What AI Tools Are Worth Your Time in 2025⏐Holdco Bros. Ep. #158

Over the last two weeks, I've talked to 30 entrepreneurs who are using AI in their business. Because of that, I now understand the use cases that they're using it for. ChatGPT as a general thought partner, Claude, coding partner, Perplexity Research. Everyone I spoke to who actually uses these models, they're like, it's as good as deep research and it's real time. Gemini workflows within the Google suite. There's a ton of good ideas, not a lot of good execution.

Stay close to the money. Build something that people can use to directly make more money that will have the highest likelihood of succeeding.

The number one use case for it right now.

You know what I love about ChatGPT? I just realized this, and you might have had a different experience. And it may change because incentives aren't aligned. It is not biased. Like, if I ask it about Republicans or Democrats or the Mormon Church or the Catholic Church, or seed oils or health or carbs versus fats versus proteins, I really feel like it's unbiased. Like it is a superpower. So here's one example. I saw this Instagram reel, and do you know what ghee is? G H E E. No. It's like a clarified butter. It's like a different version of butter that we've been using for hundreds of years. It's whole food, it's healthy. It's like an alternative to seed oils or canola oil or butter. So I saw this stupid Instagram reel that was saying basically, ghee is unhealthy actually, and canola oil is healthier than ghee. And I'm like, oh, my gosh, what are we doing? Like, there's so much info out there that, like, literally no one knows what to eat because you don't know who to believe, because you don't know who has the incentive to say whatever. So I went to ChatGPT4O and I said, who paid for the study that says replacing ghee with canola oil is better? So this is what I said. Who? Who paid for this? Right? And it's like, great question. Here's the deal with most of those canola is better than ghee studies. Many are funded by big food or ag companies. Example studies in the past that have been backed by groups like the Canola Council of Canada, American Heart association, or Unilever. Like, literally, like Coke and Pepsi are funding the American Heart association, right? So then they put out these studies that are like, Coke is actually good for you. It's like. But like, ChatGPT has no bias because there's nothing to click. The incentive is there. Whereas if I Google the same thing, like the incentives are for me to click on stuff. So I still can't believe what I Google. Right. As optimistic as I am, I don't think that ChatGPT will stay like this forever because money. Right. But I love that it's like this for now. Have you had the same experience?

I have, but a little bit different. So I think it's pretty unbiased when I'm asking it semi factual questions. But I do think it's biased when I'm asking it for feedback. And this is either it's biased or I'm insecure because whenever I ask it for feedback on something that I'm doing or about me, I feel like it's, it's overly praising me, if that makes sense. Where I'm like, hey, take a look at this post I wrote. What do you think this post is? Fire. Fire emoji, by the way. OpenAI. Can we cut the emojis? Like, what's happening? Why?

There's so many freaking emojis. I hate them.

Anyways, this post is fire. And I'm like, how would you rate it? I would rate it a 9.5 out of 10. And it's like, okay. And then I have to go back to it. I'm like, you're a neutral third party. Your job is content management.

Saying what it means it wants you to hear.

Right?

Right. Right.

So to me, that's where I am seeing bias is like, it says what it thinks that I want to hear. And there are times where like, I want it to be totally unbiased and like, give me an actual opinion of what it thinks of my business plan or this post or the argument that I'm making with my lawyer. You know what I mean? And so anyways, that, that's the only piece that I would say. I'm. I do see it biased.

It's funny because I'm. I'm gonna directly contradict myself and it's like, it's not that it's unbiased, it's just biased in a different way. There, there's always incentives. It's impossible to remove incentives. OpenAI's incentive is to keep us using ChatGPT. Right?

Yeah.

So it's like the bias is not clicked on this study. It's like, make sure Nick never churns. Right. We want to keep Nick happy, you know? Right.

I have a custom GPT. Where in the GPT prompt I'm like, you're A hard nosed, no nonsense content specialist. I need you to give me direct advice because bad advice is terrible. I give an analogy about being on the warfield. Warfield. I'm so stupid. Being on the battlefield and you know, field generals need to have not only real time information but accurate information to make good decisions, otherwise bad things happen. So I give it that analogy of like, this is why I need you to give me good information, because I need to make good decisions. And so that GPT is great because it gives me sort of unfiltered responses, but it's when I'm just like trying to ask it a generalized question that it's definitely biased.

I'm kind of debating if I should say this next part or not.

Say it. Say it.

I don't think I want to. I don't think I want to. Relationship advice. Have you like with your spouse, have you used it for that?

No, but like, what did you use it for, dude? Like I'm going to say this. You and I have talked about, we each have superpowers. When you have a superpower, you have to know when to use it and when not to use it. And you also have to recognize that you can use your superpower for evil. To me, the reasoning that ChatGPT has, like, let's say you have an argument with your spouse and you're trying to get your way where it can help you frame an argument is out of this world. And then it's a very unfair fight.

Because you're not citing to your wife. ChatGPT according to. No, you're not doing it.

No, you're not citing. You're like, well, that's an interesting point, honey, because I would think that you would want to do X, Y and Z based on your previous. You know what I mean? And she's like, yeah, that's a really good point. And meanwhile you're like looking down at your phone, you're like, she said it was a good point. What do I do now? What? This has never happened. She said I'm smart.

What's happening, dude? I have been very impressed with how ChatGPT is handled, like relationship advice with my spouse. And I've, I've never, we've never seen like a marriage counselor or a therapist or really talk to anyone like, we've got a good marriage, right? But I've just been asking it some things and it has been so helpful and it, I feel like it's advice that like a very expensive, very, very good marriage counselor would give. And it's literally Made our marriage better. And she has no idea. She has no idea.

But it's the perfect use case for it, Right? And if you're like. If you're trying to come at it from a genuine way and you're like, look, I'm going to try to give you all the facts that I can. Help me understand it, then you're getting like this impartial third party, and you're like, well, it's A.I. so, like, if A.I. thinks I'm being a prick, yeah, I guess I am being a prick about that, Right. Or vice versa. And it's. It's pretty incredible with the way, if you give it enough context, right, because you and I might have a disagreement, I might give it all the context in the world and be like, yeah, Nick, you're right. And then you might have the same disagreement.

Put, it doesn't want Nick to churn. It doesn't want Chris to turn.

Well, it doesn't want. It doesn't want me to turn. But also, I'm just giving it my context.

Yeah.

And then you do the same thing.

Your side of your story, right? Like, if you see biased info, it's going to give you biased info.

Yeah. The trick is I would like to start using. Figuring out a way to use this for mediation in business disputes, I think would be really fascinating if everybody just had. If everybody just had the Limitless, which I ordered, by the way.

Oh, awesome.

And it's like, it's like, okay, upload all your limitless data. Now, you'd have to have two parties who are like, okay with everything being uploaded. But that in and of itself would be a tell, right? Where you're like, oh, this guy doesn't want it uploaded. What's going on? But it's like, all right, take all of the side of their story, enter all the information, because that's essentially what an arbitrator is doing anyways.

Yeah, that's really good. You just create a fresh ChatGPT account and the two people could sit side by side. So here's my side of the story. Here's my side of the story. And then they both agree, like, what? The prompt is right, and it's neither one of their accounts, and you just hit enter and see what happens. Now, the hard part would probably be agreeing what the prompt is, right?

Yeah. And just like any other mediation, the hard part would be actually agreeing on the output. Because even if we're in a mediation and it sides with you, I'm going to be like, what? Well, it didn't do.

That's stupid. It doesn't know what it's talking about.

Yeah. But I do think if you're a mediator and your business is mediation for those types of businesses, your bottleneck is you. It's your time, essentially, because you're an agency and you can only bill as much as you can actually go out and meet with clients. If you could use ChatGPT to build a wrapper or to build a prompting system to get that information quicker so that you save an hour per mediation, like, then all of a sudden, you can take on 20 more clients a month. Like, you just. You just grew your business in a way that's AI enabled. That would be a really, really cool use case.

That is cool.

You and I talked about this idea, I don't know, a couple weeks ago, where it's just like, buy limitless pendant and give it to somebody almost like a lead magnet to get into their business. And you're already doing this because you're recording everything. At some point, the ability of these models to digest more data is going to happen. So what becomes the limiting factor in how well these models will get to know you and be your own personal assistant? The limiting factor will be data, context, windows, compute, power. Yeah. Or the ability to process or. You know what I mean? Or the ability. Because what's happening right now is it may have this massive data set, but it's optimizing for speed, and so it's only looking for certain. Whatever. At some point, it's going to get way better. And the limiting factor is going to be the amount of data that you have to input it into the system for it to then tailor itself to you. So I don't have a business model off of this. I just think using ChatGPT, using a limitless pendant, where you're recording every conversation that you have, it just creates more and more and more data so that when it's there, it's fed in and you're accelerated even faster.

Yeah. Can we barbell this for a minute and go offline? Yeah. So you see my tweet about the podcast billboard.

Yeah.

Putting up. So that's been a lot of fun. It won't go live for like, two or three weeks, but I used AI to generate a bunch of phrases. Right. So you saw that I shared my screen. I put the podcast art on the left, Kerner office in the middle, icons to YouTube, Spotify, Apple Podcasts, and then like a catchy slogan at the top. So I had Simon make 34 of these. They all look the exact same, except for the catchy slogan at the top is all different. Two to five words. Like, you only have so much time to read a billboard. Right? And so the catchy slogans that I'm testing are, your dream is dying. Sell stuff, not ours. Nine to fives, kill dreams, own your time. Never job again. Jobs or jail. Delete LinkedIn, work sucks, question mark. Escape the grind. Fire your boss. Still rat racing, question mark. And what was the last one? Then there was best business podcast, and then quit your job with an exclamation point. And then there's a bunch of others. But I started Facebook ads that are all the exact same. It's just like the Kerner Office podcast. Title, text description, they all say the same thing. The Kerner Office podcast, the audience, the four counties in and around where the billboard will be. And then I exclude anyone that's ever interacted with my content so they don't follow me, they've never seen my video, and they live in those counties because that's biased. The people driving by don't know who I am. Right, right, right, right. And then I just run the ads on any meta platform, meta things, Messenger, Marketplace, Instagram, it chooses. And then I just look at the clicks. Lowest cost per click wins. And then I get, because it's 950 bucks to print that billboard. The image that I'm using on, on the Facebook ad is literally a picture of the billboard with like trees in the background instead of the image, you know, and it's 18 plus any gender in those counties that don't know who I am. What do you think?

I think it's amazing. I have an idea for it as well. I know we're offline, but I'm going to take it back to AI.

I used AI to generate most of those phrases.

By the way, what I like to do with AI when I'm like, creating a title for a video is I'll be like, generate me the 10 best titles and rank them and then I'll like iterate from there. I would take all of your pictures and I would go to ChatGPT, Grok, Claude, Gemini, four of them, upload them, and say, rank these from what you think would be most to least effective. See what the four models spit out, and then compare it to the actual.

Results from Facebook ads.

Yeah, that would be interesting.

I'm just gonna do it with the phrases because all the pictures are the same.

Oh, okay. Yeah, yeah. So, like, I think that would be a really, really cool case study of, like, all right, I'm trying it with real humans. This AI has been trained on the totality of human whatever experience. Which one of them wins or which. How accurate is it? Because then in the future, if it's like, oh, they all said the same five, and those were actually the five that won, it might save you time running Facebook ads in the future.

Yeah, I like that. It's interesting. I didn't exclude my audience at first, and I started running the ads, and the winning one was Best Business Podcast. And that surprised me. And I was like, I wonder if that's because, like, they recognize me, they see me, right? And I don't yet know it's still too early, but it just made me think, like, how much power there is in marketing to call your thing the best. I had Greg Eisenberg on a couple weeks ago. His story was he was walking around Miami. He had just moved there. He walks by a dentist, and he sees a sign on the window that says, voted Best dentist by who? What? I don't know. I don't care. He walks in because he saw that sign, became a customer, been a customer for eight years, given them tens of thousands of dollars, referred 20 friends there because he saw the sign. So his idea was like, there needs to be some sort of, like, a marketplace where you could get, like, certified by whoever. It's a racket. It doesn't matter. You pay 200 bucks, you get a sign. And so it's like, I just put up this billboard. Yeah, I think it's the best. My mom thinks it's the best. Whatever. Best Business Podcast according to Chris's mom. Find print asterisk. Right? It doesn't matter. So maybe people click that the most because they knew who I was, and I'm going to find out that's actually bad data. Maybe not. Maybe it's the winner again, and people just. It's confirmation bias. They're like, I believe you, Internet stranger. I will now click and listen because you said it's the best.

Hearing you talk about Greg Eisenberg makes me think of that meme that's like Internet guy discovers rating agency. It's like these. These. These things exist, right? Like, Best business. Like, you can talk. They're already businesses in and of themselves. But, yeah, this cracks me up.

Well, there needs to be, like, for SaaS companies or for AI rappers. Like, you could do a rating agency for any number of things, right?

Yeah, I mean, that's what no BS crypto was when you were launching it, right? It's like, hey, this is a rating agency. I don't know. Looks like we made it. I just like that idea a lot because you're like, you're currently competing all of the tools, all of the advice. If you go and look at, hey, how do I grow a podcast? Is do it on someone else's newsletter. Go do paid ads, go to other podcasts. Brokers, go to other podcasts.

Right.

It's like, okay, well, everybody else is doing the same thing. They're all doing the same exact thing. And that we know that's a strategy. Zig while everybody's zagging.

Yeah. So, like, either this is going to be a total waste of money and I'm prepared for that to happen because I don't see other podcast billboards probably for a reason, or I'm going to unlock something really cool. Right, Right. Aren't you going to ask me how I'm going to track if I'm getting new downloads from this? Nick, how are you going to track softball? Me? Here, set me up.

Chris, let me ask you a question. Like, I get it, but how do you freaking track that? It's on a billboard, probably with like.

A QR code, right? No, QR codes are not allowed by Clear Channel, the company I'm using for this. Okay. Okay. So Buzzsprout is the hosting platform that we use. You can export your downloads by location by every little city. So, like, when I export mine, there's like 8,500 cities that have listened to my podcast. Sometimes once, sometimes a thousand times. Right. Now, what you don't know is how many unique listeners there are per city. You just know how many downloads. So 25 downloads could mean one person downloading your podcast. 25 episodes of it could mean 25 different people once each. Right. You don't know. So what I did was I went to the bail board will be in Tomball, Texas, which is a city of 10,000 just north of Houston. I took Tomball, the Woodlands, Spring, all the cities, and Houston itself in and around Tomball. And I exported every city by download. And then I found a percentage, like a ratio. How many downloads have I had all time that were in Tomball and or Houston? 0.012% of my downloads. And that's been consistent. And so I'm just going to run that export every week and see the raw number and the percentage as a ratio and see if it's climbing or not. Do you see any. Any fault in that?

I do see a couple of faults. Like, you don't necessarily know where it is as a percentage over time that as it's mapped.

Well, I'm going to run the export every week because that's the really the only way with Buzzsprout you can do it.

The other problem would be it's just all attribution related.

Right.

Like you're looking at this geography. Did you capture the right cities? What I would do, this is just a suggestion, is I would ask ChatGPT the best way to track this and the best way to analyze it. You can throw data into ChatGPT and then like ask it to run a regression model. And the regression model will tell you whether or not that's statistically significant or not. Because you could just have, you could just blow up over the next month. Let's say you have a couple episodes that are amazing and it's like everybody saw a bump. Right. So of course Turnbull would see a bump.

It would be proportional in that sense. Right. In that instance, I would have to look at the percentage, not the raw number.

You have so many downloads, you have so much data that it would be easy to say with confidence whether or not it's statistically significant. And I would just ask ChatGPT.

No matter what I do, it won't be exact.

No. But I think it's amazing.

It would be too coincidental if a disproportionate amount of people in that part of Houston suddenly started downloading my podcast. Right. Right. What I wish I knew is how many, because you just don't know how many raw people it is. How many loyal listeners. Because there could be 100 people that listen to five seconds of it once and then bounce. And it could look really good, but not actually be good.

Why didn't you do the here's the website and then have like one letter off?

Right. I just don't want to mess with like, URLs. I. I kind of want to go more for brand awareness. That's why I put the image of the podcast and one of them has tkopod.com in it. I just want people to search for it. Another thing you can do is you can put up Facebook ads based on a latitude and longitude point of the exact billboard. And it will show your ad to people that drive by that billboard. And so then you get double exposure, triple exposure, the billboard even more effective. But it will make it even harder to find the attribution because you're adding another variable. Right.

At that point it's like, okay, I'll spend $780 for a billboard and I'll run $1,000 worth of Facebook and you know, Facebook ads. Who cares about the attribution? I know it works. You know what I mean?

Yeah. Because any Joe Schmoe can run a Facebook ad, but you don't see a lot of billboards. It just adds another layer of legitimacy to it.

Right? Yeah. And that feeling of like, dude, I'm seeing this guy everywhere.

Right. I kind of had the thought, like, wouldn't it be interesting if one of us, anyone ran a test where they took like a random city in the middle of nowhere, like Sioux Falls, Iowa, and they ran Facebook ads like saturated for your podcast, your brand, whatever, just in that city. And like, you just became a freaking rock star in that one city and you just put them Everywhere. Instagram ads, TikTok ads everywhere. Only geo fenced to that city. And just, I don't know, go speak, have an event there. Like, it'd be an interesting experiment. Like no one else in the world knows who you are except for Sioux Falls.

What if you went to ChatGPT and you said, I want you to run a contagion analysis or model only of cities less than a million people in the United States and tell me which city would have the highest likelihood of spreading this contagion the fastest. So it's like, you know, it's going to go look at Des Moines. Like, well, Des Moines actually has the most connecting flights of any airport in the United States.

So I don't know.

Right.

I'm just like, yeah, yeah, it's gonna.

And then it's like, okay, cool, I'm gonna go target Grand Junction, Colorado, because it says that of any city under a million people, that's the one that would have the like worst contagion. And that's the city I'm gonna go target.

So like Vegas, like a city or Orlando, where people are coming in and out of all the time.

Right.

Like New York City has 20 million people anyway. Right. People are coming in and out, but proportionally speaking, there's just a lot of people there.

But for New York City, it's like you're. Or LA or San Francisco. There's just a lot of competitors. So standing out is gonna be hard, generally speaking.

Right.

But a smaller market like Orlando, I don't know, you probably have a better chance of standing out. That'd be a cool idea.

So video is cool, but you know what's better? Long form audio via podcast and my newsletter, tkopod.com go there to subscribe for free to my newsletter. It's one email a week Very tactical. And then go to my audio podcast, three episodes a week, stuff like this. You're gonna love it. All free. No sleazy sales pitch. Tkopod.com so over the last two weeks.

I've talked to 30 entrepreneurs who are using AI in their business because I feel like you and I have talked about this. There's a ton of good ideas, not a lot of good execution. Like Replit. How many things have you made on Replit that you haven't actually finished?

7.

So it's like, oh, this is cool. I can build an app. This is great. And then you get to the final point and you're like, oh, how do I actually complete it? It's like, that's that last 10%. So I've been having conversations with people and it's been awesome. I've learned so much. My brain hurts every single day because I'm learning so much. But there are some really cool insights that I've gotten out of this. So I'm going to share some of my findings with you because I had ChatGPT take all my findings and put them into summary for me. Question number one, Chris. What do you think is the most popular LLM for people to use as their primary LLM?

ChatGPT it is.

What do you think is the most popular use case? Like, what are they using ChatGPT for?

Writing emails, like rewording emails, mainly they're.

Using it as a thought partner. So it could be rewriting emails generally. Not that it's usually like, hey, this is. I have an idea for this how would I do X, Y and Z thing? And it's like, just to get them started, which was fascinating to me. Second most used LLM. What do you think it is?

Gemini. No. Grok?

Nope.

Perplexity?

Nope.

Claude. Claude. Dang it.

Claude. And the surprising thing about Claude is it's actually used more in aggregate because I asked them two questions, like, what's the primary? Like, what's your primary favorite? And then what are the other LLMs you use? And I mapped it all out. Claude is actually used more by more people than ChatGPT is, but ChatGPT is the most preferred, if that makes sense. Okay, what do you think that they're using Claude for?

I have no idea. That's like the only one I really hardly ever use.

So I have it broken down in three tiers. Non technical, technical, intermediate, and then expert. For ChatGPT, the majority of the distribution is in the non technical and intermediate. For Claude, the majority of the distribution is in expert and intermediate. And the main use case that I see people use it for is coding. It's the best coding tool, like, to help them actually write code. And then the second best thing, this one blew my mind. Its organization structure is fantastic. So you know how ChatGPT has projects?

Mm.

Claude, you can set up projects, but it. You have two fields. One is, what is this project about? And so you can write a prompt to say like, this is my content project. And then it has files that you can upload for context.

How's that different from ChatGPT?

From what I've seen, ChatGPT doesn't have the same functionality with the uploading of files and the reference for each one of the threads. Maybe I'm just not using ChatGPT.

That's how I use it. So you think you upload instructions? They call it files and instructions.

And so you use individual files for each one of your projects?

Yeah, I don't always use files, but it can.

One of the people I spoke to, she's a writer and she's like, claude does a much better. It's a much better writer, apparently. And she has this whole setup where it's like, this is my brand voice. Here are recent projects that I've done. And so she's like constantly changing the context so that it's up to date. But the interesting thing is, I didn't know this. You can only upload however big the context window is. Does that make sense what I'm saying?

Yeah.

So that was interesting. What was the least mentioned tool?

Well, I would say Grok, but I think that a lot of these calls came from people on Twitter. So maybe Gemini.

Gemini, yeah, it was the least mentioned. And everybody, like, dogged on it. They're like, gemini, it's not good. So that was just really interesting to me.

That's interesting. Are they dogging on it because they've actually given it a fair shake and they don't like it, or they're just. It kind of has that brand around it because I dog on it, but I don't really use it, so it could be great.

They're dogging on it. Not that it sucks. They're dogging on it because it's not as good as what else they use, if that makes sense. They're like the Claude model, ChatGPT. And I think ChatGPT just has a head start because so many people were using it. The two use cases that I found out about Gemini are if you're trying to create workflows within the Google Suite, Gemini is the best because it just knows Google so well.

Makes sense.

I thought that was really interesting. And then the second thing that I thought was really interesting is the context window. I didn't realize this, but do you know how big the context window is for Gemini?

I looked this up a while ago, but I can't remember. I think Claude was the biggest that I saw.

Let's back up. The context window is essentially how much information you can give the model for context of a specific question that you're asking it. For ChatGPT, it's 120,000 tokens. And what that is, it's like a 300 page book. Okay. For Claude, it's 200,000 tokens. So it's a lot bigger than ChatGPT. For Gemini, it's a million.

Wow.

Yeah, so it's a much larger context window. I thought that was just really, really interesting. Next finding. What do you think was the most used tool that interfaced with the LLMs.

That they worked with, like feature of the specific LLM?

No, let me ask it a different way. What do you think is the most used AI tool that was not an LLM?

Something that I would have heard of? Yeah, I have no idea.

Cursor.

Okay.

Cursor. And from several of them that like multiple people said, I think Cursor is going to be bigger than Google. They were like so. Yeah, they were just like so in love.

You must have talked to some coders then.

I did, I did, I talked to a couple of coders. It blew my mind though. Like on the call with them, they were on cursor and they were showing me you can natural language prompt into building an actual website. All of them who had like actual coding knowledge crapped on Replit, which was really interesting to me. They're like replit well. But I'm like, yeah, in order to use Cursor, you actually kind of have to have some coding experience. Like you can still, you can use it using natural language, but it would help if you had coding experience.

Here's my push pull framework for, for this Friday. Usability and functionality. That's the push pull of today. Replit is more approachable to more people, but it's less functional as like a coding assistant. Whereas Cursor is harder to use for more people, but it's a more functional coding assistant. And the magic happens like the unicorn status happens when you have like an Airbnb with beautiful design, beautiful user interface, like easy for anyone on the planet to use, yet so freaking functional. It just works. Everything Just works and it looks simple. Like from the outside looking in, you assume that it's very simply built, but on the back end it is just a hairy mess of complexity and it's just really good engineering and design. Right. But most companies, 99% of companies have one or the other. They have good UI or they have good functionality, not both.

I agree. So I'm going to go through generally. Broadly speaking, this is what I found to be the use cases. ChatGPT as a general thought partner, like someone they're collaborating with and like trying to figure out ideas. Claude Coding in particular, it's like that's their coding partner, Perplexity Research. Everyone I spoke to who actually uses these models, they're like, it's as good as Deep Research and it's real time, so it can go and crawl the web in real time. Which fascinating in and of itself. Grok, there was no like best use case, but it just seemed like generally speaking everybody was impressed with Grok. The speed and responsiveness. I don't know if you've messed with Grok. Like it spits answers out very quickly. They're very good, has access to real time data. It's, everybody's like, it's amazing what they've. They've built a, a top tier model in such a short period of time. And then Gemini is just, hey, it's really good for the Google suite. Those are like the general use cases that I'm seeing people actually use. And I thought that was. Those things were fascinating.

Can I give you my personal takeaways from what you just said? I need to be thinking of using Gemini more since I'm so embedded in the Google workspace. Right. I just don't think of it. We've talked about NotebookLM a bunch. I like the contact window size. I just don't like what it spits out. I don't like what it gives me. Right. So who cares how big the context window is if it's not working with it? Well, in my opinion, I'm using Perplexity more and more every day. As of today, I don't see a need to keep my $200 a month pro subscription because Perplexity does Deep Research, which is the only thing I'm using Pro for now. I still like a second layer of redundancy. I still take my prompts, my Deep Research prompts, and I put them in Deep Research and Perplexity at the same time because it does give me different answers and I still kind of like Deep Research a little better. I think it's more thorough. It gives me a longer answer with good, relevant info.

Have you tried Grok's Deep Research?

I've barely even tested it. I can't really speak to it.

It's very, very good. Very good. Surprisingly good.

I should. I want to make a YouTube video where I. I take the same prompts and put them in all three for. Specifically for deep research.

So curse you think you need to use cursor more.

I need to use cursor more and I need to use replit less just because I'm not able to wrap up these projects.

This is one of the things that's been fascinating to me is like, I wanted to learn how to do these things, so I did a project and have you ever heard of. I hate bringing these things up to you because I feel like a high schooler who's in Chemistry 101 or whatever talking to a PhD chemistry, and it's like, have you ever heard of gravity? And they kind of.

You did ask me what a lead magnet was a while back. That was offensive.

Yeah, his offense. Have you ever worked in Google Colab, Nick?

It's a good thing you asked because. No, I haven't. Okay. I haven't.

Do you know what?

No, I don't.

Okay.

Genuinely.

So I have all of these transcripts from my podcasts, and specifically, I just wanted to focus on the AI podcasts. What I wanted to do, and you know how this works, is I want to take all the conversations, aggregate them, and then be able to pull out information. Okay, but how do you do that? The context window's only 120,000 tokens, and, like, four of those transcripts are more than 120,000. It doesn't work. So I needed to clean the transcripts, get rid of all the dead space, ums, oz, timestamps, et cetera, and then get them into a format where they would all be combined. And I had them all saved. Yeah.

What about. Does. Do the transcripts identify who's speaking?

Yes, I kept the speaker.

Okay.

I did keep the speaker because I wanted to know by person and I'll show you the output that I got, but I wanted to know by speaker, who's talking and what are they saying and why. Anyways, I don't know what to do. And what do I do? I literally just go to ChatGPT and I'm like, hey, this is what I want to do. I want to take all these transcripts, I want to combine them, and I want to clean them, and it's like, okay, where are they stored inside Google? Okay, perfect, you're going to go to Google Colab. Great. Open a new project, attach your drive. Great, I did that. And run this script. So Google Colab is an environment where you can run Python script for it to actually execute a function. So I'm like, okay, so it gives me the Python script. First couple times it doesn't work. Then I go to Claude, Claude fixes my Python script and it starts working the first iteration suck, second iteration suck. But by the, you know, a couple hours later, I've got some really, really good data. In fact, I'm going to share my screen with you again. It took time for me to get to this place. Once it was spit out, I took it back to ChatGPT and I was like, hey, please clean up this data for me, make it look presentable, etc. And this is what it spit out. So now I have the episode with the person, okay, interview, name, their business type, the industry, their technical knowledge, what their primary LLM was that they used.

Oh, this is cool.

The main tasks and use cases.

This looks like a lead magnet to me, Nick.

On that motivation for using these LLMs, strengths, weaknesses, the ideal user profile, and then the key insights and like the key insight. I was like, hey, write me a 500 character summary of this conversation. But then I had to do a couple of other things. Then I had to actually go to Zapier and create a zap so that it would interface with OpenAI. And then I learned I could actually just attach the API key. I'm going like, real technical here, but whatever, the API key into Google Colab. So I didn't need Zapier anymore. But because of that I now have this amazing spreadsheet with tons of conversations with entrepreneurs who are actually using AI and I understand the use cases that they're using it for. So now I can go to this and just ask it questions or I don't know, make a quiz that says or tell ChatGPT, which I've been working on. Create a quiz for me. I'm a business owner and I want to know which one of these models I should use. So create a quiz that I can decide which model to use. And it uses this information to create a quiz for somebody so they know which model to use. Dude, it's been just so fascinating to learn from these people and see how much application there actually is, even in the limited state that we currently see it.

That's amazing. I love that you did that because we know that I Hate hopping on calls with people. So you're able to do it and enjoy it and then relay what you learn.

Yeah. The number one use case for it.

Right now, sales for AI in general. Was there a selection bias? Did you happen to talk to a bunch of sales guys or.

No, they were all either owners or coders and they were like, oh, I'm using it for my sales team to follow up with them. I'm using it to track leads. I'm using it as an after hours phone service. I'm using it as a chatbot for customers. I'm using it for my sales team to prepare for calls. It was just all sales related. I thought that was really interesting.

You know what framework that is?

No. What?

When you're launching products or services, stay close to the money, right? Stay close to the things that directly generate revenue. And you can tie it to revenue easily. Right? So let's say you want to start a SaaS, a B2B SaaS app. It's like, oh, let's do HR. Should we do, you know, administrative assistant? It's like, no, like, build something that people can use to directly make more money. And they know that your thing helped them generate more money that will have the highest likelihood of succeeding.

Yeah, yeah, it's a great point. And that's exactly what I found when talking to these people. It's all sales related.

All right, what'd you think? Please share it with a friend and we'll see you next time on the Kerner Office.