DAY 5: I ❤️ AI Challenge: Build Your First AI Employee

the agent that I wanted to build. So for VIPs, I have added, I think, 2 or 3 of them. I have more, so I'm going to add all of that in, uh, but that is part of it. Now what we are also doing, which is a new thing, um, but some of you know I teach classes on Maven, uh, this likely— not this class, but I— the, the next step of this, which is how do you actually use this in a business, um We're going to do a business accelerator. And on Maven though, I also have the ability to kind of share a bunch of these resources. So it's not the last time. If you didn't join VIP, that's okay. You will have another opportunity if you choose to use a lot of this. Let me go ahead and admit all these nice people in. Welcome in. See, I don't want to say too much before we get going. Welcome everyone. I first and foremost want to congratulate you. Like, y'all have shown up every day. Every year I do this, I'm like, Are people going to spend 10 hours like of their week with me? But it's really hard not to have 10 hours. So I'm so, so excited that you are still coming and joining me. Today we're going to focus on what happens next for those of you who are in this either as an individual using this technology for your business or as an AI consultant or someone who's going to go and help other businesses do what we've done this week. It's really important to actually start with what we're gonna do today. It's not as fun cuz you know, one of the easiest, I think one of the more fun things to do is just build stuff. But when you start talking to an organization, you'll, in my keynote presentations that I do all over the world, the way I end my keynote is letting every one of the members of the audience, I tell them like, today is the day. And the first thing I will say is like, today is the day where you have to build a strategy for your company. Around artificial intelligence. Like, you have to have a plan. If you don't have a plan, right? What do they say? Like, fail to plan, you can plan to fail. Like, my dad tells me that all the time. Like, that is like, you have to have a plan now. And I will always ask my audience, raise your hand if you have a plan for artificial intelligence. Just let me know. And about, I don't know, 1 to 5% of the audience will raise their hand. And then I will let them know, like, don't worry, good news. You actually all have a plan. And this too for you all too. Like if you have a strategy, a document that describes how you're gonna use AI for your business, which is what we built at least a little bit on Monday and Tuesday, right? We, we said for this one thing, whether it was, you know, managing code or managing an inbox or creating social media, we created a plan. This is the plan for how I, I'm gonna use AI to solve this problem. Now, if you don't have that plan, you do have a plan. It's just awful, right? It's like not written down and you're not really sure what it is, but you have a plan. Your, your plan is no plan. Um, and it doesn't really work very well. And so I'm gonna encourage you, it just depends on your client or you how you choose to go about this. I like learning by doing. I like implementing and being a doer, not just a talker. So I don't typically start with, let's write this down, um, unless it's specifically focused on a business outcome. So the method to the madness this week was let's build a strategy. But you'll notice on Monday I didn't say we were building a, a strategy for how we were gonna solve a problem, but that's exactly what we did. And so today I just wanna kind of give you the, the framework for how I think about these projects. I deliver projects every week. I get new clients, uh, that I deliver these projects for., and I do the same exact thing. One, and this is covered for those of you who are, uh, have taken my AI Solution Architect class on Maven or through the Institute. Um, this is AI solution architecture. It's how do you get someone to go from like, I'm not sure about this AI stuff to wow, this is cool. How do I get started? And most of my audience starts that way. They start with, I'm not sure about this. And then I move them to, Okay, this is pretty cool. So how do I do that? The first thing I do is I don't even talk about artificial intelligence. I don't even mention technology. I just ask them, what hurts you? Like, what is bothering you? What is frustrating you in your business? To give you an example, a few years ago I got an opportunity to go to, the Metropolitan Museum of Art. Some of you know this story. And I went to the Metropolitan Museum of Art. My job was to go in and bring 9 data scientists from MIT and help the curators, art curators at the Met build, like do better work. Like that was the ask is can you come in and bring data scientists and solve problems of the curators? Now that meant that the data scientists were in the room and the curators were in the room and it was like an awkward high school dance, right? Like they didn't talk to each other. I don't even think they knew each other's language. It was very awkward. And I remember we, I spent all day kind of talking to them about how we could solve problems. You know, I was trying to build some trust. And at the end of the day, the producer, 'cause of course I was there with Microsoft. Microsoft recorded the whole thing 'cause they wanted to record evolution of, you know, data science at the Met. And the producer of the, like the, the production team came to me and was like, it doesn't look like it's going very well. And I was like, easy, like I got this, like hold on. But the reason why is 'cause the teams weren't connecting, they weren't trusting. And it reminded me, again, candidly, of what my life was like in the early days of Amazon Alexa. Now, some of you remember from Monday, I have 6 kiddos. I was pregnant with my third during most of the Alexa launch. So if you know a little bit about Alexa history, Alexa was born on November 6th. My third child was born on November 1st. And so I had two babies that week. Yay! But needless to say, it was a little crazy, right? Because I'm having a baby and I'm having— and I'm launching Alexa and I was very passionate about both of them. And it was my third child. So for those of you who have lots of kids, you're like, eh, like, it's fine. Like, I basically went right to work before, after. There was no really pause. Judge me if you will. Uh, but that being said, I had a baby, I had Alexa. Then my dad, who was a Marine, he, uh, would run across the Mercer Island Bridge in Seattle, um, every day. And I had an app on my phone, MapMyRun, where I could watch him run just because I would be worried. He's— at the time he was 65, perfectly capable of running. But anyway, I, uh, 2 weeks after Alexa launches, he gets hit by a car as a pedestrian walking across the street. By a distracted driver texting on their phone. Not a young person, just so you get that like ageism out of your mind. It was, you know, like everybody can fall victim to this. Uh, but anyway, that person hit him with a car. He was all busted up, right? 60-something broken bones in his face, and there's so many bones in your hands and feet. Anyway, so he was all a mess. So now imagine, if you will, I'm at work 2 weeks after my baby is born and in intensive care with my 3-week-old baby. And so I go to work and people— how many times have people come to you and been like, hey, um, you know, how's it going? Like, I was just on a phone call with somebody and I was like, how's it going? They were like— it was so refreshing— they were like, ah, it's not great. And I was like, no one ever says that. Almost everyone answers, how's it going, with it's fine, or it It's good, or whatever, like some very, very quick off-the-cuff response. And so I similarly, people be like, how's it going? I'd be like, fine, great, uh, moving on, let's do our thing, let's get to work, even though I had all this nonsense happening in my personal life. And then one of my good friends today, she wasn't that great of a friend at the time, but she stopped, like, I— she was an acquaintance at work. She stopped and she put her hand on my shoulder and she was like, no, like for real, didn't you just have a baby? She was like, "Are you okay?" And I was like, "No, I'm not okay." And then I just unleashed all of this stuff because she just stopped and was like, "No, for real, I'm here for you. I'm here for it." And I laid it out. I was crying and my dad and my son and da da da. It was a mess. Thank goodness we are still friends. But every so often I'm like, "Gosh, I'm so glad you asked, but at the same time, I'm sorry about that." But what ended up happening is that in that moment, I felt like I could share all the things that were wrong. And what we were missing that day at— when MIT and me and the curators, we were missing that moment where that team was even willing to say what hurt them. I didn't want them bawling their eyes out, but they weren't say— they were like, we would ask them, how can we help you? Data scientist ask curator, how can I help you? And you know what they would say? We don't need help. We're fine. Like, everything's good. We're, we're good. And so it required us to spend a full day talking about all of the different apps. And so we ended up, in order to get— all we had to do was stop and actually put our hand on their shoulder, proverbially, and say, like, no, really, like, what if you could whisper to me one thing that you could get rid of in your day? What would it be? And as soon as we unlocked that, it was like You could fix this, you could do this. We had 17 use cases in that first 24 hours. We built out all 17 in the next 24 hours. Some of them got so popular. You can go to my Twitter, my Twitter, my X profile, and you can see one that's pinned there. My favorite was we built a talking, a talking AI, like it was an Alexa app that allowed you to walk through the Metropolitan Museum of Art. And I built it for my son, and it was built so that it would be more inquisitive of you than you were of it. So you weren't like, show me art. It would say like, what mood are you in today? And my son, he still says this to the day today. Uh, I don't know if you know, on Spotify there's a Feeling Blue, um, Spotify channel. Anyway, so he'll be like, I'm feeling blue. And so you say that and then the model would go and find art that would resonate with how you were feeling, um, or maybe uplift you, or maybe it would give you what you wanted and then say, maybe you wanna see something else. And it would just cause a conversation to happen that would allow my son to experience a museum without him having to walk, which he won't, through that museum for hours and hours on end. And that came out of this conversation where they're like, "Gosh, we really wish we could create a more accessible experience." That was the pain. "We want more people to see the art we've curated." And we're like, oh, that was, I immediately was like, my son. So I moved to that project. One of the others, was a GAN, a general adversarial network, that created new art from old art. Now you can imagine curators of the art were like, there was a bit of a visceral reaction to building art that created art from old art. And we're seeing that now, right? Where people are like, wait, you ingested Picasso, you ingested, you know, Degas, and you're creating new art from that. The half the room, not even, 3 of the curators were like, "That's new art. We should showcase that in a new part of the museum." The other curators were like, "This is blasphemous and this is not art." By the end, we got everyone to agree and it was an abstract artist that actually came and talked to us and explained how creators create and that all of us in our brain use what we've seen, what we've experienced, what we've learned. Like it's not, that is the, the way you get to your creation is not what makes it unique. It is your perspective, right? That's why they say beauty's in the eye of the beholder, right? Because it's all about perception. Anyway, we got people to allow us. So that was another one at the Met Gala that year. That was what we focused, what we showcased. Um, actually, uh, all of these, like the walking, um, the walking Met, which was the one, uh, Art of the Day, um, and the GAN were all showcased at the Met Gala that year. And it was really fun.. But I, I just always think back like, oh my gosh, if you don't build trust, if you can't get someone to actually tell you what's broken, what bothers them, you'll build the wrong thing and then you'll actually create distrust. So you remember on Monday I showed you a slide and on that slide it said the 5 reasons why people, why, uh, 95% of pilots are failing. The number, or maybe I showed you that slide actually on Tuesday, but When I showed that to you, why 95% of pilots— this is projects. So you might do 100 projects in a company, 95% of them won't work. And one of the main reasons they won't work is because the people you build them for won't use them. And the reason they won't use them is because you didn't have this initial conversation. What hurts you? And let me reverse engineer how AI can meet that gap in your life. And that's the best possible way, um, to do it. I have no idea what that is. Oh, thanks. Um, thank you. Uh, but that gives us so much power to actually create an empowering experience for people when it comes to AI, as opposed to the dystopian negative Skynet version of AI that you see a lot of on the internet, um, and on social media and even on 60 Minutes. I was super kind of disappointed because, you know, you have an opportunity to tell millions of people about AI and you're, you focus on all the bad things it can do, which is valid, but you should also be like, oh, and we now do more knee replacements with higher levels of success than we've ever done on the surface of the earth, right? Like, and we've been able to, you know, reverse cancer diagnosis at a higher level than we've ever been able to because AI can track cancerous cells in the body better than it's ever— you know, like there's all this good stuff that's happening. So I encourage you as you start using AI to always see both sides. I'm not asking you to only see the bright side. That's where I try to live all the time. But at least make sure if you are sensing that you are seeing like it's all doomsday, all negative, like search hunt for optimistic use cases. And so this is what I do. I have two— I have a lot of Google Alerts, old school Google Alerts, but two of them. One is AI gone wrong, which is the dark side, and one is AI for good. And they both come up the same day, time every day so that I can see that there are examples of both. And that empowers me as a leader right now. I can see how people are failing and I can avoid those failures, but I can also be like, oh my gosh, I never even thought about using AI in that way. And then I use those examples. So as a solution architect, our first job is to figure out what hurts people, what hurts the human doing the work today, build the problem statement. And I will ask people my— I do a, um, an AI solution design workshop for clients. As a matter of fact, a bank just asked me yesterday if I would be willing to do one for them in April. And I was like, of course I will. Uh, but they're asking because They right now are adopting a bunch of baby tigers and no one's really crystallized the problem those tigers are solving. As a matter of fact, I'm not even sure they're solving any. Like maybe they're solving note-taking, if that was actually a problem worth solving before. Um, but there's so much more opportunity if you actually ask an adjuster or an actuary, what could I do to make your life better? They have, they're full of ideas and they've been thinking about it for a couple decades in many cases. They've been wishing that, oh my gosh, I wish we could have something like this. So when you ask them and you are trusted and you are, um, you know, empathetic, they actually give you exactly what you need to define the use case. So that's the first thing, problem identification. First step, any AI project, even before it's AI, but right now you'd be hard pressed to go to a company and not have AI leading the, the conversation. Um, but I say, as soon as we're like, yes, we're gonna build an AI solution for you, but first thing I need to know is where, what are we gonna do? So that's what we did on Monday. What problem are we going to solve? Then as a result of that problem, now I'm gonna do something, sounds crazy, but I'm gonna write business requirements, like old school business requirements. I will tell you, humans don't even like writing business requirements. They'd rather skip right to implementation, which your AI would like you to do too. But it is so important because this is where we put in, right, the guardrails. We put in the boundaries. We identify scope, right? This is incredibly important to building an AI system that actually works. The 5%, this is an ingredient, is you identify the problem and then you build a requirements document. You store that requirements document outside of the AI system, which is exactly what we did on Tuesday. Right? So granted, I didn't call it that, but I just wanna give you a sense of like how this actually works in Microsoft, Amazon, IBM, the companies that I've worked in where we've built out AI solutions successfully. We had lots of failure too. So problem determination, business requirements. Now it's very easy to slip into functional requirements when you're doing business requirements. I'm pretty meticulous from an operational excellence perspective to separate those two activities. Only because functional requirements start to take on assumptions in the environment, and I don't want my business requirements to have any of those assumptions. So I separate the two. So, and, and again, you can always use ChatGPT to facilitate this, but this is the thought process that goes into before you even selected a model. You have a problem, a requirements statement on how your, what requirements are needed to solve that problem. Functional specification of like, what does good look like? What does bad look like? What data do you need? What architecture choices are you gonna make? And then you're ready to do the fun part, which is model selection. But now when you go to choose that model, you are scoped to the documents you've created. So you're not gonna be— this is where a lot of people are like, I'm overwhelmed with choices. And I'm like, that's cuz you don't understand the problem very well, right? Cuz not all models are good for every problem. As soon as you do this pre-work, your selection of models, it shrinks considerably. It's still a lot. I have empathy. It's still a lot. But it's way less than all the things, right? So it gives you the ability to use problem definition to scope and shrink the amount of options available to you. And remember what I do for model selection. I'll just share with you 3 things. So this is all going to be— this is all recorded, so I'll document it and pull it out. I'll share with you at the end. How you'll get this, all this information given back to you in a really interesting format, I hope. But when the 3 steps I go through, or 3, I guess, the decision points I make first are what kind of AI I'm going to use immediately following understanding the functional specification. So what I think to myself— thank you, thank you— what I think to myself first is, Do I want to buy or rent, or like, what kind of model am I going to use? And I'm gonna characterize it in 3 ways. One, I always do hypothesis testing in big models. ChatGPT 5.2, Gemini, right? Its latest version, Opus 4.6 on Claude. Like these big public models, I do my hypothesis testing. I don't do them on a free like in a free experience. I have my own business application, business usage. But I still use a big model to test the hypothesis and see, can I at least get there? That's what— we had a couple questions in VIP yesterday. Like, that's why I use Lovable, which is attached to a big LLM. Like, I use these tools to help me just see, is end-to-end, is what I want to do possible? Not is it perfect, not does it have the right access to the right data or the right ERP? Like, is it just possible? It is the cheapest way to test out any hypothesis is to use a big model because they're very cheap at a token level, uh, from a tokenization perspective. Then the next thing I do is I then do what's called boosting or fine-tuning, right? So first I use a big model, then I fine-tune a small model with the learnings from the big model. In machine learning, this is called transfer learning. But in just layman's terms, you're going to take— remember, we wrote all of our prompts in a document. We're just going to take the prompts that we built in the big model, and we are going to— it's called instruction fine-tuning— we're going to pass those to a smaller model. Which smaller model? You'll have to evaluate them. You'll go to, remember, Helm. You'll go to Holistic evaluation of language models. You'll go to Microsoft Foundry, you'll go to NVIDIA's developer lab. They all have leaderboards that will tell you if you're in fintech, here's the leaderboard. If you're in medtech, here's the leaderboard. Like, it's not— you're not doing something novel here. Like, we know what models are good at which things. And so do that research to figure out which model. And obviously these fine-tuned or specialized models they're going to be smaller than the gigantic model, usually by parameters, like the number of data points they're measuring is much smaller. But there's also a technique, and I'm only, you know, for the 30% of you that are like gonna geek out on this, um, there is a methodology called PEFT, P-E-F-T. Some of you might know it, right? It's where you actually take the parameters you care about in the model, like the data points in the model you care about, like if you're doing customer analysis, maybe it's the customer, their buying propensity. Like there's certain data points you need and you freeze everything else but those things you care about. So it's called parameter efficient fine-tuning. It's a little bit more sophisticated and definitely more technical than instruction fine-tuning, which is you just tell it what you want. But both are used typically in a production scale system to narrow the focus of the model so small that now your model is just kind of a tiny microcosm of a big model. Those models tend to be very cheap to run, highly, uh, like from a latency perspective, have very low levels of latency, sub-millisecond response times. Like they're just very, very efficient. The challenge is, is a lot of people like to skip all the stuff I just talked about and go right to a fine-tuned model. The chat, and this goes back to what I mentioned earlier in the week, when you get a fine-tuned model that you didn't tune, it's fine-tuned to different business requirements, different functional requirements. It's built for a different person asking a different question. And while it might be similar, it won't be as good as it can be. And so I encourage you to do this work so you get a customized, I mean, this is the value you get to offer a client. If you're a consultant, you can deliver not just a FinTech enabled model, like a model that's good for finance companies, but a model that's good for Regions Bank or Truist. Like, it is their model. And that— I was just on a, um, podcast this morning. I know it's a lot. I'm— tell me it's Friday. But I was on a podcast this morning, a security podcast, and they were asking me kind of these strategies around like, how do you make sure— so I, I, I I provide this not to increase the complexity of the conversation, but just to add another layer that in addition, when you get this small little model that you own, you also now get to own the data sovereignty decisions, right? The— meaning where does the data sit and where is the model going to sit and who's going to manage it and how do they manage it? Like your control surface as a cybersecurity professional becomes more manageable. It's still hard. But the attack surface shifts when you own that. So the future— I said this on the podcast, I've been saying it for years— the future is that every one of us will build what we did this week, a custom agent, which is really a fine-tuned instruction, fine-tuned model, right, that allows us to narrowly focus on a single problem and identify very specific results that we can now measure. Imagine building 50 of these. We built 6. Right now there's not— I didn't really build any agent that would oversee them. I'm the overseer. So now every time you build an agent, you're gonna have to make a decision. Am I going to human oversee, meaning human in the loop, human oversees it, or am I gonna build an agent that's gonna monitor it? Human on the loop. Now I'm gonna look at my agents and see their data and just kind of watch what's happening and hope nothing bad happens, but I'm there in case I need to pull a kill switch. Or it's, is it human out of the loop? There is no human oversight whatsoever. Now, some of you might be like, we should never do that. But I will tell you, there are use cases where I don't want to be involved, like sweeping my floor with a little robot. If that thing could actually do a good job, I would love to not be involved, right? If it would actually get into the nooks and crannies and get the kitty litter that popped outta the kitty litter thing, like I would never look at it again. I would just be like, good luck. If, if it could go from my kitchen pick up its own canister and throw it in the trash. Like, I would never ask it a question. I would never, I would never micromanage that process. So there are use cases that we would just like let go of if we trusted them. And most usually if we see them happen in it, it's, it's a sequence, right? First, I don't trust you at all. Zero trust. I'm going to watch. I'm going to be involved in every decision. I'm going to gate every email that goes out. But over time and with our requirements, it gets better. Now I can step away and I can just watch you and correct you, right? Now I can give you positive or negative feedback based on your behavior, but I'm not pausing the decision and forcing me to be the decision maker. Eventually you'll build enough trust, which is what we're seeing with our little— sometimes with our little robots that clean our house. Now I can let you go, right? Ring camera is a good example, right? Like I'm not micromanaging whether it sends me a message about someone being at my door that I don't recognize. Like, it just sends it to me. I don't have to— I could, I could have a gate that says all pictures go into this box and you have to approve them before you can see them. But that's— we're past that now. We trust that those images are accurate. And a lot of times this happens inside the product team before we see it as consumers. But if you're the product team, you should be creating these stages and evolving those or— and some of these, uh, products that you build will take years. So this is where I enter into how long does it take to build a really good AI system? And I will just look at history. So ChatGPT, OpenAI, right? OpenAI took 7 years before it launched at 3.0. You might remember, right, in October, November of 2022. 7 years to get them to that point. And it was still kind of bad. If you remember when we first got it, it was good, but it was like, in hindsight, we're like, wow, that was actually not great. Now it's even better. But back then we were impressed by it as humanity, but it actually still went to market with a bunch of defects in it. Specifically, if some of you remember, it would give you instructions on how to build a bomb, right? It would tell you how to manage a prisoner of war. This was the early days of these models. So that's bad, right? And I will— I'm always very empathetic to the data scientists because I can assure you, I've seen this happen on Alexa, that the data scientists are like, why would anyone ask that question? No one's going to ask that question. Like, that's their mindset. They're like, we don't need to protect against— they're not even thinking about protecting against these scenarios because they don't think you would think that way. They're not thinking about bad actors. It's not the job of a data scientist to ask those types of questions. It's actually the job of us, the leaders, the business leaders, the people who understand the domain. And it's also the job of an AI red team, right? Is to like adversarially attack these systems. Oh my gosh, so much going on in here. Okay. Uh-uh-uh. Let me go down, make sure. Oh gosh, there's a lot in here. Uh-oh. Oh, good, good, good, good. Oh, good. Thank you for dropping that in, Alonzo. Amazing. Thank you. So there's some good resources that y'all are dropping into the chat. So thank you for that. Because I named them, and then there's definitely data at Stanford, MIT, as well as Microsoft and Google that you could go and ground this in to get some actual, like, reports if you want additional information. There's also entire classes on it. Deep learning, AI. Of course, I love teaching. I don't teach all the things, though. But if you want to know where Noelle goes to learn, like Stanford's my favorite school. I went to the foundation. I actually helped launch the Foundation, uh, Model Scholars program at Stanford like 4 years ago. And so they're my favorite people now because they taught me about what we're doing now and what companies are just now starting to implement. They taught that to me when it was just research and they were like, this is a crazy idea. Uh, one of the crazy ideas was, um, this concept of FrugalGPT. And so it was a nat, uh, natural language, right? An LLM. And its job was to sit— we now have this implemented in Microsoft's products, but it sits in front of your LLM and it, like, you get a request from someone to do something, like us. Like maybe we, we send a message to our executive assistant. We're like, hey, can you do this thing? Can you reschedule this? What it would do is it would take that request and analyze it and then search for the appropriate LLM. But FrugalGPT, you can tell, what it's prioritized against. Not the best, not the least toxic, the cheapest. And some CEOs care about this. They're not asking what's the best servant to humanity to solve this problem. They're like, what's the cheapest LLM with the lowest level tokenization? The reason that was such an interesting project back in 2022, 2023 was that tokenization, which I didn't know until that time, is completely arbitrary. Like how companies choose to break up tokens, completely arbitrary. And there are significant trade-offs to how they do that, meaning the integrity of an answer is bound by how they broke up the word and how the model interprets the broken-up pieces. And so you get very different behavior. That's why you might wonder— I always encourage you, go to Poe.com, right? I mentioned this to you yesterday, I think. But if you go to Poe.com, I always encourage you to, for any, I think, I think I can just write this in here. You can go and just test out every day. I'll go in and take a prompt that I'm thinking about and I'll just put it in 4 or 5 models just to see how it's different. My favorite is, is like image generation models, right? Stable Diffusion. Because you go in and you provide a very simple prompt, give me a cartoon-like cat eating a burrito, and you put that in 7 models and they will be vastly different in how they choose to render that image to you. And all that is, is a reminder to you that the model doesn't know what truth is. None of them do. They're just guessing. The job we have is in our instruction set, right? In our strategy to define what is— how are we going to define truth and maintain truth in our AI systems over time? So there's lots of ways to solve this problem. In my organization, like I said, I create persona repos or a prompt repository is another way to think of it. Let me show you an example. And you should do this yourself. You did this work already. But I'm going to show you an example. Maybe— hopefully I won't show you some proprietary thing. Here we go. So this is Microsoft Office. This is one of my documents. I think I shared this. Let's see. Uh-oh. Is it? Oh, there it is. I see it. Okay. So you can see here, I— this is not the source. This is just what I created so I could turn it into a PDF. Mine is in a very ugly notebook, but like literally a digital notebook and it's just like, blah, that I can copy and paste when I need it. But I formatted it a little bit for you all. But this is an example, right, of the instruction set, but I have a repository of them. For every problem I'm trying to solve. This is a distinction in what works and what doesn't, is a lot of people never even— they don't start with the problem. They start with the tool and work backwards. How could I fit my problem into this tool? And while it can work, it just won't work as well. And you may end up in the 95% of AI products that were good, but not good enough to hit production. If you go in the reverse, which is what Microsoft AI— I mean, I was a principal product owner at Microsoft AI in the early days. Um, 17 products and all of them I did the exact same way. And I learned that from Amazon cuz at Amazon we had single-threaded owners and our job was we were ultimately responsible for the success of whatever domain or whatever, uh, product or whatever skill we were building for the platform. And so I needed to have end-to-end resolution. If something happened at the end, I needed to be able to point exactly to the words that I used or the definition or the scope or the specification details that cause that problem to occur. So I'm very, um, I find there's a lot of comfort in operational excellence that ties everything you see in production back to something you actually planned for. Another word for this in engineering is like code coverage, right? Like we want to make sure that we don't write code we don't use and Someone mentioned this earlier this week. If you start using these code generators, they write a bunch of code you don't use. But that's not why we do it, right? We trade off inefficient code to build code in lightning fast time. And I will do that. I will actually trade off efficiency of how the code is written over the ability to deliver something very quickly as an MRP, minimum remarkable product, not in production, but to get in, in front of someone so they can say yes or no. I will absolutely build bad code to do that. The difference is, is 5 years ago when I built bad code, that was tech, tech debt that I would acquire, right? I would now own that gross code that I wrote to quickly deliver for my client, and I'd have to undo it, refactor it, rebuild it. Now I don't have to do that. I could just say, okay, I'm an engineer. I know what I want my code to look like. Let's refactor this now. And I can use natural language to build better code. This is, I think, a really important point. Most, like, it is possible to ask a machine to build good code for you, but you have to know what good looks like, right? Just like we did on Monday. We have to define what good code looks like, what well-architected looks like. Interestingly enough, don't judge me all, but I have, like, maybe I'll just, I'll pull out one example. So this is how I learn. I'm a paper person. This is old paper though. I don't reprint anything. I've had some of these for years. But this is the Well-Architected Framework at Amazon. Like, how do you make sure you build a cloud application that doesn't fail? Elastic compute, all of that stuff. And so I keep this because I want to be able to go in. And this is how I do. I literally will take a pen, a Galinda pen, and I will check it off. Like, OK, that— I grew up in aviation. That's where I went to school, Embry-Riddle Aeronautical University. So checklists are amazing. As a matter of fact, Scott Guthrie, he's the head of Azure at Microsoft. He introduced me years ago, 7 years ago maybe, to a book which I'll encourage. This is, again, it's not traditionally AI, but this has helped me build incredibly sustainable AI systems called The Checklist Manifesto. Anyone heard of it or read it? It just is a mindset that says— and, and yes, right, Javier? I would imagine you would as well because you own a consultancy. Great book, thanks Dave. Yeah, so, so good because it shifts your mindset to don't do anything you can't check. And I, because I came from— I grew up in aviation, everything is a checklist. Some of those checklists are like Bible books, like huge, huge checklists. But you miss one thing and it's a life or death situation. And I actually have attributed a lot of that importance to what we do in AI, right? You don't do that checklist of, are we building something that satisfies this in these guardrails with these boundaries? You let one of those boundaries go, and now, like OpenAI, just candidly, right? They have 9 wrongful death lawsuits coming against them, and I can guarantee, yes, they're, There is a part of their organization that is worried about the lawsuit, but many more of their data scientists that are worried about the human impact that they had that they never thought they did. We say this at Amazon, like all of the OG Amazon developers, when we found out that there were certain Amazon skills that hurt people, number one, naming a product Alexa when there are humans that have Alexa as a name. Who knew? Like that ended up being a big pain for people. That they came after me specifically. Developers on Alexa that were public, they were like, there are humans named Alexa and you have damaged their life. I mean, it sounds funny a little maybe to some of you, but it was very real. But more importantly, they talked to us about their moms came at me in Twitter and Instagram and would tell me in my content that I hurt their children, that their kids named Amazon Alexa were getting bullied, bullied so badly, like girls named Alexa getting bullied so badly that girls wanted to change their names. Like, I'm a data scientist. I never— I didn't even ask. This is the challenge, right? I never even thought like Alexa was the product name. Like, we didn't even ask questions about what the impact would be. I ask those questions now, but like, I think this is part of each one of us has different lived experience that will bring a lens to writing the definition of what good looks like for these AI systems. And if we do it, we actually avoid Skynet, which is a question the guy asked me on the security thing. He's like, "What do you think's gonna happen? Is it dystopia or utopia?" And of course, it's never black and white, but I definitely think we're the only ones who will write these plans and build AI that accommodates those plans. I'm a golden age of science fiction girl. Thanks, Dad. I read I, Robot. I understand, you know, I'm a big fan of Asimov. There is a concept called the Three Laws of Robotics, right? I mean, you should know them. If you don't, go look them up. But the Three Laws of Robotics, it was fiction. It was 40 years ago, maybe longer. It was fiction, but it was saying like, hey, if you're going to build something that can get smarter than you, you should probably build into its DNA Rules that allow it not to hurt you, of course. And that's kind of an extrapolation of the work that we're doing here in today, like in, in this week's session is we had some fun, we built some agents, but the way all of this works out for businesses is if we start with these definitions, right? Is if we— thank you so much. Cool. Yes, that's, that's them. Yes. So thanks for dropping those in the chat, but must not harm humans. Yeah, why wouldn't we put that in? I will just tell you right now, the biggest companies in the world are making very baby tiger decisions, and worse, they're laying off the people that might have built these kind of rules. So now my mission is to equip you potentially as a consultant who can go in. I mean, I have so many, so many reasons for this, but you go in as a consultant, as an advisor, as a trusted voice, and you're just like, hey, maybe we create some rules here, maybe we create some boundaries, maybe we build a plan. So I'm going to give you now the 3 documents that I start every project with. I do it via a workshop, but you can pre-build it and deliver it. There's lots of ways to address this, um, but the 3 documents that I create— 1942, right? Oh my gosh. So like You can see where I'm stuck. Yes, thank you. Thank you, David. 1942. 42, right? The meaning of life. Anyway, if you know, you know. So the 3 documents that I build, and again, there's a lot of flexibility in how you build them, but they mirror the work we did here. The first one I do is I build a responsible AI strategy, but I do not use the word responsible. Responsible in it, nor should I, because of course it's responsible, right? Like, if you, if you have to call it out, it's fine. But like any strategy you build for a company that has AI in it, it should be layered with responsibility. What those responsibility tenets are, I'm just going to drop them in. There are a few of them, but these are rules that you're going to accommodate. You'll see them, uh, for those of you who get access to this, Um, you'll see them in my— all my prompt, um, guides, all my personas— is there has to be an identification of how you are going to approach transparency. Now I have to talk and type at the same time. Transparency. Um, and of course if you're VIP, you're going to get this all documented. But, uh, transparency— I'll just do them one at a time. So transparency And this one goes together because it's not enough to see something. You actually have to be able to, uh, understand it. Explainability, which is almost Webster's Dictionary. Explainability. Um, that means that not only are you building a system that you can audit— remember we said in our agents, you have to be able to look at and audit the behavior of the system— we not only want to be able to see what's happening, we also, maybe more importantly, want to be able to understand it. So this is why we now have the vehicle, right? If, um, uh, let me stop sharing this thing and share a different thing. Um, if we look at— let me go to my Arc browser. Uh, let's see. Oh, sorry, I'm just trying to find— I've dirty desk, um, thing going on. You're going to see all my research today. There we go. Um, I think that's it. Oh no, I'm still on Word. Argh, that looks like Word. I swear. All right, I'm gonna stop sharing again and start again. Let me try that again. Um, yeah, all right, let me try one more time. There we go, there we go, there we go. That's— that looks better. So when I build, you know, a responsible AI strategy, one of the things I can't remember, was this what I was gonna show you? These are basically the questions that I'm looking to answer as early in the process as possible. Now that AI strategy, oh, that's what I was mentioning, is that there is explainability and transparency go hand in hand. So I want to, and you could do this again as part of your solution you provide to a client or a gift you give to yourself in your own business is creating a dashboard that gives you visibility. And oh yeah, that's what I was going to do is let me see if I still have it up from yesterday. I don't, so let me go to see if I— there we go. So here, everything I build now, I add in a cute little LLM that allows whoever's using this tool to get help on it, understand it, This is explainability, and that person can ask these questions from their perspective and get answers in their perspective. So this I think is a big part of that, that solution. Yes, exactly. Sophie, all of this is embedded in the EU Act, uh, EU AI Act that was put into motion 5, 4 years, 3 years ago, 3 and a half years ago, put into motion. It finally passed and they put timestamps that said, okay, I get doing things responsibly is going to take some time. But those timelines are expiring this year, one big one in August. So you're going to start to see— that's why I think the timing of your education on this is so critical. Because, and of course, I don't know, y'all probably saw there's some craziness happening in the US, and there's— we'll see more Supreme Court decisions rolling out. So who knows what, what's going on over here. But at least the EU Act, AIEU Act, is a good example to see kind of how it rolls out. It is not instantaneous. You are not like far behind. You are not, you know, like there might be a little FOMO, but most organizations are just now starting to implement the things that we're talking about here. So it's a great opportunity to help people. So those are the first two is transparency and explainability. Now for me, in my control plane, let's see here. I think it was this one. Yeah. So what I wanted to do was provide an audit trail. So I created an application that if you fed it in an AI system, it could track day by day by day what was happening in that system. So I had an audit trail. I built that. You want to have just, you know, document as part of your process that you have some way to provide transparency, which means auditability and explainability, meaning someone can understand what they're seeing at any time. Some other ones, there's 3 more that I want to talk about. I'm going to leave the one I was about to say for last. So fairness, right? So fairness is another one. How do we make sure our system is fair? The only— the best way, not the only way, the best way to find out if your system is fair is to give it to people who might think it's unfair, right? I always say that, um, you know, empathy begets like bias detection. In other words, like somebody who's empathetic to— like me, I'm— I don't have Down syndrome, but I am very empathetic to a person who does because he's my son. So see the world through his eyes from the perspective of like, that's not helpful, that's helpful. Um, it still doesn't— like, the best solution is just to get him. But empathy allows you to see when you're looking at things, how can you see through the lens of others? And that's why I think it's really cool, because a lot of us have, you know, especially Mindvalley people, right? We're on a, on a path for self-awareness, and self-awareness creates right? Human awareness. Like, now I'm, I'm able to see through the lens of others. I have higher levels of empathy. And I think that's a critical ingredient for AI success and, and being able to tell, is this model fair? Another way to do this also is fairness is a, is a metric that's tracked by HELM, the Holistic Evaluation of Language Models. And I'm just dropping these notes into the chat more to train the model. So when it gives us our synopsis, it will include these references. Um, the next thing, right? So that's fairness. Um, the next thing of course is privacy, uh, or I can call this trust. It is your responsibility to architect— we did this on Monday— architect trust into your application. So what does trust mean to your client, to your industry, to you? Architect it in. There is a concept called constitutional— I hope I spelled that right— AI. It means when you build AI, and there are again classes on this if you need a more in-depth approach, but specifically, how do you design your AI system to adhere to certain rules? A constitution, if you will. Constitutional AI means you're going to build out a set of rules and make the— make sure that that becomes the lens upon which all your other AI systems work. I am a big believer that when you build 6, 10, 12 agents, you're gonna wanna build a system. Like imagine, right, your agent is about to build you a new social campaign, but before it does that, it needs to create a brief about that campaign and pass it to your constitutional agent and say, is this in alignment with all the things? I could do that human in the loop, or I could let my constitutional agent do that review and surface anything to me. Human on the loop, or I could have done this a long time, feel really good about my constitutional agent, and let it go on its own. Make sense? All right. So constitutional AI, um, and that's, again, that is, no one's gonna build a constitutional system to manage your integrity, your core values of your company. Like no one is going to say, is the AI you're using, does it match who you are? Is it doing things in alignment or disalignment with the way you wanna run your business? The only one who will do that work is you. So it's really, really important to provide it. Oh yes, great question, Gabriella. Of course I will, of course I will. It's a great, great question. I can't remember it right now, but I will definitely tell it to you. It's like something— I can't remember, but I will, I will definitely share it with you so you all can listen to this. It's all about security though, so it's very much for cybersecurity professionals. So if you're interested in that. Okay. Okay. So the last one I'm going to talk about is a little bit of a surprise, but I do think it's important as you round out what needs to be included in your plan, and that is robustness. So people use different terms, but robustness means how are you planning again in the design of your system? This would fall in the functional specification. How are you planning for the day that your model becomes a big tiger, right? Or for more importantly, your model is just one aspect of your architecture. What happens that day? I'll give you a very real example. Aviation, right? Airlines. Every so often you might know that the airlines have a weather incident that occurs and it messes up everything, right? And all the flights get canceled and Everyone's sleeping in the airports, and at that moment they get an extremely high peak of activity in their contact center. Well, we, of course, I worked with a team on a specific airline and we gave them instructions and guidance on how to build an AI system that would deflect a lot of work to a virtual system, meaning that a bunch of people wouldn't have to sit on hold for 3 hours. Um, they could just get an automated system. We, it worked beautifully. However, the, uh, estimations of traffic coming into the system, I don't know if you've ever sat on hold for a really long time and then the, especially in the airline industry, and then all of a sudden you get hung up on many time. I don't know if that's ever happened to you. It's happened to me a few times. Uh, it's super frustrating, but many times it's not because, which is what I used to think, an agent picks up and is like, mm-hmm, I'm, I, I'm done. I don't wanna talk to you. Click. And they hang up on me. Literally, their system crashes due to the fact that they did not plan for that level of utilization. Elasticity. Many organizations, that's like cybersecurity. Many organizations are still running on like fixed hardware, fixed network settings, fixed, and, and AI creates a level of unpredictability in your architecture. So robustness means you now need to either acquire the skills, which is what what we do in AI solution architecture, acquire the skills, or at the very least, right, get someone to help you think through what happens on your best day and how do you pick an architecture for your AI solution that can scale with demand. It can go up as much as it needs to, or it can go down as much as it needs to. Many times, this is a very, like, a very specific architectural choice you have to make, and you have to make it. At the beginning. Otherwise, it's extremely expensive to change later. Tech debt, right? So robustness is something that most people never even think about because they're in baby tiger mode, right? They're like, oh my gosh, it's so cute. I can't imagine any more than, you know, 100,000 people using it. And then it goes live, ChatGPT, and a million people use it a week. The AI systems I built at Microsoft, we were getting about 100 requests into those AI models a day in research. Today, those exact same models, they've evolved of course, but those same models we were configuring for production now get 5 to 7 billion inferences, billion per month, but still billion. Like you have to think in billions of inferences on day one, baby tiger mode. And this robustness is, you don't have to create that infrastructure. That would be a waste. But you do wanna be in an environment that's flexible to grow as you grow. Otherwise you're gonna be doomed to repeat. The development process. So those principles should govern, right? So the first document I create is a strategy that provides governance on how I build my AI, which is the process we use this week, business requirements, functional requirements, right? Model selection and evaluation. And then an application, a literally a checklist for how I ensure that every single model adheres to these principles, right, to these. And that's only 5. If you Google responsible AI tenets, every company has decided to create their own set. So you can see lots of variations here, but you want to resolve how to do that in design, not in production, certainly not production, but not even in development. The second document is maybe one that would surprise you, but most companies don't have, which is an AI usage policy. Give guidance, prescription to your employees, or to your clients, here's how you should use this thing and here's things you should never use it for. You'll see that in my guide, in my, um, prompt repos. Never do these things. You wanna give very clear descriptions. So I have a usage policy template, um, that I use for my clients that just allow them to send an email to their employees to say, here's how we use this, here's how we don't. I always say like, It does not go without saying. So sometimes we're like, ah, everyone knows this. I would just beg to differ. It doesn't hurt to tell them. So create a usage policy that clearly defines your expectation for AI use in your company. And then the final one is what I would call like a COE, right? Or a center of excellence, a group of people that usually represent— I said this on the podcast this morning, I was like, you wanna include in your COE, in the people that are gonna kind of manage AI early on, you wanna include in this, uh, group all the people that you would likely call last in a software engineering project, right? So who do you not want to talk to first? Legal, security, DevSecOps, operation. Like, who do you not want to talk to? Because likely they're going to be like, well, you have to do this. And like, they're gonna make it hard. If you invite them earlier, the good news is, is AI's gonna make the hard easier. So you, you can welcome them in, in earlier. The bad news is if you wait till later, the cost to change things, to do the things that they would've told you at the beginning to do is much higher. So it's just better to build your COE with all of those people, legal, security, um, uh, HR. I mean, anyway, include HR. Why? Because they know the humans., and they'll be able to help you really make sure your AI system meets their needs. Um, all these people that often we don't consult because we're whatever in our mind, we're like, oh no, if I talk to security, they're gonna ruin everything, right? If I talk to HR, they're gonna take out all the stuff that I like. Um, but if you invite them early, you can have a conversation of how you meet everyone's needs together. It requires human connection. Aw. Um, it's, it's foreign. Many times when I build these COEs in companies, like, Who was I at? A financial company in Charlotte. I'll just say, um, Charlotte, North Carolina. I bring these people together and they say candidly they have never been in the same room together ever. So that is a sign of success. If you bring people together at the executive level or at senior leadership and they've never been in a room together, you've got the right people. Um, that ingest a little bit, but you should look around and people should be like, wow, marketing has never talked to security before, or You know, uh, HR has never been in the room with, you know, DevSecOps. Like they, this is a huge opportunity to give, give you some really good indications that this is the trend that big companies are moving in, which means we should be thinking too, is that like Boston Consulting Group, their HR division and their tech division combined, cuz in the future HR is gonna own the human side of the HAX equation, right? Human AI experience. Human AI experience means HR and tech will live together, not the entire org, but there will be a new kind of organ, a part of the company that will manage the human understanding. What do I know my humans are capable of? And the AI amplification of that humanness. And that is what we're seeing big consulting firms, uh, double down on is how do I harness my human potential and then focus my technology teams on only building things that amplify that. And that's what's winning, right? You want to know about the 5%? That's what's winning. Let's see. Okay. Oh yeah, great, great question. So Sue was asking about the difference between a constitutional AI and AI governance charter. So a governance charter, and often a governance charter is kind of part of this COE. It's the rules upon which you're going to run your AI systems. AI governance, like data governance, right? It's how do you do the thing? How do you manage AI? It usually includes rules for how, uh, use cases will be triaged and how you'll measure success, how you'll, uh, manage impacts of failure, how you're gonna mitigate risk. Constitution is refreshingly different in that it's gonna resolve, uh, to your core values. How do you know that this AI project will, uh, lead with integrity, will always be accurate, will, um, you know, be like, what are your core values of your company? The constitution. And again, it could be anything like constitutional AI just means you're giving a set of instructions to a model and its job is to review things and observe those instructions. So you can write your constitution on anything, but the spirit of constitutional AI is that you can now use it to instill behaviors on a model that aren't part of its governance, right? Aren't part of its data strategy or privacy strategy. It's more part of the ethos of the company. And you can now weave into the DNA of every AI project an ethos check. Are you the, like, just like we would with, um, at Amazon. I hired a lot of people at Amazon and we did a little bit of a functional check to make sure you can do the job. But the biggest thing we were interviewing for was, are you Amazonian? Do you adhere to our leadership principles? And I want to know every time an AI project is about to do something new, does it adhere to our leadership principles? Does it adhere to the way we want to do business? And I could count on one hand the companies that are building those types of agents right now. They want to, but they're not there yet. So this is a huge opportunity, I think, for you all. Oh, I don't— I'm not going to reclaim host, but so you're the host. I won't hang up though, I promise. Um, anyway, so that is the, the way— the encouragement I give you as you move out of the nest is to actually kind of reverse engineer what we did this week. We did as a microcosm of a single unit of work. You're really just going to extrapolate that to higher levels. But I start when I go into a company, that's why I call it a minimum remarkable product. I start with one thing. And then I, it's a 4-hour workshop that I deliver to them. I call it an AI agent accelerator. Y'all can do it. You just did it, right? Where you build first the problem statement, then the requirements, then the guardrails, then you build the agent. And then at the end of 5 hours or 4 hours, you've built a responsible AI system end to end. But you did it in a very specific prescribed way. And now the documentation that you build is an just an extrapolation of that, right? It's how are we going to manage, right, all of these systems? What are the rules we want in place? What is a definition of done? It's kind of like, do you know the concept of OKRs? Another great book, right? Measure What Matters. So OKRs start at the top of the organization if done correctly, which many people don't— it doesn't get done correctly. But if it's done correctly, it's the large-scale objective. Why are we using AI in this company? Right? How are we using it? What are the things we're worried about? How do we mitigate risk? And then everyone in the company, the marketing team, the sales team, right? They then create derivatives of that big OKR into their OKRs that feed directly into that larger statement. This is hard work in most companies because these teams don't talk to each other normally. But it's what we, what, what is necessary when you build these AI systems that are going to span across domains. So documenting this process, like I said, I use 3 documents. I build a strategy document, right? I, and again, I called it a COE. It's actually an entity, but I document here's what managing AI looks like at scale within your business. Now, of course, some of you must know this, but I wrote a book on this. And so all my checklists are in this book. It is available to you. You can go get them. So they are also, on responsibleaibook.com, I think. Yeah, responsibleaibook.com. So like I wrote the book and I documented all the checklists that I use. So I offer that to you as part of kind of the culminating experience that what we did today was what I do with my clients. It's the tip of the spear. It gets people leaning in. They're like, oh my gosh, this is pretty cool. But then the responsibility sits in, right? Then my job now is to be like, isn't this awesome? Let's actually prepare this for production. And the appetite for doing the work that's required is much higher when they have a minimum remarkable product that's staring them in the face. Their dream realized is right there. It's just not ready yet. And so then I prescribe these steps and they're much more willing to show up in those meetings and do the work cuz they can see the result. And as a matter of fact, as a habit, I demo that product before every meeting. I'm like, this is our goal. This is the North Star. If you want this though, we need to do, you know, the needful things, right? I mean, data privacy conversations, not super exciting. Data governance, data, um, even just understanding your data landscape, doing a data inventory. Like these are things that normally I couldn't get people to show up until I flipped it to, let me build it first. Thank you, Lovable. And let me show it to you. And now this is gonna drive your interest in getting those activities done. And it's worked really, really well. So thank you all for being here. I'm, I'm a little early when it comes to like leaving y'all. I do want to mention very briefly that we have the second step of this. For those of you who are building a business, we do have an AI Business Accelerator. It's a 6-week program. I'm going to throw the link into the chat. I created a few years ago this cohort, and the reason I did it was because I was launching a business at the time. Which now is doing very well. Like, you know, my goal is always to hit 7 figures before I talk about it. So I built this company to— I built this program to kind of go through exactly the process I used to hit 7 figures in this company. Um, so you can go and see what it is. It's actually very— I don't know, I feel like you could ChatGPT it, but what I have uncovered is that it requires discipline and daily activity. And so for 6 weeks, you will— it's not homework like we did here. We're like, hey, have fun, go build an agent. Like every day you're building a company, right? So you're gonna build lead magnets, you're gonna build targets, you're gonna build sales, uh, sales scripts. Like it's very, very detailed, but it's fun. I like it. I've done it 3 times now. So I wanna invite you. Our next one happens at the end of March. AI, uh, oh goodness, let me— I can't talk and type at the same time. So hold on, let me get it out. AI Startup Studios. So we do, uh, not everyone, you know how they say, not everyone's gonna be wanna do this, but you might wanna do it and you just might not be ready. It does require some knowledge of what you wanna do in the world. Um, so if you go to AI, uh, let me click on it, make sure I put the right thing in there. Oh yeah, that's me. You could see my face. Um, so 37 days from now we are kicking this off. This will show you all the things. That we're doing. As a matter of fact, da da da da da da, where are you all? There you are. As a matter of fact, you can see the 6 weeks are built out. The first thing we have to do is figure out what you wanna do and crystallize that into a message people care about. Then we create a lead magnet, then we create a sales funnel. And so it goes through this process. So by the end, you end up with 10 clients. I will tell you, getting 10 clients requires you to say out loud that you're doing something. So we're gonna use AI to accelerate the entire process. But at the end of the day, people don't buy from AI. And I know you might hear people on Instagram tell you different. They buy from humans. They buy from people they know, people they like, and people they trust. So our goal in this program is to find 10 people who know you, like you, and trust you and are willing to buy from you. And I'm not gonna say build a $10,000 product though. Hey, do that. We're just going to show you the sequence so that it works, and then you can redo that sequence over and over and over again, which is what I did. I started with a $27 product. Now my high ticket is $25,000, right? So like, but I didn't own a $25,000 offer. I didn't even believe I could ask for $25K for anything that I did, right? Like that was a mindset challenge for me, but I started at $27, and I use the exact same process. I just increase the value of what I'm delivering and increase the price to match that value. And so that's what we'll teach you is we'll teach you the, the process so you can rinse and repeat it to do whatever you'd like to grow the business as big or as small, uh, as you want to. So take a look at it. If you're interested, you can apply. Um, I think the price is on the website. It's not a secret. It's not super expensive. You can also break it up. Um, it, it does cost something. I try to give as much value as I can in this session, as you can tell. But at the end of the day, like, it's just a piece of the puzzle, right? Building an AI employee is great, and having a social media agent is great, but it's way better if it's all driving towards something that creates, I don't know, a financial asset for you, a business, right, that you can live off of. And that's, that's my goal is how can I help small businesses become big businesses? And then you can choose to hire people and create a good culture and be a good manager. Um, something that I didn't have as I was going through my journey.