$100M AI Investor Reveals What is Coming in 2025

We are in the middle of a technical revolution. There was a time when people built machines to improve clothing manufacturing. Now, we are experiencing a revolution in knowledge work. I felt that there needs to be more investors who are genuinely passionate about technology. I believe the biggest breakthrough for me personally is that it's really simple.

I use it for about half of the stuff, and it buzzes. It says something like, "You want to read this." But then my phone just starts blowing up with notifications. I'm like, "What is going on?" I do play with Claude, and I've noticed this one small, weird trick that makes it so much better.

Hey crew, what is going on? Today we have Didi Doss with us. I'm very excited to have him on board. We are going to be talking not only about his builder hat, but also his investor hat—how he's evaluating companies. The coolest part, which I love and I'm excited to share with you, is that he's actually going to screen share and show us how he builds with Claude, computer use, and MCP. So let's jump into it. Didi, you're a builder and you're an investor. What are the cool people or what are the smart people in AI talking about right now? I think the one question on everybody's mind is: how far does this go? Is it about scaling, or is it just whether—? I think scaling is more of a tactical thing. It's just about how good these models can become. Do we get to a world where this solves problems we could never solve ourselves?

Yeah. I think that's the frontier. Because when people talk about AGI, it's quite an amorphous concept. I love to explain it. Exactly.

You know what blows my mind is that, as an industry, we don't have a solid definition of intelligence that we formally rely on. It's wild that we're still arguing about what intelligence actually is. It's like a strange side product of this whole revolution. I was telling a buddy yesterday that if we asked someone four years ago what AGI was, by all practical purposes, we’re there now. We would probably agree that if you open up GPT-40, start chatting, you're like, "Well, yep, we’re here."

Yeah, okay. So how far does this go? It sounds like you're talking about the application side, like value providing, right? Well, there is the value side.

I'm less concerned about that. Obviously, as an investor, I think about it. As more of a technical nerd, I just wonder how good these models can get. They still have flaws today, right?

Like there are still limits. You have the Arc Prize and everything. There are limits to what it can do, but what are those limits? How far does it go? I love the AlphaZero and AlphaGo analogy. Hopefully, by 2025, we will see a world where these models can handle general tasks.

It will figure out the rules. It will determine what it needs to do. It will take some time— as much time as it wants. It just achieves superhuman level intelligence on arbitrary tasks. In some ways, depending on how far you want to take that, I think that is reasonably close to AGI (Artificial General Intelligence).

Yeah, it's pretty wild. That sounds like your builder hat definition of AGI, which is pretty cool. So while we're on this topic, I have to ask about your investor perspective on AGI. How do you approach the next five years from an investor's standpoint? I've only been investing for a year.

The reason I wanted to get into investing after a 10-year career in engineering and product is that I felt there needed to be more investors who truly nerded out on tech. There are a lot of investors who are really good at what they do, analyzing businesses. But I wanted to be that voice of, "Just show me what's cool and show me what's possible." Yohei comes to mind. Yeah, fantastic. He's an investor-builder.

It's so cool to say, "Here’s my pipeline, here are the tools I built for myself, and here’s how they’re actually providing value." Exactly. I love that. As an investor, I also take at least 50% of that view, which is to show me what’s possible.

And we'll figure out a way to make it work. We'll figure out a way to make money. Sure, that's much easier than groundbreaking technology. The other 50% is, how do you take what exists today and apply it in useful and meaningful ways? It's also nice because you can call out a lot of the tech when you need to; you can just challenge them and ask real questions. I was literally, I don't want to name the technology—maybe I can't name the technology—but I was talking to a couple of other investors the other day. They mentioned a pitch where the presenters claimed they could do "rag but better." They threw out all these fancy terms, and I thought, "What do you think?" I said, "I feel like that's almost every pitch I've had for the last three months." Especially with things that are currently hot, like AI, there’s going to be a lot of snake oil. So, yeah, it is pretty important to call out the BS on a lot of these. I get a lot of pitches myself, too, and one of the easiest ways to go is to ask the tough questions.

Well, I get a lot of pitches myself too, and one of the easiest ways to go through them is to focus on the value rather than the technology itself. Don’t just tell me what your tech does; show me the value or demonstrate that a customer is paying for it. Just provide me with something tangible that comes from it because it's hard to fake true value—it's almost like proof of work right there. I would argue that this is particularly challenging with AI investments because it's easy to misrepresent value. I've seen many startups claiming a million in annual recurring revenue because CIOs and CTOs are being told to buy generative AI tools. But does that mean they have actually created any value? The revenue might be there at the beginning of year one, but the renewals may not follow. Furthermore, when you look at user metrics, like engagement and retention, if those numbers aren't strong either, then you're setting yourself up for trouble. So, I love the question: how far does this go, and what are the obvious gaps that limit us from reaching a true AGI perspective? There are several ways to break that down. Let’s talk specifically about text models. When I look at text models today, I see that they have significant potential, but...

I use Claude almost all the time, and I have played with all the other models too, like GPT-4 and LLaMA, and everything. What I really want to know is, okay, I have seen that people have all made this comment: when you use it for a very specific task—like a task that has a correct outcome, an objective task—it goes pretty far and can amaze people, but it doesn't go 100% of the way. I love to call this the full self-driving problem, where it's kind of easy to get a demo of full self-driving to work in most situations. However, it's really hard to get an actual full self-driving vehicle that can do everything on real roads everywhere. Let’s just take coding, for example.

That's one example of a very objective concept. You can also consider math, or any other form of logical reasoning, as well as biology, which combines memory and reasoning. I think about those things, and I wonder, okay, we're clearly facing a gap. People love cursive writing, but there are also some who feel, "Okay, cursive is awesome, except I hit a wall."

It can't go beyond a certain point, and I don't know what went wrong. Imagine how entitled or selfish those people must be to think that, oh, this AI is not good enough to do my job for me. You know what I mean? It's just funny that they're critical of AI when it's literally so helpful for so many things. But because it's not throwing you a touchdown at the very end of the end zone, you're giving it a hard time. Oh yeah, they're infamous. I think Louis CK, of all people, has this great joke about how people complain on flights. They're like, "Guys, yeah, flying in the middle of the sky, this is the most incredible thing that could ever happen! Have some gratitude." So I think about that too, obviously. But we also want to think about the frontier, what's possible. Even when I've been coding with Cursor and tools like this, there does come a time when I'm like, "Damn, it did a lot for me."

But now I kind of wonder; there's clearly something it can't do no matter how much I prompt it, and it goes through five cycles, and it's just not working. Yeah. And then I'm in this weird spot where I don't even know where to look. It's generated like ten files, 500 lines each. If I were writing it myself, I would know how to debug this.

Now I don't. Like a vocabulary word for a shared feeling of, I just had the AI generate a ton of stuff and now there's a bug, and I have no idea where it is. I don't know what that word is, but I hear people talk about it all the time. We need to come up with a word for it. I have a hypothesis that GPT-4 can do almost any task as long as it's sufficiently scoped down. Absolutely.

And we'll maybe get to this later on, but I love that anecdote because I play with computer use all the time. I wish I could do it with GPT-4 and all the O1 stuff as well, but I just don't. They don't have the computer use version yet. However, when I do play with Claude, I noticed one small, weird trick that makes it so much better. It was a really simple trick, and it was exactly what you said: I just went to Claude first and said, "Here are my high-level three prompts. You, Claude, have to now give instructions to a computer to break this down into tasks."

Yes. And then I feed that huge prompt into the computer, use it, and say, "Now you do it." It's a huge improvement. One of my, um, when I work with clients and we talk about prompting, one of the most common problems I see is that they stuff too much information into the same prompt. They want five tasks. I wanted to create a mini tool, and it sounds like we have the same thought.

It's like, feed them the prompt and output what the individual prompts should actually be, based on your main one. Just scope it down even further. Because then that makes evaluations a lot easier too. It does. You know where it's messing up.

It does. They call it meta-prompting, and I think it kind of alludes to this fundamental computer science concept, which is that everything is solved by one additional layer of abstraction. In this case, it's up to you to build that layer of abstraction. Then, you can create a prompt that generates more prompts, and just do it. That's exactly it.

On the investor side, you're seeing a lot of startups' pitches come through. What are the trends in the past, let's call it three months? Not the early stuff, not 2023, but what have the last three months looked like from an investment perspective? Well, a lot of people are talking about agents, right? Agents are one of those terms that make people wonder, what does that really mean when you say "agent"?

And so, when it comes to AI companies, let me break down the field. There are foundational companies. Foundational companies can be tough because sometimes you're really asking the question: are you truly trying to compete with people who have thousands of dollars? What are some examples of these? Well, a good example is actually this company, Descartes.

They're a bold, brilliant founder. He's just one of those guys. As an investor now, I've learned this: just trust your instincts. You meet this guy, Dean, I think his name is.

Just whiz. You're like, "Oh my God! This guy's so small." It's almost like you're back in the person. Go get the idea.

Yeah, you're almost always actually backing the person. But some people just have this energy where you're like, "This guy is a genius." And he was one of those, tackling a problem way out of his league, which is this idea of generative gaming. You know, it's hard enough to generate good images; it's even harder to do it in real-time and then consistently.

So the question to ask is: You have a game that involves thousands of game designers, engineers, and various aspects of graphics and how things behave in that game. What if you could just throw generative AI at it and have each pixel show up and interact within the game? It seems like this has no business being done—it's almost childlike in its simplicity and has been tackled by a small team. These are things that Google can't achieve currently, and while they have developed something similar, other companies like OpenAI and Anthropic haven't been able to do this either. However, there’s a company called Cartesia, which is working on a foundational model for audio. They came onto the scene after 11 Labs, but they took a different approach, resulting in a reasonable outcome so far. Meanwhile, 11 Labs is excelling at what they do right now as well.

They're currently shipping and building a ton of stuff. Okay, so let's talk about foundational companies. What are the other trends we are seeing here? Then you have the companies that are very clearly on the application side, which I like to call workflow automation. I know that sounds boring, but I believe we are in the middle of a technological revolution. There was a time when people created machines to enhance manufacturing, like making clothes and other products more efficiently. Now, we are in that revolution for knowledge work. The question is: what machines can we make? We refer to those machines as agents. People often take extreme positions, arguing that their jobs are not automatable because they require a lot of intelligence. However, it's clear that there are certain jobs on the other end of the spectrum that are dull and tedious, and no one wants to do them. Many of these jobs involve looking at Google Sheets, Excel files, or PDFs to gather data, writing that data down, sending it to someone, and then reviewing the responses. Most of these tasks are just about moving information around. If your job is repetitive enough, it's possible to automate some of those tasks today. A lot of people are working on automation solutions that could involve voice tools or straightforward text interfaces. When we talk about vertical AI, are we referring to platforms like Lindy and Zapier, which cater to specific industries? In this case, I think most of the companies in this category are quite vertical. Initially, you need a way to approach a company and say, "Hey, we know you have 100 people doing this job. We believe you don’t need that many, and we can do it for you." That’s the high-level pitch. It feels like there is currently an arbitrage opportunity for selling AI combined with your specific industry expertise.

Like, people will eventually get over that, and it's already starting to get diluted. But if you go to a law firm and say, "Hey, I'm making an AI lawyer," they respond with, "Great, I need one of those," without even knowing what it actually is. Yes, there is this idea or phenomenon where people buy tools that overpromise. Yeah, totally. And it doesn’t work a lot of the time. But on the other hand, I've seen it in insurance, healthcare, and finance.

There are a lot of pretty tedious industries, especially those with back offices, where it's just labor that you don't want to pay for, basically. So, you think, "Automate that away, and let's move forward." As grandiose as it may sound, we want to evolve as a species and not have to engage in mundane tasks. We can focus on other less meaningful work. Every generation has its share of unfulfilling tasks. One of my favorite consulting books—though I forget its title—conveys the main lesson that there's always a number one problem to tackle.

Once number one is solved, well, number two gets a promotion to number one, right? So there's always going to be dull work. There's always going to be interesting work. It's just a big, long cycle that comes from there. True.

But to your point, there's a whole aspect of workflow automation. How do I automate people's jobs, assist, or automate whatever needs to be done to make it a little bit easier? The third category I will touch on is quite challenging because there's such a long tail of AI startups. However, especially in the last three months, I've noticed more companies discussing the concept of what they call AI SREs, which is a common area that people want to address, and I will explain what that entails. The second aspect, which is also semi-related, involves programming agents. This isn’t a new concept; it has been around for a while, but in the last three months, the evolution has been pretty interesting. Regarding competition, there's a competitor to Cursor emerging now as well. I enjoy competition; it’s great to see it. Cursor is doing really well—who is the competitor you were referring to?

Windsurf. Yes, that's correct. It's awesome to see. It'll be interesting to see how this all plays out.

All right, so there are two categories: SREs (Site Reliability Engineers). There are tons of companies working on various versions of this. When I say SRE, I also mean that support is somewhat included in that—both technical and non-technical support. The idea is that if I have a problem in the company, such as code breaking in production or things going wrong, support is there to help.

That is money on the line, especially from a big company. Like everything, you can do the math; every hour can be worth tens of millions of dollars. It’s kind of crazy that right now the best solution we have is to send a person to look into the issue and try to figure it out. I've done that at Clean and at Google, and it is one of the most nerve-wracking experiences you can have as a software engineer. When something is down in production, time is of the essence, and there are literally petabytes of data to sift through. Which dashboard do I look at? What do I check first? What should I query next?

How do I fix it? When does the release go out? It's like my brain is going bonkers. This could be a third, and it's like, oh boy. It's crazy.

Uh-oh. It's crazy. It's a really bad, hard thing to do, but there are ways to make this easier. A couple of companies are handling this pretty well.

There's resolve. There's a company traversal. There are a number of other companies trying to do this. It's a super interesting category of problems, at least to me.

And then the last one, that's what I was talking about. It's coding agents. Nice! Yeah, there's just so much more stuff we can build.

There's just so much more, and it's so cool. It's a place that's ripe too, because everyone talks about how code is easily verifiable. Does it work, or does it not? Right?

The hardest part is still just with the planning piece. Do you get the intention right? Can the human adequately express their intent to the coding agent in the first place? Exactly. It's amazing.

One of the side thoughts I've had is that I bet the field of epistemology, which is the study of knowledge, must be gaining a lot of momentum because of the current developments in AI. People are increasingly thinking about how to express language, how to formulate their thoughts, how to approach planning, and how to improve communication.

All that wouldn't have happened unless we had these text-based LLMs that we now have a financial incentive to really dig into how we're dealing with these things. Absolutely. I mean, I think it's incredible where we are in the world. So let's move on to the builder side. Yeah.

So, what technologies are you personally building with right now that you're excited about? I think the biggest unlock for me personally is really simple. I use cloud computing for about half of this stuff. However, there are two caveats. There's a version of cloud computing that came out, designed for safety. Anthropic is all about safety, as they should be. I don't think this is something that you should just let everyone willy-nilly go and run amok on their computers. But I like to take some risks, so I cloned the version they have for Linux.

And you have to make a couple of edits to that. I can go into as much detail as you want, and I'll show it to you later if you’d like. But you can make it work on your actual Mac, and it will make things a lot better.

That's caveat number one. The second thing that I've realized over time is that, as cool as this is, its cloud computer use did this really neat thing: it could control your mouse and view your screen, then take an action. However, as impressive as that is, it's kind of the most lame part. In the sense that this is something a human could do in a second, but now it takes about 20 seconds. It's really, really frustrating when it can accomplish all this amazing stuff but then gets stuck trying to find a button and click it on your computer. So when I prompt it, I make a few adjustments and say, "Can you not do that? Just try to stick to the command line." Sure, okay. You can pretty much run anything on the command line—just stick to that. And I feel like the progress has been incredible. That's cool. So, is cloud computer use going to be the product, or is it going to be the platform that other people can build really cool products on top of? It's unclear. Right now, they have MCP, which I like. I don’t want to speak for Anthropic, but it feels like MCP is the version of safe cloud computer use. Can you define what MCP is?

Could you define it in layman's terms too? The way I see it, and this is probably not doing justice to everything it can do, is that MCP is a way for you to safely say, "Hey Claude, you can use these tools." I can give you some memory, and you can use these tools to customize how you interact with Claude. However, it’s very developer-centric and not really consumer-facing yet. There’s a whole specification that comes with it, right?

Yeah, so I don't know. I don't know where Anthropic goes with this, but I do know, coming back to the start, it's a big thing that startups are thinking about. What is that? Computer use.

How to use computer technology to automate a wider variety of tasks. The first era of task automation was primarily browser-based. The general approach involved scraping the Document Object Model (DOM), which includes everything in the HTML of a webpage, and then attempting to identify various methods for interaction, such as clicking elements. If you reflect on this for a moment, you might realize that it's a relatively simple process.

Yeah. And you know, you could think about that for two seconds and realize that it's a pretty poor strategy because that's not how humans work. It would be insane to read thousands and thousands of lines of HTML code and then figure out what to do on a page. The whole point of the page is that you see something and you do something, and computer use makes that a lot better because you have the visual aspect. But it's slow.

But there is going to be a time when the vision latency comes down, and you can start doing actual actions on pages that automate a whole other boatload of tasks, both on the web and on your desktop. So it's an exciting sort of high-level technical thing that startups can do now. Um, one last question before we dig into computers. We'll pull computers out here, so there's an open question about the future of automation. What's the interface for an AI type of thing with, call it, third-party software? Will it be more like cloud computing where you're actually interacting through the UI, or is the API layer going to get built out, and that's where the automation is going to happen? I think the jury's still out. What's your opinion on which way it's going to go? There have been a couple of startups I've seen trying to approach how that layer should look. One is called Composio, which is a way of saying there are a bunch of tools out there for you to use. I'm going to "toolify" it for you so that you can go ahead, and the LLM can see a prompt, understand an API spec of, say, GitHub, and then figure out what to do. You can think about it almost like technical documentation for LLMs.

And then I will teach an LLM how to use the tools that it has. I will keep it up to date, and you can just feed it to the LLM so it can use my tool. That's one angle. There's another company called Anon. What Anon does is more akin to Plaid for APIs. Just as Plaid enables bank authorizations via a protocol, Anon aims to extract authorization to the user while allowing an agent to act on your behalf.

And the funny part about Plaid is that it used to extract authorization directly from the user, but now it allows an agent to act on behalf of the user. Initially, Plaid operated primarily through scraping, but now it's done properly. This transformation could serve as a metaphor for the direction in which things are heading. I still think tool use has not achieved mainstream adoption yet, so it's to be determined where that will lead. There are obviously many startups and enterprise companies doing a good job in this space; however, end users have not fully embraced it. Most end users are still just talking to Claude or ChatGPT and asking, "What the heck's going on right now?" They're simply waiting for more builders to create tools they can utilize. To me, it’s crazy.

Maybe it's a good aside, but when you step out of the Silicon Valley bubble, nobody knows what's going on. It can be almost baffling sometimes. When you spend so much time on AI Twitter and you're literally in San Francisco, right in the heart of the city, it's hard to escape it. You go to these meetups, and it's like, oh my goodness, I'm so far behind because there are so many people who are so far ahead of me. But, like you said, when you step outside of it, it's almost as if you're speaking different languages when it comes to all of this.

Yeah, there’s a social and cultural element to this too. I don’t know if you’ve seen this, but a lot of the youth of the world today, like Gen Z and Gen Alpha, have this interesting hatred of AI in a huge way. It’s like it’s not cool to talk about it or use it. It’s really strange. I feel like an uncle saying this, but when I was a kid, all new tech was cool. It was really cool.

Nerdy, but cool. What is that reaction? Maybe it came from the world of art, where I know AI has received a really, really negative reaction from all artists. But now it's seeped into AI at large. People are just like, "Ugh. Yeah."

All right. Well, whatever. They're lost. Cool. Well, I tell you what, let's do some building.

So, Didi, one of the cool things is that many of the projects you have revolve around diagrams. That’s great because they're not only visually appealing for humans to look at, but they also help AI understand the flow of information in ways that language may not convey initially. So, why do you love diagrams so much? I think diagrams are just way more expressive. They bring structure to unstructured data.

And I think the caveat is, look, it’s not going to be perfect. You’re going to get things wrong. But just the fact that I'm a huge book nerd and I love reading books has always created this nagging thought in the back of my mind: I wish I had a character diagram because my memory is terrible. I watched *Game of Thrones*, and every episode, I’m like, "Well, who’s this? Who’s that?"

Why do they know them? I don't want to read Wikipedia again. I want to see the relationships. That's cool. And that kind of inspired that viral Harry Potter tweet back in the day. Sure.

Where I was like, "Let me just throw all of Harry Potter at— in that case, it was Gemini for the context window— and just get a graph out of it." That's cool. I can get a little diagram out of it. So maybe I can show you that. That's one.

And another topic I'll discuss, which we will address, is generating diagrams from code. It's literally the biggest problem in the world when you look at a new codebase. It's not the time you spend reading it; it's the time you spend not doing anything because you're overwhelmed. Sure, you're looking at it, and you're thinking, "Nah, I don't know."

That's because it's just a bunch of text files. It's like, where's the insertion point? Where do I start? Yeah, it's crazy if you think about it.

Where do I start? So, I will do a demo. What I did was dump the entire repository into a plain text file, put it into an AI admin, and asked, "Hey, tell me what the code looks like." I love it; it's beautiful. Let’s start with something simple. What do you want to summarize? What would be cool? Okay, I have a good one. Let’s start with this. Can you see my screen? Yes, I can. It’s funny that you mentioned the work of the Glern thing because I think I might have that already downloaded.

Oh, that would be nice. Cool, let's see what we can do here. This is all live. I have not done anything beforehand to prepare for this, and I think that is going to be the beauty of it. So, where are we?

So, you like working in the terminal the most? Yep, I am a purist, I guess they would call me. I like to use Zed. So, sorry, cursor guys, but this is awesome. This is a VTT file, and what this means is that I use this tool called YTDLP.

Ytdlp not only helps you download YouTube videos, but it also helps you download transcripts. I'll show you a couple. That's just the transcript that comes straight from YouTube itself, right? Yep. Okay, so we got it. What I did was generate this previously with Claude.

If you look at this code, it is basically trying to remove the timestamp artifacts, which consume a bunch of tokens, and it provides me with some plain text. This plain text is not great, and there is some repetition, but it will serve our purpose. It will get you there.

It will serve our purpose. So let's see if this one fits. We're going to go ahead and put it all in. Oh great, it fit! I'm going to say, draw a graph. I know Mermaid is the default, but I don't like using Mermaid, so draw a visually aesthetic graph. This is just what I like to do. There you go in a DOT file. Why don't you try Mermaid out of curiosity? I think Mermaid is too simplistic, and I'm just more of an old school fan. Mermaid comes from something called DOT, and DOT or Graph is what I used to use in college, so I'm just more used to that. Yeah, there you go. For those who don’t know, Didi was a hardcore engineer before becoming an investor. I know.

We actually had our first conversation when you were still at Glean a long time ago. Yeah, I was an engineer. I was an engineer creating subgraphs demonstrating the relationship between various topics discussed in this podcast. What about prompting techniques? Do you like thinking out loud? Do you appreciate tips?

What's your take on those? I wouldn't say I'm a pro-prompter. My advice is: don't think, do. Just go try something out. Just go try something.

You'll figure it out in the end. A couple of these things that I've done, I've just figured out over time. So, here, it generated a bunch of content. Now, what I do is go online to Graphviz, paste it in, and look at that—it's nice! Yeah, it's just so cool. We have some personal background on hearing impairment, social isolation, and reading focus that talks about rabbit holes and patterns. It really gets you somewhere, and it looks pretty nice. This is just such a better way to look at a long body of text. It makes so much more sense.

This makes it so much easier to do. We'll explore this a little bit more because I think this is a question I get asked a lot: how do I get out of this context window limitation? One of my techniques, or prompting hacks, is to assume all of these nodes and edges exist. Keep going. Add new ones and make it exhaustive.

And so you want it to go. You want to double-click here, so you want it to go; you want it to double-click here. I just want it to go deeper. I think this graph was a good start. Let's see where it can take me. It does regenerate some of the boilerplate here, but it keeps going on and on.

Oh, look, we have some GLP drugs, some China, and some great stuff! So now, what I'm going to do is not copy the whole thing; I'm just going to copy the subgraph. Let's see if it worked. So, going back to this—oh, look at that! It got even more complicated; that's nice. So that's cool! Now it's like, if you're a hyper-productive guy, you don't have that much time. You want to listen to this podcast, but you don't have three hours. Wouldn't it be great to download it and have timestamps for each segment? You could just click on it. Great! Now, you would just know.

It'd be so nice to have timestamps for each one of these and just click on it to listen. Great! It's funny that you ask. It's like, let's say I'm interested in AI and intelligence. Can you...? Oh, well, it can't do that; it doesn't have the timestamps, right? But it would be really cool. Yeah, in a world where we could fit the timestamps, and I've done that before too, which is just to give me quotes with timestamps for these areas that I like. That's cool. So, other than just looking at the diagram, will you use that diagram schema anywhere else? Right? Like, does that make its way to an input for a prompt for you? I haven't tried that yet.

Sure! Here’s the reformatted text with correct capitalization and spelling: "Yeah. One example I talked to was another person in this interview series called Arjun. He wanted to evaluate how well his support team was following the call script they had been instructed to use. So we first started with the diagram flow of the call script, outlining what should happen at each step: 'If this, then go here,' and so on."

You give that, pair it up with a transcript, and then the AI, LLM as a judge, will assess how well he followed that script. I've seen some voice-action companies do that. It was very cool. It's pretty cool. Okay, so that's on the diagram side.

Let's discuss computer usage. Let's get started. So, we're going to—uh, and again, this is my fork of Anthropic's computer use. It's totally unsafe, so do not use it for anything important, okay? Yeah, do not use it; it can do all sorts of weird things. One issue I encountered when I was trying to make this do something was that it started a rogue process and created like a million files in a directory, which then caused the computer to stop loading because it went completely bonkers. So, let's try to accomplish something here. What I'm going to do first is go to Claude, and we're going to say, "Instruct a computer to..." You know, there are some people who like to save prompts and are very meticulous about it. I'm not that guy.

Instruct a computer in as much detail as possible. Do not write the code. We need to instruct Claude to perform the following tasks because it loves to write code. First, create a directory called "dev." Within that directory, create another folder named "greg_demo." In this new directory, we will develop an application with both a front end and a back end. Now, let's brainstorm ideas for the application. Greg, what do you think about focusing on simple restaurants in San Francisco? This concept gives us the opportunity to create a user-friendly interface where users can browse, search, and review local dining options. If you have any pushback, please share your thoughts. I think creating an app that could help users find restaurants in San Francisco is a better approach than something overly complicated or generic like venture capital inbound. What do you think?

Oh, okay. I'm thinking about that one: Apis. Apis will be tough because you have to— Oh, you want to pull some of my grill data too? Oh, we could.

I'm trying to think. You know what? I'll do one now. I don't know how complicated you want to get. I want to scrape OpenTable to find out when reservations are opening.

Oh, interesting. There's this one spot in San Francisco that, no matter where I look, it's always booked, and I'm never there early enough. You know what? It might not be as visual, but let's see how far we can go. All right, let's figure out how we can do that.

In that directory, create an app to scrape OpenTable. I have the restaurant reservations in San Francisco to find. Do you have a restaurant? I do. I just put it in the Riverside Studio chat. It's called Four Kings San Francisco.

How do you spell that? I'm just four. Okay, number four. To find Four Kings' availability and let me—I’ll make it interesting.

And let's keep it simple for now. Tell me when it opens up. Yes. So, let's see how this goes; it might say something about it. Okay, you're creating the prompt right now that you're going to give to the computer user.

Yeah. Okay. Yeah. Man, this is a lot. I don't even know.

This is a lot. This is a lot. So this is going to be a lot. This is a lot. This is a lot.

So this is going to be fun because I have no idea if this is going to work. We'll see how far it goes. Okay. Da-da-da-da-da. Would you like me to proceed with everybody?

No, I don't want you to proceed. Try to use this another way. Try to use the command line only as much as possible and build an MVP. Yes. Boom.

So now, okay, let me open up what's going on here. This is how I typically look at it. Da-da-da-da-da. Now, let's see what's happening. This is my usual perspective.

Okay, it is trying to... It's creating all these files. It's trying to create a config. It's trying to download all the requirements for Python. This is where it's running at. And then, real quick, let me just go into that demo.

But I think for a lot of people, if you prompt it the right way, absolutely. I have a friend who works at one of the companies I mentioned before. Let's pause there and bring it up for a second. You had a tweet that said, "This is an exaggeration, but there are so many software engineers, not just FANG, who I know that literally make two code changes a month, and they're getting paid $200,000 to $300,000 a year." It went viral.

It popped off. You listed the names of companies. So, who's this friend at the company? There are many friends at many companies. Andre Karpathy tweeted back at that, and I love using other people to validate that this is a real thing.

And he's like, "People brag about this all the time." Yeah, there are entire subreddits dedicated to it. Yeah, "Overemployed" is an entire subreddit focused on it. So I think, for me, as someone in the software engineering community, I don't think this was a completely new idea.

Everyone knew this happened. I think outside that community, maybe people didn't know as much; perhaps the executives didn't know as much either. Now, maybe some people on the street are aware without question. There are people who don't take action, and is it their fault? Is it their fault? No, I mean, they aren't sometimes.

Well, it's about where the incentives are, right? Incentives? And those incentives go up the chain, right? The managers aren't giving them work; they don't have anything to do because the managers don't expect them to do anything. At any big company, managers never have the incentive to get rid of you.

They always have a follow-up tweet about that. They always want to increase headcount because that's the only thing that gets them promoted. So, no matter how badly you perform, they have no reason to get rid of you. And look, the nature of work is that you're not going to have enough work for five people. Let's say you have a team of five.

You're not going to have consistent work for five people all the time. The problem arises when you essentially have no work for anyone, and then you're in a position where you think, "I can't tell my manager that I don't have any work for these five people." So, you end up pretending that you have tasks to ensure you keep your job. This issue then escalates up the chain, and ultimately, no one accomplishes anything. To get back to the original topic, there was a follow-up in that thread where someone asked, "Does AI have anything to do with this?" The answer was no. The people doing these jobs have no idea about computer usage or tools like Claude or ChatGPT. They are practically indifferent; they are in Mexico or Hawaii, often on vacation. I can give one example involving a couple of friends who work in a similar role. I asked them, "What do you actually do?" One of them replied, "We maintain a large Java service." He mentioned that he only had to do one dependency upgrade each month. I thought that sounded ridiculous and said, "Let's talk about that dependency upgrade. How much time do you spend on it?" He said, "Oh, I spent about ten hours on it." I was taken aback and asked, "Can you show me what you're doing?" It was a friend of mine, so we met up, and when he pulled up his computer to show me, I thought, "Why are you doing this manually?"

That's insane. You're trying to resolve Java depths, and I know that this is overwhelming. Are you thinking of going to Stack Overflow or something? Yeah, he's basically Googling a bunch of things to try to figure out how to use this dependency system that he doesn't understand. I'm like, dude, you can run this through Claude, and it will fix it for you.

And his response was, "Well, they don't let us use AI at work." I thought, "Okay, well, that's a pretty poor response because you're the person who's going to have to do it now. Exactly, not me." And so that's kind of it.

That's kind of where it is. All right. So where are we now with this thing? Let's see... There is a problem that occurs. Here's what happened: the scraper ran, and it blocked the thread. It waited four minutes and then timed out. Well, there you go—that's the work in progress with it. We'll see how it goes. There’s a log, at least a little bit of one, and it confirms that we are at the right URL, which is good.

Started the reservation. I encountered an issue while loading the reservation range. Yeah, oh, look at that! I mean, they're probably pretty good about scraping. Yeah, yeah, if any OpenTable engineers are listening to this, it looked like the error was a timeout, and then it schedules a check every five minutes. Therefore, the computer use timed out and is scheduled to check every five minutes. There you go! Nice, man. But look, it wrote a lot of code. It somehow generated this notification sending code, which writes an email to you. Sometimes, when I look at AI-generated code, I just feel overwhelmed.

It's like, man, I don't have any mental bearing or anchor points to lock myself into this, and it makes me a little nervous sometimes. But I tell you what, when you click run and it works, it's like, well, all right. It's true; I feel a little bit better because I sometimes understand what's going on.

I'll give you one example while we're at it. And I know we're at the time, but I can go to 1:30 flats, don't worry. Beautiful. Here, I will do another one of these. And actually, you know what? A simpler one.

And while you're doing that, I would love to hear about what stack you're using right now. Where do you like to deploy? What's your database? What's all that stuff? I keep it really simple.

This was true when I worked at companies: do what you know, and don't do what you don't know. Sometimes I want to learn something new, and I will learn something new. I know Google Cloud really well, so I focus on Google Cloud authentication.

That's one command after one of my command lines, and then I say, "Go deploy it with Google Cloud." I ask the computer to tell me what to do, but it just doesn't respond. Nice. So, GCP.

GCP. And then for what else? What other? Let's see. I mean, GCP, you're using Zed.

You're using a lot of cloud. Oh, I use Zed. I use Cursor 2. I use a lot of cloud in computer use. Sometimes, when the cloud doesn't work, I will go to 01 and I'll try to go back and forth to get different answers.

I use Gemini when I need long context. Gemini is pretty useful in that. Yeah, that's kind of my stack. When it comes to languages, there's another really cool project that I was working on in Rust. So sometimes I'll use Rust when I view it to be super fast. But I'm mostly doing stuff in Python.

And then if I do frontends, it'll be in React. That's mostly it. Beautiful. Love it, man. I want to tell you about the coolest example of something I learned.

Usually, one of the hard mental constructs with AI is that you don't know what it can do. So, there's this "unknown unknown" problem: I know what I can do, therefore I ask it to do things that I know how to do. I don't know what I can't do, so I can't ask it to do those things. However, there was this one example where I asked Claude, using meta prompting, to give me some random ideas. The ideas it generated were quite out there.

And then I took some of those ideas, and one of those ideas was to create a song. In my head, I thought, "Oh, that's stupid because obviously computers can't do that yet." But I was curious—what would happen if I did that? This idea blew my mind because I didn't know it was possible. I’ll show you this example of how you can create waveforms from scratch simply with code. It built that from absolute scratch using a library called SoX, which I had never heard of before. I did not know how it worked, but it downloaded it and used it. This is a really simple example, but I managed to create some really amazing results. Okay, so let’s say we were going to take the OpenTable example we just did. It works, it's cool, it’s all there, and you’re about to package it up and post it on Twitter. Since you're very good at this, what is the mental process you're going through? What do you think about as you package everything up and prepare to click "send" or "post"?

Okay, so step number one is to create a GitHub repository. Usually, you can get computers to take care of most of that for you. Put it on GitHub and take care of a couple of general sanitary things, such as getting some documentation going so people know what it does. You can simply spit out all the code and say, "Now create a README for it." Make sure to have a nice README; that’s step one. Step two involves thinking about videos. There are two ways to approach this: one is to create a video of the process, which appeals to a certain audience, and the other is to create a video of the output, which appeals to a different audience. Sometimes, just a picture works better than a video. It can be pretty laborious because I have to do a QuickTime screen recording. I'm not super fancy with it, but I set up a screen recording and run the entire process back.

I literally load it into iMovie, but sometimes I use other video software. I clip it up, make it faster, and cut out the "ums" and pauses. I usually don't add audio for this because that's a whole other thing. I just want to show the demo, as it's usually intended for Twitter, not YouTube.

Cut it up, make it four times faster if it's a computer-related task, and export it—boom! I use an app called Typefully to schedule tweets. I put the video on there; you never put a link in your first tweet; instead, you include the link in the thread. So, I link to my repository in a follow-up tweet and then I just compose my thoughts: "Okay, what do I want to say and how do I want to talk about this?" One example of that was when I built something on my friend's computer. I always thought this would be a really nice concept for people to understand; everyone in San Francisco, especially, is like, "Make me a bot that swipes on Hinge." When iPhone screen mirroring came out, I connected the dots and thought, "Let's make this happen." That was a cool project. So, I took the whole video, automated the process, packaged up the demo, removed any personally identifiable information, and boom—tweeted it. So, you have an idea, you build it, you share it, and then it goes viral! Every other post seems to go viral. In fact, during the pre-call for this discussion, I mentioned I wanted to talk about your viral post, and you asked, "Which one?" It's clear you have so much great content out there!

So, how do you build up so much? Like, what's the secret behind it? Firstly, I just want to say, well, I'll tell you the story of how I got into Twitter, because I think people misconstrue a lot of the online stuff. I hate the word "influencer" because that's not what I try to do at all. The reason why I'm able to output so much is that I genuinely love building things. In the past, one of the biggest constraints, I think mentally, as I reflect on it, was the question: what's the point of me building this?

No one's going to see it, and no one's going to care. It's just going to be a little project that sits on GitHub with no views, and I can tell one buddy of mine who cares, while nobody else gives a damn. Now that I have an audience, I feel genuinely grateful because it allows me to go ahead and build it knowing that at least some people will see it. That’s really all I wanted to do.

And the reason I got into writing online in the first place was because the story is that my friend and I were out for drinks. We were having a great time when I gave him an idea. He looked at me and said, "Dude, that's a terrible idea." I argued back, saying, "What are you talking about? No, it's a great idea!"

And he kind of flips the situation on me and says, “Look, this is before I really tweeted, but I've been online a little bit.” He continues, “I trust your judgment. If you, sober, can go online and write about this publicly, I’ll believe it's a good idea.” I'm like, well, that’s an interesting thought right there. It’s literally just framing the idea.

Exactly. And then I realized that the gap between having an idea and thinking you're awesome versus pressing the send button is so big for so many people. It's so true. I realized that this was stopping me from doing a lot. I thought I had great ideas, but I didn't put them into production or put them out there.

And the second thing is that everyone does this. On the internet, you go into rabbit holes; you learn things, but you don't remember much. I found a good way to address that. I promised myself that I would go online and write one thing every day—nice, one thing that’s somewhat useful to other people that I learned—and that's how it all started. I just promised myself to do that, and it gave me a reason to go and learn. Aimlessly exploring, I would say: I learned a thing; here’s a thing. And that’s kind of how it evolved. But then, obviously, you know, it goes from there. I was just listening to the Darkech Goran interview, and I think Guern said that one of his north star metrics for himself is to maximize the number of rabbit holes he goes down. It's literally just getting into a flow state on an esoteric topic and just having fun with it. So when you said "rabbit hole," it made me think of that. One thing I would love to do is inspire people to share more here. Going from the journey of having no audience all the way up to hundreds of thousands of followers on Twitter, I'm sure there are a lot of cool serendipitous moments that happened along the way. Maybe somebody famous reached out, or an opportunity presented itself, or maybe there was a new career development. What are some of the cool things that happened along your journey that are due to having a bit more exposure? Oh, this is my favorite one. I was on the first date with the girl I’ve been seeing now for a while. I remember we were driving in my car together across the Bay Bridge, going from San Francisco to Oakland.

And it's the first time we ever met. My phone starts buzzing, and I see a notification. I'm driving, and the phone is on the dash. It buzzes, and it says something about Elon. But it wasn't the usual "Elon tweeted something."

It was—I couldn't tell what it was. I wasn't really looking at my phone, but then my phone just started blowing up. It was notification after notification, and I was like, "What is going on?"

And I probably shouldn't publicly admit this, but then I put my car on self-driving, and I'm like, "Oh, what's going on?" You're like, "Babe, I gotta look at this." No, it's because my phone's blowing up, and I saw an email notification that it blew up, and I looked at it, and it's like Elon had retweeted something and written to me. In my mind, I'm like, "Oh my God, what is going on?" To her, it's just meaningless; she's like, "I don't know what you're so obsessed about." But I have friends reaching out to me saying, "Congratulations!" Obviously, I'm a huge fanboy of a lot of things Elon does. So that was one moment where I'm like, "Damn, the richest guy in the world knows I exist." It's kind of nice. Those are the positive moments.

Like, yeah, Elon is... I think we've had ten replies and stuff now. And Paul Graham, Marc Andreessen, and a bunch of these pretty esteemed tech people have weighed in. But I think it's worth noting that you have to take the positives with the negatives. The hardest part about writing online is that there are going to be times when people tear you apart, and that has happened a ton of times as well.

But what percentage of the energy, when it ends up blowing up like that, is negative? It depends on the post. I think there are some threads where you're like, "Okay, 90% positive." There are a couple of people who are kind of ignorant and saying nonsense, and that's fine; you build thick skin to those things. The hardest part is when you get a full-blown cancel situation.

And I don't think there's anybody I've met who has some scale online who hasn't had at least one of those experiences. Almost. But when it happens for the first time in a big way, it messes with me. Yeah, I'm like, as much as you want to say I have thick skin, you're just looking at people on the street and wondering, "Do they know? Do they know? Do they know that I'm considered a villain online to millions of people?" You eventually get over it, but that's happened a couple of times too. I will share that there are so many younger people, especially, that I've met who have literally had one of those experiences happen and have stopped posting online for years.

It kills people's energy and motivation. And what's the advice to them? For me, it's like the same advice: you know, in real life, when people say you care too much about what other people think of you, they don't really care. It's kind of that—people aren't going to say, "Yeah, just keep doing you; just keep doing it." And that's the other thing too: the algorithm wants to promote good content. If you post something that's not good, don't worry; nobody's going to see it. They're only going to see the good stuff. There’s only one caveat, which is to take feedback. If there is something you're actually doing wrong, then ask yourself, "What am I actually doing here?" But in many cases, that feedback is simply, "Just keep doing what you're doing. This is fabulous." Awesome. Thank you very much for chatting!