Transcription
I was talking to this guy the other day who was doing web scraping. He spent $80 with Claude. I showed him Gemini. It cost like five bucks to do what he had done.
I was helping a friend. He was building dog parks near me, and he was saying that it was going to take him 80 hours to gather the data needed to build this website. I was like, "Wait, I can automate this for you, man. We got it running." So, I saved him 80 hours, and he launched his website.
I know a lot of people who are like, "I have this idea for an app," and it's like, "You're never going to do that cuz you would need an engineer or coder to create an MVP for you." But you can actually do that now.
Men, I'm very excited. I'm excited to learn some AI today. You are coming to me from the capital of the world of technology, the Bay Area. Here's my first question: What tools do you use?
LLMs. And for what?
I use all the LLMs, and they all kind of have different excel at different things. So, I use them for different reasons. So, for coding, anything coding-related, I'll use Claude because it's the best coding model. For general-purpose things, but also just like research, I'll use ChatGPT, especially like deep research that's been super helpful. Like I'm actually trying to learn a bit of like quantitative trading right now, like some side project with a friend, and I know nothing about it, but like the deep research has been super helpful with that. And then actually, I feel like this is the one people sleep on: Gemini. So, for 80% of use cases, I actually feel like Gemini—people who code, it does—it'll get the job done just as good as the other models, but at a tenth of the cost. And people, I feel like they don't know about Gemini. They don't know how cheap it is.
So, I was talking to this guy the other day who was doing web scraping. He spent $80 with Claude in one day, and then I showed him Gemini, switched over, did just as good of a job, but cost like five bucks to do what he had done.
No way.
Yeah. So why do you think people sleep on Gemini?
Because almost everybody who I've spoken to has said like Gemini is not even on their radar. Two people have said—one person—one person it was their main thing, and then another person was like, I just use it for the Google suite. So Gemini is actually good at two things.
Okay.
One is it has a really big context window, which means it can consume a lot of information. So, for example, I think it has a million context window compared to others—
Oh, really?
Yeah. Which only can do up to 200,000. So it can do like—so I guess if you translated like, let's say ChatGPT can ingest 100 pages, then Gemini can ingest five times that: 500 pages. Like that's just like a—well, I just know like Claude's context window is at 200,000; ChatGPT is around 120,000. I had no idea that Gemini's was a million.
So, I think it actually can even go higher than a million, but I think like the cost kind of skyrockets after—or like it goes up, and I haven't even—I've never needed above a million. So, I haven't tried it.
Yeah. So, you are an entrepreneur, you're a coder, you've built a company.
That's right.
Which—Tell me just briefly what that does.
Sure. Yeah. So Cash On, it's a—it's a Chrome extension that kind of hooks into Zillow, and usually what real estate investors have to do is they—they kind of browse Zillow, they copy and paste the data—like the price and the mortgage and interest—all that stuff into a spreadsheet, and they run their analysis. So it probably takes about like five minutes per property, but you can only find like—only maybe one in a hundred real estate investments or properties would be a good real estate investment. So what my app does is it will grab all that data for you, analyze it for you, and then tell you which of—which of the properties on the map are the best ones. It takes a process that can take usually months and compress it down into like a week or two. Best success story is just like this guy was looking for real estate investments for like two years, couldn't find anything. I worked with him for—we basically identified a market in two weeks, and then within a month he put in and got two offers accepted on two different properties.
That's amazing.
And it's a Chrome extension.
Yeah, it's a Chrome extension. And I have a free tier that like is pretty generous, and that people can try out.
It's pretty awesome.
Yeah. The cool thing, I didn't realize this, but there are some companies that are like eight, nine-figure companies, and they're just Chrome extensions.
Yeah.
No, I mean, which is insane to me.
Yeah. Chrome extensions, I feel like I don't want to say they're slept on, but they—they—they have a specific kind of purpose and like superpower, which is that they help you pull in data from different websites really well. And yeah, no, I think definitely could probably be a lot more Chrome extensions in the world.
Very true. Now, I'm curious—the money question: How are you utilizing AI? Utilizing it for work? Are you utilizing it personally? Did it build or at least help build Cash On? Like, curious to see your use cases.
So I pretty much use AI like all the time now. I kind of mentioned in the beginning, but yeah, like to learn new topics, to do research, to code. I like use cursor to code, but maybe the thing that I kind of want to show you, which I'm kind of excited about, is I use AI now. I built an AI-powered web scraper.
Yeah.
So this is kind of cool, basically. Okay. There's this concept called like directories that's kind of getting popular right now. Like Greg Eisenberg, John Rush, they're talking about it. The idea is if you consolidate information and put a website up, then people will go to your website, and that traffic can actually generate you quite a bit of money. So I think the most—like here's an example: Have you heard of AllTrails?
Yeah, the app.
The app AllTrails. Actually, I guess that's not a good example because it doesn't aggregate across the internet. But let's take for example Yelp, right? Like Yelp is like—people most often use it as a directory for restaurants. I built—it's called pickleballcourtsnearby.com—and so people like they search it, and then you know there's like—I think the search volume is like 200,000 plus per month, and with that kind of traffic you can generate like easily like in the order of like 5 to 10,000 a month. And then the idea is—wait, so are you making money off of pickleballcourtsnearby.com?
Not yet. I—I'm just starting it up. That's why I built this web scraper. So, the idea is I built a web scraper that will help me scour the web for pickleball courts and tell you information like is it indoor, is it outdoor, is it can it be reserved, does it have lights you can play at night, and a bunch of other stuff. And actually, I can show you what it looks like.
Yeah.
While you're pulling that up, I just found out about a tool. You're going to laugh at me. I'm sorry. It's called Browse AI.
Browse.ai.
Yeah, Browse.ai. So apparently what it does—it's a little bit different—like yours is like a general use case of scraping. This is like, hey, this is the website that I want you to go and scrape specific data for.
Yeah.
But I started thinking about this use case of using AI to go out and scrape information from the internet as opposed to our current methods of going out and—and scraping information. So yeah. So actually the context is I was helping a friend. He was building dog parks near me. And he was basically saying that it was going to take him like 80 hours to kind of gather the data needed to build this website to hopefully make something like 23, $5,000 a month. And that's kind of how I got into it. I was like, "Wait, I can automate this for you, man." And then like it took me about a day or two, but we got it running, and yeah, so I saved him like 80 hours, and he launched his website, and we'll kind of see because SEO is kind of like a long game. So we kind of see like the traffic that you're able to acquire. But yeah, so there's—there's a lot of like potential with these directories, but also just LLM-based web scraping because they can browse like humans and automate like what you're doing and extracting the information you need.
Is it similar to programmatic SEO?
Yes. Yes. So actually, let me—this will be just easier if I share my screen.
Yeah, I'd love to see it. I'm like really new to all this stuff. I mean, I'm trying to learn as much as I can, but programmatic SEO has been really attractive to me because I, like you're saying, it's a long game.
Yeah, SEO is a long game, but programmatic SEO is really interesting because it's a very long game, but you're also—you're like taking advantage of a bunch either different geographies or different verticals. So, I've been thinking about AI use cases around programmatic SEO.
Yeah. So, this I guess is like a pretty good example. So this is my website: pickleballcourtsnearby.com.
Oh my gosh.
The programmatic is like, you know, for every state and for every city. So this is California Irvine.
Mhm.
There will be like a page, and this is the kind of data extracted. So like this is actually Denver, Colorado.
Yeah.
Yeah. You can—the idea is you go to whatever city, and then there's like a URL that's already been created, right? So this is Texas Dallas, and then all the data is here. And so then the idea is when you go to Google search and you know like you search pickleball courts near me, this will say pickleball in Dallas, Texas. Okay, I like—update the SEO and text—AI generated this. But so how are you using AI to generate the information for these websites?
Okay, so that I can show you. Can you still see my screen?
Yeah.
Does it look like a table to you?
Okay.
Yep. So what I first did is I use this platform called Outscraper.
Speak my language, bro. I know Scraper.
Okay, awesome. Yeah. So, I scrape Google Maps for all the locations that—that have pickleball in the name or in like the description, and then that gives me a list of all the locations, and it also usually comes with a URL, and so that's actually—let's just start here. So, it'll come with—okay, not exactly what this table is, but it'll kind of look like this where you know you'll have the name of the pickleball court and the URL for like the—and where are you—sorry, where is this table? I can't quite read it. It's a little bit small on my screen. Let me see if I can make it—just for me. Just for me, it's small on my screen. But okay. Is this like an Excel doc, or is this—what platform is this in?
So since I code, this is in my database, but usually it's given to you as a CSV that you can open up in your Excel sheet.
Yeah. I didn't know if you'd put this in like an Air Table or something, but this is in your coding database. Probably could work with Air Table, too. But yeah, this is in my own database because I'm trying to build like a whole website and like—
Sure.
Yeah. But the idea is you get the pickleball court or the location that has it, you know, like these sports complexes, and then you get the URL of these pickleball court locations, and then what I do is I use this open-source software called Firecrawl, and what it'll do is I send it to this URL. So let's say activityreg.com, and I say, okay, I need—and let me show you the information I grab. So, I say, "Okay, I want pickleball courts, and I want it to look like this." So, I want the URL. I want the email and the contacts. Uh, I want to know the phone number. I want to know the address. So, this one has a bit of more like address and stuff. And so, I gather all this data—or so I tell it I want this data. So, then what it does is it goes to the website, and then it crawls it until it finds the data needed, and then it gives it back to me in the format I need to show on my website.
So the software you're using is that AI?
Yes. So it's AI-enabled. Okay. Previously, if you were trying to scrape, you'd have to hardcode everything. Like, for example, you could only like go to ebay.com and say like, okay, for this product, this is where you get the email, and this is where you get the address or like whatever. But now with AI, it can—no matter where it is on the page or even if it's on like a different page in the website, it can decide like which one to go to. So usually actually what happens is it goes to a website, and then I get it—all the links on that website, and then I'm like, okay, this is the homepage, and these are the 100 other links on this website, and actually you can see it in here—Firecrawl maps. So—okay, again, I'm really stupid, and I'm just trying to like—yeah, process through this. The reason Outscraper is able to get all the information that it does from Google is because Google probably has some standardized way of gathering and displaying information.
So yes, because—because Scraper had all this information before AI, right? And they were able to do that because it was predictable. You could hardcode and pull all this stuff.
Exactly. Now what you can do with that website information using context is say like, go find me the phone number or whatever.
Exactly. And you don't have to tell it where it is, but it knows what a phone number looks like, so it's going to go through and like make a decision.
Oh man.
Yeah. How cool.
Exactly. Actually, yeah. So, you nailed it. Basically, since people—they like, you know, they either submit their business. Well, they usually—
Yeah.
They usually submit their business, and they fill out the information, right? And then Outscraper can scrape it from Google Maps because it's always in the same place in the same format. But what happens if it's not in the same place in the same format? Then you need a human to go do that. And so previously like a job like this would have taken many people many hours—like months, you know, to like scrape it—like in my database I have 16,000 pickleball courts—if like—if you say, okay, let's see 16,000, and let's say it takes five minutes per—that's 80,000 minutes, so that's 1,300 hours, so that would be like half of a working year—it's more than half of a working year—exactly—and—
So yeah, with AI, I kind of just let it run in the background. It'll take me about 7 days, I'd say, but I don't have to do anything.
Okay.
And actually, what I've noticed is it actually does a lot better than a human, too, because it's just like—with AI, you know, you're paying—you're basically getting like a college-level graduate, and all you're doing is like pulling data out. So, it's not hard, right? It's just if you read it, and then if it exists on the page, the AI can definitely grab it. Whereas a human might miss it or a human might be like, it's good enough.
Exactly. That—that phone number might work, and they're just trying—they're just trying to get to the next thing as opposed to getting it right.
Have you messed with Operator?
I have. I have. So, you know how slow Operator is, right? When it's like—it's going and executing a task, and it—it's slower than the human would actually do it. Is it the same with this technology? Is it—is it just slow? Is that why it takes seven days, or is it just a massive data set?
It's just a massive data set.
Hey, I don't know if you remember this, but when we started this podcast, we entered into a social contract. I would spend time, energy, and money producing this podcast, interviewing these individuals, and giving you insights into how to build, buy, start, grow your business, and you would like, subscribe, and leave me a five-star review. Now, out of that, we both get to talk to really cool people and hear really cool insights. We both get a ton of value, but I just want to help you keep your word. So, would you do me a favor? Will you go leave a five-star review for me on Apple or Spotify? It would really help, and if you want, even share this with a friend.
So I think with Operator, it's—I think it will get faster over time, but it's—it's still slower now because it has to be kind of a generalized use case. Yeah. But I also bet Operator is like in a sandbox where they like—they don't want to give it total reign to do whatever. But actually my guess is it actually could run faster, but it would cost OpenAI more. So I think they purposely rail—
Oh, that makes sense.
That's my guess. But yeah, so since like for me I just need to get the job done as fast as possible, I can let it run for as fast as possible. I'm actually more limited by the computer I run it on, and so I could technically like, you know, rent a computer in the cloud. And how much coding does it take to do what you're doing? Like if I wanted to go—not scrape all the pickleball courts, I wanted to go scrape all of the cupcake businesses in the United States. You know what I mean?
Yeah.
Okay. Okay, I know the software now, Firecrawl, like could I just go do this, or like what would it actually take for me to do it?
So, Firecrawl handles the scraping for you and gives the—the website in a format that the LLM can do. And if—I guess if you were going for like an MVP, let's say, you know, like I just want to grab the data only for cupcakes. Don't need a generalized kind of scraper. You could probably get that done if you're an engineer, a software engineer. You could probably get that done in like a day, a day or two. If you didn't—kind of depends on how good your prompting is, but I think you can still get it done. And I would say like, you know, a couple days.
Yeah.
I think—well, I think the thing with prompting from a non-engineer's perspective is just like you could probably try everything you could within like three days, and then if you—it's more like either you'll get it done in three days or you're not going to be able to get it done. Like you don't know how to prompt it in the correct way. I feel like—
Oh, I see what you're saying.
Yeah. I'm just thinking for a business owner who's like, let's say I have—I have a PE firm, and I'm rolling up cupcake shops, and I want to know what—who all the cupcake shops are that are owned by mom and pops, and it's like, oh, okay. I went and got all the information from Google, but now I want you to find the specific owner. I—I don't know, some use case where you're scraping data. Like, how would I do that? Would I just go to Firecrawl and say this is what I want, or do I have to build an app essentially to run Firecrawl on?
I think specifically for this use case there isn't a product that kind of exists yet that can do end to end. I'm actually—I am planning to release this as like kind of a paid product, but it will cost a bit of money cuz—just for reference—so for the pickleball data set—that will cost me like at least $150. I'm only halfway done. And then it also cost me seven days of running my own computer.
But that's nothing.
Yeah. No, it's nothing compared to like if you had hired someone like a VA for like 10, 20 bucks, like 20 bucks an hour, cuz that would take them 55 days, which is like—be way more money. But yeah, I think for someone just trying to get it done, I would say give it a shot. Like I think with AI, you definitely have a good chance at getting it done. This is just maybe on a little bit more on the technical side, and I'm not—I'm not 100% confident, but I feel like you get as a business owner without any coding knowledge. I think you get pretty far.
No, it's—it's weird because this last week like I had never—
Done anything even remotely technical? And in the last week I've liked JSON; I know what it is. I don't necessarily know like the definition behind it, but I know what JSON is. I've written a Python script in Google Collab, which I—I was then able to like take my uh files and condense them, and I—I—I was able to get my transcripts cleaned up, and I was like, "This is freaking amazing." I don't know what I did, but I know that AI wrote it and it gave me the output that I wanted. Do you know what I mean?
So, have you heard of Boltne and Lovable—and that Lovable?
Yes. Yes, I've heard of Lovable, but Boltne, very similar to Lovable, products that will help you build websites without any coding knowledge. But my physical therapist friend, he's been trying to pitch me on this idea for a while to help like people kind of self-diagnose themselves. And I was like, "All right, like I kind of gave him a couple steps to do beforehand before I'm like, 'Okay, before we can work on this together.'" But he was like, "No, I don't want to do all that." So he just went on—I—he was like, "I want to just do this myself." So I was like, "Okay, if you want to do it, try using Lovable." He used it and he built like a product that, without AI, probably would have taken me like a solid month; with AI, probably like uh—like a couple days, but he was able to build like 80% of a working website with just like a couple bugs to help people self-diagnose themselves, like for physical therapy—like kind of pains.
No way. And that blew my mind because it really—he's had like a bunch of ideas over the years and never been able to pursue them, and now he can. And so yeah, I mean how many people—I—I know a lot of people who are like, "Dude, I have this idea for an app." Like what would you think if—and they tell you the app and it's like, yeah, you're never going to do that because you would need an engineer or coder, somebody to create an MVP for you, but you can actually—you can actually do that now. I know it's crazy, actually. It's like you guys—beyond an MVP now. I would say like there's like small kinks here and there, but it's definitely working. It definitely gets the job done.
What's the difference between Bolt and Lovable? I think they do the same thing. I think they're just competitors. What about Replet? Bolt and Lovable. I haven't actually tried Replet. I've only seen videos of it. I think it falls into a similar category. My guess though is that Replet would be a little bit more powerful because—
What about cursor? Cursor is definitely like—I think okay, so if I put on a spectrum for software engineers, for people that don't know how to write any code—level and Bolt are on the side of not for people who don't know how to code; Cursor definitely on the side for people that do know how to code, and I think Replet's in the middle, I think, because I feel like with Lovable, like I—I—I kind of had to help my friend debug a bit. It didn't let me edit anything in the UI; it—it—it's really designed for people who just cannot code and just want the AI to do everything. But with Cursor on the other hand, like you need to know how to like set up your environment. You need to like know like what to tell Cursor, but it's way more powerful.
If I wanted to build an app, would I—would you recommend I try Lovable or Bolt? I would just say Lovable because—and now I actually have experience with it, but from everything I've heard, they're equivalent products, or like they're—
I want to try it. I want like next week I want to try to build something.
Yeah, dude. I think you're gonna kind of be blown away, I'd say. So, yeah, another friend had like another idea that he's been like kind of trying to pitch me to build like this like Reddit analyzer, and then he just built it without me cuz he didn't need me anymore, and it was like this is awesome, right? Like it's hard to say no to your friends, but yeah, now that they can go do it themselves, it's awesome.
Okay. If I were to build some product for myself,
Yeah.
Is there—are there like some parameters you would give me or some framework you would give me to identify like—yeah, like what's a good—what would be a good use case?
Yeah. Okay. The best projects to start are app ideas where you don't need to aggregate data from across the internet. Let's say these calorie tracker apps, for example, where like the user can like input the data themselves, and you can kind of like do everything within the app. Then those would do well. But the moment you kind of need to reference other data sources, you're kind of getting into a part of software engineering that is harder because you need to rely on other services. So anything you can build that's like completely wholly encapsulated in your app.
Okay. So I have like 130 podcast episodes, and—and let's just say that of the 130 podcast episodes, it distills my philosophy on many things. That would be a good use case for building sort of a closed system app, right? I don't have to—it's not going to go out and pull a bunch of data, but it's the Nick Deal analyzer, you know, for—for all intents and purposes.
Is the idea to like kind of advise like if it's something is a good podcast idea or not or—
Yeah, it could be like, oh, I'll give advice on your business if you're trying to scale or if you're—if—or, hey, I want to buy this business. What do you think of the multiple? Or hey, I want to start a business. What do you think of this idea? Potentially, it just gives you feedback, right? But I guess that's just a chatbot, right?
That would be a chatbot, but you would be surprised how helpful that is. I feel like in this kind of age where you can build anything, you still need a lot of knowledge to do things well. So, I'm working with like a—like a bit of like—he's an—he has a YouTube channel and he helps with like advice on like programmatic SEO and like trying to decide which websites to build, but he has an a lot of knowledge in his head that would take a long time for you to kind of comb through YouTube videos, a lot of articles to kind of understand. So, I'm actually building something like a chatbot, but basically you give it an idea, and then it'll just ask all the questions that he would ask and then assess your idea, and you know he doesn't have the time to meet with 100 people, but now this chatbot can.
Right. Right. So I still think that's really useful. So how would I do it? I want—I want to like walk through this cuz my problem is that the context window, even if it's a million tokens, is—call it 20 podcast conversations, and I've got 130 of them.
Yeah, that'd be probably not feasible. So I would say—I would say—I would start with—and this is kind of how I'm approaching it—is you—you probably have a framework to analyzing if a business is going to be a good idea or not. Like probably like what's the monetization model? How are you going to do distribution? Like all these things and then also you can like then assess like if that monetization model will actually work out or not. I'd say take your framework—the way that you would like define a framework if you don't have one already. But you take your framework and encode that into a chatbot—some—
Or I could take 10 episodes that I really liked.
Yeah.
And like clean the data, get rid of the timestamps, get rid of like ums, o's, get it down to a point where I can fit it into a context window.
Yeah.
And then just say like, "Hey, help me codify my deal analysis." So that's actually what I did for my friend that I'm trying to build like a chatbot for. I took like five of his YouTube videos. I threw it in ChatGPT. I threw the transcript in. There's like—there—these free services that like YouTube to transcript. I threw it in and then I was like, "Okay, tell me all the best practices and tips that I need to follow when trying to choose a new SEO website to build." And then I spit out like pretty much like everything you said. It was like 20 to 30 things. I then encoded those 20 to 30 things into questions and like criteria to analyze. And then that's how I made the chatbot. And like—
Yeah.
So you could apply that same thinking. Give an AI your audio files. Say like what are the most important things to analyze and how do you analyze it? What are the criteria to make something good or not good? It needs to be improved.
Yeah.
See what it spits out and then codify that into the app.
Yeah, that is really interesting. That would—that would be really—really interesting. I love that. This was amazing. This was—
Yeah, this was my last call of the week and definitely of the 30. This is like—this is a top three. This is a top three. Very, very good, man. All right. Hopefully you like that episode. And if you've made it this far, you're either really committed or you're stuck doing yard work and you can't actually skip on your phone. So, while I have you, the show is growing, but I—I have a favor to ask of you. Will you please help me grow the show? I want to reach more people. There's a couple things that you can do. Like and subscribe is the simplest thing. Obviously, you want to get notifications for when the next episode is coming out. But if you go the next step, will you leave me a review—five star on Spotify or Apple? What that does is it tells the algorithm that, oh, hey, this is a high-value podcast because more people are leaving reviews for it. And it then pushes it out to more people. So that's why when people are like, will you like and subscribe and put a fivestar rating? It's not just to make themselves feel better; it's actually to get more exposure for the show. So if you do that for me, I would greatly appreciate it. And uh—I'll see you next