The Useful Architecture + Siri AI

Welcome to intentional AI, the podcast where we chat about becoming more productive through the intentional use of artificial intelligence. I am here with my good buddy, old pal, Mr. David Sparks. How are you? I'm doing well. Chris, looking forward to doing episode two of our new podcast together. We made it to number two. Yes, we're double podcast is officially. Yeah, we doubled. So next four, eight, we're doing the doubling integers. I guess. Yeah. Yeah. What why not? It's our podcast. We can do it however the heck we want. But we probably syntax is important. But welcome back to the show everybody. And thank you so much for all the good feedback from the first episode. We didn't even talk about much practical. We just kind of gave a lay of the land for the show. Yeah. But still the the feedback was so fun to receive. And people seem to be really into this direction where it's not about hype, not about doom. It's about the practical tactical stuff you can do with this AI technology. Yeah. And it's evolving every week. Today we're going to talk some basics, but we're also going to talk about the news from Cupertino with Apple's new Siri AI initiative and what we think about it and how that fits into intentional use of AI. Yes. Yes. You have in the show note here, David, David clarifies quote unquote robot. I was curious what that meant. Yes. Chris, I have the very first point of clarification in the show. Oh, nice. Yes. I do. I so this is the thing guys. I don't like personifying artificial intelligence. I know a lot of people they name their AI. They give it a cute name. Jarvis, I know is a popular one based on the Marvel thing. It's not a person. It's an algorithm. So I try to always keep that in mind. I think people you can fall into a slippery slope where the AI starts becoming useful to you and you start thinking it's a person with actual judgments and opinions and you start giving it too much credit. So one of the ways I fight against that is I don't call it a cute name. I call it robot. And so I say robot do this robot do that. I made a course about it called the robot assistant field guide. I refuse to give it a name. Unfortunately, a lot of people listening to the show are new to this stuff. And several of you wrote me very kindly saying, Hey, I signed up building about intentional artificial intelligence. Why do you keep talking about building robots? Is this a course about making robots or teaching AI? And I realized, well, you know, I didn't really clarify that. So when I say robot, that's a term I'm using to talk about the AI construct that is running to help me out. Yeah, I love it. That's that's a pretty good transition to do the topic today because we're talking about and I feel this is a leap that we've done ourselves as the technology has progressed where it used to be this chatbot interface. And I still remember the feeling I had when I first used chat GPT. It was magical. You know, you type something in and you get a response, which is really just generated in response to to what you said, you know, where it predicts the very next word. And that was kind of useful. This chatbot era where everything happened inside of this chat window. But the applications were limited. You know, they were largely generative. And I have my own qualms as a professional writer about generative text and generative images and stuff like that. But it was mostly stuff that happened in response to what we said. There wasn't a lot of room for using it with intention to become more productive with it. It was really just kind of a call in response, you know, put something in, get a response out that was really just predicted text in response to what you said. Chris, what was the first thing you put into a chat window that you got the result? And you're like, wow, oh, I don't remember. I think it would have been, you know, I don't remember. I don't remember. I don't remember what I said. But I remember the feeling that I had when I got the response back, right? I tingles up and down my spine because I saw text responding in a way that it never had before to something that I said. And to believe that that was one step level, I would say below where we are today, it was really quite, really quite remarkable. Do you remember? I do. I do. I gave it a blog post and I said rewrite it like a pirate because that's what everybody was doing in those days. Yeah. Yeah. They wrote it like a pirate. I was like, wow, that's pretty impressive. And then then I actually asked it, you know, what are the four cardinal virtues of stoicism? And it knew that. I was like, okay, so, you know, in my head, it was suddenly something interesting. But the limitations of the chat are, I think we need to call them out. It's number one, it doesn't have any memory. You know, you come back to it every time you're essentially starting from zero. And as a result, it has very poor context. It doesn't really understand the big picture of anything you talk to it about. And the other thing is it was no reach. You know, it was stuck in chat. You know, if you wanted it to proofread a document for you, you gave it the text, it proofread the document, it gave you its feedback, then you had to go back to your other thing and fix it. And so it was memory and reach that it didn't have. And you know, that's just a few years ago. And now we have both of those things, but I think a lot of people don't realize it. And the goal for today, the architecture show of this is to just kind of set the foundation of how you can work with a robot or an AI now. And your own personal drives in a more powerful way. And I think there's a lot of different technology stacks. There's Apple and Theroppyc, OpenAI, Gemini. There's a bunch of them out there. But they all kind of have the same ingredients that they need to be useful to you. And today, this architecture show we want to talk about, you know, what are the pieces of this building? And yeah, I think, you know, one of the things that I struggle with personally with this AI stuff is, you know, you and me, we've gone in deep with this stuff. And we're both gigantic, you especially, David, sparks, gigantic nerds, gigantic tech nerds. So we're steeped in this. Yeah. In a good way, in a great way, that'll be helpful on the podcast. Well, at least it doesn't get me beat up anymore. So I'll go with it. No, no, it makes a living now instead of the beating up, which is, you know, it all, it usually works out well for the nerds at the end of the day, doesn't it? But this is something that I think I struggle with as somebody who's a gigantic nerd about big dorkess about this stuff is what level of technical complexity do we talk about this stuff at? Because there's a lot of interconnected parts here. And so the thing that I want to say before we jump into any of these parts of an AI system is we're going to dig into these in the forthcoming episodes. And we're going to go into more detail across all these different points from the saddle to the context, the skill files. So if it isn't clear right off the bat, that's okay. That's okay. We're going to go in deeper, but I think it kind of does make sense how these components fit together, especially if you've been using AI, you've probably dabbled with each of them, maybe without even knowing. Yeah. And I really believe a complete understanding of how they fit together is the unlock. And, you know, I've taught this now to thousands of people with that course. And I have seen so many people with no technical real skills, you know, people who know how to use a computer, but they're not programmers. They've never written a line of code. And they're making just amazing tools. And I could tell stories. I don't want to go down that rabbit hole, but it's just remarkable people are doing. But it all starts with a basic understanding of what we're going to talk about today. The real power in this stuff is escaping chat. And I can tell you from teaching this to a lot of people, you don't have to be technically proficient to do that. That's the real democratizing force of AI, frankly, is once you understand the basic structure we're going to talk about, it can be applied in a lot of contexts. People use it with cloud co-work, Cleveland use it with open AI's codex. There's a lot of different platforms. You can do it apples, just announce kind of its own version of this. But you need to understand the building blocks of useful AI. And that's what we're going to try and talk about today. Just set a very basic understanding of them. Many of these topics we're going to come back to in future episodes. Yeah. Yeah. I love it. And we're going to start with the harness, which I've been calling a saddle, but harness is the common nomenclature for it. It's really just what you use. It's your point of interaction with an AI system with a model. So everything from a smart speaker where there's an AI on the other end that you're interacting with to an iPhone app that you're interacting with Claude or ChatGPT to a Mac or a PC app to, you know, we both use the app whisper flow. I think we're both a fan of to write emails and communicate a great dictation app for cleaning up what you say. You can just dictate an entire email and it'll add punctuation and formatting and stuff. I would refer to that as a harness too, wouldn't you? Yeah, I would. I think that, you know, the harness in my mind. is the link between the model and the world. Like it's like, how are you gonna take whatever you do there and make it work some way? I think the most important harnesses are the ones that are getting attached to the frontier models. A good example is a clawed coerc. Clod coerc is a harness that allows you to take anything you do in the chat and then get it outside of the chat to work with your other tools. And in addition to having an important model, like you wanna have a powerful model that can think and do the right things you wanted to do, the harness is equally important because what good is all that knowledge if it can't take the data and use it somewhere? Yeah, would you say a clod is the harness that you use the most often? 'Cause I probably would. Yeah, the one I use the most often, yes, although OpenAI's Codex is equally powerful, somewhat argue more powerful. And I'm sure Gemini's cooking up something, Apple just announced what they call Appintens, which is a form of a harness. So there's a lot of them out there. And one thing I'd like you to understand as you're learning this stuff is whatever harness you're using now, isn't probably the harness you're gonna be using in a year or two, all the stuff's in motion. But the other components of this we're gonna talk about here are the things that you can build to be portable. So you can switch harnesses. You can switch models and still get the benefit of intentional use of artificial intelligence. Yes, I love it. The next component that you work with, and this has been around since the original chat GPT. The one we both had the tingles up and down the spine for is context. An old boy, this is a deep topic that deserves it's a whole episode. But for me, context is really just information that AI manipulates for you. It can be the conversation that you're having with an AI. That's kind of the traditional baseline context for these things where it keeps that conversation in its short-term memory, which is called the context window. And you have a great analogy for how narrow this window can be when you're manipulating the different information with the AI. It's very forgetful. Yeah. I mean, the context window is small at this point in 2026. I always call it like an amnesia employee. Like somebody who just can't remember anything, but who's very good at reading the manual. So, but context is essential to making it work. Just like I said earlier, the harness is a key element to working with artificial intelligence. So is the understanding of context. You can have the most powerful model in the world, but if it doesn't have context, the results are going to be garbage. And you have to really come to this thing holistically. Do I have the harness? Have I given it adequate context to understand? And the old days of a few years ago, people would say that's all about prompt engineering. You have to write a prompt to give it enough context. But what we've discovered with modern models is a good prompt isn't enough. That's not enough context. You could give it a thousand words, and that may not be enough context to give you the result you want. But this is a thing we're going to constantly be budding our heads into as you're learning to use AI. Some of your worst results will be not a result of a failure of the model, but a failure of context. And so data literacy almost, and information management, I think is key here when context can be any data from your conversation that you're having with an AI to CSV files that you're importing into to transcriptions of conversations that you have with people. So one of my favorite things to do is when Arden and I, my Arden is my wife, when we're chatting about a problem that we're working with AI to solve, like planning our meals for a week, what we'll do is we'll just, we won't let the good context go to waste. We'll pull up a voice memo on my iPhone, and then we'll report it. We'll run it through an app called MacWhisper, just to get that text to feed back to the model that will create actions out of it. And the folders that you work inside of, if you use Cloud Core, that gives it access to all this beautiful juicy context that you can then use to manipulate and reference. And essentially have, memory is kind of a different thing entirely, but context, you can pull that into the current working memory, the current context window to be able to manipulate it in the moment. And so interviews are a great way to generate context as well. We'll talk about this in the context episode, but a quick little thing if you're using this stuff already is have it interview you in order to generate context about what you want to accomplish when you're engaging with it for something. And I'll chat about this in a little bit with the health goals that I have AI do for me. But getting good context into the AI model, into the harness. So it has a clear understanding of what you're trying to accomplish, the intention that you have with using the model, as well as just the data around the problem that you're trying to solve. It's critical, but it does have a very short term memory. But we'll get into that in the context conversation. - Yeah, like I said, we're gonna do a show on that because I think it's important, but if you just take from this conversation and you're saying that context plus harness, cross model, your bucks ahead, that's where people need to be thinking. And that leads to the next kind of related topic that Chris just mentioned is memory. And one of the things we didn't have with the traditional chatbot was memory, but with this new world we live in, memory is accessible. And the way I do it primarily is with text files. And so because you've got a harness that can see files on your computer that gives you a mechanism to give it built-in memory. And the way I teach in the robot field guide, but there's different ways to do this, is just have a folder full of text. And you tell it every time you learn something new, write it down. And the robot or the AI will write thousands and thousands of words of in essence context into this memory. And there's two categories of it really in my mind. There's the skills, like the mechanics of how to do something on your computer for you, like how to write an email or how to check a box on a website. And then there's another piece of it that's just like reference materials. In the last episode I talked about how I moved my Max Barkey community over to Circle. The one of the first things I did was till my robot to read all of the support documentation for the Circle platform and create reference files in my local drive with these text files of how it works. So in the future, when I ask you to do something to Circle, you've got your own built-in reference area. You don't have to go check it on the web. But that's the kind of things we do. Skills plus reference equals memory. And the fact that you've got a harness that can get you to those local files means you can have a nearly unlimited memory. Makes sense? Yeah. Yeah, it's great. And the final component-- Well, one more-- Of this is-- Before we get there, one more thing I want to say is-- OK. Just to make the point, I talked about the Amnesia employee earlier. And this is the solution to Amnesia. If you've got an employee that forgets everything, you have to make a really thorough set of employee manuals. And every time you give that guy a job, you say, go read the manual on that. And that's what the AI essentially does when you build that memory. So you say, OK, I want you to go put a blog post up or I want you to go fill out the monthly expense report. It'll go and read the manual, the skills, and the reference documents you've given it. And it will do it right every time, because it will read the manual every time and follow it to the letter. So the memory is just kind of a key piece of that. Something that I find really helpful is having some reinforcement loop around the skill file as well. Because it'll be great. Well, we'll see how great it ends up being when these robots, when these algorithms, can do that recursive learning loop for themselves, when they make a mistake. And then they have enough awareness to think, OK, I need to fortify my skill file. So I actually don't make this mistake again. And I find a cleaner critical path into accomplish this objective that I have, that David or Chris or whomever wants to do. But until then, this is on us that recursive learning. And so something simple that you can do is fortify those skill files when you see it make a mistake, when you see a better way to do something. Really just have that recursive loop yourself. Ask it to fortify the skill file with what it learned. Ask it to audit its skill file to figure out if there's any improvements that it can make to become more efficient about something. Have it do that level of deep research that you were talking about to create the skill file in the first place, to go out, not just use the knowledge that it has built in, even though these algorithms do have a lot of knowledge built in. But have it learn from-- the internet and do some deep research to figure out the more contemporaneous, the better ways to do these different actions. That can be really, really helpful for fortifying these things because something that I find is it'll often accomplish the thing that I wanted to do, but it'll do it in the weirdest, most convoluted way, instead of clicking one box or one button or something on a website if you're using Chrome MCP, which we'll talk about in just a second. It'll, you know, okay, I'll read the, and analyze the JavaScript of the page to do this. And it's like, no, just do this, clothe. And so asking it to fortify itself, the recursive learning is on us right now. So if you want it to do a better job, the research, the interviews with what you want to accomplish, you're putting it to audit its skill files, those are helpful things that we can do. Okay, so where are we so far? We've got the harness, we've got the context, we've got the memory, so it has memory. What's the last piece, Chris? We got the reach, and this is that these are the tools that the AI can use to accomplish things for us. So the one that I think most people would be using, and there are other ones that we'll talk about in a second, but is MCP, which stands for, I'll try to remember, model context protocol. That's correct, right? Yeah. Okay, good. So essentially it, it gives the model context through a protocol that allows it to talk to some other app. So it's always easiest to illustrate this stuff with an example. The one that I find myself using very often is Chrome MCP, which allows an agent like Cloud to actually navigate Chrome for you. So we're looking for, we're doing a bedroom renovation here. And so I ask, and I don't know about you, David, but whenever I have to look at these pages with hundreds of pictures of lights on them, like my eyes glaze over on a very just fundamental level, it's exhausting. And so I ask, Cloud, you know, interview me about what kind of light I want for this bedroom. I don't know what the optimal one would be, but I want you to work with me to find the one that would be optimal and then go find it for me. I don't know what keywords I would need to type into Etsy is where we ended up finding it. I don't know what keywords I need to type. I don't know the different styles. I don't know the lay of the land of what is just out there. So interview me about what I want to generate enough context to be able to use the MCP through Chrome to navigate Etsy and look at the images that are on there and select the top 10 options for me. I think I did top three because I didn't even want to go through 10. But it's a very simple way that allows it to connect with a different app, but there's ones for Chrome, there's ones for email applications, there's ones for to do list is one that I use as well, which is it to do list app, where you can ask, okay, how, what, what, what, what is my day like today? How much free time do I have today? Could I move a bit to later on in the week so I could just, you know, read in a park today or something? It's wonderful at manipulating this based on the context that you have and based on the overall intentions that you have for what you want to accomplish. So reach is the ability for the to escape the chat. So you get that the thing has memory and it has the ability to do proper work for you, understands how to interact with the world for you. Reach is actually the mechanism that allows it to and Chris was talking about MCPs. There are new MCPs every day for web services applications. Anything basically you do on your computer, a lot of developers are now jumping into this. The MCP protocol was open sourced by a throppec. They're the ones that kind of invented it, but a lot of people are adopting it. And for you building your robot or your AI, you're going to discover MCPs as an excellent way to do things like see your calendar or work with your email or go on the web. So that's kind of the final piece of this that escapes us out of the chat window and allows us to make artificial intelligence really work from us. So a lot of the stuff we're going to cover in the show is going to be elements of that. Talking about what's the best model. How do you get it good context? How do you give it that reach into an application or how do you build a skill set so it can do the thing you wanted to do? And if you understand those things and start working with those, even at a basic level, very quickly you'll find you can ramp up because what we're not saying is you need to learn how to program a computer. That's the part the model can do for you. Like, it can write the Python script to do the weird thing with the text that you need in your report every week. So suddenly that problem is solved for you. You've just got to have the imagination to combine all those ingredients into a work in AI. And that is where you get intentional AI. Oh, nice wrap up there. We should talk about before we dig into the Apple WWDC stuff, which I'm quite excited to talk about. We should do a quick little show and tell which we can tie all these little pieces together David. Yeah, Chris, what are you doing? Well, have you noticed something, David? You're looking mighty sharp, my friend. Yeah, you look good. Yeah, look at that jawline. Well, people can see it on the YouTube. I'm just, this is something I've learned. If you make your beard longer at the chin, it gives the illusion that you have a sharper jaw. But that's beside the point. So health has always been a weird thing for me because I have so many different goals when it comes to my health. So I want to rebuild my cardiovascular system. I want to lose a bit of body fat every week. I'm kind of at the end of that fat loss journey. I want to gain some lean muscle mass. You know, come, I want to barely fit into the frame on our YouTube video, David. There you go. No, I'm just kidding. You know, I think we can just use a wide angle at that point. No, that would be intense. But I also want to enjoy life. I want to enjoy the ad caloric surplus. I love a good feast, but I also don't mind fasting a little bit. And I want to time the workouts that I have for cognitive performance. So the challenge when it comes to the health goals that I have is I've always wanted too much, I've always wanted to do too much. And often the different things conflicted with one another. So I thought, okay, yeah, there's this clawed thing. I've got all these different parts. I've got the heart. I'll try to tie it back here. I've got the harness. I've got the context. I've got the potential memory, the skill file, and the reach here. How can I integrate all these together to actually achieve these goals? And so what I did, and we're going to jump into a skill file here. I won't share it, but I'll share what it does. It is, I programmed and had it do the deep research on all these different goals and generate a report for me based on for what I would need to do in order to achieve every single one of these things optimally. And this is something that I think is a helpful frame often is everything that I find that I end up using AI to do for me is an optimization equation of some sort. We're trying to find the optimal path forward, the optimal solution, even if you're hunting for a product. You're trying to find the optimal product that fits with your budget, your constraints essentially. And so this is what I wanted it to do. So I programmed it to be a world-class health coach that integrates all these different goals together, how to do deep research into all the different goals, and interview me about what I wanted my life to look at, like my actual days to look like in order to achieve these goals. So that's the skill file that it built. And the input, the context that I give it with the skill file enabled so it sees everything through the lens of what I've programmed it to do is I give it health data and I give it pictures of what I eat. And so if anybody has a health goal that they're trying to achieve with clode or a similar app, it'll work across them as long as you can read data in a folder. I use an app called Auto Health Export, which automatically, so I've got this Apple Watch on all the time, it's constantly reading heart rate data, exercise data. There's also data that comes into Apple Health through a smart scale that I have. It pulls it all. It gets all that into a CSV file that goes into the folder that I'm working with in clode. So in clode, you can work inside of a folder. It gets funneled into that folder so that when I give it pictures, so my only job is to take a picture of every single thing that I eat, which is socially kind of awkward, but it clears the air when you just, it's not for Instagram, it's for clode. So it's programmed to, when I give it new food pictures, it'll look through the lens of the skill file and update all my different, I have a dashboard that I've programmed it to work with. And it'll coach me for achieving the goals. Like, oh, maybe tomorrow you should do a little intermittent fast because you ate a bit. You're a bit in a surplus territory or looking at your day tomorrow. I see you have a podcast recording with David Sparks, intentional AI. So you should probably schedule it. a run for that morning. And indeed, that's what I did this morning. I don't know if I seem more sharp cognitively today, probably not. But, you know, that was prompted with AI, which means that I can outsource the cognitive overhead of multiple health goals to a system that I trust because it's programmed around the goals that I have that often conflict with one another. So, it's a bit of a complicated solution, but it's got the different components. It's got the harness, which is clode. It's got the context, which I'm feeding it, which is all the reference information with the goals. It's got the food pictures. It's got the workout data. And it's relatively good at reading caloric information from photos. There's a margin of error, but over time, it balances out the research does show. It's got the skill file to tie it all together so that no context kind of goes to waste in this instance. And it's got the, you know, not a ton of MCP stuff going on here with the, in terms of the reach, but it's sort of this self-contained system that I'm finding works wonderfully for integrating a lot of these different parts together. And you and I like about that. Not only does it encapsulate kind of the whole widget, but you're finding ways to get the AI to help you do things that you wouldn't otherwise. Like logging your food is a pain in the neck. And I know a lot of people who will say, "Hey, but you know, the process of locking it, I eat less because rather than log that I ate a bag of potato chips, I'll just not eat a bag of potato chips." And I understand that. But I think the friction of that is, it prevents a lot of people from doing the types of practices you're doing where, like your idea, just take a picture of it and let the AI log the food for me. And it's not going to be, you know, down to the exact calorie, but it's going to be in the ballpark. And it's going to be good enough. And you've got a system that works for you. Yeah. You know, obviously you also have to look at it as it's a recommendation. It tells you, maybe you should take a run. That doesn't mean you always take a run because it tells you that's something I'm kind of sensitive to. As you begin to trust these things, you want to just, you know, accept its pronouncements and I think that's dangerous. But that's a great example. Yeah. It works. You do a bit with health with Clod, don't you? Yeah, I do. But we're running along today and I want to talk about Apple. So there's a lot more to talk about with that. I do think, like there's a whole privacy angle to it, like something we didn't mention. It's like you're giving a lot of information to anthropic about your health when you do that. And a lot of people are really allergic to that. And you've got to make your own decisions. But there's a lot going on here. But I think today we, I wanted to get a basic understanding of the lay of the land with this architecture discussion and Chris's example is a great demonstration of it. But going forward in the show, we're going to talk about some of the stuff in greater detail. We're also going to talk about just ways to use this. We'll get more practical with time because we will. But you had to understand these basics first. Now, there's a little bit of news we wanted to talk about today. And that is as we record this, Apple recently did their WDC Worldwide Developer Conference. Apple is kind of a small time player in the world of AI. They don't have their own frontier model. They've kind of sat on the sidelines for it. But they're also the company that makes the device that so many people use. And they're a company with a very distinct position on AI and privacy that makes them interesting. And they are now kind of in the game now with this new version of Siri AI. And of course, have you had a chance to use it yet? I haven't used it yet. I haven't downloaded the bit and downloaded the bit. Have you? Yeah. Yeah. I've been playing with it. I will say officially, Siri no longer sucks, which is great because it did for a long time. It's official. But there's a couple things coming out of it. We talked about, and this is where the show already is paying off. We talked about context. And we talked about harness and we talked about reach. My very first question in this series, where will the weather be where I'm going to be this weekend? Well, this happens to be the weekend. I'm going to vacation. And I didn't tell it. I was going to vacation. And it's told me, hey, the weather and Hawaii this weekend will be in the low 80s. That was his answer. It looked at my calendar. It saw where I'll be. It got the weather. It gave it to me. So even Apple, which has been standing back on some of this, they get it. And they're building context, knowledge and reach into Siri. And that's that's what we needed. Yeah. And something that I love about their approach. And there's one line from the keynote that really stuck out to me, which is that they don't want to do AI for the sake of AI. And looking at a lot of the AI stuff that's out there, it feels of that nature. But to me, this is what has always drawn me to Apple things is they seem to have taste with the things that they do. And to me, taste with AI is really about that practical, tactical layer where it makes this difference in what you do every single day. It's the little things. It's tidying up and organizing your tabs, for example, you know, in Safari that to me, that's incredibly useful. If you have 100 tabs open, it'll group them all by, you know, what you're focused on automatically updating your passwords. So, you know, I don't, when you scroll through the password list that you have, you see, oh, this one's been compromised. This one's been, that's always maybe feel guilty about not updating my passwords. But now there's one button that you can click and it'll agentically go through and update your passwords for you, even just a series voice making it, you know, more or less enthusiastic, faster or slower. I want to crank that puppy up all the way to the right in terms of how fast it is. So I can save more time describing Siri shortcut. So you can create shortcuts just with your voice. To me, that abstracts a lot of the programming and a lot of the fiddly bits away from you. When you, to me, that is what all of the announcements that they had had in common is they had that tactical layer where they want or seem to want AI to solve real practical problems for us. And to me, what gives me confidence in this is the extent to which they've partnered with Google to do the models behind it. That gives me more confidence than, you know, then it would if Apple did this themselves knowing the state of Siri for the last many years. Yeah. Well, I mean, and that addresses a problem they had. You know, they're not in the Gemini/Clawed/OpenAI ChatGPT form. So they don't have a model that can do it. So they leased it from Google. And then they provided the harness and the context through their devices. A big piece of this story for Apple is privacy because they're doing all of it either locally on your phone. So the problem we talk about really with Chris's health records going to Anthropic, that's not going to happen if you Siri to do this because it's going to happen either on the phone or what they call private cloud compute where they do it on a server, but it's completely anonymous. So nobody ever knows it was you. And that is a real issue with intentional use of AI. Like if you're a lawyer listening to this and we're talking about harness and memory and reach. And you're like, well, it doesn't matter. I can't have my client records going to Anthropic or Google. I can't use any of this stuff. Well, the Apple model would work. But they also kind of the flip side of that is the reach isn't as extensive. The kinds of things people are doing now with like co-work and Anthropic and Codex really are not achievable with what Siri does. But they're in the game now. And they're coming from a distinct vantage point. And we'll see like we talked about MCP, which is the biggest form of reach. Apple has not said that we're going to let you use MCP on an iPhone to get your calendar events into Siri. Maybe they'll do that in the future. Maybe they're going to make their own version. I don't know. But it's a little bit less bold, you know, but at the same time, they're giving us the key components that we have already covered today. Reach context memory. And it's all based on your device. So I'm going to be very curious to see how it grows out. This is brand new. So these are very much kind of hot first takes. But the Siri model works now. It does give you the kinds of things. You can intentionally use AI with Siri when this ships in September. I can tell you already. And what I'm most excited by is personal context as they put it. So for context now, it's on you to load it all up into the system. You have to be the one that exports the CSV of your health data to close to process and integrate it all together. And you know, with the privacy angle, I can see that opening up whole new vistas of things that I don't feel comfortable with using AI for. So my line is kind of on health data. If my health data got hacked, enjoy reading what my resting heart rate is, everybody. I really don't care. But if financial stuff got out there, I would never input financial data into code because in a hack, your surface area of vulnerability. goes through the roof. But if it's whether it's on a private cloud, which I would trust because of Apple's privacy history, whether it's local processing, which I'm also really excited by, the fact that they're iterating on their local models. So you can do it all, and they're adding features for developers to be able to run local models on your computer and your phone and things like that. It's very exciting where they're going, where they're expanding that local model capabilities to be very native things that developers can hook into, but also the more frontier models with, seems like a great partnership with Google where your privacy is protected. So I'm quite excited. The generative stuff I could do without. I don't need anything to generate text for me or writing for me. To me, that's in the realm of human expression. And same with photographs. To me, photographs are things that happened, which when it's generated, it's a weird kind of thing for me. And if it's not a photograph, it's an illustration which to me is in the realm of art. But I think we'll all have a different comfort zone, a little comfort area with the privacy stuff behind this, with how much we like generatives. My mom sends me AI videos of cats dancing. And I have to admit, they're quite adorable, even if I don't like the fact that they're AI. - Yeah. I would like to point out the intentional AI artwork was made by a human, not an AI. - Yes, yes, yeah. After a lot of revision rounds. Yeah, it's fantastic. JD is fantastic. - But overall, I think Apple is in the game now. I hope that they keep the gas down. I hope they keep pursuing it. But even Apple, who's more conservative on this stuff, has demonstrated the principles of this show. The things I want you to take away from listening to this episode is the understanding that in order to be effective in AI doesn't just need a good model. It needs context, it needs reach, it needs memory. And if you give it all those pieces, you can actually get intentional use of AI. You can make it do very useful work for you. Chris and I both do it every day. Lots of people out there doing it. I think the sooner you figure it out, how it works, I think the better off you are. So we'll be going into some of these topics deeper in the future and just talking generally about how you make this stuff more useful. Chris, we want to try something new with the show. We're in early days, we thought we'd try an experiment. At the end of each episode, we thought maybe we'll just announce a little experiment. We personally want to run with our AI to get something done better faster and more efficiently. And they're called our new experiments. So Chris, you want to go first, what your experiment going to be between now and the next two weeks when we do our next episode. - Okay, David. So I have a lot of printed sheets of paper in my hand here. I'm flipping through, you can see. What I had, CloudDoo, is, and I don't know how this will work out. I could fall flat on my face with this. Well, we'll see. We'll see. But what I had is I had it audit my logic that I use to make decisions across the various conversations that I have with it. And I asked it to be brutally honest with what I'm delegating to it to do cognitively. 'Cause this is, it's not a fear that I have with this stuff, but something I want to manage with AI is the cognitive offloading that I do. I'm very happy, cognitive offloading my workout goals. I don't want to think about when to do stuff like that. Or, you know, how many calories are in what, I just want to be closer to my goals over time. To me, that's what I want. But I want to make sure that I'm not delegating things to become mentally lazy. And so that is what I'm excited to review. What I'm offloading cognitively, it's a bit geeky. I understand, but I'm very excited. I think this afternoon, David Sparks, I'm gonna go to the park and maybe pick up a little coffee along the way and read through. It's called the unsophoned version, six decisions one through line about what I'm offloading cognitively, as well as any logic flaws that I have in working with it. Quite excited and kind of scared. What are you working on? - I am, first of all, I just want to compliment you for doing that. I feel like you are more comfortable with cognitive offload to the AI than I am. I am very skeptical of the AI's ability to give judgment and opinions. I don't really let it make decisions for me, almost to any degree, and you're more willing to do that. And I'm not saying that as like a backhanded compliment. I think it's something that a lot of people are doing, but taking the time to audit the process and really think about it seriously, I think that's important. I'm glad you're doing it. I'm looking forward to you sharing how that goes and maybe in the next episode, you can explain a little bit how you set that up for people that want to do the same experiment at home. My experiment I want to run is I'm getting more comfortable with what they call scheduled tasks. So we talked about the whole harness, but one of the things that Claude Cawer gives you the ability is to run a task at a certain time. And I'm getting more comfortable with it handling like some of my customer support requests if somebody has a password reset or whatever without me. And as it's getting more confident that I want to experiment with making that more automated where like at 3 a.m. it just looks at email if there's any, but it needs a password reset, do it and send it to them. And I don't need to be involved. So I'm giving the robot more power to do things autonomously, which is for me, a bit of a slip of a slope and things I want to be careful that I don't go too far down. But it's an experiment I want to run and I'll report back in two weeks. I know we ran a little long, we're trying to keep this to a 30 minute show, but there's a lot going on. And I think this architectural explanation is really important to understanding what we're talking about. So thanks for sticking with us, gang. And thanks for listening to the Intentional AI podcast. - See you in two weeks, everybody.