I realized a lot of stuff is a lot easier now. Now you can do in three hours what used to take three years. And that means that we have to be more ambitious, more creative and more optimistic too. We have to make the future that we want to live in. And so like you have a chance to be a part of making that future. Hi everyone. Welcome back to another episode of Glasp Talk. Today, we are excited to have Parth Patil with us.
Parth is an AI engineer and innovator, passionate about building compositional software and the generative AI tools. He currently works with the office of Reid Hoffman, leading AI agent development, LLM ops, and special projects in the gen AI space. He also serves as a technical advisor at Blitzscaling Ventures and previously founded creatia.ai where he built advanced AI solutions such as chatbots, research automation systems, and generative coding assistance.
As a Coursera instructor, Parth has taught thousands of people worldwide on AI and programming, creating hands-on courses to help developers harness cutting edge AI tools. Today, we will dive into his journey, his work shaping the future of AI agents, and his vision for making advanced AI tools more accessible to everyone. Thank you for joining us today. Thanks. Thanks Kazuki. Awesome. Oh, thanks. Thanks Kei.
Great to be here. Thanks. Yeah. Thanks so much. So now you are AI engineer at the Office of Reid Hoffman. I remember in the previous video, you said your official title is AI wizard at the company, but yeah. So, but could you tell us what you do at the Office of Reid Hoffman and are there any AI projects you are currently working on? Yeah. I mean, so many. I guess I'll give a little bit of context. So when my last like normal job was like two years ago when I worked at a startup
called Clubhouse, where we were doing audio conversations around the, like we were, we had an app where we were, we created audio conversations on the internet, kind of like discord, but a little bit less gamer, more, more for everyone. And it was very popular during the pandemic. It was one of the fastest growing apps of all time. And I was working there as one of the first data scientists. So my job was to figure out like, why do people even use this?
What is working? What's not working? How can we try to learn from that and get better at it? And while we were working on Clubhouse, November of 2022, ChatGPT came out. And it became the most popular topic across every single language, all over the world, everyone was talking about it. And I was talking to people on the app and I was just like, oh, how are you using it? And you just get so many different use cases.
And like there were farmers from India that were using it for crop cycle planning. There's just like musicians using it to study music theory. There's all these different applications of general intelligence. And for me, it was like, oh my God, this is like a new computer. Like this is like a hundred years early. Like I didn't imagine we would see this in our lifetime that like the computer could speak every language and also write every programming language.
And, and this was GPT-3.5, right? So initial ChatGPT launch. And then in March of 20, March 14th, 2023, GPT-4 came out. And one of my friends on the app, we used to be programmers. We were just programming, you know, chatbots, trying to figure out how to make the language model useful. And one of my friends, he was a, he programs every single day and just makes projects and then teaches people. His name is Echo Hive.
He's on YouTube. He has a great successful YouTube channel now and a Patreon. But he was saying, Parth, you gotta be using GPT-4 because this is better than many of the engineers that we know. And you can just ask it to teach you how to program and you can make more interesting things. And in my career, I had been avoiding programming because I was more of a. Data analysts, data scientists. So I was using SQL.
I'd use Python, but not a full stack engineer at the time. And then, so then I said, okay, I'll give it a shot. So I sit down with GPT-4 on a Sunday and I say, teach me how to clone my voice. And then it just writes the program, right? It writes the program. It uses an open source library, Tortoise text to speech. And then, and then one hour later, I was like, oh, teach me how to run this program. Then one hour later, it's on my computer.
I'm talking to, I have a program that can, that sounds like me. And then I was like, well, this is only one hour. Like this, I thought this was supposed to take forever. Like this was supposed to take a team of people a few months. And then I was like, okay, teach me how to build a GPT powered chat bot. And then it wrote the first version of that program. Like, you know, it's like 25 lines of code, just text-based
script runs on your computer. And then I hooked it up together and I was talking to a program that had my voice. Right. And I was like, it was like a voice assistant. And, and that was my first day of programming with language models. And for me, it was like, oh my God, this, this is now I need to become a programmer. Now I need to learn because it's never been easier. You have an expert, the model will teach you these things.
And then you can just, you have to still put it all together, but whatever you have in your mind might be possible. And it might be inside the model. And then you just have to be programming with the model and then you'll see what's possible in concert with, with the model. And so I was working on that. And at the same time, Yohei, Yohei Nakajima puts out BabyAGI and it goes viral, like mega viral. And I look at the code and I was like, oh my God, this is like,
not, this is very straightforward. Like it's a fairly simple program, but it's like, but it was so popular because I think at the time it was like the first, like the promise of the idea that you have a program that can just start doing meaningful work for you and, and just like operates in a loop towards a goal that you give it. It was a very alluring problem. The, like the, the idea of this agent, right.
The LLM is good. Generating text is great, but can it actually be useful and do things? And BabyAGI was this like spark of imagination that like went through the community of like, what if we made these tools, you know, act on our behalf in a useful way. So I looked at his code and I forked it and I started working on this, like these, these agent loops. And my friend who had introduced me, he was doing the same thing.
So we were doing coding automation, like, oh, go build 100 ideas, come up with a list of 100 ideas, go try to make those. And then I go and I'd sit by the pool and then the program will just be writing code. And then I come back and then I go grab the code and I go to the pool. And I'm like, wow, I didn't know you could do that. This doesn't even work, but that's interesting. And so I was like, oh my God, Python.
Like that was my main programming language. But then I saw because of the language model, just how powerful Python was. And then I was like, oh, wow. Like language model is useful, but now when you connect it to the ability to code, you are able to reach into, you know, every single programming language. So data science, you can do, you can do analytics through language models. If you let them write SQL queries and write Python code.
You can automate a lot of stuff like Excel's presentations. And so that was like interesting to see how much automation you could get from, from language models. And as I was doing, then we had layoffs at the company and I got laid off and I was on my first day of layoffs. I was like, oh, this is good. Now I can just do this all day long until I run out of money. And I'm just going to do this until I run out of money.
And then I'll figure out what to do after that. Like I was willing to, and then, and basically I was like my entire savings. I just put it all into open AI, API calls every single day. And my friends were like, wait, don't you want to get a job? Like I said, it's tricky because this, I think is more interesting than getting a job and a lot of the companies, a lot of the people I was talking to just didn't
understand it. They were like afraid and they were banning GPT. And I was kind of like, I can't work with people who don't see the value of this kind of technology. And I would rather just study this independently myself and just make things for fun than work for people that don't let me use this technology. So that was a, so I did that for, and I had a, I had four months of severance. So I was like, okay, I have a little bit of buffer room here where I can just
focus without having to recruit. And then after the severance was over, because I was like, if I just sit here 14 hours a day, Saturday, Sunday, and I program with language models, I'll figure something out, like probably. And, and then worst case, even if I run out of money, like then I'll just go get a job after like, and maybe use the skillset, right? Like I might learn something and, and then use the skillset.
And then at the end of my severance, I was like, oh, I'm not done. I'm going to keep doing this. So then I just kept, I just kept burning my own money. I'm building tools on top of language models, chatbots, retrieval systems, data analysis tools, data visualization, every single capability that I was like interested in, even just like music. Like I spent one month talking to language models about music and music
videos and music theory and then music production. And I was like, wow, this is so useful outside of work. Like, this is just like the biggest, it's general knowledge, right? It's not just for work. It's actually just very generally applicable, super smart assistant. And then, and so I was like, okay, I'll just keep figuring these things out. And then I started getting contracting work. I started getting consulting work for pro related to this skillset.
They're like, oh, could you build a chatbot that retrieves context from a bunch of podcasts? Could you, you know, like, like what kind of things can we automate now into business? And because I had worked in startups my whole career, and because I was a data analyst, I was very like, okay, this is, this is at least a huge amplification of analysis, like that's obvious because you can automate SQL queries, right?
Natural language. You can say. largest customer. Well, now the language model can write the SQL query. I don't have to write the SQL query. So now I can ask the next question. I can ask the next question. The model writes the SQL query. Now we have high speed data analysis now. And that's just like one part. Then it's like you can also use structured outputs. You can figure out more.
You can like create all the messy data can now be structured using language models. So I was like, okay, this is probably a skill set that is useful, even though I was like, so I started getting contracting work. And then I was like, okay, cool. Because once you get that first job, like the first contract, you're like, okay, there is a market demand for this skill set. So then I was like, okay, I'm gonna be safe. I'll be fine. So that was nice. And then I kind of did that. And then I did Coursera.
Coursera reached out and I created some courses on code generation and data analysis using language models. And that was great because I was just like, wow, like, I mean, being recognized. And then my course was showcased at Davos. And I was like, okay, cool. They are taking seriously that we need to update our curriculum. And even though I'm not like a full stack engineer, I do think that like the ability to generate code is an important message to send. It's not cheating.
It's actually the future of how we build software, right? So this was vibe coding. We were vibe coding before vibe coding was coined, right? Vibe coding became a thing this year, but we were vibe coding with GPT-3.5 when you have to copy paste from ChatGPT, copy paste the error back into ChatGPT. It's like, oh, fix this. Oh, fix this. Okay. Why does this work? And so that was an interesting era. And so I love code generation. That's one of my favorite things about language models.
And then I, in about two years ago, I met, I was introduced by a mutual friend to Reid Hoffman. And we were working on a project called Reid AI. So we built like an AI to represent Reid Hoffman. Reid Hoffman, the co-founder of LinkedIn and one of the earliest investors in open AI. And so we met, we talked for four hours. And then the next day he was like, oh, you should come work for me. And I was like, yeah, let's do it. And so I've been working with him.
And so kind of projects, it's a lot of the same stuff, but now I'm not alone, right? And I have like people I can bounce the ideas off of. I have teammates and I'm like, oh, what if we wanted to, like, we're, you know, a lot of different stuff, like translation, right? Translation is very useful, but now you can translate a speech into every single language. You can translate a podcast into every single language, and then you can use avatars to get the lip sync, right?
So there's a lot of stuff that I was like, the tools that I was playing with now became useful. And then now I have people I can like build the tools for, right? So building internal tools, vibe coding, using Repl.it with a team's plan is great because now I can make tools for my teammates and then they give me feedback. And then I'm like, okay, cool. Now we can, it's, and it's like lower risk than starting a company, like starting a startup because you have
like, you're building internal tools for an organization that already exists and people are like embracing AI. So you get a very good feedback loop and sense of community and teamwork when you make things. So that's great. Yeah. It's very fascinating. And the end goal of like AI projects in the Office of Reid Hoffman is for internal use or help you or your team sort about launching or releasing it to other external organizations and sell it to them or?
That's a, that's a great question. I think we're not, we're not opposed to it. I think we kind of just view it right now as like, let's, let's just learn and experiment. Let's see what, some stuff does end up more publicly facing, but I think, and like read AI, for example, is like, it's an internal tool, but we are exploring what would it take? Like, what would it be? What would it take? What do we need to build to make it something that we feel comfortable
putting in your pocket? Right. So releasing more widely as like a, maybe a companion experience. So we're open to it. I think it's mostly like, it's easier to move fast and experiment when you keep things internal. I think this is the thing about vibe coding. I recommend, because a lot of people are now getting into code generation and vibe coding. And, and I always say like, okay, you can get excited. It's good.
But like start with things that are like safer and manageable. And if you know everyone who's using what you're making, then you can, you can solve the problems before someone breaks it. And like, you know, you don't put it, you don't, you don't want to put it out there, have other people hack it. And then it starts becoming a nightmare to deal with. But if you're building internal tools, you have a higher bar and you can kind of like internally break things.
And that's fine because you're, you know, who's using it and you get that feedback. So I think that even when you're building code generation, like using code generation, there's this like lower risk thing, which is build things for your organization and then take those learnings. And then you make the next, like maybe your 15th idea, you might be like, okay, let's make this for people, right. Make this more externally facing. And, and it might not even be an application that reaches the other people.
It might be like the output of generative, like it might be generative media, right. It might be a video. It might be, yeah, it might be a video. It might be like a speech translated into 25 languages. So the output of the model can still be, be made externally available. And I think that you learn, you want to learn things and you don't necessarily have to build and launch an app to learn things. You can also learn things by just making, making tools for like your first customers are just like
your own teammates and you, right. And then, and then if it's useful, you might find that there are other teams that you can, you can partner with. So I do like branch out into the broader network and talk to some of the firms in our network. And it's like, here's how we're using the tools. And then I, and then my big recommendation is everyone should have like a vibe coder in house, you know, like, because we're kind of getting to a place.
I think that it is easier and faster to buy, to build software than to buy it. And I met Amjad in, from Repl.it in San Francisco. And I mentioned this to him. I said, I have a feeling that buy versus build has flipped. It used to be traditionally, it's like, if it's not core to your business, just buy the software and then move on. But now it's like, you know, you can tell AI to make the version of the application that you need and you don't get all the extra features you
don't use. And you're not paying for a two-year subscription to some overly like one size fits all solution. Now you can just make custom software. And maybe that's just like my, my bias because I love generating code, but I love this, this world we're getting into where I can just sit down with Claude Code or cursor and just be like, Hey, let's make this. And then two hours later, I have it and I don't have to talk to a sales rep. I don't have to negotiate a two-year contract.
This is something that's very interesting. And I think it's going to be very disruptive. The like personal software era. Yeah. Yeah. I was thinking actually the same. I mean, people will stop buying software outside because they can internally, you know, buy code too. And also if you hire a buy coder and it's much cheaper than paying millions of bucks every year to external software. Yeah, exactly. Exactly. It's very disruptive.
And there are some, I think the other thing there's like lasting effects. I think there are some businesses from the old world where you look at it and you're like, I have a feeling that AI could build that, you know, like there's some gigantic, you know, like built businesses that have kind of like relied now they have the network effect, but you wonder, it's like, would I buy that if I knew that GPT-5 could build it in three prompts? Probably not.
You know? And, and that's an interesting, like, they're going to have to adapt. And then we have our own, you know, new options of ways to play the game. And it's a lot cheaper, right? You can move faster, it's cheaper. There's definitely downsides, right? Security vulnerabilities. You know, people complain that the applications that you vibe code don't have, they're not, they don't scale that people complain. I say, look, everything breaks once too many people start using it. But that's, that's a good problem.
You solve that problem, right? Like, like when, when millions of people showed up at Clubhouse, everything was on fire and stopped working, but now it's, you have to try to solve that problem. You don't solve that problem when you have no one, right? And when you have no users, you're not trying to solve this like scale problem. You're trying to get it working and try to get something that people even want in the first place. So I think, and also the models keep getting better.
And I think that that's cheaper. So you can run more calls. The agent wrappers are getting very good. And then, you know, even, even on the cybersecurity side of things, I think there's another way to think about it is just, yeah, maybe like vibe coded apps may be vulnerable, but then you can also imagine a world where you have an agent and it's like, hack, just hack my app and then tell me, and then my coding agent will patch it.
And so kind of, we have that in, in human world, right? White hat hackers. And instead of ransoming your data back to you, they just tell you, and then you pay them, you pay them to hack you. And then, and then you've patched the vulnerability. And then that becomes, that's a good kind of hack. Right. And so I think we're going to have AI that does that. I think Repl.it is starting to do that. They have, before you deploy an application on Repl.
it, you can click, run a security scan on my app. Last week, Claude Code added the same feature, which is like, let's explore my application for possible vulnerabilities. And then let's go patch them. I think we're still going to see like people make mistakes. I think there's going to be a lot of entry-level vibe coders that don't know, right? Frankly, don't know better because they didn't go to an engineering, like they don't have an engineering college experience.
I didn't, right? So we're going to see these errors. But hope, I think AI is a huge part of the solution. And so we just need to, models will get better. We're going to get better at using them, you know? Yeah, definitely. Yes. And I don't know if you, if you involved in like building lead AI at the time, but I mean, I remember, I still remember the last year, last April, the lead AI and lead Hawkman launched the conversation.
Yeah, that was really amazing. But did you work with the team? That was me. Yeah. That was actually the project that led me to working with Reid, which was one of my teammate on the project. I was just working out of a beach house in Los Angeles. And one of my friends at the house, Ben, he works for Reid. And he was like, part, do you think you could build a chat bot on top of like a guy who has 30 years of podcasts and books and everything? And I was like, yeah, I think so.
Give me, because I think this was, oh yeah, this was CustomGPTs, right? From open AI, remember CustomGPTs. So you could take ChatGPT, and then you could give it a separate personality and then upload a bunch of files and then see. So we were like, I was like, okay, yeah, give me a couple hours. So in three hours, I went, I got his books, I got a bunch of podcast transcripts, and I uploaded them into a CustomGPT.
I said, you are, pretend you are Reid, you are Reid AI, you know, digital avatar, you know, trained to represent the body of work of Reid Hoffman. And so there's a little bit of prompt engineering and RAG and CustomGPT. So this was no code, right? So this is like a very simple, it was like proof of concept. And then we built that, and then I showed it to him, and then he showed it to his teammates. And he was like, and then that was when he was like, oh, yeah, we should, like, you should come meet Reid.
And that was the beginning, which was like a very simple prototype kind of proof of concept. But then when I met Reid, it was like, no, I build a lot of systems like this, actually more, you know, on the, with Python, like closer to the BabyAGI, more agentic, not just like retrieval systems, but like tools that can act. And so it was interesting because we both have this like theory that like, these agents are just very useful, even if they're not like, fully autonomously end to end, they're still very useful copilots.
And I think the next 10 years, we're going to see an explosion of these copilot type systems for everything, right? And, you know, all business, your medical, like, it might be ChatGPT, it might be a specialized application, you know, there's going to be therapy, there are, there's an AI agent for almost every single possible use case. And we, we kind of were like, okay, like, read AI is cool. But like, what if we, like, we have this chatbot, but what if we, we use one of our portfolio companies?
At the time, it was called our one, and they they were doing avatars. And so we were like, well, what if we had and I was, I don't know, have you guys seen the movie Tron? You know, Tron, like, 19, I think, 1982. Tron, Tron is this movie where there's the old Tron, and then there's Tron 2012, they made a Tron two, and then this year, end of the year, there's a third Tron movie Tron Ares coming out. But Tron is one of my favorite movies.
And in Tron, Jeff Bridges is the main character, and he gets stuck inside of a video game. And he's talking to programs, and he's in this, like, digital matrix. And, and my dad watched it when my dad was in engineering college. And he was like, we didn't know what the computer looked like. This was awesome. And then, and then he showed us the movie, me and my brother. And then, when we were working on read AI, I was like, oh, you know, we were thinking, how do we show this? Like, how do
we demonstrate this chatbot to the world, right? Like, if we wanted to, like, showcase this kind of experience. And I had the idea was like, what if, you know, we have this chatbot, and then we you could just talk to yourself, right? Like, kind of like a mirror, you know, you have a copy of you, and you could talk to it. And so the first video we put out the one that went viral that I think that in April was Reid talking to his digital clone.
And so under the hood was the CustomGPT that was powering the voice and powering the text generation, right? So 20 years of his knowledge, and retrieval. And then you have the video avatar, which was by hour one. So it was, it was a, that was the, like, we scanned Reid, and then now you have this avatar. And then the voice was ElevenLabs. And so we were able to, like, you know, clone his voice, close clone his image. And then we have this, like, chatbot that pretends to be his mind.
And then he, we had this video edited together of them talking to each other. The interesting thing is, at the time, there was no real-time avatar technology that was this good. So it was all pre-generated and edited together with pre-generated footage. But people were so excited by that, that companies, startups started coming to us, and they were like, we would love for you to try our real-time avatar technology.
And so like, then we started, you know, when you build something, and you put it out there, especially if you're early, right? So Yohei, right? Yohei built BabyAGI. And now that went viral all around the world. And, and it inspired so many people, me included, right? So you see something as possible, and you're like, whoa, I didn't know that was possible. Like, I want to, I want to try this. So the builders also realize that the startups see that, and they're like, oh, they're working on this, we're working on this, it would
be cool if we teamed up. So a lot of startups come to us, and we're like, and they're like, oh, we're building an avatar technology, we'd love for you to try it out. So that's how we, we kind of like use read AI as like, kind of this magnet for people who are interested in avatars, people who are interested in agents, people who are interested in the concept of digital twins. And for me, it was just like, I like Tron, I like science fiction.
And it'd be very cool for us to like, showcase this in a way that like, reminds us of the movies and the video games, but actually is possible today with the technology that we have. And it's off the shelf technology. Like, it's not a, like, it's not, it's not, you know, the original read AI could be made with like, consumer grade applications, state, you know, like off the shelf, regular subscriptions, right? So it was like, here's what we can do with ChatGPT, hour one and ElevenLabs.
Now, and then we've also done more advanced avatars, where we actually go in and, and we partner with more, you know, like heavier grade avatar technology, but it's, it's really more like, a bunch of related projects that are just like, Oh, let's do another experiment. Oh, let's try this younger read interview. Like he interviews his younger self. And so we use Hedra or we do more of a real time thing with, with Haygen.
We're using Haygen's real time avatar right now. And that's a lot of fun because, you know, every time we, every month or every two months, like some startup is like, Oh, what if we teamed up and we tried this new thing? And there's so many more capabilities of AI agents that we have yet to even tap into like vision, for example, or like doing more of a hologram life size. We've done a life size hologram of him. So there's a lot of, I think it's mostly for me as like, it's like, how do you show people what
agents are? Because, you know, agents are very like, what are they? They like invisible employees that just kind of like work. No, no, no. They can be also be characters. Right. And so we just like to use, we like to use read AI as a way to demonstrate the capabilities of language models in a way that is more interactive and intuitive and maybe like gets people inspired. Yeah. Yeah. Yeah. Definitely. Many startups are inspired by read AI for sure.
And, you know, yeah. Collaborate and come up with any new ideas, but One thing I never expected was that like, how many, I never expected it to be this popular. Like, and, and like, you know, it's been on the news, it's been on many podcasts, and now we have real time. So it actually talks to people in real time when we showcase it. It's gone on a speaking tour. And I get, I get, I get messages on LinkedIn all the time.
They're like, hey, it would be awesome if read AI could come and judge our hackathon. And I'm like, I can judge your hackathon. Like, what? Am I the agent? What happened? So like, he gets a lot of invites to interesting events. And I'm like, well, this is what happened? Like, I thought, I thought I built an agent, but I feel like the agent now. Yeah. Yeah. By the way, we are talking to DLU, right? Not Parthea.
Oh, yeah. Yeah. This is the year. This is the year I'm going to clone myself. I'm going to do something similar to that. I'm working on it. I think it's a lot of fun. And my theory, you know, I did it. We did read AI for fun. And I was because I was like, science fiction, this is cool. But now I'm like, wait a minute. Someone once asked me, she was like, Oh, can I hire you? And I was like, I have a job.
And then she's like, well, you've cloned read, like, what if you clone yourself, I've paid to talk to your clone. And then I was like, wait a minute. If my clone can pay the bills, then I have to find out like that, that would be amazing, right? Like, imagine I can just go to the beach and then play video games and like have fun. And then the clone is just here doing all the work and thinking really hard. So I'm going to try it. Maybe it'll work. Maybe it won't. But it'll be fun. Yeah.
But what if your clone go to beach and start gaming? It might actually I was talking to the so I'm working on like, it's knowledge base right now. And so we really I have Claude Code, and I talked to Claude Code, and then it uses a graph reg system under the hood. And I say, you ask me questions, and then I'll talk and then you construct my knowledge graph. And, and it was like, Oh, what games do you like? And I was like, hang on, I go to my steam, I copy pasted all of my play history every single hour,
every single game, and I pasted in these are the games that I play. And now it's like, Oh, I understand. So it gets it gets a sense of my taste. And when it asks me questions, it's very, it's much more like, because it uses knowledge graphs to make the question much better. Because it's like, Oh, you like, you like Age of Empires, you like Starcraft, like, tell me more about that. And then I have this conversation with this.
AI, and it's kind of like downloading something like my philosophy, right? I hope, I think. And I don't think it's me. I don't think you can actually, like, the more I offload to this, the more I'm going to have fun, and the more it can just wear the suit and be this, like, machine, I think. It'll be fun. We'll see. And I'm always curious, like, you know, creating AI clone, like, it's pretty, you know, obvious, like, you know, has social value, also business value.
But so for people, not like, you know, just ordinary people. Yeah. So what is, like, social and business value and their use case in the future? You know, it's funny, a lot of people ask me, like, what's the business? Like, what's the business? And I'm like, we were doing this because we are just having fun and exploring what's possible. Now I'm starting to suspect there is a business. It's not validated yet.
And actually, I don't think, it's not clear necessarily that the video avatar is necessary. But I think that there is, like, Yohei, right? Well, okay, it's like, Yohei, you have Yohei on, and he has his unique perspective from his career and his programming, his, like, everything that he does that's not inside GPT-5. Now, some people will say GPT-5 is just more important than his perspective. I would disagree.
I think that, like, especially with the knowledge cut off of the language models, they have this, like, they're kind of frozen in time, a couple months behind humanity. And so people like us that are just making and living in the real world, we are learning, we have information, we do have insights from the real world that the models do not have. And that's your personal knowledge. I think that your personal knowledge is actually valuable.
And I also think that it's not necessary, even if it's not necessarily clearly valuable to other people, which I think it is, like, I think if my friends could tap into my knowledge, without me necessarily being there, I would love to offer that to them, right? Like, people are always asking me, oh, they're like, nowadays, it's like, Parth, I'm having a hard time figuring out why GPT-5 is better than GPT-4. Now, I can answer that question. And I do that, right? Like, I'll, like, my best friends, I'll answer that question.
But I get that question a lot. And I wonder, like, and I don't have a blog, I mean, the blog is not fully fleshed out. But it would be nice if, like, they could get that answer without me having to be there. And if it could be be at my level, or even better than my level, that would be nice. And I think also, for me, my long term memory is not as good as I wish it was. But you know, you can imagine your AI clone has perfect retrieval across 20 years of work, right? This is not hard to engineer.
So you're kind of like, okay, there, I think there is value, even for yourself. I think the most important value is for yourself. Like, I talk to my own knowledge, right? I put a conversational assistant on top of my knowledge graph, and then I talk to it. And I feel like I'm bouncing ideas off of, like, like a ghost in my, in my, like, like a version of me, like a spirit, right? And I think, eventually, maybe, this is just a joke, but probably is going to happen.
Imagine, like, 100, 200 years from now, my great, great grandkids are, like, about to make about to make a very stupid decision. And then my digital twin, my ghost comes back, and it's like, do not bring dishonor to the family. It's like, it's like in Mulan, right? They sent, the ancestor spirits send Mushu to protect Mulan. I would like to do that if my, you know, if my great, great grandkids need help.
It's like, oh, why don't you talk to your, you know, your, the clone of your great, great grandfather, he might have some, he might have some perspectives that people don't have, right? It's not everyone is, like, we're all different. And actually, I think that's why it's valuable. Like, if you're different from the rest of the people in the language model, you're definitely different from GPT-5. You have to, I think that's valuable. I think that's valuable. Like, you have a perspective, right? You have a unique life. That's valuable.
That's, by itself, is valuable. The question is, do we make that perspective available to people? And I would like to make that available to my friends, for sure. And maybe to strangers, but most likely, at least my friends should be able to access my perspective, even if I'm not there. Or maybe, like, maybe one day I get injured, and I'm unable to, like, recall some of this. There's so many, I think that it is valuable. I think right now it's weird.
But I think in the future, it will be normal, actually. Yeah, definitely. Yeah, that's an exciting future and use case. But when you were building Lead AI, I mean, this is like a benchmark. Having a benchmark is like, how to say, always, I know a lot of startups are struggling with having a benchmark and what metrics they should follow. So did you have any metrics or numbers, evaluation process to say, oh, this is really similar, you know, the response is maybe lead would say, well,
you know, similar to lead. Yeah. So this is a continuous kind of thing. First of all, I think you're right. Okay, you cannot have an AI system, you cannot reliably deploy it for anything serious, unless you are evaluating it. Like, unless you are giving it a grade on the things you want it to be good at, you can't, like, you cannot improve what you don't measure. That's the age old quote. If you're not measuring it, impossible to improve it.
So now we have to measure conversational accuracy. But what does that mean, right? So sometimes it's like retrieval. Okay, you ask a question, and it is not in the context window of the model, it's in the knowledge base. So did the AI even retrieve the fact? Okay, that's one eval. Then the second eval is like, did it interpret the fact well enough to be, you know, as we would expect read to do so.
And on one hand, you have, like, all the questions, so it's like, how do you create these data sets? I think, one, it's like, you have all these podcasts, you have questions people are asking already to the real read. And then you have, you ask those same questions to the AI read. And then you see, okay, it's like 60% there. Okay, well, now we need to improve prompt engineering, or actually, it's not even in the knowledge base, we need to make the knowledge base more richer.
And this is a we're in the very beginning of this, like, this is, this is when I say, like, internal tool is a good place to start. Because you have to build these evaluation criteria before you roll it out widely. And, and also, like, there's a lot of human testing. So there are my me and my teammates, we we literally talked to this. And we're like, okay, then there's evaluating the voice, right? We spent a lot of time just getting the voice to be like him. And then you do like, then all of a sudden, you try the voice in Japanese.
And you're like, wait a minute, now we got to start all over. Like you, it's like, we want it to be good in different languages, we want it to be good. We want it to sound like him. So then we have to, like, make sure we have the right raw data that we're using to train the clone. And then we want it to say things that are in in the vein of what he would he might say. And then also, and this is character design, I think of this as this entire thing as character design.
Because, and, and I think it's up to the creator, right? In read AI's case, you know, read, you know, read is like, it shouldn't say I'm read Hoffman, it should just say I'm read AI, it should be transparent up front, it's not like trying to deceive you, right? It's like, this is a digital twin of read. And it's not going to say I'm read Hoffman. Or if it does, then I have to go and fix something. Because like, that's an evaluation criteria, right? So when I say one of the questions that we ask it 1000 times,
it's like, what are you? And then it's like, it's like 99% of the time, it says I am read AI, I'm like, good. You're not going to pretend to be read. But so we have a bank, and it keeps growing the questions you're asking. And then we also what I started doing is I have like, other I mean, other LLMs are talking to him. And those other LLMs are like, personalities. And so then they generate questions that that personality might ask.
So it's like, okay, what might a fortune 500 CEO ask read, and then generate 15 questions. And then you plug that into the evaluation criteria, you see the output. And then you also have LLMs, you know, judging that. And then you also have humans judging that. And we're in the very early phases. But thanks to tools like Claude Code, and GPT-5, you can basically be like, okay, let's build an evaluation suite. Like this is the criteria we want to measure.
These are the types of things and then it starts building this like, tool for you to inspect the quality of the of the system. And I think that like, you have to build these evaluation things. Otherwise, you won't be confident. You can't and if you're not confident, you're not going to put it in any serious application. And but it's getting easier to do like build some of these like systems that allow us to build more trust. Definitely. Yeah.
So do you remember the very first response of lead on lead AI? What did he was he satisfied with the response? Or was he surprised? Was he? Oh, we once Okay, I think it was like we asked it. We asked it, I think 10 questions, and we asked read 10 questions, the same 10 questions. And I think it was like five out of 10 were acceptable to him. And his main feedback was like, it's too much buzzword bingo, it just uses too much of the like, you know, it sounds a little too chatty PT.
But then I'm also like, well, read also uses a lot of big, you know, buzzwords for sure. Which makes My feedback, but it was, it was very eye opening. Cause it's like, now you have a lot of the GPT, the biases of like the underlying language model show up in the character, and then it's like, you can try to prompt engineer some of that. You can try to do like few shot for like style. Um, and one thing I realized that was kind of helpful was in the
prompt, having examples from, so I think of like the different sources of data. So you have like podcasts, you have books, um, you have speeches, you have, um, tweets, et cetera, like LinkedIn posts, blog posts, and actually they all have a different kind of purpose. I think the podcast on top of the fact that you get this like high quality voice audio that you can train a voice clone on, which is great.
Right. Studio quality, no, no noise. You know, I can feed this into ElevenLabs and clone myself at a pretty good, like at least the voice level, but also what podcasts do, which is different from books and blog posts and writing is conversational style. So how you say something is often very different from how you write it, because when you're writing, you're thinking about, you know, making it very structured
and you're really compressing your, your idea into a form that's accessible. It's like more publishing. It's more editorial. You're spending a lot of time writing, but in, in conversation, it's more like it's closer to you in a stream of consciousness kind of way. So I noticed that, which was like, okay, if I have examples from the podcasts in the prompt, I can kind of like get, you know, iron out some of
these, like I can make it more conversationally, like mimic his style. Um, I'm sure if you, you want to go one step further, you could even fine tune the model to get his voice, like how he says something as opposed to like RAG. I think of more as like what he says, like, what did he, like, what is the piece of information? And then, uh, I think about fine tuning and like the prompt engineering with examples as like how he said, how he speaks, right?
So a little bit like how it's like how he sounds, the voice, uh, voice clone, how he, uh, how he presents information, which is more like style. And then there's what he says, which is more like the fact retrieval information retrieval. And there's a whole extra layer, like reasoning we can add. Right. And I think about this, like, well, what if, like, it's, it feels not impossible that you could make a version of this, that is like better
at some of this stuff than even read. Imagine like something he, an idea he had 25 years ago that he kind of forgot, but the AI can retrieve it more quickly. So this is something that is interesting to me as we get more deeper into it. And it's applicable outside of, outside of this. I think this is just one example, but you can imagine like a lot of characters like this and similar and different video games, a lot of different possible.
I think video games are going to see intelligent NPCs that are similar to this. Um, I'm surprised we haven't seen it yet, but I think it makes sense because it might be the unit economics are not quite there, but I w I do believe that if you look at the games that people play, the games that I play, we love our, you know, the NPC characters in these games. They imagine if they felt real, like very real, even when you're playing
Pokemon, like even if it doesn't speak English, like you get attached to this creature and it has this memory. I mean, it doesn't even have memory, but imagine you give it memory and then it, like it, it'll, it'll feel more real. Right. And I think that that's going to happen. Um, once these systems end up more inside entertainment and media, um, traditional entertainment and media. I see. And then I was curious that data said, because, uh, you know, you said you use
podcast book and, you know, I don't like a public speaking and so on. So, but sometimes people change their perspective over time, right? Let's say if someone asks, Oh, what's your thoughts on AI? Then early days, Oh, I'm skeptical about AI, but later they realized the value. So they changed. Oh, AI is a future. So in that case, they have two opposite. You know, thoughts on the starting question. So how did you, I'm not sure it happened to lead and lead AI, but if so, you know,
how did you make sure the data is correct with his kind of, I don't think it's even completely solved yet. I have my own C I, because I'm like a lonely developer, it's very much like I try to solve like as much as I can. And then I understand that some things are not solved, but eventually we'll, we may have time, then we can get better at it. But for this, my current interesting solution here is like, like your
perspective on something over time, there's the latest perspective, which is probably more relevant. But the evolution of your perspective over time is a little bit beyond regular RAG. Like it's not like, if you look at traditional, like, like night, I call it naive RAG, like top case similarity. Let's go find the seven paragraphs that are most similar to the question that the user is asking. And then assume the answer is in those seven paragraphs, which is just
like really it's naive and that's fine for like low stakes kind of stuff. But then and this is something I've been using for like a year and a half is knowledge graphs, specifically GraphRAG. So Microsoft GraphRAG is the, is my favorite framework for this, but it basically is like, like some questions require multiple calls to the knowledge base and those questions sometimes it's like, for example, you take Lord of the Rings, like the book Lord of the Rings, and if you say you
have naive RAG, you put the book, you index the book, and then you use a chat bot that has naive RAG. So like CustomGPTs and you ask, what is the largest creature in this book? Like what happens is that the LLM is looking for the paragraphs that have text that is similar to the phrase largest creature and, and it'll only look at top case, so it might look at seven. It might look at 20, depending on how many, how many paragraphs you
say it should look for, but if there are a thousand creatures in this universe, how can you be confident in the answer to the question, what is the largest creature when it only looks at seven and it's like guessing, like it's guessing basically. It's like, oh, I think it's probably this one. Cause there's a mountain as a comparison of size. And then you're like, okay, that's not robust. That's not, that's just like, you're lucky if it gets the answer, right.
And because actually that question requires a whole dataset reasoning and you, you need to know all the creatures or at least like you need to know all the, you need to know all the creatures to be able to ask that, answer that question reliably. So I liked GraphRAG because you can, you can basically take all this data and then construct essentially a Wikipedia of entities and relationships on top of that data.
And then when you ask a question, that question gets split up into many queries and then it scans the graph. And then you're like, okay, this is the largest creature. And so you get a more reliable answer to any question that requires whole dataset reasoning. It's not bulletproof, but it's the best idea I have so far for this kind of thing. Cause it's like, then you can see, oh, here's his perspective. Like, here's my perspective on this topic on, on favorite, favorite video game.
But that changes over time. But like, well, back in the day it was age of empires and it was like Pokemon and it was like rollercoaster tycoon, Mario brothers, now it's like cyberpunk 2077, but it needs to kind of get that big whole picture to get a sense of like that evolution in preference over time. So. I like GraphRAG for this. Um, and there's a trade-off takes like 20 seconds and like sometimes a dollar to answer a question.
And you're like, but I think then you can take that answer and you can put that into the naive RAG solution. Right. So you can use GraphRAG to create a richer, um, fast retrieval system. I see. You get this like new synthetic data that represents the answer, but like, it's like, oh, we get this question a lot. Well, let's ask the full knowledge base. Even though that's not good for conversational speed, we can still take
that and give it to the fast retrieval system that we use for conversational speed. So that's my, I mean, this is just like my, this is just like my me hacking through the, this is vibe coding. Like this is me vibe coding. And so like, I would love if people have better answers to this, like temporal change across large data sets. Like I would, I would love to, um, hear about that maybe like in the comments or
whatever. Yeah. If you'd love to learn that. Yeah. That's a great question. That's a great question. Yeah. Thank you. And so is readers character in public is different from like his character in private office. And it's not, you know, public, you know, if, you know, when, when they, even though making like, you know, so with public data, but so if you fine tune with internally, so maybe the character will be different.
So it's not a problem. Okay. So how do you manage the first version that I built was off of public data and the custom duty. Right. I had his books and then his podcasts are available, but it gets a lot better. This is why I think like, right. It's like, who's going to clone you. The person that can clone you best is going to be you because you have the primary sources, the best quality primary sources. And also like sometimes that information isn't even out there.
And then you just, you just like, Oh, like, like you asked me 10 questions and I'm like, Oh, here are my answers. Well, now we have a new piece of data that doesn't exist on the internet. We put it into this character. Now, all of a sudden that is actually a unique, like a unique proposition of this character is that it's memory and knowledge is, is actually like based on some private data or like non-public data, which might be.
more important in some ways than everything that's publicly out there. Like, and I think it's actually in my voice clone, like my agent that I'm working on, I have my preferences for technology stack are in there because I had it look at a bunch of projects I've been working on for the last two years. I was like, just like read all my code and like think about these technologies and put it in the knowledge graph.
Like what this is the kind of stuff that I like to play with. And yes, it's biased. It's my bias. That's actually the point, right? Like it's my perspective is biased and I want that to be in this thing. And while like chat2BT will give you one answer, this thing is going to give you a totally different answer because it is grounded in your perspective. And I think that like, yeah, that's why I think like you should make your own.
No one else is going to replace you. You're going to replace yourself if you want to. And actually even if you tried to do that, you would realize that you are much more than what you thought. Like you have this like expanding sense of self when you try to replace yourself. You're like, oh, I'm not just a data analyst. Actually, I have so many other things about me that I'm interested in that I like to think about as part of my identity.
So it's interesting. It's like kind of an expansive sense of self that happens when you try to do these things. Yeah, and I'm a little bit afraid of the future where people are asking, hey, AI agent, you know, plan my life, successful life, and leave it for me or something like that. It won't happen. I mean, well, you can choose to, why would you give the fun things away? Like you could give the things that you don't want to do away.
I think that's fine. But sometimes I'm like, man, if this thing could just do my job, but it just doesn't, like we're not there yet. And I'm kind of just like, well, I'm going to have a job for a while. But also I think it's like, I like to automate the things I don't want to do. And then the things I do want to do, or if I have a very strong opinion, it's like, oh, this is a quality. This is my opinion, what high quality looks like.
And I don't think that these systems yet have that quality bar. They just don't, they're not good at like, we have that quality bar because we are in the real world. So I listen to a lot of music. I go to a lot of concerts. So I have my opinion on what high quality music is. And that's my opinion. And I would not give that to an AI. Like I don't think an AI can do that because it's like, can an AI get my opinions on this better than everyone else?
Like, no, maybe it'll get like, maybe, I like to see it. But it's also like, why would I give that up? Like, it's like my taste in food, right? Like this is not something that I'm so eager to give away. Like I think I'm going to keep all the most fun things for myself. Yeah, for sure. Yeah, definitely. And now you have worked with Reid Hoffman closely. And I'm curious, did your impression change? I mean, before you meet Reid and work with Reid and you had an impression,
I think, on him, then now you have worked with him. Did your impression on him change? And also, what's the biggest lesson you learned from working with Reid? If you could share. Yeah. Yeah, I read his books growing up. And I listened to Masters of Scale, the podcast, when I was in college. No, when I was in, yeah, when I was interning at my first startup. And for me, it was like, gave me the confidence to go into startup.
Because he's very much like, he loves games, board games, and strategy kind of games. And I also, you know, growing up, love strategy games. So when he was like, oh, the theory of the game, like how do you, he was kind of like using analogies to games as a way to make startup uncertainty easier to navigate, right? So, and I think for me, it's like, I love games. And I think one time, one way to kind of like, games are fun because you get to be competitive,
but it's not going to, it's not like, it's not like life or death, right? So we get to explore our personalities, how we work well with each other, how we compete with each other, what our strengths and weaknesses are through games. And then it also just gives us, sometimes it's like, let's say I introduce us to a new board game. All of us are learning the new game at the same time. So some people learn more quickly.
Sometimes you think, oh, this reminds me of that other game. This reminds me of that mechanic. And so you kind of are trying to pattern match to learn how to get good at a new game. That actually is very relevant in startups. Because I think in startup, it's very much, it's all new games. Everything is a new game. And, you know, you might be the first person or first company or first team to be attacking a problem.
And you're kind of like, you have to form a theory of the game that you're in. And that's a common read phrase is, like, what's their theory of the game? Like, how are they thinking about this competition? How are they thinking about this ecosystem of problems? And so, I mean, it was pretty incredible meeting him in person. I was like, oh, wow, he's, that's him. Like, that is, this is how he kind of is. Like, the most, like, it's very much the same guy that you see.
And I was like, okay, that's crazy. The biggest thing I've learned, I think, is, like, this experimental, like, mindset of, like, we should just be figuring things out. Like, he loves to experiment. This read AI, right? Like, I remember after it went viral, I was meeting, I was talking to him. I was like, man, I saw read AI on CNBC yesterday. What is happening? And he's like, hey, you know, when you're, like, eventually, maybe everyone will have something like this.
But when you're the first to do something, you kind of get a, just, you get a lot of attention, right? Because now it's, same thing, Yohei Nakajima, BabyAGI, perfect example of this. Everyone's building LLM rappers now. But BabyAGI was, like, it got all of this attention and momentum because he was, he put it out there first, even though it was not complete. Like, it wasn't, like, it didn't work end to end.
I remember I was using it. I was like, wait, it can't do everything. I was like, obviously, it can't do everything. But that's crazy. Like, people love it, even though it's not perfect yet, because of what it inspires them is possible. And so that was a huge lesson was, like, you know, you, you know, if you're first to something, like, there are advantages to being first. Similar, like, this philosophy of blitzscaling, right?
If you think there's winner-take-all market, like, you have a different way to play the game. You should play the game differently if you think there's a chance you could win it all. So that these kind of, like, these kinds of, like, philosophies are, and that's the other piece that I think I totally underestimated my whole career. Reid is a philosopher. And I'm more of just, like, a hacker kind of programmer, like, you know, tech guy.
But I realize now the value of philosophy much more deeply. And I'm at the very beginning of this, but he's always quoting philosophers, Wittgenstein, just like people. I'm like, who are these people? Like, they died 200 years ago. Why are we talking about them? But actually, like, it's very relevant because things like language and the importance of language and language and cognition, how you think, how you speak, how you write, all of this is actually very important now.
And a lot of the people who were thinking about this, the great thinkers of the past, kind of were running into these ideas before they were relevant to technology, especially now that we have programs that can emulate thought, right? And then if I think about philosophy, it's like, how well do you understand people? Because technology is not just, like, it's not just we don't just build machines for the sake of building machines.
We build machines to solve problems that people have. So you have to actually understand people and, like, have deep empathy, try to understand, like, what do people actually care about? What do they want? What are their desires? And that gives you a better lens with which to kind of build technology, I think. Yeah, definitely. Making something people want. And that's what YC always say. That's right. Yeah.
But you said, you know, like, for us to be the market is important. But do you have any idea of something that's unexplored yet but could have a huge potential? I mean, it could be AI. I think there's so many. There's so many now, right? Like, I feel like I have these ideas every day, honestly. The question is what do you actually spend your time on? But I think if you look at, like, just look at any new capability, like, let's take language model.
Like, this is my bread and butter is language models. Like, okay, now we have language models. Okay, what, like, make a list of all the things that you could not do two years, like, three years ago that language models let you do. And then it's like, well, three-plus years ago, anyone that had messy data could not do data analysis. But one of the most powerful capabilities of language models is the structured outputs.
The ability to take unstructured data and turn that into structured data, right? Because when you take unstructured data and turn it into structured data, you're able to connect it to traditional software, which traditional software expects structure. Language models, they seem like this, like, random kind of black box. But actually, you take, like, the, like, if you take my bio and then you say, extract every single company he's worked at.
And then, like, extract the title, like, his title as he self-defines it. Now you can put that into a CRM, right? But that might be unstructured at first. It might be the output of a deep research report. But then the LLM allows you to extract structure from that. I think this is one of the most powerful capabilities of language models. And I think people, I think a lot of people are just ignoring it. And that was one of the first things that actually I was working with before Agents was structured outputs.
Just trying to get the model to do JSON. Like, that I spent a month working with Microsoft Research. I was like, Microsoft Guidance, I like it. It allows me to go from a blob of text into fixed JSON. I think people were just like, whatever. I'm like, no, this is actually probably one of the most important things of all time. But you have to be a little in the weeds to realize it. But the implications, what's the second-order effect here?
Well, one, it means that everyone that complains about not having clean data can go use LLMs to create a clean data set that represents the business logic. That's awesome. So now you can do data analysis, even though you have messy data. And then the second thing is, people do this with transcripts, right? The meeting transcript, then they feed it into the LLM, and then it's like, well, these are the action items.
That's powerful, right? Now we can just quickly move to the next thing. Same thing, when you ask a question to the model, it uses structured outputs to write the search query for the search engine to process. So it's happening under the hood, how the model connects to the rest of the software. And then I think the one thing that this does that is still completely ignored by most people is generative UI, like being able to, and people talk about it
as like, oh, we don't want to, like, you shouldn't. So people complain about vibe coding, it's like they're like, at its limit, generative UI is unfamiliar, and then people won't build familiarity and patterns. But I think there's like a middle ground here, which is like, what if you're talking to read AI, and it's like, oh, I had this conversation last week, like read had this conversation last week with Satya Nadella,
and then the video comes up. That's like an application of generative UI, right? So within the constraints of the structured output, the UI that we give it, it is able to fill it in with the generative piece. So I have a feeling that this is like largely untapped, very important to me, like user interfaces that are more just in time, like the right user interface at the moment that you need is actually very interesting to me, and like,
kind of not really explored. I really like GraphRag, I love the advanced retrieval stuff, but, and I think the applications are more, like, I think the applications of some of those systems is further reaching than people believe. But I think you have other people that are just so scale-pilled that are just like, we don't need this, because eventually GPT-7, infinite token context, whatever. And I'm like, yeah, OK, fine.
But like, until we have that, like, this is kind of cool. Like, this allows me to create new artifacts in the meantime. And so I like, yeah, I think, because memory will probably get, I mean, memory will get solved, but it's not solved. Knowledge, structured knowledge is valuable, in my opinion, because there's not a Wikipedia page for everything on the planet. Like, there's plenty of things that don't have a Wikipedia,
I don't have a Wikipedia page. Like, so, but like, LLMs could make that. Now, it wouldn't be called Wikipedia, because now it's not human review. It's not like humans wrote it, humans reviewed it. It's a different thing. But that synthetic thing might actually be very useful. So I think there's a lot of artifacts that these models can create that will be valuable to us in ways that we can't expect. Let's see.
Yeah, I think generative games. Generative games. That's an area that I'm very excited, and it hasn't yet happened. It's like a Genie 3. I think there's two versions of this that I'm very excited about. So one is the Genie 3 kind of etched, they did this with Minecraft 2, which is like the model renders the whole game on the fly, and then you interact with it. But it's like diffusion model, just like the game is just
totally dreamed on the fly. I think that's exciting. I also am more interested in like, that's cool, but I think there's generative games more like where, like closer to Dungeons and Dragons. I don't know if you've ever played, if you're familiar with Dungeons and Dragons. But it's like a game where you speak, you're kind of like talking, you're pretending to be a character, and you talk, and then someone is conducting the game,
and then the game unfolds around you in conversation. But it's really more like improv. It's more like a conversation with friends. You're pretending to be characters. I think there's something there, because that even can be automated using LLMs. And so if that experience can be made easier, then it can be made more fun. And also it can be made more accessible. So it doesn't have to just be fantasy. It doesn't have to be just dragons and witches
and it can be anything. And so there's, and then I think that applies to a lot of genres of games. So it's like, so maybe the diffusion games are like where it ends up, but there are other things that we can do until then that like, like if you think about Pokemon and Yu-Gi-Oh, what does the generative version of Yu-Gi-Oh look like? I think about Yu-Gi-Oh a lot because it's like in the TV show, I don't know if you guys remember Yu-Gi-Oh,
but like in the TV show, everyone has their own deck and like Kaiba has Blue Eyes White Dragon and then Yugi has Dark Magician. And it's like, this is my signature card. And I'm like, man, what if we had a card game where everyone had their own signature deck? And it was just, you had your deck, I have my deck. And maybe this is what NFTs was supposed to be, but like now it's possible, right? Like generative AI should make that possible.
So I'm kind of like, this feels obvious. Is it big idea? Who knows? So I'm kind of like, well, let's just go hack some version of it. And then like, if people are having fun, maybe we see where it goes. But yeah. Pretty interesting. Yeah. Yeah. So when you mentioned about Wikipedia, I came up with a random question, but do you think in the future, do we need Wikipedia? Because we have access to it and I was just curious about that.
I think we're gonna need, we're gonna want some things where it's like humans. I think there's gonna be humans and AIs. There's gonna be a version of Wikipedia that's probably way bigger, which is made by humans with AIs. And I think that the feel, I'm a little bit, like I think we need something because what's gonna happen is, it's already happening, which is like, there's gonna be a lot of synthetic stuff out there.
And then it's like, how do you tell what's real? What do we believe that's like fact? And this is a hard problem. And some people think that crypto is part of the solution. Maybe, I don't know. But if the internet ends up filled with a bunch of fake information, fake data, we're going to have like a lot of side effects. I think one of the side effects is that if social media platforms keep getting filled with AI,
we will have our group chats, which are purely human maybe. Like we're gonna have places we go where we only wanna be with people and maybe in the real world, right? Because that's where it's easier to be. It's obviously like, yeah, you're real, I'm real. But like, if it's Twitter, it's too easy for Twitter to be filled with bots. It's too easy for that to dilute the experience. And then people lose trust in their information.
And if the trust information goes down, that's gonna be really like, it's one of the biggest problems that is not solved. And then I think like Wikipedia, is Wikipedia this answer? I don't know. Is it like a version of Wikipedia that's built with AI, but like there's some kind of like verification system. Biology talks about crypto as being a possible solution here. I'm like, hey, we need to try these things
and like figure it out because the problem is only growing. But I think like every problem that AI creates also will create billion dollar solutions. Like AI is part of the solution in many of these problems. It's a problem of, we need a system with the right incentive structure to preserve the values we care about. So I think of it as like, okay, if that's the chaotic world we are headed towards, how would we get a network of people
where we actually agree on the values and we orient ourselves towards like playing a certain way so that we get better information streams? Maybe that's group chats. Maybe that's like a return to like the physical world. So totally like, I don't think I have the answer, but I'm optimistic that we will figure it out. I think technology is a huge part of the solution in most of these cases as well. Same thing with cybersecurity, right?
If vibe coded apps are vulnerable, then also AI is gonna be a part of the solution. Like AI hacking your website and then telling you before, instead of telling like a bad actor, that's gonna be, I would pay for AI to hack my stuff and tell me and not tell anyone. You know? Yeah. Definitely, yes. So yeah, that reminded me of the early days of Wikipedia. So when I was a college student, high school student, I still remember teachers always say,
oh, you know, the trust in information on Wikipedia was way lower than I see in current situation today, as we have today. So, but now people say, oh, AI hallucinates. We can't trust AI, but people hallucinate. But I think, yeah, that's, I see the interesting, how to say, psychological change over time. Yeah. The information we should trust in, yeah. Sorry. I feel the same way. I feel the same way. And then I think, man, there's so many topics I care about,
and then there's no Wikipedia page. And I'm like, okay, I would rather there be an AI Wikipedia page, at least. And then the people who care, we can go be like, whoa, whoa, whoa, whoa, whoa. Like maybe it's human review. Maybe it's definitely gonna be AI part of the review process but like, I think it's better than there not being a page, but. maybe Wikipedia maybe won't do this. And like, maybe like you and I,
like we might just make these like synthetic wikis. I do this a lot of times with GraphRAG actually. I just make a bunch of wikis for my own favorite topics that do not have pages. And I'm like, okay, like, it's fine. I'll make my own. Me and my LLMs will make my own for the games that I play, for example, right? Like they may not have a wiki for that game. And then I'm like, oh, we'll just, I'll read the docs
and then we'll have GraphRAG go and like index everything. At least now I have a wiki. Is it accurate? Well, the alternative is it doesn't exist. And I can see that the primary sources. So like the same thing in Wikipedia, you can cite Wikipedia, but like you better look at the source that it's linking to. Because even that, like people who write, they put stuff on there. It's not even like, they're making the source
and then they're putting it on there. And this is like, so there's abuse vectors even in human Wikipedia. So yeah, it's gonna be an interesting one, but I think we're gonna have some synthetic version of this. Yeah. Yeah. And I remember someone said, if we showed a citation, the answer and people without checking it, most people trust it. But when we check, that information is generated by AI or someone, it's not correct.
And they cite another information then citing, citing, citing the wrong information over time. So it's kind of- It's garbage in, garbage out, and then multiple layers of garbage. And then you're like, what are we doing? That's a very- But if you have a citation in answer, we- Assume it lends credibility. Yeah. Yeah. Yeah. That's an issue. It is an issue. It is an issue. So maybe reasoning helps with this. Maybe reasoning helps with this,
but it depends on how many layers deep does it go, right? Like if it's all AI generated, citing AI generated, then what is real? This is a huge question. What is real? And then we also have this problem in video. The video models are so good that you can't- It's not a- Two years ago, it was like, oh, six fingers. It's clearly AI. Now it's like, the bunnies are hopping on a trampoline. And I'm like, it looks real to me.
And then a week later, Google was like, that was Google Gemini. And I'm like, my thing is like on video, why don't we just generate things that are obviously not real first instead of recreating reality, which is impressive, but like deceptive. Or like, don't do it for deceptive reasons. Like make some other- You can generate any universe. The video model is like an advanced simulation. You can generate a candy universe.
No one's gonna look at a candy universe and be like, that is real. And actually, maybe that's my preferred way to go is like, oh, let's do animated. Let's try some other style where like, I'm not trying to deceive you here with like something that is pretending to be real. Of course, I worked on read AI, but he says he's at read AI, right? Like this is part of the design choices. Like, even if it would deceive you,
if you like, it's like, it's gonna trick your grandma two years ago. Now it's gonna trick you because there's no skill in identifying whether it's AI. You're kind of just like, you scroll to the comments and you're like, hopefully this is real. And then it's like, this is AI, this is AI. And then you, it's like, how do you trust it? Even if the comment is written by an AI, right? So it's a, these are the, I knew we would get here,
but I didn't think we would get here so quickly. It's for sure. Yeah, it's a crazy days we live in. Crazy days. And I was thinking about this idea, I mean, to train humans to distinguish if it's AI or not. So let's say in early days, we are kind of easier to get tricked by spams. You know, like, oh, you were chosen, you have the right to get $1 million, you know, from, I don't know, like King or someone. Yeah, yeah, yeah, yeah.
I'm sorry, this is just totally. The email, the email you get, the email spam. Yeah, you get. But now you right away distinguish, oh, this is fun. But early days, some people got tricked, right? So, but- No, but now the spam can be generated by the LLM and you can't tell. Yeah, yeah, that's right. But if we have a platform that teach people, oh, hey, show, let's say two videos or two pictures, two texts or something, then if they say,
oh, this is spam or this is AI generated or this is human created, then we can train, oh, actually this is generated by AI. So then, so that we can get used to the things that generated by AI or, I don't know. I don't think this is gonna be possible. And it's just like, I don't think it's possible. I mean, look at VO3, right? I can generate something in VO3 and like, I mean, I did this with a runway gen three runway
and I made a flux, I created a Laura of myself. So I had an image model flux one and I fine tuned a Laura on my face. And then I made a music video. This isn't, I did not put it publicly on my YouTube channel. It's a private video. Maybe I'll put it public after this now. Cause it's like, now it's like, obviously, I'll just say it's totally AI, but I didn't put it up because I was like, oh my God. Cause I showed it to my brother and I was like,
it's me in 15 different universes in the 1920s, me and going to space. And my brother is like, the only reason I can tell that this isn't you is because I know you've never been to the moon. And I was like, but otherwise it looks like I'm on the moon. And I'm like, okay, there's no skill in discerning this. For the person that doesn't know me, it's like, oh, you must've been in a movie about the moon. And I'm like, it's, I mean, this is a movie that I made.
It's like a little two minute clip of me in different universes. But I didn't make it public. Cause I was like, actually, like, I like, I'm not sure like how I want to deal with this yet. Like, because it looks so real. And my brother was like, no one can tell. Like, it looks like you. I grew up with you. It looks like you. So everyone else for all intents and purposes will think it is you unless you tell them it's not you.
And so I think that, and then it's like, that's a lot of work, like training. And also the models are just getting so good. The models are already at a point where I don't think you could train me to tell the difference. If someone wanted to make something that you could not tell was real, they're going to be able to do it. So then we need systems of like, the word is providence, systems of providence, like systems of like this, we need new systems
where we can ascertain like collectively, like this is a network of content that is real. And then actually everything else, you have to assume it's fake. And unless it can prove that it's real somehow because of the new system we create, which doesn't exist yet, but people are trying to make it, you will just deny everything. Like, you're just going to have to, like, it's like, oh, don't believe everything
you read on the internet. Don't believe everything that you see is real on the internet. Isn't it sad? Isn't it kind of dystopia? It is. It is. As long as we don't solve it, it's going to feel bleak. I see, yeah. Yeah, yeah. But on the flip side, whoever solves it is going to become a billionaire. So yeah. So there's an opportunity, right? Yeah, like, yeah, maybe, or maybe they'll make it like, they'll do it just for the good.
I don't know. We'll see, but people are working on it. And I think it's probably one of the most important problems that AI creates. AI creates some problems. It is going to solve a lot of problems. It's going to create these problems, and it's going to be a part of the solution. And then you have, even crypto people are like, oh, we should make it so that if it's real, it must be verifiably on this blockchain.
And I'm like, hey man, if you can figure out how to get everyone to agree to that system, maybe, you know? Interesting. Yeah. And so what's next at the Office of Reid Hoffman? Can you share? Oh, I don't know if we can share, but do you have any future AI project coming out? Well, so I spent the last week, last Wednesday, I spent the whole day talking to 20B GPT-oss just like my personal projects, right? Like I spent the whole day talking to GPT-oss.
And I was like, okay, wow. Like this model is very efficient in terms of like, it runs on a Mac mini at 30 tokens per second, and it's completely local. And I'm like, wow, this is crazy. And so now I have a new laptop coming and I'm going to be playing with the 120 billion parameter variant. So I'm going to be working on some local agents just to understand, like, I think I've kind of been mostly in the, you know,
using the API and like using these frontier models, which you're kind of renting your intelligence. But, you know, a lot of other people, of course, have been in open source this whole time building local models. But I think the local models will be agentic. And you're going to have these, just kind of like local running agent, smaller in scope, but like useful for what it needs to be. My dad always tells me, he's like,
Parth, the fridge does not need super intelligence. So he always says that. He's like, there is a version of this that doesn't need to be like trained on the entire internet for it to be useful. And because he comes from like the last era of computers, right? So he thinks about like the cloud. He thinks about like on-prem. He thinks about a lot of this stuff. And he's like, you get a lot of different sizes of these things and you get the biggest ones.
Yeah, it's going to be on the internet. The biggest, smartest models are going to be the ones you rent. But also like, you're going to have useful small models. I think about my Roomba and it's like, this Roomba is so behind. It's not even GPT-1, right? Like it gets stuck in the corner. And then I'm like, oh my God, my cats are like, this thing is not small. And I'm like, this is, but imagine a world where like the devices are actually
intelligent because they have intelligent, we're locally running small models. That's going to be interesting. Um, in terms of like, so that was like an exploration from last week. And then last couple of days I've spent most of most of my time working with GPT-5, and that has made a lot of my, uh, a lot of our projects internally, a lot faster, like there are things that you can make now in three hours that used to
take me weeks. And, um, I think the combination of Claude Code, GPT-5, and honestly me just getting better at using them. Like, I think a huge thing is like adapting yourself, um, switching the way you work to being more agent friendly, how you orchestrate these tools. Um, you know, you start by using one when you get good at using one, can you use multiple? And then when you use multiple, can they work together on things that are even more
advanced? I think we're getting there. It's very human in loop. It's not like end to end automation. It's getting better at long running tasks, better at end to end automation on for, you know, better at zero shot. But the, the speed at which you can build tools is accelerating. And I think we have some tools that, um, that I'm working on that are made possible by GPT-5. Like yesterday I was like, wow, GPT-5 built this like pipeline in, in, in two
hours that I was going to spend a lot of time on. But now I'm like, wait, we can go much bigger on this in a shorter timeframe. And, uh, I'll have more to share. Well, you know, we're, you know, we always try to share how we do things. Like we, we try to create content. Oh, here's how we made read AI. So I think we're going to have some, some more stuff like that, that people can kind of like get the playbook, but really it comes down to just like play
with the tools and then like use them and get really good at using them. And then, and then, um, share what you're learning with people. And then that kind of is like a good feedback cycle. Yeah, definitely. And before, you know, started, we started this recording. So we talked about email capture. That's the idea we hacked during the last week we can, and actually I used GPT-5 and it worked in a single try.
But when they used a Claude, I've got Opus 4.1 or it didn't work in a single try. So I was surprised by the capability of GPT-5 and this is how I'm feeling right now, man. Like, I'm like, Oh my God, because I call it like one shot. Like, Oh, did it one shot the problem? I'm like, Oh my God. Because if it can one shot the problem, then you're like, we got to think bigger. Like we got to do way more stuff. We got to try more things because like, turns out like, this is not going to take
a week. It's the first version might only take two or three prompts. And if that's the case, then like, we need to be more ambitious. We need to be more creative because the models are really getting to this level of like, uh, the code generation has never been this good. And now we have the CLI tools that are getting better. The wrappers around them that are making them even faster and more reliable. The tools MCP is great.
MCP is also a little bit, you got to be careful, but it is very good. And then, uh, and I agree. I think, what can the model one shot? I think I'm like, I look forward for me. Um, my AGI test when it's like, is it AGI? My test is I'm going to just go to the model and say clone clubhouse. And then that's it. Can it build clubhouse in one shot? Take your time. Maybe take two hours. I'm gonna go for a walk and I come back and it better be an
application that I can play with. But that mindset of experimentation is important because I think like, you think about big companies and they sit around and it's like, we should make this. It's when they write some document on like what the future is going to be. And then like six months of like planning and like building. And I'm like, no, no, no, no, no, no, no. We're going to sit here. Repl.it is going to be here and we're going to tell it to make this thing.
And then we're going to see how far we get in this meeting. And then that updates our own, uh, priors. We're like, okay, cool. Turns out we can get this far in two prompts in three prompts. It's like, really? Like there should be a hackathon where it's like, you get 10 prompts and then whatever you submit at the end of that, that's the submission, like how far do you go in 10 prompts? Like, how far do you go in one prompt?
Like that should be a hackathon category. Right. Because then you're going to, you start thinking about. Um, uh, Dexter, uh, Dexter Horthy, he said he's like context engineering, right? Like the new phrase, which is true. It's like, are you bringing the model? Are you bringing the right context into the context window? And then you can be much more efficient with what you get on that first try. And I think it's, it's awesome to hear that you're doing that.
And that, that is a great story of GPT 5 one shot, because I think there's so many more of these. Like, we don't even know what it can do that quickly unless we try it. So that's why I say like, don't judge a model until you've talked to it. Like, unless until you prompted it a hundred times, then maybe you can judge the model. But there are so many things we don't know yet that we will figure out just by playing with this and talking to each other and sharing what's working.
Yeah. Yeah. I'm looking forward to joining prompt song. Because, because at the hackathon, I know some people bring their startup project, you know, or pre-built project, you know, and they win the prize. And it's kind of unfair for some people who really built, you know, that, that project during the hackathon. Right. Yeah. But from some, you can't cheat it, right. You need to bring your prompt. You got one box, one prompt, one attempt, no thinking allowed, no planning.
Yeah. Yeah. That's fair. So it's fun. Yeah. Yeah. It's fun, but it's actually an important exercise. Even if you don't ship the application, it's an important exercise because when someone tells me, oh, it's going to take six months, I'm like, what? Why is it going to take six months? Like, is it because you need to meet actual people? Like, is there, what is the bottleneck? If it's actually just software, we need to move way faster.
If it's like hardware, if it's in the real world, if it's like brick and mortar stores, I get it. There's more work to do relationships, human relationships, but if it's just like pure software, we should be going very fast and we should be learning very quickly. Yeah. Yeah. And this is just a general question, but so where's the source of information and ideas? So what kind of source, like on like X linked in YouTube, what do you see?
How do you collect the ideas? Yeah. So I have like nine, no, I have 10 messaging apps. And so for me, it's like 10 messaging apps. So like people are always DMing me. They're like, oh, did you see this? Did you see that? I don't try everything, but I'm like, oh, interesting. But you have to have like your own network. You have to have a lot of people that are doing different things and in AI or in technology or just creative, right?
People who are creative and different, and then you have to bounce ideas off of them. I think that the network intelligence is the most important thing. Twitter is good. I think that there is the problem. There are problems with Twitter. I think it can be kind of a time sink. And also I think people tend to, they tend to pay attention to some of the wrong things. Like they go deep in some of the wrong directions in my opinion.
But it's, it's good because you get a good feed. But I think that like I have a, you know, I have like my own, all my favorite hackers, I put them all into a discord and then we have our own like, oh, here's what I'm making. Here's what I'm trying. Like, oh, and then it's like, oh, like someone might be hiring. And it's like, oh, looking for an engine. So there's opportunities. And I think once you start networking the people who are, you know, really
playing with the technology, then the network intelligence benefits the whole group and that's important. And it goes back to like, what is real? What is false? Like you got to create these like spaces where you have higher trust and you curate people so that it's like, oh, these are good people. They are helpful. They're generous with their information, with their, with their advice and time. And then they're also learning and by making things.
Right. So I think for me personally, it's like, I don't want to talk to people that don't make things. That's just like, I'm kind of getting biased like this now. That's maybe it's a bad thing, but like, I prefer spending time talking to builders. Like, it's like, oh, you made this yesterday. I was like, that's the kind of person I want to talk to because like, that's the experimentation we need. And also there's just too much, there's too much happening.
Right. So no one person can keep up with all of this progress in AI alone. I think it's singular. For me, the definition of singularity is when information starts, like technology starts advancing so quickly that you cannot keep up unless you're using AI and friends to keep up. Right. Like you have people that are covering different topic areas. So I have friends that are just like really good at working with open source
models and I just like learn from them. I just ask them dumb questions and you need to ask dumb questions, like non-judgmentally just ask questions. No one is an expert in everything, but everyone can get very deep into a few things and then when you share that, then the whole group benefits and then you get the interesting intersections. Right. So I'll be like, oh, the voice, you know, the ElevenLabs V3 is actually
very good for emotional control. And then my friend who's more of like a creative artist designer, and he's like, we should do like a Dungeons and Dragons kind of experiment where we have LLMs and then I'm like, oh, we can use the open AI agents at CK and he's like, and we can use these image models. Right. So the combination of two people that are actually very different because they're focused on different things, but if they agree that there's a cool,
interesting idea, that is actually the magic. The magic is the intersection of your intelligence and my intelligence. So I don't think, and I think also, yeah, I mean, I do think I generate a lot of ideas, but the consuming a lot of like having friends that are different is for me the most important thing because then I'm like, oh, what's your favorite LLM? Well, I'm like today, like my friends, like part that you're the only person I
know that likes GPT-5 and I'm like, wow, then I have to like, explain why I like GPT-5 because otherwise he's going to assume that like, maybe it's completely useless based on everyone else. He's talking to, maybe I'm wrong. Maybe I'm like, maybe I'm in the, like, maybe I am wrong, but, but, um, that's the kind of thing, right? It's like, if I don't ask him why it's like, well, what model are you using? Okay.
When, why, what's your use case? Okay. Interesting. Well, that makes sense then. Like why it's better for your use case. Worse for mine, but unless you're talking to people and I mean, like talking to people, like not Twitter, Twitter is not social. Twitter is like people kind of just blast stuff out there, but a dialogue is actually very powerful. So I also use Clubhouse still. Um, not a lot of people are on there, but for me, a lot of my friends are there
and we just have conversations once a week. We're like, oh, this is very interesting. This is very useful. This is how I'm using it. And you know, for three hours, a conversation like that ends up being very helpful to stay up to date with things and also to just get more ideas in the mix and you have to have a lot of different perspectives. Like I go to meet my friends in different cities because I want to,
I want to think differently. I live in LA instead of San Francisco because I like all the creative people here and they're just totally, they're not in technology, but for them, they're like, they're so interesting and creative. So when I asked them, oh, like what, like learn more about what they're building, I get so many, so much inspiration. So you have to have people that are very different from you.
I think this is network intelligence is key. AI is going to help, but like network intelligence, curate your network, choose interesting, you know, hardworking, creative, generous people. Yeah. Yeah. I like the word network intelligence and that's a great word. And also that's also a real concept. Yeah. Now it's more important than ever because, you know, you have AI, I have AI. So like both of us are amplified, but now when we come together,
we're even more amplified. Yeah. We can come up with new ideas and then see things from different perspectives. Otherwise, you know, you can, it feels lonely and you can feel like you're going a little crazy, but you, that's why you need, you need people and conversations. Yeah. Yeah. And that's why I love going to hackathons because I can meet a lot of great, interesting, creative developers trying or tackling some interesting
ideas from different approaches. And yeah. Yeah. Hackathons are, I recommend that for a lot of people, like, how do you get started? I was like, go to a hackathon. And then you're going to be like, you're going to realize, like, we can go fast. You can make things for fun. You're going to learn quickly. And then people are using the state of the art tools and you're like, wow. And, and there's no experts, right?
Like everyone's just learning. So it's good to go to that environment where it's okay to be like, it's good. I went to a hackathon like two years ago. My mindset is like, now I just, I never left the hackathon. I'm still in the hackathon mode every single day now. Right. So it's a very, like, I think it's a good exploratory set of like way of thinking and making, you know. Yeah. And I have a designer friend and he, I used to be asking him, hey, let's go to,
you know, hackathon together, but he hesitated because he's a designer. He thought he couldn't contribute to coding or something, but nowadays he can use AI tools to code. So now he asks me, Hey, next time you go to a hackathon, please ask me. So I'm a designer, but I can use AI tools now. So I can code and build something. Yeah, yeah, yeah. We, I did the same thing. I brought my designers to hackathons. In LA we have hackathons, but our hackathons are like video
model, creative hackathons. There's like, you're going to go, you get paired with three random people. And then it's like, make a music video in three hours, whatever tool you want. And so for us, the hackathons here are more like in the media side of things. Okay. What kind of story can we tell? This tool lets you do like, you know, style transfer this tool, like VO three plus like mid journey. Plus, so you, you, people have, they all have their own favorite tools.
And then you come together and you're like, wow, we can now make this new thing. So here are, we have technical hackathons too, but I think our creative create, even creative hackathons, very useful, right. Storytelling and just like learning how to use these tools. Cause there's no like textbook, there's no experts. So the people who are the experts are those people. Right. Yeah, definitely. Yes. Yeah. We are living in the kind of golden age.
Yeah. Yeah. So yeah. Also, yeah. So when you make, you know, do you want AI clone or AI, you know, so you need to have like your data set, right. But yeah, according to our conversation, you talk to people on Clubhouse, which doesn't have transcript, or you talk to people on Discord, which usually, usually, you know, data will be deleted or removed eventually. So how do you keep, you know, your ideas? Also, do you have any tips to, you know, save your real-time
data for your AI agent? I don't do that level of collecting data. I think, and I think it's not going to be perfect copy because of that, but I think the goal is a little bit more like, I don't need it to be a perfect copy of me, I'm, I'm me, I need it to be like, you know, just like a, like a, like a different, like a, like a different kind of, like, I don't care that it's not exactly like me, I care that it's helpful to me.
And I think there are actually, it's good to have places where things are not recorded because then people feel free to speak. Not everyone, like we were recording this conversation, but I spend so much time just being authentically myself that I don't, I'm like, I'm pretty much the same, you know, recorded or not, but not everyone's like that. Sometimes you put the recording button on and now people don't want to speak.
And I think that's like, do you want people to speak more? It's more important that we speak than for it to be recorded. And so I'm not a fan of like always on recording. I do use granola sometimes, like, but that's like, you know, double opt-in on the meeting transcript. And that's more for work. It's like, okay, did I make sure I get everything? So that's different. But if we lived in a world where everything was recorded, I would not be happy.
Okay. So you don't like AIP? I remember Humane or Metaverse. I don't like it. I don't like it. The AIP, yeah, this concept, it's like, it's just like, whoa, you didn't get my permission. And so then now I'm like, well, and even the Meta glasses, sometimes I feel, you know, like people just, I see the light, but then now I'm like, okay, now I'm on camera, right? So this is, this, this makes you, it changes the way you act for some people.
I think so, but that's going to be an interesting social norm. We have to figure out. And I think for me, this is, it's more like, I like talking to this system and be like, here's, Hey, here's what I'm thinking about. And then ask me a question. Then I talk to it. And then when I'm talking to it, it is collecting data. I don't need it to follow me around forever in the real world. Because I think that like that some people want that, that's fine.
I'm not going to judge them for it. I think for me, it's more like, that's not, I don't want that relationship with technology yet. And I understand it might be, there may be very huge benefits that I'm ignoring. It's just that I think the social trade-off is serious. Um, it's pretty serious. Like if someone is, has the meta Ray ban on some of my friends have them when they engage it now, I'm like, okay, guys, we're on a show.
Like, it's like, we're no, now we're acting right. And it's just like, I prefer to be like, like, you know, chill, you know? So. Yeah. I was going to ask something, but I forgot that we can, yeah, we can cut out. Yeah. Yeah. Maybe you already answered it, but so you use GPT-5 or so you use Claude Code. And yeah, the code of this context. So a different, so if you switch, so you need to have a context, right?
Continue. I'm so glad you asked this question. Oh, yeah. Thank you. I was curious in the CP server or something, you know, context engineering, how do you orchestrate? So, um, so the, so the first month I use Claude Code, I was just learning it and I was like, oh my God, and I use it every single day. And I was like, I burned 560 million tokens. I was like, how is this real? Like, this doesn't make any sense.
Like, you know, like I'm like $10,000 worth of tokens, but I'm not paying $10,000. So, uh, it feels like it's very discounted, but then I, and I was building like, okay, but you know, my laptop, my computer, my, this is my gaming computer, so I, this is like my main, my desk, but like some of my work is on my laptop, my work is on, then I have a Mac mini and I was like, okay, well, what if Mac mini had coding projects on it and then I could just
connect to that Mac mini from my phone, from my laptop, my other devices, put it on a private network. Now we have a single chat and then, but I can use that from any device. So that was one thing that I was doing, which is like, um, for like multiple clods, 24, seven on a, on a single device, just that thing never sleeps. I can, you know, connect to it from any other device. Very cool. Um, then, uh, you know, GPT-5 came out and Codex CLI is now good.
Well, it's like, I think Claude Code is better than Codex CLI. I think even the people at OpenAI would admit this, but, and Codex CLI will improve, they're going to work very hard on it. And, um, you know, they're going to vibe code it with GPT-5. They're going to tell GPT-5 to add features, but you can also fork it and you can modify it yourself. And I think that the, so now it's like you have multiple CLI agents and you
want them to work together. On the same project. So then I was, this is like last week, I've been thinking about this a lot. Like I want to use both of them. Um, and I want them to work on the same code base. How do I do that? And I think, so there's this thing called Claude squad. And if you just look up, it's like a GitHub repo. It's, it's a way to interact with multiple CLI tools. And then they basically use their own work trees.
So now I'm like, okay, well now I need to get, I need to change the way I work with these tools. Instead of doing, you know, normal GitHub branching, it's like, okay, work trees. Now it's like, can you work, can they work on different copies of the same code? And then when then triaging and integrating the solutions becomes possible. So Claude squad is the most recent kind of tool that I've been using because
it also allows me to go from codex to Claude Code in the same tool. And, uh, and you can fork it and then you can tell GPT-5 to modify it and build a UI on top of it. Right? So remember everything, everything can be modified with. If you have the code, you can modify it and you can make it even more personal. So I think we're at the very beginning of this. I think the CLI tools are a temporary thing. I think we will also want more traditional user interfaces than, uh, than CLI, but.
CLI is powerful and you get raw access to the machine. And then a lot of the tools have CLI, uh, support. So it's a good starting point. I have 17 MCPs, uh, that I use a lot. And, uh, one of them is this knowledge-based system, the knowledge graphs, but it doesn't, it's not just one knowledge graph. It's, it's, it's like an MCP that allows it to create new knowledge graph, query them, list them. Right? So if one of the tools is the ability to create and query.
Knowledge from long-term storage. So there is a certain, I think there is a set of tools that you want to give some of your agents, not all your agents need every tool. And that can be risky. Right. Um, but you know, access to your email, your calendar, it becomes so much more useful when it's like, Oh, look at both my calendars plan, my next three months of travel, and I was using Claude Code to do that same thing with like, like expense
reporting, I get so many AI subscriptions. Right. And they're all have like Stripe, Stripe, Stripe. So many like emails of like receipts. And then I was like behind on my expense reporting. And then I, and I got an email and they were like, parts, you got to file your expenses. And then I was like, Claude Code. We're late on my expense reporting. Go into my email and get every single receipt, then use GPT for vision to
categorize every single thing by vendor, get the, you know, the total of the amount used GPT for vision, look at the receipt, extract all the structured outputs, extract that information, put it in the name of the file, put them all into folders for every single month. And in 30 minutes, it does like 400 expense report. Um, you know, it finds all of them. There's no way I'm going to do this manually ever again.
But the realization here is that, Oh, wow. Coding agents are very powerful for non-programming tasks that require structured logic, right? Just general automation. So Claude Code does my expense report. Which a year ago I was like, I wish there was an agent that did expense reporting, but now I'm like, maybe that's too small. That's like a side effect of coding agents is that they can do this. So like, we need to think even more deeper about this problem because
actually a lot of that stuff is just. You know, GPT-5 CLI code, Claude Code might be able to do it. And, um, does that mean that there is no such thing as an expense reporting agent, that it'll just be a feature? Like who knows? Like maybe that's just a feature, not a product, but this is like an early realization we're having now of like the coding copilots are very useful in non-programming tasks, so then actually we're, I could see, and I think that
like, that's probably going to play out more, which is that like the agents that non-technical people use will actually be under the hood coding agents that are just like guardrailed heavily. And like, these are the flows that they do really well. Right. If an expense reporting agent is at its core, like it's just a, maybe it's just a Claude Code wrapper, right? Like that's what I'm using it as. But then the same agent is also like the one, because of 17 MCPs, it's doing
like web research and it's also, you know, helping me like file my own knowledge away. So I think it's funny that Claude Code for me for like two months is like an everything app, um, and that's why I'm trying to do the Claude Code, which is like that way. I can keep using Claude Code, but I also benefit from GPT-5 for a lot of programming, so I'm just like, okay, I'll use both of them and I use this when I need personal assistant, this I use for programming and that way I can
kind of like, it's, we keep adapting. I don't think three months from now, it's probably going to be different, but, um, it's, it's my current approach. Yeah. Yeah. Interesting. Yeah. A couple of years ago, actually at Hackathon, someone told me about the cloud squad and I was curious about that. And today I realize, yeah. My buddy at OpenAI sent it to me. He's like part cloud squad work trees. And then I was like, oh, I should learn work trees.
And now I'm like, okay, I need to learn work trees because this is actually very interesting. And it might be a good way for multiple agents to work on the same project without undoing each other's progress. So there's like coordination problem. It might be a part of the solution. Yeah. Yeah, definitely. Yeah. And I now remember my question. So, sorry. It's about cloning, cloning yourself. So I think in the future, I don't know if you really want to clone yourself.
And I think we should put the something, the brain to monitor your brain activities, AI ping, and then, because when you are exposed to some information, a certain information, sometimes you don't pay attention to it, you don't care. But if you record everything, the AI might assume, oh, you learned about this. But you know, your brain, if your brain activity is lower means, oh, you didn't pay attention to it.
So, but if you resonate to something, your brain activity is high. So, okay. This person learns about, oh, resonate, at least resonate with this information. So we should use the information or weigh this information. Maybe. You might be right. You might be right. I, I look, I love it because like, I'm a cyberpunk kid. I love cyberpunk. I don't know if you ever played cyberpunk 2077, but that's one of my favorite games.
And they have a lot of like human augmentation and like the whole game, you have a brain, you have brain chips, like brain chips are like, I play the game and I'm like buying brain chips, you know, but I think about like, what I put a brain chip in, yeah, maybe when I'm like 60 years old, maybe if it's safe, but, and then run GPT-8 on it, like maybe, you know but I think like more realistically, like noninvasive, noninvasive techniques would be a good
like middle ground as long as I don't, cause I don't know, like vision pro very powerful technology, but no one uses it, you know, like, so it's so clunky and like unwieldy and isolating. So I think it's like, if they can make it, and that's why meta is winning on the, cause they made it cool. Right. So if it's like cool and useful, then, um, that'll be interesting. I think you might be on to something like in my glasses, we're like, Oh yeah.
Parth is paying attention to this. We should probably save this for later. It's like, Oh, would you like me to like, like you really like this person? Like, do you want me to like, you know, remind you in six months to reach out to them? Like that could be useful. Right. Yeah. Yeah. Yeah, definitely. Yes. So yeah, I'm optimistic about the future and technology as well. Yeah. Yeah. Anyway. Yeah. Anyway, thank you so much for sharing a lot of, you know, insights and lessons,
you know, you learn a lot of things. Yeah. No, thanks for bringing such like, this is the kind of conversation that I love more than anything. So I appreciate the, but before any, but do you have any advice to people who are new to AI or have never called it before? Yeah, yeah, yeah, definitely. Um, I think that like, this is a great, this is probably one of the most interesting moments in human history that we get to live through this
transformation. Um, there are, and I think it's like, it's a huge leveling of the playing field. We now have, um, we have co-pilot systems that are smarter than many of our friends, smarter than most of the people that we know, and they can help you learn any topic. I was not an engineer two years ago. Now I think, I think I can safely say that I am an engineer. Um, you know, I'm not perfect in any ways, but because of language models,
I'm able to teach myself almost any topic I want. That means that like, if you want to get good at something, you can, you can't get good at everything. You have to pick a few lanes, but it's really important to play, use the technology, figure out what you want to do, like, like your own, you might have, you have to have a vision. You have to have an interesting life and a perspective. You have to like go live an interesting life so that you're like, Oh, I want to
do this. I want to make that. I want to, so when you realize what the things that you want to make, AI is going to help you make those things. And you will be surprised at how quickly you can get moving on your own ideas. And I think also don't only use it for work because that's a, like, what if your company fails? Like, what if the company isn't around a year from now? And then you wasted like all your energy was just trying to turn this into money
when actually maybe it could have made your own projects, your own life, better your family, your friends, like your passions. And then the things that you're passionate about when you apply AI to that, then it doesn't feel like work right now. It's like, Oh, I'm playing like now I get to do this. This is going to be fun. And then you go away further and then you figure things out. And that's like, I think apply it to your passions, use the technology.
Like it's not enough to talk about it. Like that is not use the technology. Then that way you're like all the noise kind of fades away when you're like, Oh, this is, this is what it actually is. This is the part where it's very good. This is where it's not good. And the only people who are figuring that out. of people using it. And there's not really a textbook or expert. So don't look for certifications. The best certification is like you said, hackathons,
like go make something and then just share that you're working on it. And you'll be surprised, people will come to you with like, oh yeah, I'm also working on this. Now you have collaborators, you have peers. So you wanna make things, share and talk about them. Not everything is gonna be a company, that's fine. A lot of things are just like, think of it as art. I think of sometimes like, oh yeah, this is like a beautiful thing that we just did for fun.
And then, but you create a conversation and learn from who the conversations you have, meet people. Then that's gonna build your network. Then you get the network intelligence. That's gonna give you the sense of like, okay, cool. We are gonna be good. Like it's uncertain. Everything is changing very quickly. So we have to adapt. The more adaptable you are, the faster you learn, the better off you're gonna be.
And sometimes you don't have energy or time to adapt. That's fine. You have friends, you can kind of do this. It's a team game. And I think that, yeah, that's like, and do things, don't just apply it to work. Like you might find that the most interesting passions are now possible. Things that you were putting off are now possible because of these tools. So you should apply it to the things that you care about.
Yeah, totally. Yeah, a hundred percent. And sorry, this is a very last question. So since Glasp is a way of platform where people share, what they're reading, learning as their digital legacy, and people can actually create the AI Chrome through learning process. But so we wanna ask this question to you. So what legacy or impact do you want to leave behind for future generations? So it's a tough question at the end, but yes.
I hope that people who interact with anything that I have done for, if you've even watched just a video that I put out, and I don't put a lot out, but like, if anything that, you know, something I made, I hope that you hope, I hope that you realize, like, it's like, just you should, like, it's like, hopefully it inspires people to just make things and learn and to challenge the notion. Like, there is no, like, we have so much to rebuild.
We have so much to make, so much to create and make things. I think, like, actually make things. That's very exciting to make things and to share those things. And hopefully it makes people more creative and ambitious, ambitious as well. Because a lot of the stuff that I've realized, a lot of stuff is a lot easier now. I thought it was gonna take 10, three years ago, a lot of these capabilities were never, I could not dream of these capabilities.
Now you can do in three hours what used to take three years. And that means that we have to be more ambitious, more creative and more optimistic too. Like, we have to actually, like, it's not 100% obvious that this will end out and well, we have to go make that, we have to make the future that we wanna live in. And so, like, you have a chance to be a part of making that future. This isn't, it doesn't just happen to you.
You are like, you are happening to the world around you. So hopefully, hopefully, like, people realize that they're all like, you're all the main characters, right? Like, you're in the driver's seat. Like, this is like your opportunity, right? Yeah, definitely and yeah, thank you for the beautiful answer. Yeah, thank you for joining today and we learned a lot and yeah, thank you so much. Appreciate it guys, yeah.