256 - AI Is Here and You Need to Prepare (Here's How) with Amanda Orson

0:00We're right on the bubble. Right now, there is a company being born or already exists. So it will be the first single person unicorn, billion dollar valuation company. That's already happening. I'm confident of it. The amount of work that I can get done with just ChatGPT's operator is probably replacing two or three human beings. Imagine how much more efficient and how much better and more robust that tool becomes in six months or a year. The rate of acceleration is crazy. I don't think most of the things that we do in our lives are going to continue to involve manual processes that include humans.

0:28It doesn't make any sense at all. The whole world is moving in this direction.

0:36Seed strapping. Go on. I don't know that I've ever heard that. So seed strapping is basically taking outside capital once. In our case, it's because we came with a team. That's why we took capital at all. We came with an embedded team that had all worked together at a different company for about three years beforehand. So that was a heavy payroll load, for lack of a better way to describe it, we needed to have outside capital to give us runway. Seedstrapping is basically take outside capital to get your initial product off the ground and to get to revenue and then use rapid revenue reinvestment to accelerate your growth and preserve optionality and for the founders to preserve as much of the valuation as you possibly can.

1:15I think this is probably something that if we had been a year later, we probably would not had to raise capital at all because the difference between 2023 and today, 2023, we were still coding almost everything manually with complement. And now I would say something like 50 to 60 % of our code, maybe more than that is completely AI generated. Yeah. That's what I was going to ask is like, has AI allowed you to run more lean? Yeah. I don't anticipate ever needing to have more than, I don't know, wildest dreams, 50 employees, but probably closer to 20. Wow. That's pretty, that's awesome. Because a lot of these companies in the past, they needed the capital, not just to go acquire customers, right?

1:58But they had to pay for the infrastructure. And they were like, okay, well, once we get to 100 million users, that's our break even. And it's like, well, geez, Louise, you're burning cash operating and you're burning cash acquiring customers. And then at 100 million, you're like figuring out the economics, but you're saying you could be cashflow positive. Now, are you cashflow positive now? Or are you like? We're right on the bubble. You are. I would say right now there is a company being born or already exists. So it will be the first single person unicorn, like billion dollar valuation company.

2:27That's already happening. I'm confident of it. The amount of work that I can get done with operator, just chat GPT's operator, which I know that you created great content for is probably replacing two or three human beings. Like just play that out. That's today. Imagine how much more efficient and how much better and more robust that tool becomes in six months or a year. The rate of acceleration is crazy. What's your current revenue? MRR. Is that how you're measuring it? Yeah. 60K. 60K. That's fantastic. When did you launch Navigator? Four weeks ago. Wow. 60 ,000 in MRR in four weeks? Yeah. B2B too.

3:07We have B2B on the other side of it. Was it because of the relationships you already had? I mean, maybe I'm stupid. I just feel like 60 grand MRR is a lot in four weeks. So we already had some business relationships, but also I'm not coming into this completely cold. Like I have to be honest that I'm at this point, I have a network. I kind of know who I'm going to call. I'm not figuring all of this out for the first time. So it's a bit of an unfair advantage, but if you have that, you can go really fast. Yeah. I mean, that's, that's the whole game is find your unfair advantage and just hammer at home every single time.

3:39You know what I mean? Okay. I want to talk to you about operator because I've used operator a ton. I don't use it a ton anymore because for me, it's not practical yet. I have to sit, let me set a little context for those who don't know. ChatGPT has this thing called operator. It's the first iteration of what agents will become. And it's not just you asking it a question and then it responding to you. It's you telling it to perform a task or a function and then it going and executing on that. Problem with operator is like, I have to keep the tab open. I have to be there. It moves slower than a human.

4:10And so I haven't been using it. Are you actually using it for things? Yeah, I'm using it to prove concept before I actually deploy automation against it. What does that mean? Yesterday or the day before, I was saying something like you should test everything manually before actually investing in building that out. So a good example of that is understanding, can you use operator to capture all of the details about a business category in a geography? It's a good example of that. And if operator can do it, then you can infer that a script can probably do it even better. So it's a really good back of the napkin is my best analogy, like back of the napkin way to figure out whether or not something is automatable and then develop a script and then actually automate it.

4:53Oh, that's okay. So it's like a proof of concept for you. That's a really cool. That's a very, very, very freaking interesting framework. Chris and I were talking about this yesterday. We record our podcast on Thursdays and I was saying, I feel like right now operator, or even the way people are talking about agents, they're calling RPA agents, robotic process automation. Yeah, I would agree with that. Where it's like, okay, you're just being very specific with telling it exactly where to go. Agents aren't really here yet. I'm not downplaying agents. I think agents are going to be amazing. But what you just described was, yeah, I'm using operator quickly to validate whether or not I could implement the robotic process automations.

5:30That makes total freaking sense to me. I'm using deep research a ton. That to me is an agent because I'm like, go out, here are the databases that I want you to scrape and look at for this question that I have. This is the format I want you to put it in once you're done. This is the table. And this is the information I ultimately want you to derive. And then it gives me, you know, whatever, 10 page report on what I asked it. So like that to me is more of a more agentic than what operators doing, but. I agree. And I would say the big, so deep research, no question has just eliminated a host of analyst roles everywhere.

6:07When I first added a bunch of marketing information, like analytics tools, analytics outputs into that, and got the best, like, response, I have ever gotten from any analyst at any company I have ever worked with, or ever hired as an outside vendor. I was like, oh, wow. Analysts are screwed. They're just, there's no room for them anymore. So I would say, yeah, deep research. If you want to take really large, meaty data sets and understand what the wheat from chaff is, or if you want to get some kind of insight that you might have missed, it's great. I'm still having a problem, though, with the white space.

6:46I still have to be very directive about what it is I'm looking for. and sometimes you don't know when you're looking at your own data, the best insights are always the thing you were not expecting to find. That white space is still what I'm struggling with. I want to see AI kind of come in and offer up more and more overlooked insights or flag anomalies to me. So first of all, I love diving into this. This is not where I thought we were going to go with this conversation, but for people who are listening, I freaking love the framework of you using operator to validate whether or not you could automate something.

7:18I think that's genius. right? Because the operator is not there yet from an agent's perspective, but to validate whether or not you could actually run a script to automate something, whether it's, oh, I want to generate leads. Could I go and scrape this website and then put it into a database and then use that to, you know, anyways, I think that's, people should take that away from this. Figure out where it's blocked and figure out if you can unblock it. And all of the magic is on the other side of solving that problem. Deep research. I talk about this frequently now. Now, I think people need to think of AI as another employee.

7:50And many times we're like, oh, AI is just going to figure it out. But that's not the case. Like think about when you hire an employee, you've got to onboard them. You've got to train them. You've got to check in with them. You've got to figure out whether or not they understand what they're supposed to be doing. It takes some time to get them up to speed. But then you have this whole cost of payroll that's every single month. There's a significant amount of overhead. But you need to do the same thing with AI. You need to be training it and be very clear and specific with what your asks are. And so my prompts now are like very long.

8:20I try to add as much context as possible because of what you're saying is I want it to give me the insights that I want. I don't want it to just give me some random insights. So I am trying a lot more to be very directive and broad with the context I'm providing in the prompts. Are you using AI to help you create those prompts? I am actually. Yeah. That's the cheat code right there. You use ChatGPT, I'm assuming? I use every... There is nothing that I don't use. But I use different ones for different purposes. Tell me. Okay, so tell me the different purposes. If I want to avoid emojis. No, I'm kidding.

8:56I think that... That's so true. What is it with the emojis on ChatGPT, man? I'm like, this is my space. What's happening? I don't... What engineer was like, yeah, yeah, yeah. Add more emojis. That's what we need in these answers. People will think that AGI has been achieved once we put emojis in these responses. It's a good flag to figure out who's tweeting via chat GPT. I know. I would say. Well, next month, it'll be a year since I started this podcast. I've had over 100 interviews. It's been a ton of fun, but I promised my wife at some point I would decide whether or not this was going to be a business or we would just continue to pay for it out of our pocket.

9:33And I'm exploring monetization, to be honest with you. There's not a ton of costs, but I do want to figure out a way to cover the editing and the packaging and the promotion costs of putting this podcast together. So I am looking for sponsors. I don't know what that looks like. If you've enjoyed this podcast and you've ever found some value and you want to sponsor, reach out to me, nickatcofounders.com and let's get back into it. Plot is, in my opinion, the better discernment tool outside of deep research. What do you mean discernment? Like if I'm trying to do anything that requires the population of an artifact, like a database or a diagram or something like that, Claude is superior.

10:15The limitations on Claude and why I would go to ChatGPT are basically two. One is that you hit your message limit really quickly. And then I often hit my day limit really quickly. Oh my gosh, I've never hit. Oh, with Claude. Okay. I thought you're talking about ChatGPT. I was like, wow. Oh my gosh. I hit both of those two things really quickly. I find their projects, though, are much tighter and much better. So if you run projects in Claude, that's a much more native experience. ChatGPT is not there yet. Can you tell me what a project is in Claude? Yeah. So I have ones for, like, for example, just for traction and customer distribution, where I have populated over time and that update on a regular basis, the context for where we are, where we want to go, what the data points look like this week or today.

11:03from various marketing campaigns or marketing efforts that we have out in market, who I think my customer avatar is, where that person is on the web, what the segmentation analysis says, that sort of thing. Like it's a running conversation against just these things. I had one where we were deep diving into product research. So I literally had read.ai on every single one of those interviews as I was trying to understand how to build this product out, then took the transcripts from those interviews as.text and then uploaded that into Claude as a project. And then as I'm going through and iterating and thinking about different features or the prioritization of different features, I'm literally just having Claude go backward and kind of re look through all of the project or all the product interview notes, and then assess whether or not that was something that came up.

11:46Like there are so many use cases. I went through a negotiation recently and literally I thought, you know what, I'm going to help. I'm going to have ChatGPT do my negotiation. And so like all of their responses I put into it before I put their responses in, I was like, this is the context. I just gave it everything. I uploaded like email chains. I gave it the contracts, like everything. And then I was like, here's their response. What should I say? And then I didn't just take it at face value. I was like, I don't really like this. I like this. Like literally it's my assistant that I'm collaborating with.

12:17Right. It's my thought partner. Totally. And that has been so valuable for me. I don't just take like the first response and be like, Oh, this was a dumb response. Like, no, there's okay. There's a little nugget here. What could we work? And so that chain that you're calling it a project, but that chain has been super valuable because I just go back to it every time that there was an update in the conversations. So that is if you just took those individual things and populated them into a project, that's a way to tap it on a more permanent basis and then update parts of the context where it's just one chat or just one chain in a Claude, you will hit that message limit pretty quickly.

12:53Okay. So what's the difference between a project in Claude and like a GPT in chat GPT? So, cause you can build custom GPTs in chat GPT and you give a context and you tell it, this is how I want you to act. This is your specialty. These are the responses I want you to give, et cetera. It sounds like it might be similar to what a project is in Claude or is it different? It is similar. I will say the quality of the responses I tend to get are better in Claude with the one exception if I need it to tap real-time information by crawling the web, because Claude obviously does not do that. So that is a very good use case for ChatGPT, or honestly, even perplexity.

13:29So one of the things that I frequently do is if I know that I have to tap something that is live or has been updated more recently than whenever Anthropic last updated Claude April or something of last year, then I go and I actually grab the information from perplexity. And then it will take the content of that and then pre-populate it back into some conversation that I'm having with Claude. I've heard that there are tools now that allow you to aggregate those interfaces. So it's like you can ask it a question and flip between the models. So it's like, you don't have to go to a different tab. I haven't used it though.

14:02I saw that on Twitter earlier this week. That's really cool. I like it. I think that that's a good idea. And I think that at the end of the day, these become utilities. This is electricity today. And we're just in, you know, Thomas Edison is lighting up New York city timeframe. We're not even close to where this becomes really ubiquitous and well-used by everybody. So anybody who's listening that is either early entrepreneur by yourself, or you want to become an entrepreneur, to me, your thought partner becomes these LLMs. I'm not saying they're perfect. I'm not saying, you know, that they're the end all be all, but you don't have anybody else to bounce your ideas off of.

14:39I have somebody who I know that's like, I'm trying to figure out product market fit. And they're using chat TPT. They're like, Hey, what do you think of this? Oh, that's an interesting insight. And they're just having somebody to bounce ideas off of. I'm going to give you a cheat code. Oh, okay. You can make almost anybody your thought partner. So one of my very favorite, I'm a nerd and I read a lot. And when I'm not reading, I'm probably listening to a podcast about books. Oh, I like where you're going with this. One of my very favorite things is to take distillations of books or interviews or transcripts and then populate that into a project.

15:17So I literally act as if random entrepreneur, like imagine if you're Steve Jobs being able to talk to Edwin Land, who was his inspiration, his his that's the guy that invented Polaroid for anybody that isn't aware. imagine using Edwin Land as your thought partner. That's basically what I'm doing in chat GP or excuse me in a cloud projects right now is I have a few people that I've admired the businesses that they've built over time. And I have mass transcripts, distillations, quotes, you name it that populate this project. And then when I'm stuck, I want to think about it like, you know, famous industrialists of the past has thought about this problem because it was truly green field for them then.

15:58Or I want to think about it like Steve Wozniak is my technical co-founder. If you don't have a founder, like a co-founder that you can actually have a thought partner with, not only can you use AI to thought partner with you with its broad knowledge, but you can also have very specific and refined knowledge by populating the context. I freaking love that. Oh my gosh, that is so good. How do you do it? Because the context window is only so large. when you say distillation. Okay. Walk me through it. Let's say Elon Musk. Oh, I love Elon Musk. Whatever. I have one of those. I can tell you exactly how I do it.

16:33How'd you do it? A lot of transcripts and then reduce it to its lowest, literally mechanically, its lowest file size, lowest file type. So I call a lot of the information that isn't necessary, like isn't relevant to what I needed to do. How did you do that? I literally ran it through AI to say like, remove all of the information, like the prologue, the introduction, all that stuff. I just want to hear the soundbites. And then I have all the soundbites in a.text file on some topic that he has talked about. Call it his thought process on product market fit, or his thought process on overcoming a setback, something like that.

17:08And then I will literally upload that as a.text file because it's about as low as you can possibly go. And that's part of the project. I think this part's really interesting because I love data. I freaking love data. But before you can get to, hey, run an algorithm against this data and tell me, what the variables are that actually affect the outcomes the most, you have to have the proper data set. So there's a lot of cleaning of the data before you even get to the place where you can perform the analysis because it's coming from different places or it's spelled differently, or one's in a.text file, one's in a CSV, one's in an Excel file, one's in Word and PDF or whatever.

17:42You've got to aggregate all of this data to get it clean. And then you can perform the analysis. And that's one thing that I've been using AI for is help me clean this data. Do you do the same? Absolutely. Yes. I'll have like OCR is a great application. Optical. I don't know what it is, but it basically takes PDFs and takes the text out. So don't quote any of us on what OCR stands for optical something, but don't upload PDFs. It's a really big file size, especially if it has images in it, just grab the text. Like, so from a data normalization standpoint, one, you want to remove redundancy. That's a really easy thing that you can have AI do too.

18:17You want to have the text. You don't want to have all of the images and all of the other copyright data or nonsense. You just want the text. And then you want to take all of that data, make sure that it is in its lowest mechanical file size type and capture as much as interests you. And I would say even it's important to think about things that you don't think are interesting today that might be interesting to you 10 steps ahead. Because a lot of the magic is in understanding the journey. And a lot of people are not self-aware when they're in it, but they're reflective, like coming back into imagining that scenario.

18:51Well, and people think it's a waste of time. I mean, I feel this way where I'm like, oh gosh, I'm wasting time cleaning the data or I'm wasting time having chat GPT, put it in a format that allows me to then put it into a custom GPT or a project and cloud. It's hard sometimes to overcome that. But what you have to remember is there will be a moment in time where you're dealing with a problem. And I've had this already in the last year. And the problem is tens of thousands, hundreds of thousands, million, it could be whatever the number is. And you get a specific insight because you prepared, you put all that information into the system.

19:20You ask a question and it gives you an answer. And you're like, oh, wow. Yeah. That was staring me in the face. Great answer. And I can move forward. That will save you so much more time than the prep of putting things together. But people should spend some time and curate essentially these AIs to help them when they're in need of these answers. Absolutely. And I would say we knew that even before we had AI as a tool, A good example from a marketing analytics standpoint is if you aren't really good about understanding how your application, whether it's web or a mobile application, is tagged to fire events into your analytics tool.

19:57Say on iOS, you have capital case is how you tag your events. And on Android, somebody does it in lowercase. Well, then you end up with a data set that is broken and you can make completely bad business decisions with high confidence because you're only seeing 50 % of the data. So data veracity is really, really important. You want to make sure that you've actually done the homework up front, not only because you'll get better insights, but you'll also save yourself from making terrible mistakes. How do you organize Collade, ChatGPT? Because I have tons of chats, right? There's tons of conversations.

20:32I've started using the folders on chat GPT to just get myself more organized. But do you have a system? Is there like, how do you sort of keep things straight? I would say for chat GPT, and this is another advantage for people that haven't used both chat GPT will understand more contextual relevance from you throughout all of your chats. Then Claude will Claude is very like that chat specific or that project specific. What Amanda's saying right there is like, ChatGPT will understand your preferences. It's like, oh, I know Nick. Yeah. Based on your question, this is kind of where I would give the response.

21:09Correct. And so, you know, ChatGPT will know both that I wanted ideas for my niece's birthday and also that I was talking to him about Galleon for some very specific context. For Claude, I can't see all of that. So I would say for Claude, I heavily use projects. I really, really, really recommend them, especially if you are somebody that hoards a lot of data. Like it's just a game changer. On ChatGPT, I also have projects. I haven't found them to be as tight and self-referential because it does have the ability to pull things from the web. And that's both a benefit and a hindrance depending on the context because sometimes you don't want it to be inventive.

21:47And I would say there's probably a 50 % overlap in those projects, but it's not 100%. That's a good point. Back to the chain that I have of this contract negotiation. I've had to re-upload the contract a few times just to remind it like, oh, you know, these are the points of the contract. Because you get so far down, it's just not able to keep that in its head, all that context in its head for that long of a period of time. So it sounds like you kind of use, well, this is my summary. ChatTBT is like the new Google, asking it questions, getting ideas, but you're using Claude to work. It's like, all right, Claude, you're my thought partner in actually specific data sets, specific projects that I'm working on.

22:30Yeah. Like this morning, I've actually had a running conversation with Claude in a project and I had it spit me out. All right. Now's the part of the show where I feel the most uncomfortable, but my therapist says I need to face my fears. So here we are. I've started a newsletter and I want you to subscribe. And what you're going to get every single week are the aggregated conversations from that week that I have on this podcast with an overview of what their business actually looks like. I'm also going to throw in a review of one or two businesses that are listed for sale. I'll give you my opinion on whether or not the EBITDA multiple is good, or there's customer concentration, or there's red flags or green flags.

23:00And then lastly, I'm going to give you one piece of actionable advice every single week on how to buy your first business. So click the link below, subscribe to my newsletter, and let's get back into the show. A sales forecast based on individual product type and also its implication for MRR over time. and it created a completely dynamic chart. Really? Yeah. The artifacts on Claude are fantastic. The output that I sent to you via DM. So I walked Nick through my own house sale. So we didn't even say this, but I sold my own house using Galleon. And that was part of my understanding of kind of where he was and why seller finance worked for me.