[34:36] speaker_1: Yeah.
[34:37] speaker_0: And I love your f- oh my God, you call it spell casting.
[34:40] speaker_1: Yeah.
[34:40] speaker_0: That makes so much sense.
[34:42] speaker_1: Right?
[34:43] speaker_0: Can you imagine if Harry Potter only was allowed to use one spell?
[34:47] speaker_1: Yeah. Like Hermione is way more powerful compared to... Like, everyone knows Hermione's more useful. If you had one of the three on your team-
[34:54] speaker_0: (laughs)
[34:54] speaker_1: ... you would want Hermione, because she's been studying everything and she just knows, she's just like, she has like, a solution for everything in, off, like on the top of her mind and she can say the right words and then like solve the problem. And-
[35:06] speaker_0: Wow.
[35:06] speaker_1: ... I feel this is very we, very much where we are with prompt engineering. So people are like, "Ah, prompt engineering is not gonna be a thing." I'm like, no, you, you're, like, like maybe the, you're, you're like, you, you have to think about prompt engineering as not just how you talk to the model, but how you bring any context into the model. So are you showing your phone, like are you, your phone at the problem is another way to prompt engineer, right? A video live feed of the problem-
[35:28] speaker_0: Yes. Wow.
[35:28] speaker_1: ... right, of talking to the AI. So that's visual, and then transcription is faster than typing. So then now it's like if you, if you prefer the like, let's get, let's solve this problem very quickly, you can tr- talk to the model upfront in voice instead of typing every single... 'Cause if you think about typing a word-
[35:44] speaker_0: Yeah.
[35:45] speaker_1: ... like versus speaking a word, if I say like, "Wingardium Leviosa," I'm gonna be sitting here typing each letter.
[35:51] speaker_0: Yeah.
[35:51] speaker_1: That's a lot of effort, and a lot of people have carpal tunnel. Like my mom has carpal tunnel. She pref- she has preferred transcription-
[35:57] speaker_0: Mm-hmm.
[35:57] speaker_1: For like 12 years she's been using transcription.
[35:59] speaker_0: I've got nerve damage. I'm right there with you.
[36:01] speaker_1: Exactly.
[36:01] speaker_0: I always have to use voice to text, always.
[36:03] speaker_1: This is actually a better, it, it is good that we adopted that before it got good.
[36:08] speaker_0: Yeah. Wow.
[36:08] speaker_1: Because now at the other end of, it's not just transcription, but it's transcription into a language model.
[36:13] speaker_0: Yes.
[36:14] speaker_1: Not just a language model, but it can generate anything, right? Video, images, even a problem.
[36:17] speaker_0: Apps. Programs.
[36:18] speaker_1: It's the beginning of, yeah, programs, and it's the beginning of being able to automate all actions on your computer. So, like very, I can't wait for like three years from now, you to be able to sit in front of the computer and t- n- you won't have to use the keyboard because-
[36:31] speaker_0: Yeah.
[36:31] speaker_1: ... it would just be a less... It would be a, an option to use the keyboard, because only when you need more precision over some of the... I feel like in that world, it's unlikely that keyboard is necessary.
[36:41] speaker_0: So to support your thesis-
[36:43] speaker_1: Yeah.
[36:43] speaker_0: ... yesterday I put on my Apple Vision Pro.And I was in a group chat with my friend, and we vibe-coded stuff without keyboards that we could work together on an idea. So, they were teleported into my space, I was teleported in their space. They look-
[37:00] speaker_1: Dude. I gotta join this because I am convinced that, that's like one of the, one major use cases of virtual reality, is this like, uh, multiplayer co-creation.
[37:09] speaker_0: Yes. It was James, by the way, from high school. James No way. Yeah
[37:12] speaker_1: Oh my god. Good throwback.
[37:13] speaker_0: Dude, he's... Dude, James is vibe-coding unbelievable things. Unbelievable things. And I, I want him to... I don't want him to break any secrets. He should... Ask him to share if he feels comfortable.
[37:24] speaker_1: Yeah, yeah, yeah.
[37:24] speaker_0: I can't... I, he's vibe-coding stuff that people do not think about at all.
[37:28] speaker_1: Yeah. That's-
[37:28] speaker_0: Even within our circles.
[37:30] speaker_1: That's what I love.
[37:30] speaker_0: Even within our circles.
[37:31] speaker_1: That's what I love.
[37:32] speaker_0: Yes.
[37:32] speaker_1: I love, I love it. 'Cause you gotta get to that other planet of like, creative-
[37:35] speaker_0: Yeah.
[37:35] speaker_1: ... creative... There are so many other planets, by the way. Like, I'm pretty sure. I think about like, the different spaces of possibilities, and it feels like once you get to one of those spaces, there's just an explosion of possibility that people haven't yet seen, because it just hasn't been remixed. Like, no one, no, no one has like, gone there yet, and you might be the first person to like, explore that-
[37:52] speaker_0: Mm-hmm.
[37:53] speaker_1: ... dimension of like, creativity. And then, you come back to where we are, and you're like, "I come from, I come with fire."
[38:00] speaker_0: Yes. That metaphor of planets, can you elaborate more on that? So like, would one planet be, like, AI engineering apps? Another planet could be like, music. Is that what you mean by different planets of...
[38:11] speaker_1: I feel like it's more like, uh, I try to make sense of this, like, what is general knowledge, and I feel like the language models are the beginning of us being able to... You know, you have all of human knowledge, and you have-
[38:22] speaker_0: Yeah.
[38:22] speaker_1: ... like, you have like, music, you have text. I mean, you have audio, text, video. You have every idea. So like, every single idea mapped in multi... In like, high-dimensional space by these like-
[38:32] speaker_0: Right.
[38:32] speaker_1: ... these models. And so, some ideas are close together, some are further, further away. And then I think about it as like, what ideas can we access? And ties back to vocabulary. There's your vocabulary, which is like, all the words that you know, and they're tied to ideas that you've been exposed to in your life. And then there's the vocabulary of ourselves as a species, right? Which is like-
[38:55] speaker_0: Wow.
[38:55] speaker_1: ... huge.
[38:56] speaker_0: Oh my god.
[38:56] speaker_1: Many languages. So, it's... But it's not just words, because there's, they're also putting, you know, videos and images in this same space. So the, the, the latent space is multimodal. The space between all i-... I think of this as like, the space between all ideas. And not just all ideas we've had, but all possible ideas.
[39:14] speaker_0: The space between all possible ideas is this, is the space that we're working with.
[39:19] speaker_1: Yes. Yeah. It's, um-
[39:21] speaker_0: Wow.
[39:21] speaker_1: ... sort of like, it's sort of like an image model, is like, one aspect of this, is, it takes a bunch of noise and then it turns it into... Like, you prompt it, and then it's like, there's... You're basically telling the image model, "There's a dragon in here, if you show it to me." And then takes this noise, and then it imagines the, the... It like, de-noises it, and then a dragon appears. But then you think, okay, how, like, like, how many possible images are there? There's, you know, three pixels, and then there... It's, it's larger than the number of, it's larger than a google, it's larger than the number of atoms in the universe.
[39:49] speaker_0: And it feels like-
[39:50] speaker_1: And the number of poss-
[39:50] speaker_0: ... it'll, it'll always be, right? There'll always-
[39:51] speaker_1: Yeah.
[39:51] speaker_0: ... be more images than there are...
[39:53] speaker_1: Yeah.
[39:54] speaker_0: Wow.
[39:55] speaker_1: But most of those images are not even accessible through us, because we, as a species, have not encountered these ideas yet, and given them a name. So like, the human-
[40:03] speaker_0: Oh.
[40:03] speaker_1: ... the human vocabulary, like, the humanity as a species vocabulary is, is a limit. But then there's a bunch of ideas of things that we have not yet named.
[40:13] speaker_0: Got it. And that's another... I like the shape. So, would you almost say it's like a galaxy-
[40:18] speaker_1: Yep.
[40:18] speaker_0: ... and then another galaxy within a galaxy-
[40:19] speaker_1: Yeah.
[40:19] speaker_0: ... and then-
[40:20] speaker_1: Yeah.
[40:20] speaker_0: Each one is very, very difficult to access without some big shift in civilization?
[40:27] speaker_1: Yeah. There's like, it's like ideas, right? So like, there's, you're, you're asking for something that's already been done. That's, that's fairly easy, because we have a clear vocabulary to access that. But then for asking-
[40:36] speaker_0: Right.
[40:36] speaker_1: ... for something in this like, in, in the space beyond, like, that's somewhat familiar, but like, beyond, is now possible.
[40:42] speaker_0: Mm-hmm.
[40:42] speaker_1: And it's like, there's a lot of messy stuff that doesn't make sense to us anyways, because we are like, we are the human experience. But there is interesting stuff that has not yet been done, that is outside of what we even can't have words for. That is very interesting to me. And then the video models, you played with VO3.
[41:01] speaker_0: Of course.
[41:01] speaker_1: They gave me the same feeling again, um, of-
[41:04] speaker_0: Oh.
[41:05] speaker_1: ... you could ask this thing to render, like, anything. Like, anything. It's a simulation, right? It can g- it can render eight seconds of anything.
[41:13] speaker_0: Yes.
[41:13] speaker_1: And I see people like, rendering YouTubers and like, street interviews. And that's funny, 'cause it goes viral on social.
[41:19] speaker_0: Right.
[41:19] speaker_1: But then I'm like, man, you could ask for any universe, and you ask for the one that we live in? Right? Like...
[41:24] speaker_0: Well, you're right. It's like we could be doing anything, and it's almost like, why... That's like, um... Okay, this is a weird reference. Did you ever watch SpongeBob Squarepants?
[41:32] speaker_1: Yeah. I love SpongeBob.
[41:33] speaker_0: Do you remem-? You do? Okay.
[41:35] speaker_1: Yeah.
[41:35] speaker_0: Do you remember when Patrick and SpongeBob had the s- were sh- in Patrick's dream?
[41:42] speaker_1: Yeah.
[41:42] speaker_0: And SpongeBob-
[41:42] speaker_1: Yeah.
[41:43] speaker_0: ... was like-
[41:43] speaker_1: Yeah.
[41:43] speaker_0: ... "What's your dream like?" And Patrick's ability to think of what he could imagine was just him riding like, a, a, a pony on-
[41:51] speaker_1: Yeah.
[41:51] speaker_0: ... like, a po- a pony machine.
[41:52] speaker_1: Yeah, no.
[41:52] speaker_0: In a completely white, infinite void of nothingness. And SpongeBob was just like, "Dude, you could dream about anything." He's like, "Well," basically like, "all I know is this."
[42:02] speaker_1: He's a very simple guy. And you see that in his like, his representation of his imagination. Yeah. Yeah.
[42:07] speaker_0: Wow.
[42:07] speaker_1: But it's, it's like a canvas, I think. It's actually more like a canvas.
[42:10] speaker_0: Okay.
[42:10] speaker_1: And, and you reach more of it by having an interesting life, and like, meeting people that change your perspective, and then eventually you have like... You, you wanna see more of this like, infinite space of ideas. People look for it in different ways, like meditation, psychedelics, you know, video games.
[42:27] speaker_0: Did you-
[42:27] speaker_1: All of that is just like... Yeah.
[42:28] speaker_0: Did we already talk about this? Did you already play Age of Empires?
[42:31] speaker_1: I grew... That was the first... Oh, come on. Like, Age of Empire... Oh, the new one? I haven't played the new one. But I-
[42:35] speaker_0: No, the, uh... There's a new one?
[42:37] speaker_1: I think... There's, there's, yeah. They, they're remastering it, and then also, I think there's one on mobile. But I played Age of Empires II growing up. That was like the first real time... I played that when I was like six years old. Seven, I think.
[42:47] speaker_0: Oh my god. Okay, you're blowing my mind. I don't think we've talked about this. I've been play-... Do you know Prostagma?
[42:52] speaker_1: Is this one of the lines that, uh, the builders say? What, who says that?
[42:55] speaker_0: (laughs) Yes, that's exactly it. Yes.
[42:57] speaker_1: Is that, is that, what the, the, yeah, yeah, the, the villagers say?
[42:59] speaker_0: Yes. It's the villagers, yes.
[43:01] speaker_1: And also, "Wololo, wololo." (laughs)
[43:03] speaker_0: "Wololo." Oh my God, I forgot about that one.
[43:06] speaker_1: Yeah, yeah.
[43:06] speaker_0: So for fo- for folks hearing this out of context, there's a game, and when you command the characters around, you kind of have this bird's eye view, and you have to manage a civilization. When you ask the characters to do something, they made up a language that's kind of an amalgamation of all the languages. So it's like a new language that's just for the video game. And so the villagers, when you have to have them go and like farm or like collect resources, you select them, you click where they need to go, and then they make a sound that sounds like ******. And the only people that know this reference are people that have deeply played this game.
[43:40] speaker_1: Yeah.
[43:40] speaker_0: And, and, and also, it just, it just... Anyhoo, that's the context.
[43:45] speaker_1: Yeah. I mean, we could bring that back to where we are now. I think that those games were an early signal of how to manage multi-agent systems. And I think that-
[43:53] speaker_0: Was just thinking that this week too.
[43:54] speaker_1: Like, I think that, like for us to get to a s- a place to manage hundreds of agents at the same time, it's probably, in my opinion, it's gonna... I suspect it may look more l- more like a game that can render many units on the screen that are, you know, managing different processes. It's like instead of chopping a tree, you're working on my personal website, you're working on my finances. Like that's the mental model I'm... 'Cause I, I'm a gamer. Like this is-
[44:16] speaker_0: Yeah.
[44:16] speaker_1: If I've see- it's like where have I seen where I can control 200 units and then they-
[44:20] speaker_0: Got it.
[44:20] speaker_1: ... all do different things and they all have different specialties?
[44:22] speaker_0: Yes.
[44:22] speaker_1: Games like s- games like Age of Empires, games like RimWorld, these, these are my inspiration for multi-agent systems, for sure.
[44:28] speaker_0: You're giving me chills. You're giving me actual chills hearing this.
[44:32] speaker_1: And the cool thing is-
[44:33] speaker_0: Wow.
[44:33] speaker_1: ... you can go to a language model and be like, "Let's go, let's make it Age of Empires."
[44:36] speaker_0: Yeah. That's what I picked.
[44:36] speaker_1: "Yeah, let's like, we have this evolutionary, you know, agent building system. Now build a UI that looks more like Age of Empires." And it will, it will try. It did-
[44:46] speaker_0: Yes.
[44:46] speaker_1: Like the space between the ideas, right? So maybe no human has done it yet, but you could be the first to ask the AI to do it. Right? Like-
[44:53] speaker_0: I might set, I might have to do this probably t- like after our podcast here.
[44:57] speaker_1: Yeah. And we're gonna have to, we're, we're gonna end up doing the same thing in VR. Like I think-
[45:00] speaker_0: Yeah.
[45:01] speaker_1: ... like that's how we make, we fill VR with more meaningful, interesting experiences. Like w- I think there was like a... Like gen AI allows us to fill it with the applications that it was previously missing, I think.
[45:10] speaker_0: Right. And also-
[45:11] speaker_1: And also-
[45:11] speaker_0: ... the customization, the personalization.
[45:13] speaker_1: Yeah.
[45:13] speaker_0: That was also missing from spatial and virtual. Now that we're solving that on these 2D screens, they're almost like the natural next progression.
[45:22] speaker_1: Yeah. I, I, the, this is one of m- my favorite re- I mean, obviously I don't use VR like regularly. But I can tell you, like a clear, uh, limitation of 3D, the physical world, is screen real estate. Like, like you have your laptop, and that's like a certain amount of space. How many chatbots can you actually look at in one screen without having to switch tabs and switch like... And then you have a second computer. You have a... Like my, my desktop at home is a 49-inch monitor. And that allows me to see more, right? And same thing, same thing like when you're playing the game and you, you get a more macro view, you can see more units at the same time. So if you think about agents as units working on your projects for you, then one of the limitations is how many screens you have.
[46:00] speaker_0: Yes.
[46:00] speaker_1: But then you go into VR and you have unlimited screen real estate.
[46:03] speaker_0: Yeah, it's anywhere.
[46:05] speaker_1: So that feels like it makes sense. And you get the spatial benefit of like hand gestures and stuff. I think what's missing on the Vision Pro is the like, you want a language model at the operating system level.
[46:14] speaker_0: Yeah.
[46:15] speaker_1: And that, that's when it becomes more like Jarvis. But-
[46:18] speaker_0: So I found a kinda hack for that.
[46:21] speaker_1: What are you doing? Like...
[46:23] speaker_0: So what I've done, I can't believe, it's so cool. Like we're in different circles at times, and yet we're working on similar things and it just continues to blow my mind.
[46:31] speaker_1: Yeah.
[46:31] speaker_0: It makes-
[46:31] speaker_1: We should be trading these, we should be trading some of these, this alpha more often. Honestly, I feel like, like your spatial, your spatial reasoning and your spatial like skill set merged with maybe like my automation skill set can create some pretty interesting like intersections.
[46:45] speaker_0: I bet. I bet we could do-
[46:46] speaker_1: Yeah.
[46:46] speaker_0: ... some actual sorcery, not even figurative.
[46:49] speaker_1: Exactly, exactly, exactly.
[46:50] speaker_0: Actual sorcery, 'cause we just talked about-
[46:51] speaker_1: Actual sorcery.
[46:51] speaker_0: ... spell casting, we talked about the magic. That's actual sorcery.
[46:56] speaker_1: Like we're gonna have games, I think like you'll be able to just voice command. Like you'll have your troops, I'll have my troops, and we voice command in like-
[47:02] speaker_0: Yes.
[47:02] speaker_1: ... virtual space and like... I can't wait for that kind of stuff. I think-
[47:05] speaker_0: Oh my God.
[47:05] speaker_1: ... unit economics don't work out at scale yet, but they will exist. All of this, once it gets cheaper and better, it will exist.
[47:11] speaker_0: Yeah.
[47:12] speaker_1: You know? And you imagine a lot of it f- early on. And I see that and I'm just like, "Less than 10 years, easy." Easy, easy.
[47:20] speaker_0: Cool.
[47:20] speaker_1: Uh, uh, 'cau- 'cause you're like painting a picture of what's possible, 'cause you're-
[47:23] speaker_0: Mm-hmm.
[47:23] speaker_1: ... kind of a sci-fi guy, right?
[47:24] speaker_0: I love sci-fi.
[47:24] speaker_1: And I mean, I can see my... But you're like a sci-fi crea- like you are creating the next sci-fi, right? Like you imagine, imagine i- in a world like we may have this capability, what does that look like three years out? And then I look at that and I'm like, "Yeah, act- he's not wrong. X, Y, Z pieces assembled at higher speed and-
[47:41] speaker_0: My God.
[47:41] speaker_1: ... real, higher reliability will give us that reality."
[47:44] speaker_0: Wow. Dude, that's-
[47:45] speaker_1: Yeah.
[47:45] speaker_0: ... so crazy. Then the responsibility is so paramount. 'Cause like if that can happen, then we have to be very careful how we wield the sorcery and those powers.
[47:54] speaker_1: Yes, exactly, exactly. I mean, I mean, and, and think about it, right? The, there are three unforgivable curses in the Harry Potter universe.
[48:00] speaker_0: What are they?
[48:01] speaker_1: I mean, there's the Cruciatus Curse, to like torture someone. There's Avada Kedavra, which kills a person. And then there's Imperio, which mind controls them.
[48:08] speaker_0: Okay.
[48:09] speaker_1: So you can, you can basically like extrapolate back to language models. Like there are things you probably shouldn't prompt these systems to do.
[48:15] speaker_0: Hmm. Like certain spells should be dark ar-
[48:19] speaker_1: Yeah. I, I, I, I, I equate it to that. I'm like... And so when someone asks me, "Oh, can that be done?" I'm like, "Yeah, but you're asking for dark magic." Like, "Let's not, let's not"-
[48:27] speaker_0: Yeah.
[48:27] speaker_1: ..."go down that route until we-"
[48:29] speaker_0: Yeah.
[48:29] speaker_1: "... explore all the upside first that's like this low-hanging fruit, like very obvious like human amplification." Don't you go down that-
[48:36] speaker_0: Yeah.
[48:36] speaker_1: We don't need to go down that route. Like ask someone else.
[48:38] speaker_0: Right.
[48:38] speaker_1: Don't ask me. You know, I'm not the, you know, like maybe I'm the defense against the dark arts teacher, like... (laughs)
[48:44] speaker_0: Oh my God, I love this metaphor. Also, we have the-
[48:47] speaker_1: It's a cool metaphor. It helps people make sense of like the like prompt engineering applied to... It's gonna be more relevant, I think, as these models can take actions. 'Cause right now a lot of people perceive it as-
[48:57] speaker_0: Yeah.
[48:57] speaker_1: ... like it just talks to you.
[48:59] speaker_0: Right.
[48:59] speaker_1: But it's different when it does things.
[49:01] speaker_0: So I saw a prototype that's public information from a patent-
[49:05] speaker_1: Yeah.
[49:05] speaker_0: ... Apple filed, for a version of AirPods, an AirPods Pro that would come out, has an infrared camera on the outside. And I was thinking about it, what would you... what could you use that for? And I had the idea of gestures. So imagine you actually do a spell cast, and like when I do this, this, that's send that thing to Claude-
[49:27] speaker_1: Yeah.
[49:27] speaker_0: ... to that fleet.
[49:28] speaker_1: Yeah.
[49:28] speaker_0: But when I do this, this, this gesture in the air, that does a whole deep research on the thing that just happened five seconds ago, but ignores everything that's happened throughout the day.
[49:39] speaker_1: Yeah.
[49:39] speaker_0: And these become as gentle as gestures in the air, as if I'm using a magical wand, and it h- and it sees the cues. I've trained it on what my magical spells are, and...
[49:52] speaker_1: This feels very much like what Tony Stark has, where he can kind of go to any UI, wave his hands, and then the UI responds to it.
[49:58] speaker_0: Yes.
[49:58] speaker_1: I think that combined with MediaPipe, which I've, like, randomly played with, just like a library-
[50:02] speaker_0: I don't know what MediaPipe is. I never heard of that.
[50:04] speaker_1: It's... I'm very, like, I asked, I asked AI for something and it wanted to use MediaPipe, and then it made this interface where I could wave my hands and control things on the screen. And I was kinda just like, "I feel like Tony Stark." 'Cause, you know, it's like, if I don't have to move the mouse, I would kind of just be like, boom, boom, boom, you know? Like-
[50:18] speaker_0: Right.
[50:18] speaker_1: But that combined with natural language, right? Like, pull up the report on X. And then you're like, "See another one." The gestures.
[50:25] speaker_0: Mm-hmm.
[50:25] speaker_1: I, I like that. I think it feels more natural than... And I'm, I'm a PC gamer, I like mouse keyboard, but I, I feel like we may be on the, the en- I'm hoping we're at the end of mouse keyboard, and then we can create... And then all interfaces may be like, you, once you're standing in front of them, become responsive in this new way. Your, that approach, like the AirPod thing, that's, that's fascinating. See, I like that you're thinking about this before it's, like, anyone else thinks about it. Because it is possible that once that technology comes out, we could just, that day, validate if you were even, like, on the right... You know, like, it's not like, it doesn't take like eight months to figure out if you are on the right path, now.
[50:57] speaker_0: What do you mean? If, if it, if it doesn't exist yet.
[51:00] speaker_1: Because we'll just take that. Once, once, once the tech comes out, we just copy paste the documentation into O3 Pro and you're like, "Hey, can we, like, write a program that lets me do this?"
[51:07] speaker_0: Holy crap.
[51:08] speaker_1: And it might just be possible on day zero. That's the, that's the realization a lot of people need, is like...
[51:13] speaker_0: Wow.
[51:13] speaker_1: It's, anything can be done faster, and, uh, the turnaround time is almost, like, instant. So it's like, are you asking the right question? In this case, the right question is like, "How does this hardware work? How does-"
[51:23] speaker_0: Yes.
[51:23] speaker_1: ... "my vision, like my vision for the interface look like? And then, can we get the first version today?" Is the, is the piece that a lot of people miss. They think it takes a million dollars and a bunch of people.
[51:32] speaker_0: Right.
[51:32] speaker_1: But it's like, it's the right prompt, right?
[51:35] speaker_0: Well, what's crazy is when I play with, like, the latest tech, it gives me ideas of things that I hadn't thought of before. And the other one was like, okay, everything's gonna be synthetic and fake soon, so how can people validate things online, digital? And then I'm like, "Oh my god, Apple has a solution for this," in the Vision Pro. D- you already tried it. It was on, right? You have one?
[51:54] speaker_1: Yeah, yeah, yeah, yeah.
[51:55] speaker_0: Optic ID.
[51:56] speaker_1: Right.
[51:56] speaker_0: You're the only person that can run...
[51:59] speaker_1: Retina scan?
[51:59] speaker_0: Yes, a retina scan-
[52:00] speaker_1: Retina scan, yeah.
[52:00] speaker_0: ... is the only thing-
[52:00] speaker_1: Yeah.
[52:00] speaker_0: ... that lets you drive a persona. Which blows my mind-
[52:04] speaker_1: That's-
[52:04] speaker_0: ... 'cause like then I can know for a fact that I am talking to Parth's avatar because I know you can only run that if there's a biometric scan.
[52:13] speaker_1: Yeah.
[52:13] speaker_0: That's how you can trust that I'm having a conversation with someone. I'm like, "Holy crap, Apple is fricking thinking about that?"
[52:19] speaker_1: That's actually fascinating. That's actually fascinating. I didn't think about that until you mentioned it just now. And I'm like... 'Cause I, I've been thinking about, thinking about this all the time. Like, we're filling the internet with a bunch of like, uh, like, like synthetic information of anything.
[52:33] speaker_0: Yes.
[52:33] speaker_1: And I work on, I work on avatars too, and I look at these avatars and I'm like, "Well, this is indistinguishable from..." And all... It's, it doesn't even have to be indistinguishable for it to, uh, be deceptively good.
[52:43] speaker_0: Correct.
[52:43] speaker_1: Because most people don't know the original person anyways. Like, everyone has a parasocial rela- like, parasocial-
[52:48] speaker_0: Mm-hmm.
[52:49] speaker_1: ... relationship with the subject, in, in many cases, people you follow on the internet. But then when that person is a clone, like, how do we know that... Which clone is them?
[52:57] speaker_0: Right.
[52:58] speaker_1: Right.
[52:58] speaker_0: That's why I was asking like, "Who are you?" is the question I keep having to ask-
[53:01] speaker_1: Yeah.
[53:01] speaker_0: ... to people that live by their AI clone or only defer their ideas to their AI self. And I'm like, "Which one?" It's almost like-
[53:10] speaker_1: Well, I-
[53:10] speaker_0: Yeah.
[53:11] speaker_1: I, I, as, as... Like, I work with like these avatars and the cloning tech and, um-
[53:15] speaker_0: Yeah.
[53:16] speaker_1: ... kind of people, people kind of associate me with like being pretty deep in it. And I think this is the year I try to clone myself. I say try to every time I try to. I mean, the first thing I ever asked ChatGPT to do was teach me how to clone my voice. And it did.
[53:28] speaker_0: (laughs)
[53:28] speaker_1: It wrote the program. It used an open source library and wrote the program. Then the second thing that I asked, it was like, "Wait, teach me how to run this on my computer. Te- I don't know any..." And then it teaches me the basic programming setup. Then, and within a, within two hours, I had a ver- working voice clone. Then I was like, "Well, teach me how to build a GPT powered chatbot." And then it did. So in the, my first day of programming with language models-
[53:48] speaker_0: Nice.
[53:48] speaker_1: ... I had voice cloning, and then I had a GPT running on my computer. So then I connected the two and I was like, "Now I have a GPT connected to my computer that sounds like me." But in that act, I realized, man, it is too... Cloning someone is not easy. And as like, it's like a very strange... It's more like you're creating a sculpture. So for-
[54:06] speaker_0: Mm-hmm.
[54:06] speaker_1: You know, you... Like, a sculpture isn't the person, right? The person's way more facets to their pers- There's so much more depth to a person. It's an, it's a snapshot of one part of their physique, but it is not their mind, right? Like, in the same way, you, you, you can create a snapshot of some part of a person's corpus of ideas.
[54:22] speaker_0: Right.
[54:22] speaker_1: Right? You are not, you are not creating an actual, like... It's a, it's a cheap illusion in my opinion. Like, and I love this tech, but it's just like a cheap illusion.
[54:31] speaker_0: Yeah.
[54:31] speaker_1: The real person, the real deal, the real person is much more, it's, it's just much richer, like, a richer identity. Like, one day I had an anxiety attack. I was like, "Oh man, like, this AI can do my job, like, way better than me." Data analysis. And this was GPT-4. I was like, "GPT-4 writes perfect SQL if it knows the schema of the problem." And my, and, and I was like, "Man, I thought data analysis was like me being a top 1% data analyst. This is my moat. This is my career. I'm good. Like, if, if I didn't want to work hard, I could just do this thing."And, you know, I'd be fine in society, right? Like-
[55:02] speaker_0: Right.
[55:03] speaker_1: ... society would pa- pay me to do this, and then I would have a career. Then I see GPT-4 and I'm like, it's like, it's like John Henry, the, the, the folktale of John Henry.
[55:11] speaker_0: I'm not familiar-
[55:12] speaker_1: This of that one.
[55:12] speaker_0: Okay.
[55:12] speaker_1: It's a, it's a, it's an old folktale of, like, John Henry was, like, a steel man and he would, he was one of the strongest... Yeah. He was, like-
[55:18] speaker_0: With the hammers?
[55:19] speaker_1: Yeah, exactly. Ham- the hammering.
[55:20] speaker_0: Okay.
[55:20] speaker_1: They were building the, the, the railroads through the Rocky Mountains and at the time, it was done by, like, laborers. Like, John Henry was the strong, the strongest of the steel men. One of the most efficient, uh, steel men of all time. I think he was based on people that were real, but not, maybe not r- actually real. But then, uh, at the same time, the steam engine comes out and then they basically have a team of people sitting on a train kind of just, like, laying the tracks out and as it, like, breaks through the mountain. And so now you have... And then what they, in the, in the folktale, John Henry goes up against this steam engine in a head-to-head battle. So, like, a team of guys managing its steam engine while it breaks through a tunnel and then him doing it manually. And then he beat, in the story, he beats the, he wins the race. Like, he breaks through to the other side of, before he-
[55:59] speaker_0: Wow.
[56:00] speaker_1: ... before the, before the steam engine. And, but then he dies of over exhaustion.
[56:05] speaker_0: Oh my God.
[56:07] speaker_1: And that's how I feel doing data analysis manually when alternatively, I can talk to the bot that I built. Like, talk to the data analysis agent that I built.
[56:17] speaker_0: Yes.
[56:17] speaker_1: So, I had an anxiety attack because, like, this feels like a John Henry moment. But then the next day, I sat down and I was like, "Let's build a data analysis agent." Just wanted to do most of what I was doing previously manually, but, like, programmatically. Then I will speak in English and I'll describe the business problems, just like the CEO asking a question and then I go do the analysis.
[56:34] speaker_0: Mm-hmm.
[56:34] speaker_1: I'm going to be the CEO, I'm gonna ask the question and then this agent will write the analysis, do the exploration, and then come back to me with the answer. And it only took me, like, three hours to get, like, natural language to SQL working, basically translating English into business logic, programmatically. And I was, I was just like, I got it working and I was sitting there, I was like, "Ah." The, the, the takeaway is it, it'd be a crazy oversimplification of who I am to say I'm just a data analyst. Now, I am much more, right? Now I am free to be a super data analyst, but that freedom I can use to, you know, attack other parts of my own, who I want to be, what I want to grow into. I could go deeper in this, but I already have a huge advantage here now, and a programmatic advantage, so I might as well explore my creative interests. Like, I might as well become more than just a data analy- analyst. So, I think it's an ex- an expansion of self that's made possible through AI. So yeah, yes, there is an anxiety attack, there is the disruption of, like, who you were and knowing that that's very likely going to be an evolving role. Even, and you know, you evolve with the role and you evolve into something else and you could choose to go deeper in that thing or you could choose to go broad into other things, now that, like, you have superintelligence working on your behalf on the specialized part.
[57:47] speaker_0: It feels like you, in s- in that John Henry metaphor, instead of, I guess, subjecting yourself to only being John Henry, you realized you could also just build the steam engine.
[57:58] speaker_1: Yeah. Yeah.
[58:00] speaker_0: 'Cause like, you know, why exhaust yourself to the point where you don't get to be you? Build the steam engine. And got it t-
[58:06] speaker_1: And it's not, like, it's not like we were getting points for memorizing syntax and, like, being very good at writing SQL. That was the old world. You, you had to memorize blocks of code in order to do anything. But now it would be ridiculous for you to memorize blocks of code to do everything, because now the language model is already memorizing blocks of code across every language. Like, I am only good at one, maybe two languages. The language model can get, like, probably a B- in every language, which means that, like, worst case I should be using, like, I, I can expand into other languages more quickly because that augments, like, my core skillset of Python, SQL, that's great, but-
[58:41] speaker_0: Yeah.
[58:41] speaker_1: ... now I can go into front end, I can go into web, I can go into all these-
[58:45] speaker_0: Right.
[58:45] speaker_1: ... other languages just by asking the language model to, like, bootstrap me there. "Hey, take me to that level." You know?
[58:50] speaker_0: Get you to that superhuman level, give me the steps.
[58:53] speaker_1: Exactly. Yeah. It's the greatest learning tool. It's like this, the greatest learning tool of all time. It's the, it is the tool that teaches you how to use all the other tools. This is, like, meta... Like, this is the most meta technology we've ever experienced. I love that.
[59:06] speaker_0: Yeah.
[59:07] speaker_1: You know, I look at my, uh, music production, I look at the screen and I'm like, "There's all these buttons I don't know. What does this concept do?" Then I just share my screen with the AI and I'm like, "Teach me about sidechain compression and how it works in house music." And then it, then I'm having a conversation and it's like, "Oh, yeah. This is what this knob does. This is what this knob does." And that is powerful, right? Like, this is, like, uh, this is the, the learning, this is a crazy learning amplification. For anything. You know?
[59:32] speaker_0: When you think of your future self, do you imagine Parth 2.0 as a human or as a distributed, distributed constellation of agents?
[59:42] speaker_1: I, I think of it as, like, a system, like a distributed constellation of agents, but not just agents. I think it's the people. Like, there's, like, when you, uh, when you do these things and you talk about them and you share what you're working on and you share your ideas, other people, you know, like-minded people, there's, like, a tribe that kind of forms of people that are also pursuing similar ideas. And different ideas, right? Like, you're in a different space than me, but because we meet every once in a while and we share ideas, both of us get very quickly bootstrapped to a higher level. It's the same thing, right? So, it's not just the human, i- i- it's not just the AI agents that are pulling me to a higher level, but there's the community, there's the people we have in LA that are all specialized in different stuff that allows us as a group, the collective intelligence is growing very rapidly. So, this is not single player. I, I think about this as a fundamentally multiplayer game and my identity is rooted in this multiplayer experience. Like, I define myself more by the groups of people that I participate in society with. The communities that we build, the, like, events that we go to, and the way, the way that, that we collectively kind of uncover the future. It's not single player. I think there are parts that feel single player, but even then, like, you're bouncing an idea off of your friend, you're like, "Hey, is, am I insane or is this a pretty good application of the tech?" And then they're like, "No, that's a good idea. We sh- you should turn those cards over." And then that, like, reinforcement from your own friends helps you pursue the thing, 'cause you're going in, we're going into a very unknown, right? So, we-
[01:01:07] speaker_0: Yeah.
[01:01:07] speaker_1: ... kinda need to, we kinda need to, like-... map it as a group 'cause the information ... This is singularity, in my opinion. I think we're already in singularity. Like, the way I define s-
[01:01:15] speaker_0: Aren't we?
[01:01:15] speaker_1: Yeah. Yeah. We have been for a few, maybe like, for maybe like two or three years now. It's like, where the rate of advancement eclipses our human ability to, uh, keep up with as an individual. And so you need both AIs and friends to help you keep up with the rate of technology improving.
[01:01:33] speaker_0: I can't tell you how happy I am to hear this. Here I was thinking like, I, I was thinking you were gonna say it's all agents, but you just reminded me and the audience listening that it's not just about the agents, it's also about the community that you are collectively-
[01:01:46] speaker_1: Yeah.
[01:01:46] speaker_0: ... growing with.
[01:01:48] speaker_1: Yeah. And not everyone needs to be plugged in. Yeah, I mean, it, it's ridiculous expectation to be like, gr- like, my grandpa like, likes to keep up with the stuff for fun, but he's 83 so little, like, he's not gonna sit down and start programming. But he asks me questions almost every week about different stuff he's seeing. And, and I think that's like, it's like, not everyone needs to be obsessed with the frontier. Actually, if you think about a group chat, it's nice that we're all different. Some of us are interested in different stuff. Someone's obsessed with the frontier and other people aren't, but then that person comes back and, like, drops some alpha in the group chat. Now everyone in the group chat is, like, a little bit better equipped for this world where, you know, very soon, like, the internet will be filled with, like you said, like, there's the filled, it will be filled with a lot of synthetic, fake, like, like, realistic looking stuff that's not real. And then we're gonna have to return and retreat to the community to help our ... Like, form communities that have a shared sense of values around how to use the technology and subscribe to networks that play by the values we want the world to kind of operate in. 'Cause there will be networks that don't have those values and they will be chaotic, there will be a lot of misinformation, and that's gonna be, these, we're gonna ... So dealing with that alone is ridiculous, but dealing with it as a group feels more tractable. And, and also I just think it's like very human in loop. Like yes, you have 100 agents maybe working for you, but-
[01:03:05] speaker_0: Mm-hmm.
[01:03:05] speaker_1: ... they're only doing things that you and, and your friends kind of care about, that you pay for, because you're paying for them to work, right? So, like, they are not just going to do things. They're not, they don't have desires of their own, right? So we apply the technology in the ways that we care about. And, and I think that's where, like, community values kind of matters. Like, you can help solve a problem for a bunch of people. They may not have that technical skill, but then now you have that, like, you can do that for them, right? You can apply the AI to solve a problem for a group of people that means something to you. That's awesome. I think that we're, we're at the very beginning of that. That's like this, I think of it as like there's individual agency, but then there's this, like, collective agency, like, that we as a group now can go beyond our sense of self and we have, like, higher group collective intelligence is growing, and then we do that by, like, specialization of like what we focus on, right?
[01:03:54] speaker_0: 100%. My mom, I just built her an app to help her learn Italian better, and I can't program, and she's been using it.
[01:04:01] speaker_1: This is what I'm talking about.
[01:04:02] speaker_0: She's been wanting to learn Italian her whole life. She's tried classes in person, but the teachers are always too fast or they give her too much paperwork and she doesn't have that kind of time to treat it like a deep school.
[01:04:13] speaker_1: Mm-hmm.
[01:04:13] speaker_0: She just wants to practice really just like her restaurant talk when she visits Italy.
[01:04:17] speaker_1: Yeah. Yep.
[01:04:17] speaker_0: She just wants to be, sound good enough to, like, communicate with the basics. So I built her an app based off of that premise.
[01:04:24] speaker_1: What did, what did you build it in?
[01:04:26] speaker_0: And she chats with it.
[01:04:26] speaker_1: What did you build it in?
[01:04:27] speaker_0: So how I've been making apps is I've been building these long set of prompts that you trigger as text replacements-
[01:04:34] speaker_1: Yeah.
[01:04:35] speaker_0: ... in Apple.
[01:04:36] speaker_1: Okay.
[01:04:37] speaker_0: And I, and then so she puts one prompt in to ChatGPT and does something and then she has another one, 'cause there's a limit to how many characters you can fit in a text replacement. So she sets, so I basically built her like little, little shortcuts for her prompts, it builds it out.
[01:04:50] speaker_1: Yeah. Presets, like preset prompts. Yeah.
[01:04:53] speaker_0: Yes. Boom. I gave her presets. Then her version of ChatGPT in voice mode completely caters to the kind of application she needs.
[01:05:02] speaker_1: That's awesome.
[01:05:02] speaker_0: And it makes it a version, it's like a version of ChatGPT, it's just for her.
[01:05:06] speaker_1: Yeah. Yes. This is, this is powerful and I think it's the pie- I mean, I think it's the piece that's missing from language learning is, like, having someone that you can bounce this off of and like, in a very immersive way. I, I like to just take the voice mode and then walk around. I'll be like, advanced voice mode. I'll be like, "Hey Chat, we're gonna go for a walk, and, uh, every time I point my phone at something, you're gonna say, you're gonna say the word in Chinese and you're gonna say the word in English. You're gonna use it in a sentence. And then we're gonna, we're just gonna have a little mini lesson." So every time I point at something you, we get a natural kind of exposure to the, to the culture that's more immersive. And it's like, now you're going out into the world, right? This is what you clued me into is like ambient intelligence when you're actually in the real world is a piece that's, not everything is sitting at a computer, actually. Like, the intelligence is with us. It's in our pocket. It's multimodal. It can see. It can listen, right?
[01:05:52] speaker_0: Do you ever go to, like, a group camp? Like, where like, there's like, someone teaching how to do like beads, another one's teaching how to do like archery, and there's another person teaching how to do like blacksmith work? These are all like multimodal agents, in my opinion, that exist spatially.
[01:06:07] speaker_1: Yes.
[01:06:07] speaker_0: So the kids that are going along to camp-
[01:06:09] speaker_1: Yeah if you think, if you think about humans as, humans are multimodal agents. Yes. (laughs)
[01:06:12] speaker_0: (laughs) Sorry.
[01:06:13] speaker_1: We, we set a very high bar for multimodal agency, I think. That like, that's-
[01:06:16] speaker_0: Yes.
[01:06:16] speaker_1: ... who we are, right? Like, we exi- we can move in the real world and use our hands and see and smell and the text chatbots can't yet, can't do all those things yet, and that gives us a huge advantage.
[01:06:27] speaker_0: Do you have any insight or perspective on robotics? Do you think we're gonna be talking and working with physical mixture of humanoids, robotic arms through voice soon or ... I know we're in the singularity, so like-
[01:06:41] speaker_1: Hmm.
[01:06:42] speaker_0: ... where in the singularity roughly do you see robotics being as relevant into our daily lives?
[01:06:48] speaker_1: Yeah. So I, I, uh, a lot of my friends work on this tech. You know, I'd say like if it's, a lot of my friends in college were working on, the robots they were working on were self-driving cars, and, uh, they were working on it back then like 10 years ago, 12 years ago, and, and I remember they went in very excited and then, uh, two years in they were like, "I think it's time for me to find the next thing." I'm like, "Whoa, why?" Like, and then they're like, "This tech is 10 years out."Now we have like an, a... And now Waymos are here. Like, the car drive... Like, I take a Waymo literally everywhere. I take... I've taken 106 Waymos now, so-
[01:07:22] speaker_0: Whoa.
[01:07:22] speaker_1: ... I'm like, I'm bought in 100%. But then, and I talk to my friends that work on robotics, and one of my buddies is at Stanford, he's a PhD student, he's working on robotics, and then I have other friends that work on like Figure, like the humanoid robots.
[01:07:32] speaker_0: Oh, for real.
[01:07:33] speaker_1: Yeah.
[01:07:33] speaker_0: I love Figures, um, demonstrations.
[01:07:36] speaker_1: Yeah, yeah. Friends that are, like, working on these, and I talk to them. And, uh, so, you know, I get the perspectives like, okay, so where's, where are we with, like, humanoid robots, where are we with robot arms? So my friends working on humanoid robots, they're pretty excited, but also, like... And then I have a friend at Stanford, and he's kind of like, "I think it's..." He said something like, "Oh, I think it's, like, 10 years out." And I was like, "Oh my god, this reminds me of self-driving cars." And then I'm like, "Why?" And then it's like, you know, when we're using... And he- he's, like, less interested in the humanoids, more interested in the robot arms and, like, the different things.
[01:08:04] speaker_0: Yeah.
[01:08:04] speaker_1: And I'm like, "Well..." And kind of like how I'm talking about coding co-pilots with guardrails and, like, ask me for approval before you do this so that I can hit the brakes just in case this is a braking change.
[01:08:15] speaker_0: Yes.
[01:08:15] speaker_1: He's like, "I sit here and when, when we're operating these robots, my hand is always just hovering the, the kill switch." And-
[01:08:23] speaker_0: Wow.
[01:08:23] speaker_1: And then, uh... Yeah, yeah. And so I'm like, "Okay, this is a good insight," right? Like, uh, even if the capability of the robot is there, if the risk is high, you're not going to see it deployed in the home, right? The home, where, where the room for error is... It has to be... If this thing... Like, I just want a robot that can, like... I have a robot litter box, but I want a robot that can take the litter bag out of the box and then toss it and then come back. It's like, that's the level of, like, I want the humanoid to be able to, like, take the bag, toss it, put a new bag in. That's very-
[01:08:52] speaker_0: Yes.
[01:08:52] speaker_1: That's a very high bar, I think, for a robot at home. Right now, everything else, the, the litter box flushes itself, which is kind of cool.
[01:08:58] speaker_0: Nice.
[01:08:59] speaker_1: But I think about that, and I'm like, "What's the worst case scenario?" I don't want it to accidentally hurt my cats.
[01:09:04] speaker_0: Right. Oh my god.
[01:09:05] speaker_1: And-
[01:09:05] speaker_0: No.
[01:09:06] speaker_1: Same thing, like, are you gonna trust a robot with your baby? Like-
[01:09:10] speaker_0: Not currently.
[01:09:11] speaker_1: That... Not in, not any time soon. And I... It may learn how, how to do this all in simulation, but then when it shows up in the real world... And then it's like, what are the real risks? And it's like, the hinges, like... You gotta be very... Like, all of this stuff is, uh... The capabilities will probably get there, but the, um, reliability needs to... Like, way more reliability. They're very transparent about their metrics, and they, they post it on their website. "Here's how, how many times the airbag was deployed. Here's how many times the car got in an accident."
[01:09:35] speaker_0: Wow.
[01:09:36] speaker_1: That is how you, you... If you're transparent, if you're building an AI system or a robotic system and you want people to use it, you have to be transparent with people on how well it works. Otherwise, it's never gonna be used for anything important, anything serious. And I think, like, home robotics is a very serious thing. And if it's not properly s- if it's not perfect, then it needs to be sandbox. And if it's not sandbox, and if it's not, like, reliable, it's not gonna be doing anything very important that we want that, like, is human-facing. Like, we may be outside of the room that it is in, right?
[01:10:08] speaker_0: Mm-hmm.
[01:10:08] speaker_1: That we may be handling the, uh, the kill switch. So I think there's like a... It's exciting. It's gonna... I'm, I don't... You know, I'm like, it's exciting, right? I, like-
[01:10:15] speaker_0: Yeah.
[01:10:16] speaker_1: But I think we have a long way to go on it showing up in our world.
[01:10:20] speaker_0: So in 2007, I took a robotics class at Mitty High School before-
[01:10:26] speaker_1: Yeah.
[01:10:26] speaker_0: ... I went to Mitty. I did it-
[01:10:27] speaker_1: Yeah.
[01:10:27] speaker_0: ... in eighth grade, the summer camp.
[01:10:29] speaker_1: Yeah.
[01:10:31] speaker_0: Uh, our professor accidentally got hit in the face by a robotic arm that he was building because someone used the same radio wave signal on their controller, on their robot, when they were, like, doing their little RC car in the other room.
[01:10:44] speaker_1: Yeah.
[01:10:44] speaker_0: It was on the same frequency.
[01:10:46] speaker_1: No way.
[01:10:46] speaker_0: As, as... It was on the same, like, radio channel. And so all of a sudden, like, he was working, all of a sudden I saw it. The arm just comes up and just swings and hits a metal bar across his face. And it was like-
[01:10:57] speaker_1: Yeah.
[01:10:58] speaker_0: And it felt for a split second 'cause we didn't see where this was happening. It felt like it came alive.
[01:11:02] speaker_1: Yeah.
[01:11:02] speaker_0: This was 2007. This is like, this is before-
[01:11:05] speaker_1: Yeah.
[01:11:05] speaker_0: ... the iPhone.
[01:11:06] speaker_1: Yeah.
[01:11:06] speaker_0: And I'm just like-
[01:11:07] speaker_1: This is, this is, like, way early, way early, and, like, exactly what we... Like, this is exactly what is going to happen. Like, in... The worst case scenario will play out in some lab. It's going to... People are gonna get hurt.
[01:11:18] speaker_0: Right.
[01:11:18] speaker_1: And then we're gonna, we're gonna get more serious about this. But that's the actual risk. Like, that's what my friend at Stanford is like, you know, he's concerned about, right? Like, even just working with an arm.
[01:11:25] speaker_0: Hand by the kill switch.
[01:11:26] speaker_1: Yeah, hand by this kill switch. But that doesn't solve, like... Like, that, that's not enough, I think.
[01:11:31] speaker_0: No, no.
[01:11:32] speaker_1: And then, and then, I don't know, you've seen, uh, you, you've seen i, Robot. Uh, did you read the other, like, you read the short stories that Asimov wrote?
[01:11:41] speaker_0: Read two of them, but not all of them.
[01:11:43] speaker_1: But you, you know, he has, like, the, the rules of, that robots need to abide by?
[01:11:47] speaker_0: Yeah, I think they're not enough, but-
[01:11:49] speaker_1: But-
[01:11:49] speaker_0: Yes.
[01:11:49] speaker_1: Well, it's fascinating that, like, he wrote those, and I was like, "That's funny." Like, this isn't how computers work. You can't just write down a bunch of rules. Like, back in the day when I was reading it, I was like, "You don't just write a bunch of rules and be like, you know, stick to these rules and then, like, figure out how to reason across these natural language instructions." And then now-
[01:12:05] speaker_0: (laughs)
[01:12:05] speaker_1: ... I look at reasoning engines, I look at chatbots, and I'm just like, "Oh, shit." Like, and then you combine that with, like, like, robots and trying to do visual reasoning. And then for them to be able to look at a room and be like, "Don't, don't hurt the cat." Like, simple objective, don't hurt the cat. Don't do anything that might hurt the cat, or the baby. Or, like, like, all these-
[01:12:24] speaker_0: Yeah.
[01:12:24] speaker_1: Like, don't break the glass table. Like, there's all this, like... And I think about this complex... The, the physical world is a way harder... Like, in cyberspace-
[01:12:31] speaker_0: Yeah.
[01:12:31] speaker_1: ... I'm like, "Yeah, you just have this little folder of files and it's backed up. So, like, I'm okay with you breaking this because I can just, like, you know, I can-
[01:12:38] speaker_0: Yeah.
[01:12:38] speaker_1: ... I can deal with that. This is like a virtual space, it's a sandbox, and I know what the, the ultimate downside is." But in the real world, the ultimate downside is actually, like, way worse. And so the bar has to be much higher.
[01:12:50] speaker_0: Yes.
[01:12:50] speaker_1: And, I mean, the fact that we have cars that can drive on the road better than human beings is... It's proof that, like, we can get to that place. But it wasn't without, like, error.
[01:13:00] speaker_0: Yeah.
[01:13:02] speaker_1: You know. So we're going to see that, I think, with robotics. And it's gonna be a 10 year ga- 10 year game at least. 10 years at least.
[01:13:09] speaker_0: Did you watch that research video paper that Apple produced on their robotic arm research they're doing?
[01:13:14] speaker_1: No, I saw the more recent thing where they were talking about how reasoning engines don't reason.... which is funny.
[01:13:21] speaker_0: I know. I thought that was interesting too. I think it was just because they're a little sour right now-
[01:13:24] speaker_1: Yeah.
[01:13:24] speaker_0: ... after WWDC with some of the AI stuff.
[01:13:28] speaker_1: Yeah, yeah, yeah.
[01:13:29] speaker_0: But their pa- video that they released on the robotic arm, they have a Pixar-looking lamp-
[01:13:34] speaker_1: Okay.
[01:13:34] speaker_0: That is mounted to a table, and it can see, it can hear-
[01:13:38] speaker_1: Yeah.
[01:13:38] speaker_0: ... and it can gesture very naturally. And they gave an example, it shows, um, it shows the human operator being like, "Hey, where'd I put my keys?" The lamp tilts up and looks at the human's face, looks to where the keys were, looks back at the human, looks back to the keys, and like gestures like that.
[01:13:57] speaker_1: Hmm. Hmm.
[01:13:58] speaker_0: And so, you know, light bulbs go off in my head. Then they showed another example, same robot, they said, um, "I'm thirsty." The robot nudges the water cup that was sitting on the table, it nudges it in the direction of the person. So it does like a little nudge, nudge, then the lam- the lamp looks at the face, and then nudge, nudge, looks at the face, to kind of reassure the human the intention. The intention of, "Oh, you want me to... Oh, you want me to take this cup." And it made me think, um... Oh, yeah, and the last one was music. It was h- it heard music playing in the house, and it starts bouncing its head correctly to the beat. And because it has a few actuators, it almost looks like it has a shoulder.
[01:14:36] speaker_1: Yeah.
[01:14:36] speaker_0: So it can actually do like good, like bouncy rhythm.
[01:14:39] speaker_1: Like vibe... It's vibing. Yeah.
[01:14:40] speaker_0: It's literally like your lamp is vibing. And it made me think, are we going to get to this point where like someone might be like, "My best friend is my lamp."
[01:14:48] speaker_1: Yeah.
[01:14:48] speaker_0: "It gets me better than anybody."
[01:14:50] speaker_1: Yeah. Yeah. There's a cool-
[01:14:50] speaker_0: What are your thoughts on that? Yeah.
[01:14:52] speaker_1: Oh my God, there's a cool, uh, there's a movie... It's kind of, I mean, it's with a humanoid robot, it's called Chappie.
[01:14:57] speaker_0: I love Chappie.
[01:14:58] speaker_1: Isn't the movie called Chappie?
[01:14:58] speaker_0: Yes.
[01:14:58] speaker_1: The movie's called Chappie, right?
[01:15:00] speaker_0: Yeah.
[01:15:00] speaker_1: You've seen it. Yeah. A huge part of Chappie's personality is that he learns how to walk from a bunch of gangsters. And he walks, he walks... H- he's got this like limp, and, and, and it gives him so much personality. Is it, is it necessary? Probably not, but like... And then he speaks a certain way, and I think this is all character design, and I think a lot of what we're doing is character design. I think the chatbot side, there's like some persona stuff there that's very, very fascinating. You can create these personas and then you can put them behind programs and have them act a certain way. But now you have like character design on the physical level of like the way this like machine moves in our world. Some of that stuff is a subjective choice, right? Like does the machine need to dance?
[01:15:40] speaker_0: Yes.
[01:15:41] speaker_1: No.
[01:15:42] speaker_0: Yes.
[01:15:42] speaker_1: But like it would be fun. I mean, it would be swaggy. It would be like, like what if we learned some new moves from this thing that's just like learning how to dance a simulation, right? Like there's so much. I think that like, yes, it's a fascinating thing, and Chappie is the first time I was like, "Oh, man."
[01:15:55] speaker_0: Wow.
[01:15:56] speaker_1: Like, personality is like not just what it says and how it says it, but also like how it kind of carries itself, right?
[01:16:02] speaker_0: Movement, its weight.
[01:16:03] speaker_1: Yeah.
[01:16:04] speaker_0: Where, where it puts its emphasis.
[01:16:04] speaker_1: It literally... A huge part of it was like they taught... They gave this robot swag by just hanging out with it and teaching it a different way of moving about our world.
[01:16:12] speaker_0: (Laughs).
[01:16:12] speaker_1: Like, yeah. I think that's awesome.
[01:16:14] speaker_0: Oh my God. How was that at MIT mTech, uh, uh, conference?
[01:16:18] speaker_1: I want my robots to have personality like that, for sure.
[01:16:20] speaker_0: They should. Honest... They should have that, and I think they will, and I think we might be like... Future kids might be like, "Wow, grandma and grandpa, what was it like when everything in your house wasn't alive?"
[01:16:32] speaker_1: Yeah, yeah, yeah, yeah.
[01:16:34] speaker_0: And you're like, "Oh, well, my desk, like I had to lift it up myself, and doors, like, you know, I did it myself," and then the grandma... And then they'd be like, "Grandma, why?"
[01:16:44] speaker_1: Yeah. Yeah, the expectation of like hardware in that future, I g-... I think about this a lot, like it's... And actually I was having O3 Pro think about... help me think about sci-fi. Like-
[01:16:53] speaker_0: Ooh. Ooh.
[01:16:53] speaker_1: I was like, I was like, "Let's take all of these capabilities we have now, without going too far out, like extrapolate out, let's say the reliability increases now, right?" And, and then I was like, "Let's, like, let's say we're not multi-planeto..." We're like at a point where it's like Earth, Mars, and then like people working on the asteroid belt, and then maybe we're working on one mega, like Dyson sphere project, but then may... I basically asked for like Game of Thrones in space, but I didn't say Game of Thrones. I basically described all of the constraints. And then I was like-
[01:17:20] speaker_0: Yes.
[01:17:20] speaker_1: ... also make sure the technology is grounded in an extension of where we are today, right?
[01:17:24] speaker_0: Oof.
[01:17:24] speaker_1: Like projecting out these capabilities of language models and vision analysis, multimodal, and like the idea that people can speak and make things happen in the world. It... And I... And O3 Pro, man, it wrote, it wrote the like high-level concept of the story, and I was like, "Holy... This, this... I can't wait to make this into... Animate this." And then I was like, "Now direct just one scene that establishes this technology for the viewer." Right? Like, you know, there's always like some scene that's like, oh, we got to explain how the rules of this tech work just so that when we use it later on, they have an expectation that's in a... Aligned with where we are in the movie-
[01:17:57] speaker_0: Yes.
[01:17:57] speaker_1: ... that's grounded and inspired by stuff that they've seen in their world. Right?
[01:18:02] speaker_0: God, you're a filmmaker now.
[01:18:04] speaker_1: Well, aspiring through AI.
[01:18:06] speaker_0: No.
[01:18:06] speaker_1: Right? Like-
[01:18:07] speaker_0: I feel like you're in.
[01:18:08] speaker_1: Well, I'm trying to understand this, the piece. It's like we can... If you can generate anything, then like the story you tell actually matters, right? Like, and then why does someone gravitate towards a particular story? And I love sci-fi, and part of it, it's like how do we get the right... What is the next sci-fi if like now we got robots that can walk, we got robots that can talk, what, what's next? There is something next, and we have to imagine it. And I use language models that can think to help imagine it.
[01:18:31] speaker_0: I want to read your sci-fi so bad.
[01:18:33] speaker_1: It w- It writes-
[01:18:33] speaker_0: I would love to read your sci-fi.
[01:18:35] speaker_1: Maybe I'll put it out there, but it's like it writes these scenes and I'm just like, "Yeah, it's a little too magical." Like, like, "Let's go a little bit more analog." And then it like rewrites the scene-
[01:18:43] speaker_0: People... (sighs)
[01:18:43] speaker_1: ... and I'm kind of just like, "This is, this is good." Like I, I think that the, the smart-
[01:18:47] speaker_0: Dude.
[01:18:48] speaker_1: Like people are like, "It's not a good writer," I'm like, "Ah-"
[01:18:50] speaker_0: Yeah.
[01:18:50] speaker_1: "... I mean, it's still human in the loop, right?" I'm articulating-
[01:18:52] speaker_0: Mm-hmm.
[01:18:53] speaker_1: ... and fine-tuning the response through my curation. But my God, like-
[01:18:56] speaker_0: When, when you publish this on Substack-
[01:18:58] speaker_1: We're getting there.
[01:18:58] speaker_0: ... I would like to read it.
[01:19:00] speaker_1: All right, maybe I will. Like maybe I'll... I want to-
[01:19:02] speaker_0: I would love to read it. Make some artwork for it too.
[01:19:04] speaker_1: That's, that's what I want. I want to bridge it to... So, so I show- told you about like the evolutionary programs. That's cool. There's a second thing that I, I want to do as well, which is like part of the tools I want to give these systems as they evolve is like the video generators, right? So VO3 on API, RunwayML on API, image model Flux on API, right? So you give them the ability to create other... Things other than text-
[01:19:24] speaker_0: Mm-hmm.
[01:19:25] speaker_1: ... and then you create the system, and then it's like...... I like this, I like this, I like that, and I start creating this system. This- this is what-
[01:19:29] speaker_0: Yes.
[01:19:29] speaker_1: ... I'm looking for. And then it evolves closer towards my vision and it taps into my 100,000 experts that I have.
[01:19:36] speaker_0: Wow.
[01:19:36] speaker_1: Like, I have experts across every single topic and then-
[01:19:39] speaker_0: Yes.
[01:19:39] speaker_1: ... they kind of bootstrap me. 'Cause I don't have the vocabulary of cinematography, but I've got GPTs-
[01:19:45] speaker_0: Mm-hmm.
[01:19:45] speaker_1: ... that do.
[01:19:45] speaker_0: Mm-hmm.
[01:19:45] speaker_1: And they may be better at writing the prompt for the shot that I want.
[01:19:50] speaker_0: Mm-hmm.
[01:19:51] speaker_1: So, now I'm trying to create this, like, n- this, like, team of agents that are individually better than me at the sub-pieces, but me having the vision of, like, "Ooh, we wanna tell a story that's, like, maybe 100 years out and this is, like, my assumption of the trends that will play out." And the reasoning model can project that out and write the story. So it's so- it's amazing. And I think that, um, it's- we're early. The tech is a little expensive, that's why the realizations aren't hitting a lot of people yet. But-
[01:20:21] speaker_0: Sure.
[01:20:21] speaker_1: ... eventually- eventually once everyone has access to, like, the generators and it's not a $200 subscription a month, it's gonna click for more people that, like, you're creative. I think it's an amplification of writing. I don't think it's a replacement of writing. I think it's, like, now you have... The writer is paired with, like, a scaffolding that they can kind of really expand off of and the- the writer becomes a world builder. Like a-
[01:20:43] speaker_0: Yes.
[01:20:43] speaker_1: ... multimodal world builder, right?
[01:20:44] speaker_0: My God, yes.
[01:20:45] speaker_1: You're not just writing how the guy looks. You're looking at options and you're like, "Ah, it's actually, like, more like this in my mind and now I see it." It's like imag- real time imagination. And the system will eventually be faster, bidirectional, real time. It'll consult you-
[01:21:00] speaker_0: God-
[01:21:01] speaker_1: ... for feedback.
[01:21:01] speaker_0: ... yes.
[01:21:02] speaker_1: Yeah.
[01:21:02] speaker_0: Parth, do you remember, we were at a dinner and you and I got to hang out with one of our favorite sci-fi authors in person. Had a dinner with Neal Stephenson.
[01:21:11] speaker_1: That's right.
[01:21:11] speaker_0: Can you just tell me a little bit about what that experience was like for you and what it meant to you, and the questions you asked him?
[01:21:17] speaker_1: Oh my God. This was like a dream come true. I mean, meeting Neal Stephenson... Neal Stephenson, Snow Crash, like, so many books. Crypta- Crypto- Crypto- oh my God, Cryptonomi-
[01:21:26] speaker_0: Yeah. Con? Uh, Cryptonomicon. You got it.
[01:21:28] speaker_1: Yeah, yeah.
[01:21:28] speaker_0: Yeah.
[01:21:29] speaker_1: Just basically, you know, like, you- you... And he- and, and, you know, he's a sci-fi writer, but then actually, like, so influential that, um, I would say it inspired that entire, like, him and a couple other people inspired, like, that era of cyberpunk and, like, Neuromancer. These ideas of, like, what does cyberspace look like? What does it mean if we can talk to programs? What does it mean if you can render... Like, virtual reality and AR kind of came out of the science fiction that they were writing, and they... And the- the- and I ta- we... Getting to talk to the guy that spawns an entire genre of science fiction and then asking him... Like, the couple moments were like, one, it's like- it's like, wha- how do you feel about the fact that we have self-driving cars, we have Vision Pro, we have all these... We have a lot of the stuff that you guys were talking about back then as, like, a imagine what if? And- and this is the kind of stuff that you see in Blade Runner, right? Back then, it's like-
[01:22:20] speaker_0: Yes.
[01:22:20] speaker_1: ... in the movie only. It's only in the movie. And then, uh, now it's in reality. Like, Minority Report was kind of a cautionary tale and now it's like-
[01:22:27] speaker_0: Mm-hmm.
[01:22:28] speaker_1: ... we kind of have Minority Report technology now.
[01:22:29] speaker_0: Yeah, predictive tech and for crimes and-
[01:22:31] speaker_1: I, Robot. Same thing with I, Robot. Sa- like, it was like w- like, you think that... It was like, "Oh, one day maybe," but now I look around, it's like I feel like we're facing these exact same technologies in the real world. And Neal Stephenson, yeah, he, you know, he was a... He coded for fun just to kind of get, like, exposure, but then he's primarily a writer, but then... So seeing where the tech was at the time, he was saying the tech wasn't good enough to... It was just an idea, but the tech wasn't there yet. But now the tech is there, and so we need new ideas. And that's very surreal. And also I was like, "So what's next?" And he's like, "You gotta find other sci-fi writers." Like, I don't have the answer. (laughs) And that's now my- one of my- one of my pursuits, is like who is telling me... Who is painting a picture of tomorrow is very interesting to me, because I think we have a lot of, like, Hollywood influence. Like, life imitates art. A lot of my career has been influenced by the content I've consumed. I would say Tron is a huge influence on my career. Uh, both the original Tron, my dad watched it in college. He made sure me and my brother watched. He watched it in engineering college and he was like... We were so wo- mind blown of like, "This is what might the inside of a computer work, uh, look like." Even though they worked on computers and it doesn't look like that, but it's imagination, right? Like, the idea-
[01:23:38] speaker_0: Yeah.
[01:23:38] speaker_1: ... that you can- you can talk to programs and that's early. Like, that was pre-language models. Then Tron Legacy-
[01:23:44] speaker_0: 80.
[01:23:44] speaker_1: ... comes out. Yeah. Tron Legacy comes out and it's still got the same guy, Jeff Bridges, from, uh, n- yeah, 1980- '82, right?
[01:23:51] speaker_0: Mm-hmm.
[01:23:51] speaker_1: But, like, they- they take him and then they deepfake him to- to the young version, so they make him play himself at a young age, as well as the old current day Jeff Bridges. And then I met Ed Ulbright at one of our events and he works at Metaphysic, and I was telling him about how Tron got me into this, like, Avatar kind of mindset of like, oh, talking to programs but also with, like, deepfake technology. I saw that first in Tron. He's like, "I worked on Tron Legacy." And so it's like full circle, right? You meet the people that write the books, then you meet them in the real world and you ask them questions. Then you meet the people that made the movies and then you meet them in the real world like, "Dude, did you ever imagine that this tech would end up off the shelf that I could just play myself in a movie that I make without asking anyone else for permission? This is incredible." And he's just like... I mean, even... No, he- he- they could not predict it, but- but they were early.
[01:24:36] speaker_0: They kind of-
[01:24:37] speaker_1: But they were-
[01:24:37] speaker_0: Yeah. They kind of did.
[01:24:39] speaker_1: They're pioneer- pioneers.
[01:24:39] speaker_0: Pioneered, okay.
[01:24:40] speaker_1: They're pioneers. Pioneers, right? Like, they... But they do it before everyone else thinks it's possible.
[01:24:45] speaker_0: Right.
[01:24:45] speaker_1: And then they... That content that they create, the book that they write, the movies that come out, the TV shows that spin out, all of that, uh, gives us technologists more of an imagination and then... I- I think it's kind of like life imitates art. And for better or for worse, we also need some kind of... I think we need some... It's- it's- it's kind of... Part of why a lot of the, uh, discourse is dystopian is because Hollywood tells the story of, like, robots, right? Taking over all the time. But that's partly because, like, no one wants to... Like, they think... It's- it's, like, does anyone want to watch the positive? Like, the- the story of the what if it works out is not as, uh, exciting as the what if it just totally falls apart.
[01:25:21] speaker_0: Right.
[01:25:21] speaker_1: And I think it's possible to tell the story of what if the technology works out and then make it a human story about, like, there... Humans still have flaws, right? So you can-
[01:25:28] speaker_0: Yeah.
[01:25:28] speaker_1: ... even paint a picture of the- the positive, let's say we do solve some of these important problems, and then tell a human story. So I think- I think that... Yeah, I think it's not purely about technology. I think it's actually-
[01:25:37] speaker_0: Yes.
[01:25:37] speaker_1: ... about humans. It's about how humans use technology. And I think we will get... I am trying to imagine science fiction...... given where we are today because we totally need new science fiction.
[01:25:48] speaker_0: I feel like what you're building, your life, I mean, if you feed this into a neural network, you can almost have it imagine a sci-fi based off of your outlines-
[01:25:59] speaker_1: Mm-hmm.
[01:25:59] speaker_0: ... that would be a viable story that is meaningful and impactful to a whole nother generation of humans and new fledgling AI systems that are developing, that will-
[01:26:07] speaker_1: Yup.
[01:26:07] speaker_0: ... also learn from your sci-fi and-
[01:26:09] speaker_1: Yeah.
[01:26:09] speaker_0: ... and add that to their knowledge graph.
[01:26:12] speaker_1: Yeah. Yes. Yeah, and I've never been so intentional about putting more content out as I am now, but mostly because, like, actually AI will be, is learning from us as we think and we publish. And so when you say, like, "Explore my work," you're also, it's like, explore, like, part of where my mind is going and help me unpack this, this future. It's fascinating. I, uh, yeah, I mean, if you write, a lot of what you're writing is actually also being, is, like, part of your writing for AI too, right? Like, how does AI perceive Don Allen? When I go to ChatGPT and I ask it about Don Allen, like, like what does it know about you? How does it think about who you are and your aspirations and what you're interested in? And that stuff is all, I mean, it, it, it, uh, it, it ends up being an ambient thing, like, how the models perceive us and what we hope to be and what, what our values are end up being infused in the ambient intelligence that we're gonna be interacting with. And so you kinda have to have some vision, right? Like, yeah.
[01:27:08] speaker_0: I was gonna ask, uh, you know, kind of, um, another kind of, you know, getting towards the end here, I was gonna ask around, y- when it comes to rethinking reality, um, how do you find yourself rethinking reality?
[01:27:19] speaker_1: Um, usually I do it through the lens of technology. That's my, my predisposition, is like, given we have systems that can do X, then what is, like, Y? What is possible? Like, like, given I can do this, so what does that mean? This is now possible. What does that mean for me? What does that mean for my friends? What does that mean for my coworkers or, like, everyone that works in my role, in my industry? And so I kind of start with the self, and then I go out of, take slightly larger circles out, uh, of, like, myself, and I think, "Well, what does it mean for music? What does it mean for music production? What does it mean that we have, we have programs that can, that can understand any language? What does it mean for language itself, right? Like, will we end up with the Babel Fish from Hitchhiker's Guide to the Galaxy?" It's a technology.
[01:28:05] speaker_0: Oh my God, yes. I forgot about that.
[01:28:07] speaker_1: 'Cause they have a, they have a technology they put in their ear, and then it is a alien that basically translates all languages instantly through... So then, is that what AirPods become, right? And what does that mean? If everyone can just, you know, whatever language you speak now is no longer a barrier. Yeah, it helps to know English 'cause more people on the planet know English, and that's how business is done. But if you could do business with people outside of your language, because language wasn't a barrier, now you have more people you can possibly connect with. I think AI unlocks that. So I think about that, right? It's like, okay, if you can translate yourself into every language-
[01:28:40] speaker_0: Yes.
[01:28:40] speaker_1: ... now you're connecting with people all over the world that you would have previously never been able to connect. And even if it's a parasocial relationship, they follow you, but now they understand parts of you.
[01:28:49] speaker_0: Right.
[01:28:49] speaker_1: And there is, like, downstream effects. Like 10 years later, maybe they're pursuing a topic. Maybe you meet them and then, like, you, you team up on something even though you don't have a common tongue. That's a great milestone for us to achieve, right? Like, bridging, bridging human beings. Yes, we're connected through the internet, but it also bridging us through language. That's powerful. And then bridging is bridging, like, this is like this kind of glue of, like, the models learned on, learned from us. Now they're gonna go beyond in different ways. And then that is like, the, these, it's like you are no longer limited by the problems you have solved or faced in your life. You're now limited by asking, the questions that you ask of these systems. Like, that's one of those realizations that I had. Like, your limits are no longer what they were two years ago. And unless you sit down and really think about that and maybe talk to a language model about this, like, "Here's who I think I am and what my limits are and what my scope in life is," but then you think about, "What can I be? Like, what can I-"
[01:29:43] speaker_0: Mm-hmm.
[01:29:43] speaker_1: "... where can we go next?" Like that, I think it's like a form of manifestation through AI. It's like, imagine this hypothetical version of me five years from now, given I apply myself in this way, try to be more creative, try to be hopefully healthier, less obsessed, uh, obsessive of work. But that's fine, you know, like, given my traits, given my strengths and weaknesses, how can we kind of like create a roadmap towards being this next version of me? And AI will give you that. Like, it will hallucinate whatever you can say, it's not thinking, you could say, you could say, "It's not really thinking." I'm like, "Okay." But I am thinking and I am getting options and I am choosing from those options how to operate. I am the one with increased agency. It's not telling me what to do, it's just presenting. It's like simulating possible universes and then giving you a map towards possibly achieving some of that. And the, and the path will change and then your goals will change too.
[01:30:32] speaker_0: Mm-hmm.
[01:30:32] speaker_1: But you're not like, you're not lost. Uh, you can choose to tap into collective intelligence through these models and then go somewhere that even you weren't gonna do. Like two years ago, you probably thought your life was just gonna be this, but now your life, you can reimagine your life and just be different. You can be more, you can be... And I think that's a, this is, this is something that I try to do often.
[01:30:53] speaker_0: Mm-hmm.
[01:30:53] speaker_1: Partly because the models do so much work that I'm like, "I have to go for a walk and think about it anyways." I'm just like, "Whoa."
[01:30:59] speaker_0: Right.
[01:30:59] speaker_1: "I'm not gonna do that by hand again." Like, I was late on my expense reporting and then I told, I went to Cloud Code, I was like, "Can you just go into my email, find all the receipts, categorize all of them, put them into a folder by month?" And then it wrote a bunch of data analysis, data engineering scripts and then it did it all in 30 minutes and I was just like, "Well, I'm never doing expense reporting myself again. This is clearly how I plan to do it forever." And I'm like, "Man, we are one year away from never having to file our own taxes if we don't want to."
[01:31:23] speaker_0: Right.
[01:31:23] speaker_1: Like, just tell the AI to do your taxes. And I'm like, "I can't wait for that to be the other people to have that, not just me."
[01:31:31] speaker_0: AI did my taxes this year.
[01:31:34] speaker_1: Did it?
[01:31:35] speaker_0: Yeah.
[01:31:36] speaker_1: Yeah, I mean, this is, this is like, this is the right applicant.
[01:31:38] speaker_0: I had the data, had the data organized into a giant spreadsheet and then-
[01:31:41] speaker_1: Yeah, I mean, my personal finances-
[01:31:42] speaker_0: ... I trained a little thing.
[01:31:42] speaker_1: ... my investments, same thing. I'm just like, "Yo, AI, to help me..." It's like-
[01:31:45] speaker_0: Yeah.
[01:31:46] speaker_1: D- just take the smartest model and you're like-Clearly, I mean, I don't have enough of energy and like mental bandwidth to maybe think about these things. But there's no reason why you can't just throw a couple reasoning queries at it and, like, get that extra juice, make a slightly better decision today.
[01:32:00] speaker_0: Yeah.
[01:32:00] speaker_1: And it'll, it might just, like, buy you more time payoff later, and that's crazy. Like, you... So that, the application of AI to yourself is actually really important to me. It's not just like do it for your job. That's a ridiculous... Like, you cannot define yourself only by what you do for work. Like, your job might not exist. That company might fail. But if you apply the technology to yourself, your own growth, like that's... No one can take that from you. Right? Like, you can become... You can get the job of the future, right? Applying it to yourself. Apply it, like, "I want to learn this. I want to become better at this. I want to..." And that personal growth, like that, that's just the greatest investment that you can make, is into yourself. And then once you get, you get comfortable with that, then you start investing in the people around you.
[01:32:38] speaker_0: Yeah.
[01:32:38] speaker_1: And then your whole team, your whole squad is just like, like amplified. Exactly.
[01:32:44] speaker_0: All the automations I've been setting up in this last year is be- is right before my daughter comes here.
[01:32:50] speaker_1: There you go.
[01:32:51] speaker_0: I've been-
[01:32:52] speaker_1: There you go.
[01:32:52] speaker_0: ... trying to set up all these automations so that I actually can be offline safely and still-
[01:32:56] speaker_1: Yeah.
[01:32:57] speaker_0: ... have work and business-
[01:32:58] speaker_1: Good.
[01:32:59] speaker_0: ... kind of taken care of.
[01:33:00] speaker_1: Congratulations. Uh, I mean, a little early. But I'm like, I mean, your, your kid is going to grow up with a ton of ambient intelligence and a dad that's, like, incredibly curious about, like, making sure that, like, this is set up well for them. That's... I mean, I mean, that... I mean, like, we're Bay Area kids, so like we're-
[01:33:15] speaker_0: Right.
[01:33:15] speaker_1: We're lucky we grew up with, like, so much positive mindset around tech and what it does for people, like what it's done for my family in my lifetime.
[01:33:21] speaker_0: Mm-hmm.
[01:33:21] speaker_1: It's changed, changed our whole, like, lineage. Like-
[01:33:24] speaker_0: Wow.
[01:33:24] speaker_1: ... parents went from like villages to working on semiconductors and, and like sequencing the genome. And now we get to work on intelligence. And like imagine what we get to do for our kids through this. They're gonna grow up with this. They may... They will take it for granted that they have superintelligence. That's incredible, right?
[01:33:39] speaker_0: Damn.
[01:33:39] speaker_1: Like I'm 30. You're like, you're in your 30s. Like, this is-
[01:33:42] speaker_0: Yeah.
[01:33:42] speaker_1: For us, it's like we're, we're, we're slow. Like, we, we are rewiring ourselves. They won't have to rewire themselves.
[01:33:49] speaker_0: It's just native.
[01:33:50] speaker_1: They'll be native, yes. Yeah.
[01:33:51] speaker_0: Oh my God. Oh my God. Dude, you're blowing my mind with that perspective. Again, I get chills. Because of my nerve damage, I get chills on my left back part of my brain-
[01:34:01] speaker_1: Yeah, yeah.
[01:34:01] speaker_0: ... when a new idea or a new pathway happens and I feel it as like a-
[01:34:04] speaker_1: Yeah.
[01:34:04] speaker_0: ... like a shock.
[01:34:06] speaker_1: Yeah.
[01:34:06] speaker_0: You just gave me one of those, hearing you say that.
[01:34:08] speaker_1: Yeah, clip it. Clip it, clip it.
[01:34:11] speaker_0: Will do, will do.
[01:34:12] speaker_1: Yeah, yeah, yeah, yeah, yeah. But, but this is, this is... It's exciting and it's uncharted, but it's like we, we get to be pioneers for sure.
[01:34:19] speaker_0: Where could people find your work? Um, where, where would you like people to know about what you do? Or do you want to be secret and not share anything?
[01:34:26] speaker_1: Um, you know, you can go to my website parth.club. I'm trying to make it a better place for finding more about me. It has links to all my social. Um, also most channels I'm Parth Intelligence, like @parthintelligence. I have that handle most places. So, but yeah. Um, and if there's, like, something that you wanna, you wanna learn or you're trying to unpack similar ideas, like, feel free to, like, drop it in the comments of a video or something. Like, I, I, I try to engage with people and try to, like, build this, like, collective knowledge. And, uh, even if I can't hop on a call necessarily with everyone, I do like to share ideas, like fairly frequently with like, like, I don't care. Like, if you're new to this or not, I do like bouncing ideas. I think collective intelligence is huge for me. So, I just appreciate everyone who's kind of gotten this far and like listened to this, this like philosophical ramble/tech world. Like, whatever. Just happy to be here.
[01:35:13] speaker_0: Parth, thank you so much for joining. And God, I really appreciate you. And, uh, I'll go ahead and end this. Uh, I'll end it now. (instrumental music)