ai personal productivity - ed watal

transcript.txt

We all know about user personas. You've heard about personas, buyer personas. So imagine if you as your business created a buyer persona agent that you could ask customers to. It's like literally hiring a customer and having them sit in your business all day. And they'll answer all your questions. And just imagine the power of doing that for your business. So now, and that's something we recommend any business should do. Now the question is, how many do you create? And literally, we could do this exercise today if you want to. And let's do that. If you have your computers, let's pull them up. Let's go and create a buyer persona agent. And let's see if it works. You want to do this exercise? Yeah? OK, let's do it. Let's pull up ChatGPT again. Yeah, you need the paid version. Yes, you need the paid version. So yeah, you could use Claude. Yeah, again, you could use Gemini. Yeah, I just keep gravitating to ChatGPD because that's what I use all the time. No, I have faith in Sam Altman but no stock in him. That's good. Yeah, so you agent mode. So the one way you could do it is you could create. Have you ever tried to create an agent? Okay, let's see. Yeah, create GPT. MyGPTs, if you see on the top right corner, create. Oh no, I think it's, are you in your premium account? Plus account, yeah. You should have, if you go to new chat, I think, on the top right corner, it should show you, oh. Group chat. Your interface looks different. No, I don't think so. Click on explore. Click on explore. So, yeah, so strange. I don't see library codecs at this. Click on my GPT's. My GPT's. Here, on the left. Yeah, and then create, yeah. Create GPT, right there. Yes. Where do you want those? Sorry, yeah, you can go to my GPTs or GPTs on the left where it says GPTs. No below under Codex GPTs. Yeah, click on that and then create. On the right. Create GPT. Yeah, you got it. Everybody got it. Everyone's there, yeah. Yeah, explore and then create. Yeah, create. exactly so now you can start to describe what your ideal customer is like and then profile this it this GPT to be your customer you could give it a document that you have saying oh and give it as much information as you can then you can ask it questions and over time And you can specify. So if you wanted to create a specific persona, a simple exercise would be you take your buyer persona. If you have a buyer persona document, you just upload it in there. It's the easiest way to do it. Because you would have your marketing would have produced a buyer persona. You just put it in there and you're done. It's as simple as that. And then you have to give it an instruction saying, hey, every time I ask you a question, behave like this buyer persona that I have attached. That's it. No more complicated than that. And then you can train it over time. In agent mode, you're changing your behavior to that because then it's going to think that's you. But if you create a GPT, it knows it's clear separation, it's a separate thing from you. Otherwise, in agent mode, it'll start to think that's how you think, that's you. No, you're not doing agent mode. You're just doing a regular GPT, yeah. Yeah, on the left, if you're here, you click on explore. Yeah. Yeah, you could even try it on your phone if you wanted to. Yeah, yeah. You can. I mean, definitely you can do it out of Copilot, but right now I'm just trying to get... Yeah, you can do it out of any tool, out of Gemini, out of Copilot, any tool. Yeah, it's not specific. Yeah. Okay. And you could think of, we're thinking buyer persona in terms of someone you're selling products to. You could model any persona within your enterprise. Imagine you're a recruiter, you're trying to hire somebody. You're gonna have a conversation with a potential candidate. How do you discover the right person that'll fit this job? You have a conversation about that. And that'll help you discover how to write a better job description. That'll help you discover... Like all your outcomes can be accepted. So I wanted to try something like a candidate. Yeah, you could. Yeah. So are we writing this in the first person? Like I am this? Or is it better to prompt it with more like pretend you are? You pretend you are. Like you have to pretend you are. It's like, you know, assume you are. I use the word assume, not pretend. I assume you are, right? Does this use a different large language model? It's the same GPD algorithm. It's the same model. So I'm trying to understand how you use this mode to interact differently in the chat. It is just containing that knowledge into that bucket. You already have one? So if you were in chat mode, then tomorrow if you ask a different question, it'll think it's you. Now it knows that it's not you it's having conversation with, it's this extra thing called GPT, it's a separate persona. So it boxes all that conversation and interaction with that. So you're going to get different outputs? You're going to get different results and different outputs. So if you ask the same question on your normal chat versus under this GPT, you'll get slightly different answers depending on how you customize this GPT. Why is that? Is that because in the chat mode, it's trying to please you? In the chat. So everything boils down to context. So it's going to respond based on the context. And the LLM has a context window, a certain context window in which it is responding to you. You might have seen that if you keep asking a lot of questions, sometimes it deteriorates. It kind of forgets what you told it in the beginning. So you have to go and remind it. because it has a context window, a length, that every time you're asking it new stuff, it's sort of compressing whatever it knows. It's kind of like even how our brain works, right? You kind of forget whatever's from the past. You keep only the useful stuff. So it's doing that compression, and so if you want it to not forget a certain aspect, you're going to create a separate GPT. So then that retains all the context over a period of time. And again, even that'll have deterioration, but then, you know, within a given chat. Yeah. You want to create a GPT that others can use. So the whole point of creating a GPT is then you can publish it. So today your chat, you can't really, your projects, you can't really publish to anybody. And if you did, people ask questions, then it becomes part of your memories. So when you create a GPT, whatever questions people are asking, that doesn't mess up your memory, your context. Sorry? Yeah, yeah, GPT, once you create it, you can share with anyone. So you can create one and share with whoever. Within a group, you can even make it public. Put it in the GPT store, like an app. Yes. So... And then you describe your persona and everything else. And if you click on configure, you can configure it further. So are we able to make one? Everybody? Anyone struggling with making a GPT? No? Very good. What happened? Yeah, I think you're going to have to just ask it to do it again. And I think in some of these, let's look at click on. Let's click on configure. You updated all this already. Okay, perfect. It's all in there. Okay, yeah. Then, yeah, I think. How do you know if you created a GPD? No, you have, because I think it says whole life customer persona, right? It says draft. You click on that. Click on the back end. And then do you, under MyGPTs. And then you should see it there. And then you can go edit. Yeah, that's ready to go. And then if you go back there to MyGPTs, you click on that three dots. Sorry, not here, only me. I think you can change who can see it. Right now it's only you, you can take a URL and you can share it with others and so forth. Yeah, and then you can prompt it more. The more you prompt it, not here, because this is actually the customer mode, but here is where your creation, under creation. Whatever prompts you give, it will keep learning. And then you can click on configure. I think you already did that. Yeah. And if you want to add to that. You can add to that. You can add, you can upload files. If you go down recommended model, you can choose the model you want to use. Yeah. Yeah. No. Yeah. Exactly. Exactly. Yeah. Yeah. Wait. Are we able to all try it and it's working? Okay. Cool. We did it? Okay. Oh, perfect. Did you make more money yet? Yeah. No, that's a great question. That's a great question. Yeah, yeah. Yeah, and I think this is a great use case, and now you keep this in a sandbox, because once you've created a GPT, it's like a sandbox. You keep training it, you keep asking it more questions, and now you've got a customer in-house for free. You know, 25 bucks a month, you have a customer working for you. I don't know if you remember the days before this, that you would actually have to send a customer in, focus yeah i mean there's some maybe some soft aspects to that yeah but yeah Everybody able to walk through this? Yeah? You did? Nice. That's cool. Now you could run agent mode on that GPT and ask it to pull all the competitors' pricings and then ask it to recommend your pricing. Yeah, exactly. I mean, that's a little more involved conversation. All we need to think about is what data you want to expose to who. The best way to do it is, yeah, the best way to do it is, it goes back a little bit to governance and guardrails. So you do need human in the loop, because if you don't have human in the loop, it's going to say something, right? I think I would highly recommend using something like this internally first and abstracting people away from the chat GPT interface. Otherwise, it seems like you're promoting chat GPT and then it locks you into one solution. But whatever your app is or whatever your interface is, you can go against the APIs of this tool and then build it in there you know you can use the the value that the tool is providing through the through its apis yeah yeah more than likely yeah thank you you're welcome Because they don't care. All right. So now we all feel good about this? Yeah? We feel like we learned something in the process? Good. So now we're going to try the The Delta Learner, yeah? You want to try that? How would we do that? Again, you can't really try Delta Learner because it needs to know everything about you. But we can run an experiment, another experiment, like what we did. Because I think getting a little more hands-on will get you also more comfortable with using it in other contexts and use cases. So let's try another exercise. How many of us have actually gone and Googled our own names? We all have, right? We all have at some point, yeah? Yeah, we've all Googled our name at some point, yeah? So how about this? Let's find out how much does GPT know about us. Don't do anything yet. Don't do anything yet. Let's find out how much does GPT know about us. And you'd be surprised as to what it knows. More like scared. No, I mean, you can Google me. That's fine. You can actually go to my website, edbattal.ai. You don't need to Google. But that's probably easier. So let's try something else. Let's get out of that agent, go back to your normal chat GPT so we're not stuck with that agent and we end up talking to that agent. So let's go back to our GPT, yeah? Are we back to the GPT chat, yeah? Okay, all right. I would recommend that you, unless you intend to share it with others, I would recommend you create a new project so that you don't mess this up with whatever else you're doing. So just say create a project and put your name on it so you remember that that's you, yeah? create a project with your name on it easy to remember so everybody should have a project with their own name on it yes everybody has a project with their name and if you already have one then you can try something else yeah you got it everybody in it yes okay next step go back to the chat and let's do these steps are you in the chat okay let's type Let's find everything. You can just type this. Let's find everything possible about, and put your name. right and your title and say everything possible let's find everything possible about but don't hit enter yet yeah no no let's let's right it's a collaborative effort with gpt let's find everything about and then put the full name and don't press enter yet right yeah you've got it huh yeah and your name and your title and your company right so a new project and inside the project a chat Yeah? Company as well? Yeah. Name, title, company, whatever you want to say, right? Company is not needed, but I think company is fine. Yeah. Actually, it'd be easier if you give the company, right? Because there could be many people with your name. No, it won't. It won't. Because you're saying let's find everything about, right? Then you click on that plus button again, you remember? The plus button, and then go and say deep research. Okay? And then . Let's find everything about your name and company and title. What kind of information are you looking about, Jeff Ross? Could you please clarify the following? Oh, did you do a research mark? Yeah. I just hit all in the question. Okay. Yeah, okay. You're kind of... Give me everything. That's it. I think this is just Sam trying to save tokens. Just say please. Yeah, just give me everything. . Yeah, well, that's a lot about you out there. . And whatever you find is strictly confidential to you, you don't have to share with anybody. It'll be interesting how you find a combination of, it'll find stuff about you, it'll present it in an insightful way, and sometimes it'll just imagine stuff about you, because it may hallucinate. Or may find someone else with the same name. And it could die as well. So it's just an exercise just to see the art of the possible. And the reason I encourage you to do it with your own name because that will give you a lens of how much to trust the answers when you do research on some other topic. and it'll give you a sense so because and that's the point of the delta learner in the context that you also know what the machine really knows not just that the machine knows what you know technology Thank you. Yeah, I think it's going to be a while before it comes back. So we were just saying, like, where is the setting for you to give all the niceties, you know, how it kind of flatters you? Like, say you just want it. So, yeah, so we can do that. We'll do that after this exercise because otherwise we'll... So in GPT, there's something called custom instruction. So you can go and write a custom instruction, which essentially tells it what to do and what not to do. Now, there's some best practices around that I can share with you. I can share with you what I use. I don't recommend using that necessarily, but if you think it's useful for you, you can try that. . That's fine. One year. Is it halfway down for anybody or not yet? About halfway? Yeah, you can keep clicking through on the right if it's showing you the... How come we don't see the activity panel? No, you're not. Okay, interesting. Do you know a friend of mine named Lisa Molina? I just saw that you're 50-50 Lisa. Is she in Chicago? I just had dinner with her last week. We've known each other for 30 years. Is she a military person? No, not at all. But she has a deep affinity for war. The military. She went to business school with one of my closest friends from high school. And that's how I got to meet her 27 hours ago. She's awesome. Just given what she can do. It is scary, right? It is summarizing it for you. Google search wouldn't summarize all this. It's finally going to give you a bio, like a report, and probably be the most comprehensive bio you've ever had. You could. Then you could take it and say, now turn this into a resume. We'll do that. That's the next thing. And then you'd have a really nice resume. She's an amazing human being. She was the vice-chairwoman. Did we get anything? Not yet? It's still working through? She's definitely connected to this. She's like a master at that work. Yeah, I can see her connected to this. Not a lot of them. There's a lot of competitors. I mean, Salt, I put her at the top, 0.1% of humans on this planet. She's smart and good. Everything. And everything. Does she know you think that of her? What's that? Does she know you think that of her? Yes, she does. She's been in a number of boards. I think she was head of... She's head of one of the subcommittees of the board for the DTCC. Okay, okay. Oh, I know the legal general counsel. Then my guess is that you will find people. She definitely knows. She has an intersection also with government and sovereigns. So you want me to give you right from there? Yeah. Sure. What's your email? It's a whole different world out there. Everybody knows each other. Where's that? Where's that? Where's that? I briefed years ago, there's an organization, Wall Street, 85 Broad. Oh, of course, yeah, because that's cold. So I briefed probably a hundred. Oh my God. They asked me to talk to them. Is it still called 85 Fraud? I don't know. I think it turned into something. 85 Fraud. Well, because they left 85 Fraud. Yes, yes. Right? They built their own. I love that title. Yeah, it's great. 85 Fraud, right? Although, I was like, we may not be set by the... broad term. Isn't that somewhat derogatory? You can say that. Does that have your arrest reports in it, too? Exactly. No, I got those in Spudged. They're in Spudged. When I got on the recent . Unbelievable how deep it was. Like it's just, they wanted to do, they wanted me to be showing cash flow analysis. Okay. Done for anybody, no? Not yet? You're finished? And it's done? And you've got a report? Okay, perfect, you've got it. And how long is it? Oh, you can open it? You can download it? and all the other . The OCC by far, although the fact is, the OCC has this granular level, and it comes down to sort of personalities, because we would talk across all the things, not in a way that disclosed anything, you know, about MRIs, MRIs, although some did, but I've never heard of them. Anyone? You've got a dossier. You've got a dossier in yourself. Perfect. Okay. Everybody got a dossier in themselves? Oh, you've got it? You've got yours? Everybody has a Google account? Gmail? Yeah? Now you can simply throw that dossier that you've got into a Google Doc. And we'll do something else with it later. Yeah? All right, let's do that. Yeah. The only other time I've seen managing director partners. You can throw that into a Google Doc. Everybody, yeah? You do not have another chance. Yeah, you can throw it into a Google Doc or you can download it onto your computer as a document. Yeah, you could even download it as a doc on your computer or a PDF. And then we'll use that document later. Once we have it, yeah? So, you were not around. You were not around. And it's a way of going back and saying to the regulators, oh, we didn't like him either. Oh, that's not great. Right? It's gone. But you have to align. You can download that as a document on your computer. or put it in a Google Doc. No, you can just select the whole text and then put it in a Google Doc and then download it on your computer or just save it in a Google Doc. Ideally you can download it on your computer. You've got it? You've got your file? Still loading? You've got it? You've got your bio? Yeah. I'll send her personally. Okay, perfect. Yeah. So whoever's got it, please copy it into a doc and save it on your computer. Thank you. You already know Randy Snook and John Edelman. No, I don't. But you should go back to Randy. Yeah. I think whoever has a massive online presence is going to take a while because it's looking through everything. I'm going to give you one of my blurbs, Patrick. Melville Feinberg exemplifies the modern board director. There is a wealth of financial acumen risk management strategies. If I said her that, she'd be like, leave me alone. She'd be like, get away. You got it? Okay, now you can just save it on your computer as a doc, like either download it, I think it should allow you to download it. Or you can just copy the text and put it in a rule doc and just save it on your computer. It works. And I did him, and he did me. Eventually, Claude was saying, you're missing a lot of personal information about Michael. What are you going to want to do? That's funny. This thing has given me a new skippiness about how good I've been. Yeah, it's good. A little motivation. Pop up your Nikes. You didn't prompt it and say, please only speak truth. Calm down a little bit. Oh, I put it down. Distinguished. How do you think? Please. You think that's your career? Okay. That's right. I'll give you a little hint. I think that's bullshit. That's funny. It was really nice, but I didn't see it. So here are some 10 tips on custom instruction that you can give to GPT to not get some random answers or random behavior. So this just guides GPT to say, hey, never mention your NAI, for example. If your custom instruction doesn't say that, then you feel like you're talking to a person. Otherwise, it will make an excuse. Oh, I'm an AI. I can't do this. So this is a guardrail. These are just guardrails that you can put. And you can come up with your own guardrails. But these are like 10 commonly used guardrails by people across the internet. I use them pretty much. You can add your own. Never mention you're an AI. Avoid language that constructs or could be interpreted as remorse. Because you don't want it sitting and regretting and being sad and unhappy. Because that's not going to be good for your own mood. So you don't want it to sit and regret and be remorseful. Because it does that if you don't tell it. I'm so sorry. So you want to create a positively vibrating GPT, which thinks abundance in its head. Sorry, apologies, regret. So you ask it to discard all that. Then refrain from disclaimers. Oh, I'm not a professional. I'm not an expert. It does that all the time. So you ask it to refrain from doing all this. So you know it is GPD. And it knows what it knows. It doesn't know what it doesn't know. So it doesn't need to tell you that every time. Yeah, so if you want to do this, you can literally take a screenshot of this. I'll have to get out of the way. But you can, and then you can extract the text first. And then you know how to do that. Actually, that's a good exercise. Take a picture, throw it to GPD, say, hey, give me the text. And then take the text. Take the text, or I could email this to everybody, which would take longer. Probably easier. Take a picture. Say, hey, take the text out of this. Take that text and go to your preferences. You click on your name on the bottom left corner. And then when you click, it's a personalization. Click on that. And then under personalization, if you see, there's something called custom instruction. There's a box called custom instruction. You literally paste all that text in that box and save, and you're done. And then all of a sudden, you realize, next time you use GPT, you feel like, whoa, this is so much better. Yeah. Just literally these 10 custom instructions will suddenly elevate your GPT game at least by 2x, if not 10x. Do you recommend doing something similar for the agent? Yeah, absolutely, for the agents as well. You can give it a tone. A tone, exactly, a tone for the response. And you can play around with this, but by and large, if you go too fancy, then it just gets messy. And if you don't give anything, then it's too basic. So this is sort of like Goldilocks custom instruction. I would add to it one thing I found a lot is that because it's doing a lot of web searching and content data, it has data that's over a lot of different time periods. So the problem is you want to constrain it to say prioritize more recent or now data than older data so that you get the most recent information because sometimes you'll get answers that are less. I'm sorry, Manish. What's the thing to put in? He's saying that you want it to prioritize new information over old information, but it's It depends on what the use case is. Like if you're a historian doing research on history, then it doesn't matter. But if you're trying to look for news, then you probably want to. But then sometimes you may miss out on important, like that's a signal, that's the alpha bear question. Because the moment you start to do all this, then you're getting in the alpha bear zone. You know, what is relevant and what you find is useful. Yeah, so we got the custom instruction everybody. Yeah You don't have to you can add just one but if you add all 11, you know, it's useful You could add definitely the top 10 the 11th one used to be relevant back in the day when GPT would keep saying Oh my cutoff date is September 2021. So it's not as useful to add number 11 now because earlier they had they weren't connected to the internet and So that was useful when it was not connected to the internet. Now it's connected to the internet, so 11 is kind of redundant. That's why the top 10 are the most useful. And 10 is useful only if you're asking it to research. It's not useful for other things, but you can leave it in there. It doesn't hurt. Personalization, then custom instruction, and you go drop that under custom instruction. And what this does is it doesn't matter what project you're in, or what chat you're in, it's going to apply every time. Where's that personalization? Bottom left corner, where you click on your name, where it says your name. And then personalization, and under that, custom instruction. And you just paste it in there? Yeah, you just paste it in there, and that's it. And then just save it. I mean, it automatically saves it. I don't think you have to click a save button for that. Yes. What is a personalization? Yeah, here you click on your name. Personalization. Yes, you've got it? It's better to keep the numbers. Yeah. If you don't give it the numbers, yeah. So when you're putting your custom instruction, if it's not numbered, then it may not have the same outcome. Do you guys understand why? Yes. So you just copy this? Yeah. No, you first give this image to GPT and ask it to extract the text, and then it'll extract the text. copy and you just paste that to a new chat in GBT. You go to GBT, create a new chat. get out of that, create a new chat, and then just go paste it in there, that image, and say, give me the text. Perfect. Numbered, with numbered bullets, right? Text with numbered bullets. And then we'll give you that, and then you paste it in there. Yeah. Yeah. use these 11 instructions and update my custom instructions. And it actually just did it. I didn't have to do any copy-pasting. Brilliant. Yeah, you can do that, too. What did you do that? Oh, wow. Attach the photo. Yeah. Yeah, that's smarter. All right, so now we've done custom instructions, yeah? Everybody got it? Yeah, everybody got their research back about themselves? And we've all put it in a doc and saved it on our computer? You never showed up? Okay. All right. We'll work with what we have. Okay. Everybody else got there, saved on their computer? Yes? Everybody? Yes, yes, yes, yes. Yeah. Yeah. You just copy the whole thing and put it in a document and save it to your computer. Yeah. As long as you have it in a Word doc or a PDF on your computer, that's all you need. Maybe somebody with your name, if you weren't. Okay, we're all good? We got it? Yes? You have it? Okay, perfect. now for the next part you're going to need a google account gmail account you have it yeah i think everybody has it so you can go to your google account log into your gmail you logged in i think everybody is yeah logged into your gmail yes okay OK, if you're logged into your Gmail, now what you can do is go to this link on the screen. That's at the top. I don't know if you can see it. I'm going to make it bigger. Yeah. You got it? You got your document, custom instructions done? Yes. Now, you can go and say, create notebook. Create new. If you're there, then upload that file that you just created. Oh, okay. That's fine. You can just follow along with somebody else. You just see it and you can do it later. Yeah, create. It should be like this. No, no, no. You go to notebooklm.google.com. A new tab. Yeah. Notebooklm. Google.com you got it yeah yeah okay create notebook and upload that file you do that okay perfect now what you could do is click on you can do video or audio but video won't show your face because there's no video there but you can do audio overview Oh, mind map, we'll create a mind map, but let's just do the audio overview. Yeah. Got it. It's generating. I think it's generating. Yeah. And then you can do audio overview. What's the next step? Yeah, so are you on that URL, notebooklm.google.com? Yeah, Google URL. You've got to go to that URL, notebooklm.google.com. And a new tab. Don't close your GPT window. Go to a new tab and... Notebooklm.com. Notebooklm.com. lm.google.com yeah and then just say okay create new create new and then upload that file that you just had word doc PDF doesn't matter the one that you just created about you you just upload that file You got it? And then click on audio overview. Once you've created it, that's it. And then you've got to wait for it to generate. You can create audio overview, video overview. mind maps and now you can picture doing this for anything like you would take a document create a mind map out of it you can take a document you know you could take a group of documents and create a mind map out of it very quickly so imagine someone sent you a whole bunch of stuff and you want to quickly mind map it you can take 50 documents throw it in create a mind map in like five minutes or less you could do all three by the way yeah you can do of that yeah you do all of that yeah yeah yeah yeah I don't think enterprise I don't think enterprise restricts you from doing that Oh, yeah, yeah. Then it just, yeah, yeah, yeah. Anyway, and the reason we chose this exercise is because you're taking stuff that's about you that's already on the Internet and you're giving it to Google. Like, there's no loss of data here, right? Because you're not losing anything in the process. I put the file on the chat on there. What? Not on the chat, on the road book alert. I think it's just getting comfortable with the idea that your ability to use AI tools right and your ability to sort of work through between different tools as a person that's all we're trying to do right and it's more a personal exercise and then the use cases you'll naturally discover because it's a personal exercise it's easier to remember it becomes harder to forget yeah and i mean as nyu instructors a few of us here i mean we use this a lot Yeah, yeah, I'm pretty sure you do. Yeah, yeah, yeah, yeah. Yeah, yeah, instructions, yeah, yeah. Yes. Oh, perfect. And then you click on that. And then you wait for it to generate an audio. It was not bad, that team. I think for businesses, the value could vary depending on what you're trying to do. For example, like Stephen here is saying that we use it at NYU a lot. But for businesses, often people are learning. Learning happens in different modes. Some people are visual learners. They prefer a mind map. Some people are auditory learners. Some people are audio, visual learners, they need a video. So depending on the learner, and the key here is, and this just elevates the means of communication within your business. So if you want to communicate something even within a group very quickly and depending on You could literally create all these three modes and put it out there. And it just elevates your outcomes. Thank you. That's the delta learning. Exactly. That's the delta learning and how much of your knowledge you want to share. So, the point of doing this exercise and Stephen is asking this question, what's the point of this exercise and how would you leverage? Obviously, one is just getting familiar with the tools and using different tools. But within the enterprise context, sharing your GPT with everybody else may not be practical. But what this coordination does is you can take pieces of what you want to share with others. And instead of putting a blog or whatever else, you're creating these different formats. And you can even have somebody do it for you. Because at some point, it's mechanical. So you can even automate it. And if you have a software engineer, they just automate. this for you. You just drop a file in a SharePoint folder and then boom, all of that is published wherever you need it to go. And I'm sure at some point there will be tools that will emerge that will do this for you, GPT tool, whatever. Yeah, I mean, Notebook LM is actually a fantastic product, but the product manager who actually did it left Google. That's kind of sad, but if she was still around, I think it would be a pretty incredible project. I think she's doing something on her own. OK, all right. So I think we've all got a feel for this, yes? Everybody able to generate mind maps or audios or videos out of Notebook LM? We know what it is. We know GPT, custom instructions. And you kind of get a sense of how we would apply it in business. So what we'll do now is just go around the room and try and ask everybody to sort of give their lens on what they learned in the process of this exercise. And think of one area where you think you could use this on Monday without breaking the law. blowing up your enterprise. So think of one way, and we'll just go around, we can move the microphone around, and just, you know, whoever wants to start, just want to go around and say, hey, and again, no right or wrong, even if it's just an idea, feel free to throw it out, that's conversation here, trying to learn from each other. To say, where would you use this learning, and how would you use this learning, that will give you like an immediate aha, like an immediate bang for the buck on Monday. Yeah, please. I think the buyer agent was very helpful. So for Central Park Conservancy, Josh used the persona of a donor, and I used the persona of an ice skating rink, somebody who's interested because we just opened an ice skating rink. But because it was so new, it was renamed, it was clear that the public domain didn't have enough information because there's two ice skating rinks, right? And then it's been renamed to Davis Center, but was previously called Lasker, so it was confusing itself. So I think this is going to be very helpful to our marketing team to figure out how to get them to use the ChatGPT because you need the public domain to have the information. It just doesn't have the information right now. That's helpful. Thank you. In the same company, obviously Michael and I are on the it side. We're talking about doing a chat bot, not so much a chat bot, but a automated process for people to open up support tickets. But in looking at this, I we've also been talking about on a top level personas for staff. And I think it would be interesting to start creating personas for our staff because we. Central Park Conservancy, as you imagine, we've got some business end people, fundraisers, human resources use technology a lot, down to like guys who mow the lawn and trim the trees who barely use technology at all. So creating personas where we can then kind of understand the best way to interact with each of those users. Yeah, that's great. I think for me, it's not an aha moment. I use these tools all the time, but I think this exercise of literally showing people how to go from one tool to the next and do simple things is really powerful for some senior executives who are not maybe necessarily using the tools. I hear this frustration from my C-suite all the time. Where's the information? I can't find the doc. It's in my SharePoint. It's not in my SharePoint. Why can't I edit it? And I'm like, what are you guys doing? So I think tools like this, bringing a simple use case like this to senior people will open their thinking to how to use a technology for, I would say, business productivity. Thank you. Okay. Anyone else? No? Yes? All right. So now let's think about data. So we say, okay, this was, you saw, you tried to get data about yourself on the internet. How many of us felt that it was accurate? What was the degree of accuracy? So let's say, anyone who believes it was 100% accurate? Wow, okay, these two people, that's good. How many of us believe it was about at least 75% accurate? Almost all of us. Okay. That's pretty good. So about 75 percent to 100 percent accuracy, which is pretty darn good. This is just based on public information. How many of us had hallucinate to give something that was just completely wrong? no yeah well i did research we did each other okay she said he was on some kind of task force task force task force for new york city and he looked he's not well now you are you've just been nominated I'll take my 100% back because now that you said that, I did say that I had a MBA from Columbia and I had gone to like a week-long Columbia management program and thought that was an MBA. I'm not going to argue with it if you want to give me an MBA. Well, now you're a task force, you have an MBA. Who else had a great accomplishment so quickly? Timelines are wrong? Okay. Yeah, so obviously, you know, some of this is likely to happen because what GPT is actually doing is not really doing, I mean, it's all method. Sorry, yeah. Well, I was going to say, but I mean, isn't the point of this, what we were saying before was that even if it's a little off, it's still far better than if I had had a staff member, if I was saying, you know, build a review of my career so we can post it up on the website or something like that. they would be probably more off because they wouldn't find as many things and that, so it's still better than what you have gotten if it were from just a person doing it, right? Yeah, yeah. And again, I think that's where also the importance of human in the loop comes in. So if you were to just ask GPD, hey, create this and publish it all day about your enterprise information, think about it, right? And there were some slight nuances. You could end up with a massive litigation at the end of it because you published false information or factually incorrect information. And we had an airline case. I can't remember what airline it was. The airline chatbot offered a refund. And then the refund was processed. And then in the end, it was taken back. And the person was pretty annoyed, saying no. But you process the refund. You can't take it back after you process the refund. And then there was litigation around that. So again, how far you use it in business, what guardrails you put are critical. So back to data. Here is a lens on how you can think about what data does it make sense for you to train a model on. And when we say train a model, we're not talking about GPT. We're talking about making an expensive decision to create an ML Ops pipeline, take a lot of data, and then train a custom model, a custom trained model, a transformer model, for example. Could be any model. Doesn't have to be a transformer model. And where would you use that? Should you be training it on public data, internet data? Would it make any sense to do that? Because you have GPT. It's already been trained. And you add zero value to your business by training the internet data. Unless you're trying to compete with Google and OpenAI or Anthropic, then sure. By all means, do that. But if that's not your business, then training on public data makes absolutely no sense for your business. Now, that is not to say that you don't want AI to know your public information. So the value in public data is less about training on public data, but going what General Kauffman said yesterday, fix every SharePoint link, meaning go look at your public information. So there's this thing called SEO. We're all familiar with that, search engine optimization. And there'll be people who will sell you this idea of LLMU. I think that's just garbage, nonsense, but there's no reality to it because finally that's all based on SEO. It's all based on what's on the internet. So if you have, and so the biggest service you can do to yourself for being optimized for an LLM is making sure the information on the internet about you is accurate. And that's the hard problem because you can only limit it to the extent you put it out there. But when others put it out there, and that's the whole world of misinformation, disinformation, and that's the hard part in dealing with that. But again, training on public data probably not ideal. Training on client data, should you be doing that? Training on what? Training your models on client data. data that your customers put on your website, like Facebook, they created Llama, pretty much using client data, because all the information you put on Facebook is technically your data, but you click that terms of service. But the question is about if it is ethical or I can do it, because maybe if I put my information in Facebook and I accept the terms and conditions, There's nothing wrong there, right? You hit the nail on the head. It goes back to what your terms of service are with your customers. So as long as your terms of service with your customers allow you, for example, if you have taken third party permission from your customers, and depending on what country you're in, certain countries, even if you can't take third party permission for PII data, there's no such thing. So GDPR limits that. So depending on where you are, you would have certain limitations. For example, in Jamaica, the citizen owns the data. So you have no choice but to take citizen's permission for every last bit of data. You can't just take blanket permission and say, yeah, I've got all the data. So that's client data. Limited access data, meaning data that is largely third party, or limited access, meaning it's not necessarily data that you've generated, it's data you've got from somewhere, you know, someone you're licensed from. So it's essentially still third party data, or second party data, it's not your data. And then the last bucket is proprietary data. What is proprietary data? It's data that your business is generating through the course of its action. But how do you think about data is the question. How do you think about data in buckets? Because data, what is data? More foundational question. What is data? One of the biggest assets of the company. Sorry? One of the biggest assets of the company. One of the biggest assets of the company, yes. One of the most valuable assets of the company. It is a valuable asset, for sure. But to the extent it is proprietary. Yeah. So what? How else can we define data? What is data? Yes, Jefferson, what is data? Experience, experience. So data is simply a record of something. And by and large, it could be a record of an activity. It could be a record of an entity, meaning a person or a place or a thing. And usually a person has a name or a date of birth or whatever else it is. And then it could be a record of a transaction, buying and selling of goods. So in a business, you have three big categories of data. Entity data, which is all the products and services and people and customers and things that you deal with, which is also called often reference data or static data within enterprises because it doesn't change as frequently. Then you have what is called transaction data because you're in the business of transactions. That's what any businesses, you buy stuff and you sell stuff. and then you have what is called activity data everything that happens in the middle meaning customers coming scrolling clicking on a website you know didn't check out etc etc right negotiations conversations everything else you have every transaction has many conversations leading to a transaction so those are the three buckets and the easy way to remember is humans eat food machines eat Data. Data. And that's what we just discussed. Entity, activity, transaction. It's an acronym. EAT. Entity, activity, transaction. So machines eat data. So you can remember it like that. Entity, activity, transaction. So that's, what does the machine eat? Data. And it's entity data, activity data, and transaction data. And in your business, think of what entities you deal with. People that you hire, your customers, your partners, your products that you sell, all the bill of materials, the raw materials that you get together, and so forth. So those are all the entities. Activities, everything that happens with these entities, the processes and everything else, the workflows. And then the transaction, which is your final purchase or sale of a car or a ticket or whatever it is that you sell. So how do we, so should we just take all this entity activity and transaction data and just throw it into an AI model and like how we just dump stuff into GPD or will it work, will it not work? What are the guardrails? And what do we need to know about data when we are feeding data in? Data, some people will say data is the new oil, right? Oh, data is the new oil because machines eat data. And we said yesterday, the scaling laws of AI, model architecture, you know, AI is going to do a better job than us. Compute, it's either limited or unlimited depending on what you're doing. Yes. Somebody told me that data is not the new material, it's the new uranium. New uranium, yeah. It's actually interesting. It has a decay. It has a half-life. It decays pretty fast. So it's interesting. Nice perspective. Thank you for sharing that. So data has a value chain, and it's refining, much like oil. There's raw data. What is raw data, anybody? Raw data, it's like a text message that you type. It's like a document that you type, anything. It's raw data. It's raw information. Like you're having this conversation, if someone created a transcript of this conversation, that'd be raw data. It's just raw. Because when you see the transcript, sometimes the names are spelled wrong. Sometimes the information is inaccurate. Mostly it gets spellings wrong of people names, entity data. So that's raw data. Then comes clean data. What is clean data? You open the document, the transcript, and you go and you fix it, right? And that's cleaning the data. And in an enterprise, you could have to do this with information that comes from different sources, different places. Then enriched data. What is enriched data? you add more stuff to it for example you know if you had a product and you say hey you know what the color is missing or the dimensions are missing on the product or you know texture information is missing you know we had all these clothes in our store we never added a tag called texture and what the texture is and we now have a tag called texture that's enriching the data so you enrich the data Then there's aggregated data. What's aggregated data? Yeah, aggregation is simply grouping information, summarizing information. That's the simplest way to think about aggregation. It's just summarizing. Like how you can take GPT, take this big document, summarize it in three lines. So that's aggregation. And any information can be summarized. Instead of looking at, oh, buyer persona has 3,000 parameters. You could have 3,000 parameters in your buyer persona. But you can say, hey, just summarize it in one sentence or give it what really matters. What's going on? That's Howard calling me. Just one minute. Sorry. I'm going to have to let him know that I'm calling. OK, so aggregated data. I think we talked about that. Precision data. What is precision data? Anyone? Precision data. Accuracy in numbers, maybe? Financial computations or any mathematical formulas? Yeah. So actually, I should use a picture. Yeah. Maybe easier for everyone. I don't have it in the deck. I'll just run a Google search. I think that should do it. So can we see this? I don't know if it's visible. You see it? Yeah? Can we all see the chart on the right? So accuracy versus precision. You see, precision is if you hit the dartboard 17 times, you're kind of hitting in the same place. Doesn't matter whether you're hitting in the center of the dartboard or not. Accuracy is you're hitting your goal. Meaning you're hitting the center. And accuracy with precision is you're hitting all the darts in the center. So obviously we want accuracy with precision. But precision data generally means whatever the objective is, the objective doesn't have to be the center of the dart board, it is always precise. So essentially what it means is repeatable outcome. So you have the same input, you get the same outcome. So it's not like you ask, like GPT, you ask the same question 17 times, you get 17 different answers. Not precise. So when you need precision in data, and then now you can now think of what you could do. You could go to custom instruction, and if you wanted precision, you could say, hey, I would like precision in all my responses. And then you would get consistency every time. You ask the same question. It wouldn't give you different answers. So that's precision data. Then of course data flows, because data needs to move from point A to B, that's what most businesses do, you know, you move data around and businesses that are good at moving data will have a significant advantage over other businesses. And what's the challenge? What do you need to know about moving data? Huh? Yeah. So the hardest part in moving anything from point A to B is fidelity or loss or translation, loss in translation. If I say something in one person's ear and by the time it reaches the last person's ear, maybe completely different. It's called Chinese whispers. So in data, the same thing happens. As data moves around, it can lose meaning and context. For example, when the data is stored in a database, It has some meaning. It has some context. You export it down into a flat file, and you give it to somebody, half the information is lost. So when you're moving data across an enterprise, one of the key things is to think about what format and semantic is carried with the data. And invariably, organizations don't do this really well. If you're doing point-to-point file integrations between systems in your enterprise, you have this problem and it is not ready for AI. So any enterprise that is ready for AI has gotten rid of all file transfer systems from one system to another. So if you still have this enterprise that does file transfers between systems, and sometimes it's done for a good reason because we have a batch process that happens at the end of the night, you probably still aren't ready for prime time AI. Not to say that you can't implement and have operational efficiency, but if you're really looking for transformation, then that will be the limiting factor. Data signals, a lot of times you don't need to move a lot of data around. What you need to do is look at what matters in real time. For example, let's take an example. Let's pick anyone who wants to volunteer. We'll use a use case and parse through it. Anybody? Take a business use case that you have, and we'll talk through, so it'll be easier and ground everyone in your business situation. Like what do you sell? Let's think of a product that you sell. What do you sell? Insurance, who's that? Okay, so you sell insurance. What kind of insurance do you sell? Is it property, casualty, home, life? Let's say life. Life insurance, okay. So you sell life insurance. Who do you sell it to? What's your target market? Let's pick one. Let's say we want to sell it to affluent people that can afford multi-million dollar insurance policies to protect their Family. To protect their what? Family. Family, okay. So multi-million dollar insurance policy is to protect their family. It's a great use case, right? And let's say you were that insurance person and that was your job, to sell multi-million dollar insurance policy for? Protecting their families. Protecting their families. Now, who might want that? Everybody, but who might be able to afford that? That's the question. So the key question is what's the data you're looking for? You're looking for data on affordability. and buyer's ability to pay, because that's what it boils down to. So where would you find a data signal that will tell you where to find that person? And if you can capture that signal in real time, you have instant lead generation in real time. You don't need to hire an army of people, make cold calls, and do all that. And you can just generate an automated campaign to go behind that signal and then get the highest possible conversion rate. But the question is, where do you pick the signal from? Where is that signal today? Where is the raw data so you can go through all these steps and generate the signal? That's the question. And where's the raw data for that use case? It could be many places. It could be neighborhoods and zip codes. Neighborhoods and zip codes, sure. Someone bought a multi-million dollar home. potential candidate. We wouldn't know. Maybe they already have the insurance policy. So, you know, the signal would come from a partnership with a real estate company that sells multi-million dollar homes, and they'll say, yeah, every time we sell a home, we will cross-sell your insurance, or we'll upsell your insurance. This is how you would do it traditionally, but imagine in an AI world, you just create a data signal. you create a data pipeline, they would, and again, now it goes back to the ethics and all the other data sharing agreements and those questions. But that's a data signal. You could have a data signal based on a company being listed in a stock exchange, for example, and so forth. So thinking of what those raw sources of data are, and all of these are public sources of data, and the more scalable and repeatable are, and specific to your target market, I think, Then the question is, how many data signals do you need to triangulate and find your ideal buyer persona? So that's essentially the idea of data signals. And then comes the question of the data decision. What's a data decision? who makes the decision on whether to chase that lead or not and the expense of chasing that lead. Now, you could just completely automate it saying, yeah, if there was a person who bought a house and also their company was listed on the stock exchange in the last one week, let's go after them right and then you have two data signals and then that decision is automated and it goes through an automated pipeline now how likely is that signal to appear probably very extremely rare so think of what other such signals would be that you could triangulate along with the purchase of you know a house so That's, in essence, looking at the pipeline of data. But then the question becomes, how do you play in this space, offense versus defense? Because everybody's playing this game now, what we just discussed. Everybody's trying to refine this data. It's not just you. So then there is signal-to-noise ratio on the size of the customer. Because the customer's going to get, OK, now everybody's built their data pipeline. Now I'm going to get seven emails today. So how do you decide where to play? And how do you decide how to win? And it goes back to these five areas. And it depends on where your business is focused. And this is essentially depending on your leadership level. If you're the owner and founder of the company, it's really your decision to make where you want to steer the organization, where you want to focus. Versus if you are in the leadership team, you can encourage these questions with the board of the organization. to say, where are we playing in AI? And this is an important decision because that drives all other downstream decisions that you make. Some organizations are heavily profit-oriented, and that's, you know, it's all about quarters and bottom lines, and that's all that matters. Others are focused, like private equity firms, there are no quarters and bottom lines, it's more about, you know, productivity. And as long as you can show, it's still measured by quarter or even by day, but it's really about measuring productivity and operational efficiency, for sure. Just a quick question. If you're still behind on the data journey and the data governance, where are the opportunities to leapfrog? Like is it, I don't know, is it in the systems and the flows and how you manage that? I mean, I just don't know if there's any suggestions. So help me, if you don't mind, help me understand, when you say behind, can you characterize behind? Okay, well, and then Josh, I mean, this might sound really sad and pathetic, but you know, we still have systems that don't talk to each other, data sitting siloed, so it's really, unstructured, it's clean, they're sitting in systems, whether it's in Workday or it's in Salesforce, right? So it's maybe is it going ahead and investing in a data lake? So I mean, that's one example of the behind. And I think another example is just not having the focus on data being predictable, right? Still confirmation data, right? So some of that's just more to do with transformation internally. But I'm thinking more about the technology and the AI in a system site, how we curate it. So then we actually can work towards enriching the data or aggregating it properly. Do you want to add something? Yeah, I was just going to say, you know, as a nonprofit, obviously, where you spend your funds, you know, you have to kind of spend it towards the mission. And I don't think historically a lot of departments have seen the mission as cleaning up their data. Recording data, yes, but then cleaning it up and making it accurate so you can use it in the future has not really been a focus. So as a result, I think we have a lot of, before we aggregate it, we've got a lot of cleanup, and that's kind of an effort that I don't know if, you know, the company is behind. I guess the question might be, how do you, how do you cut down on the cost, using this process, how do you cut down on the cost for that so that the company can finally recognize it as essential and also not terribly costly. Yeah, so I think the cost of data management is probably one of the biggest costs within an enterprise. And if you look at any enterprise implementation or project, they almost tell you that mapping data between systems is what takes the biggest effort because sometimes you don't understand what the data is and the domain knowledge around the data is the biggest constraint. It's a hard problem. There's no easy way to solve it. And I can't tell you enough how many years I've been in rooms with people talking about master data management and data warehouses and data lakes and big data, which we don't talk about anymore. So surprising, right? There was cloud, we still talk about cloud. AI, we're still talking about AI, at least now. Big data seems to have just disappeared from our conversation. There used to be a tool called Hadoop and all these other companies. We don't talk about them anymore. It's like almost fell by the wayside. What happened? All this idea of structured data, unstructured data, there was so much conversation, but we don't really talk about it. And of course, there's AI now, so maybe it's not that relevant. Few things to consider. Number one, When we think about what we call the AI power play, think in three steps. Foundation, unification, acceleration. Get the foundation right, unify your data, and we'll talk about what that means, and acceleration. Acceleration is all your use cases. Now that is not to say that you can't do some use cases while you get your foundation and unification right. And how do you do that? For your enterprise to go to the next level, you do need to connect all the data. You saw a simple example of taking data from GPT to Notebook LM. And you saw the value of doing that. And you think of multiplying that n times over. So the silos is a tangible problem. And breaking the silos and bringing data together is important. But it doesn't have to be a boil the ocean exercise. Because the first thing you do in foundation is what is called inventory or an assessment. Like imagine we're right now supporting disaster recovery in Jamaica. We're not just saying, oh, let's ship all the medicines and all the clothes and everything, and we figure out what goes to who. We actually send people in the field. They go there to do an assessment, a need assessment, and we go have a conversation. And earlier, they used to do an iPad, and people would go punch data on iPad with a clipboard. Now we use AI. So we go there with a recorder and we say, OK, tell me what your problems are. And the AI just generates a very detailed need assessment that would never be possible through any clipboard. So the first thing is assessment. And that's what foundation is all about. You first assess, do you have in your enterprise somewhere a data catalog or a data dictionary? A catalog is simply a list of all the systems. And what data do they have at a high level? Without even getting into the details, what fields do they have? What tables do they have? That's a catalog at the highest level. And then you have a data dictionary, which goes down to every table and every database and every column and what data it contains and the size of every table. Now, Intellibus, the company, this is all we do in the beginning when we get into any organization. Because without this, all the AI conversations are sort of meaningless and pointless. We call ourselves an AI company, but like General Kaufman said, we like the military, go and change every SharePoint link. Because that's grunt work, hard work, and guess what? Nobody wants to do it, because it's just painful. This is the most painful exercise. But without that exercise, all this talk about AI, can only get you halfway, never gets you all the way. So that's foundational work. There's no escaping that. It's going and cleaning up, not even cleaning up, it's assessing. It's figuring out what you have. Now comes AI. Because once you have made this assessment and you bring that data catalog and the dictionary and you throw it into AI, it's going to do wondrous things. It's going to tell you immediately how to unify, how to rationalize, what data sets need to go away, where the duplications are. And you don't need an army of data scientists or data engineers to do this. Today, that's how it's done. You have an army of people doing this. And it's kind of not very rewarding even for them. They know it's grant work, but it pays money, so they do it. They could be doing better things with their time and life. And so that's the first step, assessment. So that's step one. Make a catalog, make a dictionary. Now use some responsible AI, don't throw everything in GPT, and put some good prompts around it to go and say, help me rationalize this. and rationalize the model. And then it's going to present to you what is called ontology. Now, that's an exercise. We're not going to do it today, but I'm happy to point you to people who know how to do this for a living so they can help you. And what you need is not a traditional ontologist, because then they'll put you in a rabbit hole of data models and you'll never get out of it. But you need an AI-friendly ontologist who knows how ontology works in the AI world. And they can give you very simplistically what your data model of the future needs to look like. an ai world then comes the hard part which is all the mapping because now you have to take all the data that's across all these disparate systems and that's the unification and bring it into the unified model and that's a non-trivial problem to solve because in some cases you don't own the data in some cases you have constraints on who can access the data And then, even if you unified the data, how do you make sure it's real time? Some systems don't support real time data. It's just a file. It's a spreadsheet. Like all kinds of stuff that comes in. And that's the other hard part of the journey. But these two done correctly, you will be the winner in your category of business. There's no stopping you. But I can assure you that most people won't do these two steps correctly. Or do the hard work. Because that's what it's going to take to win in AI. And that's the bottom line. So, question, I mean, to make Josh feel better, having been on Wall Street for a while, data there is still challenged. There's millions and millions of dollars to spend. And I guess my question is related to that. Can AI take care of bucket one, right, the foundation, and help sort of rationalize your data, make sure it's harmonized, understand where the golden source of data is? Can it help sort of get you to the next step of unification? Or has it not arrived in a material way? Because that's where organizations get stuck. That is absolutely right. AI can enable the person doing it, but AI by itself cannot do it. And the reason is there is context involved. And that context often is not documented. It's in heads of people. Like a spreadsheet was created. And it's there in a SharePoint. And that's the one that everybody uses for pricing. And we have no idea who created the spreadsheet and what it was in it. And then you even call the person, they've forgotten what it was in it. Because they didn't document it well. They're just like, yeah, yeah, let's look at the formula. And spreadsheet heaven is one of the limiting factors. Everything else? AI can do it, databases AI can go against. The spreadsheet heaven is the hard one to go against. So if you have a lot of spreadsheet heaven, then you've got some challenge. But other than that, I think you can do it. These days there are enough tools available from enough vendors, and you can write your own, you can roll your own, it's just a bunch of scripts. I would not even recommend investing a lot of tooling there. Just use simple things on AWS like Glue and Airflow and whatever other open tools are available to pull the data, inventory the data, catalog the data. The hard part is someone then going and looking and updating the dictionary because, yeah, the AI can generate a dictionary looking at your data, but how do you know it's right? And that's the hard part. Someone's going to have to sit there and review every line, every field. It's like, I don't know, if you're a large organization, we're talking 12 million fields. How many people do you need to do that? Someone's got to look at it. Yes, and then maybe a third one, just a proposal, a third toffee is Maybe you do this work, but then how do you maintain it? Because you need the body of people that will continue to maintain it. And some budget. It's not evident. Maintenance is a nightmare. But again, this may seem quite hopeless. So let's think of where the hope is in this. And the hope is you don't have to boil the ocean. You just want a bottle of water, you gotta take a bucket, and you gotta desalinate, and you're done. That's the winning strategy for every enterprise. So you just gotta figure out how much water do you need, where it is, and what buckets of water do you need to desalinate. Yes? So when you're unifying the data, would you also perhaps think as part of that unification, how to make the data so it's more secure? Like if I unify my data as an all-nice, organized thing, easy for someone to take my data and steal it, as opposed to do I break it into some sort of thing, or you take any part of it, it's useless without the other pieces. Absolutely, so you kind of hit the nail on the head. The question is, in an ideal world, if you're an OPC, meaning one person company, right, you would just put all the data into GPT and be done. Because you don't need any authorization, any controls, any governance. Who has access to what? The challenge is, organizations have roles and silos. And so the silos have been created for a reason. Because sometimes it's regulatory in nature. Back in the day, you couldn't have sharing anything, even today, between brokerage and research, or brokerage and investment banking. There were laws, Graham, Leach, Liley. or whatever that was. So they had this distinction that you couldn't share information. So there'd be some hard rules on who has access to what. I mean, in the military, of course, you have different levels of secrecy. All that has to be followed. But even within that, I think the key question is, what's the problem you're trying to solve? And for that, what is that data value chain? We talked about the value chain going all the way from raw to the signal. So what you can do is isolate who the signal needs to be shown to. You could put all the data with the app. That's not a problem. But who should be alerted with that signal? Or who has the right to enrich? So you build guardrails in that value chain. as you unify the data. You could do it at source. And again, there is a broader question of do you bring the data in one place, you kind of have to, to connect the data, and the models are federated data, meaning the data stays where it is, you just create a view of that data, you don't actually copy the data, it's a read-only view of the data, no one can edit or modify the data, versus the reverse, where you create the centralized system, where everything is central and it's kind of distributed out. But it depends on what your use case is in your enterprise. I don't know if that answered your question. Thank you. So it really goes back to, as an enterprise, where are you playing? Are you playing on profitability versus purpose? And there are all these layers in the middle. think of what's happening in the industry open ai we all know great company it has so many offshoots right now you know it's almost like you could you know i don't use the word mafia but some people used to say paypal mafia now there's an open ai mafia out there but and and all these companies exist and they're all people trying to build different solutions in the ai space and and again Anthropic I think we all know perplexity we all know X we all know when there are some of these other other firms as well Eureka pilot and others that are there in the industry, so the industry is reshaping itself But that's happening at more of the foundational level. The foundational models in AI will change. But that doesn't necessarily change how you're going to run your business. Because you're going to assume that there's a constant. You're going to assume that today there is GPT, tomorrow there could be something else. So the key thing that you want to think about is you don't want to build anything too locked in and proprietary. So build, and what we recommend as IntelliBus, again I'm going to wear an IntelliBus hat for a second, is building a framework, an abstraction layer. So you can unify your data, but you build an abstraction layer. So you're not going and you build whatever agents, personas, whatever you're building, you build in-house. Because if you started to use GPT, I think the question, Stephanie, you were bringing up earlier the point, if you build everything into these proprietary tools, a vendor may come in, a vendor may come and say, oh yeah, I've got this mobile app container ready to go. And it's got everything in it. You don't have to do any hard work. Sure, you can do it for time to market. But then now you're pushing all your data away into yet another silo. And you have no control over it. So think of how you would want to end. We live in a vibe coding world. SAS companies are going away. So there's absolutely no reason for you to not invest in building tools in-house. which will pretty much get you whatever the vendors bringing to the table unless they have some real secret sauce and these are the only companies that will have any real secret sauce because they're going to push the boundaries on model architecture or cost of compute data is you so unless you're talking to someone who's giving you better model architecture or better value on cost of compute it's your data And software is becoming a commodity. Anyone can write it using AI. So there's no reason to lock yourself into third-party software. It goes back to our build versus buy conversation. So we'll take a pause here and take a quick break, and then we'll come back.