Kevin Riggen

Speaker 0: Jul beans. So I think I'll start with I wrote A.I, booklet recently, short book. AI I created a workbench to build this thing. And then I basically, I worked with agents to figure out how I should organize what I've been working on for the last few years. That kind of culminated in workbenches. Speaker 0: So a workbench is a temporary space that you use to make your life better as you're doing something with AI. So for this book, for example, this is my workbench. This is the introduction. I'm able to write here. So actually AI write these things. Speaker 0: Sometimes I'll copy paste and ask Chachi CPT to help me, articulate it in the same AI, you know, like trying to say this A.I AI read it and I'm like, okay, that's good. Okay, it's not. I don't use this workbench to actually generate the content more just to organize the chapters and then to export it into a API that I can print at the place that I like to print pdfs. Right? Loops, a loop is what it sounds like. Speaker 0: The the reason that it is so interesting now is that we have these things called agents that are doing something called inference. Inference is the thing that's happening with AI. Right? It's kind of, if you see AI as an engine, it's like the the combustion that happens. Inference technically is just giving you text out when you give it text in. Speaker 0: And now there are some other ways of doing inference AI image, sound. But the magic of it is that when you give it text in, you can get text out that's more meaningful. We'll get to that in a moment. Agent orchestration. So agents are things that do inference in a way that's really smart. Speaker 0: So it's not just that they're running inference as large language models, but that they have memory systems, they have knowledge Graphs, sometimes they have user interfaces, they have tools and they have ways of communicating with each other and other things. I'm gonna go way back Date talk about why I do this and why I care. Right? So, AI imports the majority of what it consumes AI% of our food and even our fresh water too. That's kind of silly. Speaker 0: If if you understand agriculture and regenerative agriculture, it just doesn't really make sense. Right? And you start to learn about the history of Hawaii and you're like, okay, now I get it. It doesn't doesn't make sense from the perspective of an island that, you know, you wanna have food security on or that you wanna raise kids and grandkids on in a way that you feel comfy about in this day and age. But, we got here from a history that is rooted in, you know, exploitation A.I, like, extreme industry A.I it is what it is, but we can use AI to start to flip this around. Speaker 0: And the way that we can do that is is the same way that most of you here, expressed what AI is for you right now. It's a way that you can go Folder, you can learn more. It's helping you learn. You're using it as someone who is a sounding board, something that is a sounding board. Why should I go to this event? Speaker 0: Should I is it going to be helpful? Right? Saving your time. It AI to agriculture too. So someone like me, I like agriculture. Speaker 0: I like mushrooms. I've grown some mushrooms, but if I try to put a farm together, it's not going to go that well because you need to Knowledge lot of things. AI can get me there. I'm certain. I can use chat 2026 right now and it can help me with a process for learning how to get a working mushroom farm started. Speaker 0: To understand the process, you know, what is a mushroom? It's mycelium. You know, the mushroom is the fruit of the mycelium. 2026 City knows all of those things extremely well. What I think we should be doing as people in Oahu or Hawaii that that care about making things better in general is coordinating, communicating, finding people that know what they're doing, and there are a lot of people here that are doing wild shit. Speaker 0: You find them randomly. And and the institutions here, the media here, they they skip over these people. There are reasons for this too, but but it doesn't really matter now. Like that's the thing about AI, about agents, about knowledge graphs. You can skip over all this legacy bullshit. Speaker 0: You can find the things that are really valuable A.I you can just tune into that. You can magnify it. So to me, artificial intelligence is a it's a bad word. It's a bad name. It's not really helpful. Speaker 0: AI, a good name is like amplified intelligence or programmatic intelligence. The way that that works is because our words have meaning And all of us underestimated this, but language has meaning baked into it. So when we throw A.I text at something that is a very, very dense neural network with billions of parameters, whose job it is to take text and then transform it into other text and Event it back to 2026, it turns out that we can actually, like, have a amplified meaning machine. So I think that's profound. I don't think it's hyped. Speaker 0: I think that whatever you're here for, Notes, it's pro, you're probably correct. Like if you have a hunch that AI is going to help you A.I then and then there's this thing you're seeing, this silver glint you're gonna start going towards, I highly recommend you do that. So these are slides Date off of that book. I'm going to kind of go back and forth on things. I'm very curious what you all want out of this 2026. Speaker 0: Please interrupt, please ask questions, raise your hand or just shout it whatever you want or need to do. We kind of went over this. AI imports most of what it consumes. I want to give a quick example of how how much potential there is for us to transform that. I live on a farm in Waialua, and there's this weird refinery next to it. Speaker 0: I had no idea what it was. You can tell it's AI compressing something, has these big compressors and stacked on top of containers. It looks exactly like you would imagine some, like, mini Folder punk refinery on the North Shore Of Loops. I met the guy who runs it because his car was broken down. I helped him, and it turns out this is a hybrid pyrolysis refinery that has been sponsored by, University of Hawaii for about 20 years. Speaker 0: They have Agents. They've proven that it works. What it does is it turns waste into commodities and biofuel. So it heats up and compresses oxygen, to the point where the the system is oxygen lead. It doesn't have oxygen anymore. Speaker 0: And it's heated and compressed to the point where it's running on its own energy, then it turns waste into a commodity, in this case, activated carbon. It can be another another commodity too. It depends on what they throw into the thing. So they throw, macadamia Notes shells in, they run it, they get activated carbon out A.I biodiesel. The activated carbon is a filter for water A.I it helps with soil health. Speaker 0: And biofuel, you know, that's cool. It doesn't generate that much, but it generates a bit. So anyway, this guy is just working out in North Shore without any, like, real attention. No real media. No, sponsorships from the institutions here who get government funding. Speaker 0: He's smart as hell. His dad is smart as hell. They've proven it with patents. In A.I few different ways. It would AI, like, $2,000,000 to put this thing on-site somewhere, and Oahu has a pretty bad waste management system. Speaker 0: AI guy from the government came back from touring Asia and is like, hey, we need to start burning our waste. Okay. You have these 2 nodes now. The government is noticing that we have a big problem with waste. No 1 likes our landfills. Speaker 0: And then we have these people over here who are just too busy. 1, making a living and then 2, actually, like, doing really an interesting experimentation who've proven a pilot program, and it's been through the university system. There's no 1 who would get in the way of this City. So you just connect these people and start blasting media here, you can transform something. Speaker 1: So Speaker 0: that's that's my Publication, I suppose. And I'm seeing it all over on this island. I can give like 5 examples just off the top of my head API people I've met in the last year through these meetups and elsewhere. The coconut rhinoceros beetle problem is also very interesting here. You know, we have an invasive species, has no natural predator, population growth of 50 x every 3 to 4 Agents, it's going to hit carrying capacity that hasn't already and our native coconut trees are probably gonna go extinct. Speaker 0: What do you do? I don't know. No 1 in the university system is carrying that much. There's not much funding to City. But then you have this guy who's file A.I rogue farmer who found this fungus that Asia uses. Speaker 0: He tried City. It kills them. Right? And many places in Asia, Malaysia, for example, Guam, and I think Thailand, Vietnam, they use this regularly. Why don't we in Hawaii? Speaker 0: They have some regulation issues. What regulation issues? I don't know. No 1 knows. Okay. Speaker 0: Hook yourself up to data.honolulu.gov and tell your agent, hey. We need to solve this problem. What do we do? Who do we talk to? How do we get funding for it? Speaker 0: How do we talk this, government funded organization into not levying lawsuits against us because this is, like, almost a AI? There's a lot of stuff that I I think personally we can start to seriously improve by throwing smart agents at it, but it starts with data and knowledge. I'll get back to that in a bit. Let's get through these slides A.I see if there's anything that I've missed so far. It's an intelligence amplifier. Speaker 0: Truly, you can think of it like a guitar amplifier or A.I keyboard AI. And you have a microphone A.I instead of sound, it's intelligence or meaning. And sometimes you give some little bit of something, you get a lot back. You know, you say, hey, build a website for my business. This is what I do. Speaker 0: Build a slide deck. Here's the PDF. Other times you give it a lot of something. You want something, you know, tiny. You say, hey, all AI legislature, all of this stuff that's happening this year, how much of it matters for Kevin Waialua when there are more Loops, right? Speaker 0: So, yeah, intelligence amplifier. Programmatic intelligence is another thing to really keep in mind. So, you know, as an example, we have never been able 2026, A.I software engineers, write in our code something that says, if there is a car with realistic accident damage in this photo, say yes. And if it looks fake to you, say no. We have been able to, you know, 10 years A.I, maybe we could train a small machine learning model over years and years and millions of dollars to give us a yes or a no there. Speaker 0: But there wouldn't be any versatility to it if we, tomorrow, had to say, hey. Also, if there's a dog in the photo, call this number so that I know immediately. Right? Or, like, if this is fraud, email this person. This is our fraud person. Speaker 0: We want them to, you know, be right on top of that. That's programmatic decision making, programmatic intelligence. It is brand new. It's a new invention. Think of it like fire or something or like the written word AI I think it's probably that meaningful. Speaker 0: We have machines that can think for us. Right? So the ways that this goes wrong. Yeah. So it's not comprehension. Speaker 0: Amplification without comprehension. Comprehension is a is a thing that these machines aren't really doing. Right? Like they we have reasoning models large language model that was introduced in, like, 2018. That's the that's the Publication version of things. Speaker 0: It probably existed before that at DARPA or something. And that is the kind of model where you send a text 2026 sends you text back. And and its purpose is to basically complete what you sent it. So you'd say, hey. I really like AI, and then it would return author. Speaker 0: You know, something like that might be what it's doing. We figured out how to stack these things and stack them and stack them and build routers within them so that, you know, the 1st step is what kind of query is this from this person? Is this a question? Is this a joke? Is this person AI trolling me in in the language model? Speaker 0: Are they asking for advice? Are they is this a coding problem? And you have all these different ways that it can be routed through. So the comprehension thing was a really big problem a year ago, 2 years A.I. When you were trying to do programmatic intelligence, If you weren't careful, you might get really weird outputs. Speaker 0: You know, things might go really awry inside of pro, automated systems. You know, for example, is is there a dog in this, car? Well, what if it's a sticker of a dog or AI a fake dog or something like that, you know. There was no reasoning to 2026, what is this person actually asking me? The latest models are absolutely incredible at what is this person actually asking me. Speaker 0: It's it's really strange how good they are at it. How many of you have tried Fable? Yeah. Okay. Yeah. Speaker 0: I think that's that's like the next level reasoning model A.I and the GPT AI 6 Folder model just came out. It it seems to be just as good. Yeah, they're scary good at reasoning. So Honolulu, the confusion part here like, sorry. The confusion part here is now very rarely from the models themselves. Speaker 0: It's often from the humans. I see this a lot. You know, people might people might be speaking to each other through Notes documents in a way that's AI, copy this thing this person said or wrote, give it to ChattCPT, and say, hey. Make me look smart or whatever, AI then paste it back. A.I then the other person takes that, copies, and says, hey. Speaker 0: Make me look, you know, useful. A.I it's an expert who pays that. So people aren't actually understanding each other, but there's a ton of communication happening. In ORGANIZERS, I've seen that. I've seen serious problems with taxonomies, the meaning of things A.I like how you should label systems. Speaker 0: It'll just go absolutely haywire if you let someone spend 2 weeks defining some AI AI into what they're actually trying to accomplish. A.I they hand you something to implement Event you're like, this is, you know, you can't even necessarily put your finger on City, but, like, there's something wrong here. What are we actually trying to do? So, yeah, confusion is a real problem. Malice, obviously, I don't even think I need to talk about it. Speaker 0: I think that, you know, I don't know what to say. I mean, mean, people are using AI 2026, target individuals to kill. You know, that's happening right now. Our military does that. Other militaries do that. Speaker 0: They're using robotic dog like things to go into tunnels and do things. You know, I'm really happy we haven't seen it in our, public spaces. I hope we don't. It gives me a lot of AI, actually, like A.I lot of optimism that we haven't really seen that sort of thing yet. And I don't know what else exactly to say AI it. Speaker 0: I don't have strong opinions on it except that, yeah, it's just another thing that's kind of scary about this this, time that we're all in. It's a it's a pretty wild time. Agents, yes. I've been careful enough for the most part here, but it's, you know, you give an agent access to your database that's valuable and it has right access and you tell it to do something and there's a miscommunication or you you made a AI. Or it is hitting some issue in the entropy system AI file, it can do a Date, it can malform data. Speaker 0: You have you will not be able to understand what happened compared to how much damage it can do in milliseconds. Gating these agents is, you know, it'll probably be a full time job. Speaker 2: Yeah. Speaker 0: Yeah. A.I there are plenty of, you know, public cases of this happening. Workbenches. Okay, I'll keep this brief. I decided I'm gonna write this this book. Speaker 0: I wanted to print this book. Speaker 2: AI? Speaker 0: Wahoo EI Field Notes 2023 to 2026. The introduction is what, you know, my spiel on how all this matters to Oahu. Yeah. What AI is, I I just covered all of that. We have workbenches. Speaker 0: So field notes wise, workbenches are relevant. I showed you this earlier, but there's a bit more to it. This made my life really nice when I was was writing that book. And by the way, I'm gonna help my friend write a book about agriculture. Okay? Speaker 0: He has this idea called, like, the 1 pig revolution. He says that in Hawaii, everyone should have 1 pig. Most people should have a sow. 1 person should have, I don't know, hog or whatever the the male equivalent is. And from there, you get healthy soil. Speaker 0: And from healthy soil, you get regenerative ag. Right? So now I have this this workbench that was temporary to help me do this thing that I just wanted to do that I'm I'm going to turn City into something more permanent. And while I was using it, I was building it. So it's kind of like flying an airplane as you're building it. Speaker 0: This is just for internal use. The UX doesn't really matter. It the UX is almost funny sometimes and, like, how bad it is, but it's okay. I I encourage you AI LLMs, like, how bad it is, but it's okay. I I encourage you to do this. Speaker 0: You can tell an agent, hey. I need a surface for what we're doing here. Because initially, what I was doing is, you know, I was just talking 2026 author. Here we are. I have threads like this over and over and over again AI, hey, let's organize this book. Speaker 0: Here's all of my projects. Look at all of my projects. There are 12 directories for you to look at. Look at the City logs. Look at the database schemas. Speaker 0: What is it that I'm really working on and care about? And City the agent knew. Yeah. You know, like, yeah, you you use workbenches a lot. Loops, orchestrations, knowledge Graphs. Speaker 0: Okay. If I wanted to stay in this environment, I could, but this is so much nicer for me cognitively to to create this thing. So if you're finding yourself in a situation where it's AI weighing on you how raw all of this stuff is and disconnected, start building interfaces. Tell the agent AI, hey, I need a working surface, make it a simple web application, a local database, use SQLite, use Python or Notes or whatever, doesn't really matter because you probably don't have to look at the code A.I you can tell your agent to run the server too. You know, if you've never run a web server on your local computer, tell the agent to do it and have it explain to you what should I know about this? Speaker 0: You know, as a lawyer who doesn't really care about becoming an an engineer, what should I know about this process? Part of this this book that I made is I wanted, you know, reasonable illustrations. So part of this workbench is me, being able to just do a shotgun approach of, hey, create SDGs for me. That might be cool. I don't like that. Speaker 0: I'm just gonna click a button, generate 6 more. Okay. That's reasonable. I take City, I press A.I button, it bakes it into the PDF. I can print the PDF again. Speaker 0: It's really raw. I wouldn't be able to sell this. I'm I'm trying to if I had a friend use use it, they they'll probably figure it out. But mostly, this is tuned to me for this very specific work. Any questions on workbenches? Speaker 0: Anything else? How are you interacting with it? Just on a day to day flow? I Oh, and sorry. Part 2 of that question would be, would you want the interaction to be different depending on what you're working on? Speaker 0: Like, this is a book. Text text is the needed medium. If you were working on a coding product, for example, like how would it change? Yes. Let's look at another workbench that's totally different. Speaker 0: City would change completely. So for example, if I wanted to generate a video for my friend and I was file, I have this idea of a good way to generate a compelling video. We have a drone image. Let's see if I have this workbench running. Okay? Speaker 0: That way it'll be lead. I don't I don't know AI I'm looking at right now. Well, okay. Different workbench. We'll go with this 1. Speaker 0: For Hawaii Tech Week, I'm working with my friend AI to use his database with, you know, thousands of participants and some sponsors and things like that in a way that is actually interesting and smart for everyone involved. Of course AI went straight to lead build a knowledge graph out of this. Now his, he has a bunch of recordings in Google Drive of the different presentations and keynotes and things like that throughout the years. I was able to spin up an agent and say, here's the Google AI, ingest the videos, transcribe them, label them. Only take the long videos because they're a bunch of short videos that are cuts of those that's going to be confusing. Speaker 0: With the transcriptions, run through it and find everything we need for a knowledge graph here so that we can do things like connect people who are working on similar things. We can recommend events to people that are interested in things A.I we can also AI from those previous presentations, we can start to organize what I call claims or statements that people make. So let's see how this goes. So these are the claims of that system. So initially this was just, I think, it was probably codex. Speaker 0: I go back and forth between the smartest models and that usually the highest thinking between codex, Opus, now File A.I the other ones. It was basically figuring out writing Details, Date the Agents, I was talking with it in order to steer it to AI tools 2026 ingest those videos, categorize them, transcribe them, and then we figured out a schema together. It's AI, okay, the schema should be claims. We wanna know what people are saying throughout these keynotes. And and from that, from an actual part of the transcription that's interesting, we can start to build relationships. Speaker 0: Oh, that's an entity. Right? So, like, do we have, we have entities and claims. Yeah. So from so this is my workbench for his knowledge graph. Speaker 0: Right? So and this is a pretty shallow A.I pretty shallow representation of of those videos Date transcripts. So we have entities that have been mentioned. Right? Mayor Rick, that's my favorite 1. Speaker 0: It's so funny. Uncle Mayor Rick is just so funny. We have organizations here. Now the this is at the moment, we're just reading this stuff. Right? Speaker 0: For it to be really useful, what I can do, which I think is baked into this, this is in progress, so we'll see. I like these notable statements. So, let's find a good 1 here. Okay. Sure. Speaker 0: After reaching 2,000,000 ARR, Community should triple triple again A.I double repeatedly as a rule of thumb. Okay. I I don't know, like, if that's true, but it's a cool strong statement. What is the evidence of this thing even being real? Okay. Speaker 0: We have a transcript here. We have this highlighted portion within the transcript. What's the context? They're talking about series a or b. So they're talking about raising. Speaker 0: Okay. Here's the video. HEW lead from the front. Let's open the source. So the question is how accurate is this? Speaker 0: That's what this workbench cares about. So I can tell this system LLMs up or thumbs down on this, which I don't have implemented yet. So that the Agents part of this loop can use that as A.I grounding, really relevant to AI. I want it to only be relevant to Hawaii. I don't care about, you 2026, relevant to Hawaii. Speaker 0: I want it to only be relevant to Hawaii. I don't care about these secrets to raising money unless it's really grounded here. I could write that here in annotations and have this system start to get smarter according to what I want to happen when it does those transcripts A.I then the, the thinking and organizing over those transcripts. So, yeah, the workbench, it's always different. Also, this 1 is, it's a different front lead. Speaker 0: Notes svelte. It's svelte svelte on top of, something called next JS. This workbench is, sorry, these, this workbench is all AI fast API for knowledge graphs. I'm so sad to say this, but I found fast API and AI to be too slow. So I started to experiment and next JS plus Svelte on the front end was just so nice. Speaker 0: So that's what I use now for knowledge graph interfaces. For this book, it's so simple, fast API is AI, you know, like totally fine. Yeah, I mean, I I kind of touched on this, but every serious AI workflow begins as an experiment. I don't I don't know about you all, but I kind of have my modes of, AI help. 1 of them is AI where I put it on highest thinking on pro Notes, and I create projects like, like health, you know, like Oahu AI, AI, lit scenes, another project. Speaker 0: And I treat that like, typically, I treat that like a product owner or some kind of, I guess, CEO, maybe a CMO. And I give it the gist of what I'm trying to accomplish, and I'll give it author a lot of context or maybe I don't have that much context. A.I I'll say, what should I do here? Right? A good example is with knowledge graphs. Speaker 0: It's like A.I year ago or whatever when I really was becoming obsessed with them. It's like, you know, we have these different knowledge graph vendors coming up, open source projects. What should I pay attention to? Is it really this complicated or should it just be a simple Postgres table? And eventually in that conversation AI realized, yeah, I'm right. Speaker 0: Just a simple Postgres table. I don't need to pay for, you know, whatever it's called. I can use open source software with this pattern. It scales. No problem. Speaker 0: Cool. So now now I have some way to guide my experiment. So my next step usually is a coding agent A.I then I give that a ton of context. Initially, I'll take context from tractionpt.com, pro, max, etcetera. Copy it and I'll say, hey, this is our architects plan. Speaker 0: You need to come up with A.I implementation plan for this. I put it in plan mode. I get that plan. I copy that. I paste it back to AI. Speaker 0: AI I say, what do you think of this coding Agents plan? Is this a good place to start? It'll typically say, it's pretty good, but you need to make sure of these things. So you need to change straight things. Copy that. Speaker 0: Paste it. Depending on, you know, how perfect I want that 1st approach, I'll do that a few times A.I then I'll go. I'll let the thing code for 15 minutes to 30 minutes. Now with File, it can sometimes go longer but also it's pretty fast. So author, 30 minutes is the sweet spot A.I then I have a code base. Speaker 0: I have a code base with, you know, for example, I could have this. So so this is an example of what could be generated at that point. If I say to chat GPT Pro, I say, hey. I have Honolulu or sorry. Data.honolulu.gov. Speaker 0: They have a nice API. I just looked at it. I trust it. Let's use it. Let's put this on this map, and let's start creating a knowledge graph that makes it easy for us to show things that people are talking about at community meetings on the map. Speaker 0: So chat should be probably like, okay, Speaker 1: crunch Speaker 0: it all, ask me questions. I usually tell it like, hey, ask me questions, ask me clarifying questions, make sure I'm not misguided. Put it in the coding Date, copy that plan back to chatcpt.com, copy that result back to the coding Agents. Let's go. 30 later, we have something that really works. Speaker 0: Now what doesn't work initially is the data integrity, right? Because it's going to cover gaps in order to hit those AI, and you have to work through this in a loop to make the data have high integrity, which means that you need to communicate with the agent what that means to you. So for me, what it means is AI these statements have to be accurate. A.I if we don't know the person who's speaking, we need to be, you know, truthful about that. We can't hallucinate a name. Speaker 0: And we also need to get like Hawaiian words AI, Hawaiian places, right? So how do we do all of that? There's there's ways to cover all of that stuff. Okay, any questions, thoughts, concerns? 2 questions. Speaker 0: 1 was, you mentioned going from the coding agent out to highest level Yeah. Chat gbt and then back. Yeah. Do you not have access to chat g b t in the coding agent? So part of this is keeping myself sane. Speaker 0: I have Folder and boundaries with the interfaces that I use. Right? So I I am comfortable having a chat box A.I a browser tab be my strategist. I'm not so comfortable with a coding agent being my strategist. They're very biased towards code and like really getting things written. Speaker 0: So when I'm thinking and architecting, thinking about any sort of like business strategy or approach, it's it's just kind of nice for me cognitively to stick to that browser window. And then when I'm coding, it's nice for me to stick to the terminal window because I'm always checking git anyway. A.I then, you know, if if I'm trying to design something, if I'm having trouble design for me, I open up Lead. The Claude API for me is like my design partner. Sometimes it's Claude AI, sometimes it's just the ClaudeChat. Speaker 0: This is kind of at the moment AI how I've found all these tools that work well for me A.I it shifts over time a little bit. You know, Claude comes out with this thing called AI. Okay, I'll try that. Sometimes it's Date, but other times, I have to go back to the lead chat. And then sometimes I don't like what it's showing me. Speaker 0: It's AI, I don't know if, you know, the the performance is degraded or it's just not good at this thing. I have to go try Gemini. So so yeah. Does that kinda answer that? Yeah. Speaker 0: I think so. I guess AI just wonder why not start in chat GPT with the original plan, I suppose, or and and it sounds like it's just it works. This workflow works. Maybe chat to continue. It doesn't spit out those thoughtful plan that is congruent with the way the coding agent is prepared to operate Right. Speaker 0: Or something like that. So Yes. Okay. Carstack is starting in the right context, taking it out for refinement, and then bring it. So the CLI agents, codex A.I lead code are really good at understanding code bases. Speaker 0: And for me to try to let the ChatGPT browser session understand my code base is just too much guessing and too much uploading and copying and pasting. So I keep that at a high level. I I treat it like a business stakeholder or an executive that doesn't wanna know about the code base. You know, they they have very serious and and valuable AI. But they should probably never really look at the code base or at least stay out of the executive function on the code base. Speaker 0: Right. It's file, give it the Riggen it wants to know about the code base, but it wants to know, like, what you're already planning to do. Right. It's not think about it by itself. Yeah. Speaker 0: That makes a lot of sense. Thanks. The other question was about making Hawaiian more accurate. I haven't put any effort into this, but I do have an interest in Yeah. Just because you mentioned it. Speaker 0: Yes. So you have so for transcribing, for example Yeah. I was making some flashcards or whatever. Okay. Well, I'll tell you my solution for transcribing. Speaker 0: Right? Because you there there is the open AI whisper, service for transcribing chunks of audio. It works really well. There are competitors or open source solutions. They all will fail with Hawaiian words. Speaker 0: But if you give a system instruction and not all of them support this, but 1 of the chat, sorry, 1 of the open AI models does. You can give a system instructions that we're talking about Hawaii right now. Here are some of the places involved. This is what we're doing. Here are some of, the words that I know have already been said A.I it will get it pretty freaking accurate. Speaker 0: Yeah. Speaker 2: So Yeah. Speaker 0: Transcription team wants to help with that 1. Yeah. I mean, it it was I started with whisper, took it back out or something. No. I used the OpenAI Whisper service for for this in particular. Speaker 0: So there's this 1, there's the HTW 1, the AI tech week. Both of these are using that OpenAI AI transcription service A.I there are 2 different models there. 1 of them labels different speakers and 1 of them doesn't. 1 of them allows you to give system prompt and the other doesn't. So you have to, you know, there's a cost benefit here. Speaker 0: So I use the 1 that allows you to have the system prompt. I give it a bunch of flying words and more context about what's happening A.I the transcriptions are much better. Yeah. Yeah. No worries. Speaker 0: Okay. Okay. So I think, you know, workbenches, I'll AI tell you about the other 1 for videos. I don't have it ready to 2026, but another example of a workbench to me that I'll use is AI, I have this idea. I have this drone footage. Speaker 0: I want to take the drone footage, send it to an image model, have it transform it into something that looks like a sketch. And then I wanna splice that throughout the video. So there's like this pulsing effect. So it's the real footage and then it's A.I sketch. AI I wanna do that over and over again. Speaker 0: So I could make a AI script out of that. But I learned that running the python script and then opening my files and then looking at the video sucks. So I built a, I had the agent build a web interface that, you know, I can just click buttons now, upload AI, go, Yes. No. Modify the prompts A.I then just kind of repeat. Speaker 0: So the workbench thing is AI, how do you speed up your own, process for transforming things or building things on your own computer. They're not really going to be hosted on the cloud or anything else. So this gets into loops. So So like, you know, for example, with the the frame AI thing, the initial implementation is me talking to an agent, showing it the video file, telling it what kind of transformation I want. It starts to write some code. Speaker 0: It's like, hey, you wanna run the code now? Let's run the code. Let's see the output. Okay. It's okay, but it's choppy. Speaker 0: What can we do about the choppiness? Here's 1 idea I have. What ideas do you have? You know, here are 3 ideas. Okay, let's go with that 1. Speaker 0: Rerun it. Okay, it's good but you know, there's a black spot in the beginning. How do we fix that? You do that over and over again, right? That's the loop between you and an agent. Speaker 0: It is a good idea to think about what you're doing with AI as loops that you can start to automate. So, for example, it's ambitious, but 1 way you could automate that, is is by, for example, saying you want it to look smooth and impressive and good. So I want to build some system that that does the video modifications, writes the tools and the scripts, generates the video based on like 10 different sources I have, and then rips open the frames and then analyzes the frames to make sure there's no black space. Right? And you can do that over and over again until you have a process in the code itself and the tools that the agent is using that is foolproof A.I then you can take the agent out of that system. Speaker 0: Now you just have a script that does it. The feedback mechanism is very important here. It's something to really think about. You know, we have the inputs and the outputs, but the thing that surrounds those is really worth thinking about. Now this is where the term harness harness engineering came from. Speaker 0: People realized that when you're, especially when you're writing code, you have this harness that writes the code and then tests it by like looking at it in a different way and then runs it and sees what those outputs are and then gives a yes or Notes and throws it back into the Loops, then we have a way of generating code that's quite trustworthy. And if you 2026 some things AI really high quality and you can even bring the model way down. So you don't have to use Fable with a proper AI, you can go down to SONNET or whatever model is cheaper, an open source model. Yeah. Any questions? Speaker 3: Can you describe how to build the harness a bit more? That was something that I've heard people just nonstop using that word. Speaker 0: Yeah. Speaker 3: It like came I'll tell you. Fast and furious, but I don't really file mental model. I don't really understand. Speaker 0: I wish I had I wish I would share this more. So this is a really good way to start. Let me see if I need to find the right Agents, and it's not this. K. So let's let's just find it real quick. Speaker 3: Is there another way to think of it like a steering file or is it Yeah. Speaker 0: Yes. Speaker 3: Okay. Speaker 0: Yeah. I think that it would be in this 1. Let's see. Yeah. Okay. Speaker 0: So So agents dot m d is a file that if it exists in the folder you're running cloud code in or any other CLI agent, the agent will listen to it. A.I then so, like, caveat, cloud code might Notes. You might have to change it to cloud dot MD. City kinda shitheads about that. There's been there's been an open issue on GitHub for, like, a year. Speaker 0: We we have all these ways to AI get around it. It's easy to get around, but it's also annoying. We'll just talk about agents.md. You might need to name it plot.md. You might need both or whatever. Speaker 0: So the the harness that I use all the AI, it's super simple. They work that is so so nice is basically this loop to follow in my Agents. Right? So I say so 1st of all, upon initialization echo reading agents.md Date way AI know that this thing is being listened to. So that's just A.I sanity check for me. Speaker 0: Every time I see it saying that I'm AI, cool, it's listening to my instructions. AI have a docs folder and I keep something in there called vision dot MD. Dot MD is a file extension for markdown AI. They're great. It's kind of like a word file if you haven't used it. Speaker 0: It's just a really nice way of, having human readable and machine readable thoughts and notes and logs. So I say, check vision dot m d and if it doesn't exist, make the user articulate their vision here. I found that to be really helpful and important, for a few reasons. Okay. So let's see if there's a vision.md for this project. Speaker 0: There is. It's simple. AI is a neighborhood signal map for Oahu. It turns public traces such as neighborhood board meetings, etcetera, into evidence backed map of what is changing, what's broken, what people care about. Perfect. Speaker 0: I don't think I've read that before but it's perfect. That's great. So the 3rd thing is help the user make wise principle level modifications. Principle level means if this was the highest paid engineer at some org, principal engineer, you need to think like that person. You need to help this person. Speaker 0: They're not that person. This is a solo dev. You need to guide them to do the right things. You know, don't let them, create a database in HTML. AI, that's silly. Speaker 0: That's not gonna be good for them or for you or anyone who takes this project. Use SQLite or use something else that makes sense. I say, we're gonna keep going. I tell it to keep a log. So I've experimented with different ways of doing this. Speaker 0: But in this case, just keep a log, log. Md of everything that's happened. Let's see what that looks AI. Okay, so here's the log on this project. Just 186 AI. Speaker 0: 2 of them are not even committed. So I can see the last things I did starting a 20 minute technical presentation deck for the Oahu AI Civic Knowledge Graph. That's the 1 of the decks I've been using today. Okay? So in this project, I said, hey, I have this book. Speaker 0: I have this thing. It's probably a good, it is a good demo for Knowledge Graphs, create a deck for me. I don't think I use this deck actually. I use the another project, the book 1. And the last thing, Autocomplete City. Speaker 0: Cool. So now I know what the agent has been doing. There are other ways of doing this that are more surefire like checking git Loops, but this is nice and also it's a good memory for the agent itself. So you can say, hey, remember AI when we refactored the map from this 1 thing to that other thing A.I it's way faster now, but we missed, this this way of like zooming in and I wanna find that again. The agent can read through this and realize, oh, he's talking about this project. Speaker 0: I know exactly what to do. Yeah, I have, I have finer grained logs in here too and I have a thoughts dot MD in here. Okay? So these are the thoughts of the agent as it's running through things. I wanted to keep this because of big decision points that happen. Speaker 0: I didn't wanna lose the reasoning behind them. You know, file, for example, why use Svelte versus React? Okay. We're looking for highest performance possible. Svelte, it it doesn't matter, but, like, now the agent knows highest performance possible is why we're using this this other project. Speaker 0: And if we come into some problem with it, like, hey, react would be better. It'll be like, oh, but you know, there's a reason for this. AI is 1 that's cool. This is for humans Notes agents. So really, if I had to, I should be able to recreate this entire thing by copying this file. Speaker 0: There there have been some memes about this recently with, AI think Karpathy, just a legendary AI guy who figured out AI or, you know, he he wrote a post about how you can just have an instructions repository A.I someone can copy and paste it into their coding agent and get exactly what you already coded, but it's not your code, it's their code. This is similar to that. AI, you can copy and paste this A.I now you know exactly what needs to be built technically. Yep. And then the plan. Speaker 0: Plan. So I use plans a lot. So that means like this thing is in plan mode, right? And I ask City. So shift tab down here, how it says plan mode. Speaker 0: It can't change code or AI to my file system in plan mode. So now I'm I'm very comfortable just letting it run and going about my day without being concerned or AI thinking too much about what it might be deciding. So I can let it think for 15, 20 minutes on something and then come back with a plan A.I I can read it or more likely just copy and paste it into my strategy agent A.I say, hey, how does this plan look? Should we go with it? I'm happy to share this Agents. Speaker 0: Md with you AI. But this is like a very lightweight harness. Now, if I wanted to so this is like the lightest weight harness. Right? There are other there are other types of files you can use for your agents to follow. Speaker 0: There are skills where you AI very comprehensively about some way to do something AI scrape Date, and it'll find that skill and use it. You have tools where you have actual code that for example is really fast at scraping data A.I it'll find the skill with the tool and it'll use those things. Harness engineering mostly is concerned with coding agents writing code that you can trust. So they have things like linters and testers and ways to spin up code that's been generated into some new container and test it through there. Maybe Event get screenshots of it for the agent to look at. Speaker 0: There's a lot written about this. There's a company called Human Lead, I think. No, it's not that. Sorry, let's find out. It's really quick. Speaker 0: Human layer. Yeah, they have an AI out that is like, you know, harness engineering, maximized. I haven't exactly tried it. I installed it and I didn't really spend time on it, but I'm sure it's good. They were very smart. Speaker 0: Okay. So before I just continue, I'll ask you guys if there is interest in talking about agents A.I orchestrating them, scaling them, or not. Is anyone having to worry about that right now? I mean, Yeah. Yeah. Speaker 0: Yeah. Anyone else? Yeah. Yeah. Okay. Speaker 0: I'm gonna grab a sip and then think about this. There there are a handful of patterns here. So I have this coding Agents. It writes code for me. It can do a lot more. Speaker 0: It can open up websites. It can scrape them. It can use databases. It can use tools. It can use my file system. Speaker 0: It can generate images and videos. So if you are a marketing agency and you start writing skills and you start having a harness around a coding agent that's sitting on a computer, you can generate marketing materials in a way that's automated. If you have, you know, 12 brands that that are your customers A.I you have their brand briefs, their brand identities A.I Agents, you can tell the coding agent for each brand generate 3 ads based on their brand identity. Right? Now AI can do that on my computer. Speaker 0: It can spit out the files. But what if I want that to be in the cloud so that my computer doesn't have to be running or so that I can scale this to 12 people that replace me? There are a few different ways. I have tried all of the ways I can think of. I'm gonna find the chapter in this book. Speaker 0: It'll just, it'll just keep me organized A.I I'll AI. I don't know. Okay. Yeah. We got some lag here. Speaker 0: So 2 of the ways that I built full solutions to a few months ago. 1 of them is a local running Agents, and it is being controlled by by other people's inputs. And it's syncing with other local running agents via a shared database and signals back and forth. So this computer, his computer, her computer are all creating ads for our marketing agency. We don't want them to be competing with each other and generating the same ad. Speaker 0: We want them to know about which ads are being generated. We want them to have the same data that they're working off for those Graphs. And when they're finished with the ads, we want them to all have access to the final product. So if you have a local web server that or scratch that, okay? If you have a local agent that has a connection to a database and so does he and so does she, and you have skills and instructions that say how the database should be Tuesday, and you have an API that the agent can hit in order to queue things like image generation unless you're just generating on your own Community. Speaker 0: A.I you have a way to actually coordinate multiple agents and scale something like this without having to set up really any serious cloud infrastructure. You have a database that's hosted, maybe an API. In this case, no API. Just decentralized coding agents with the same GitHub repository, looking at the same skills and ways of keeping, duplication of effort down and keeping statuses synced across them using simple Date. The other approach is a cloud running suite, right? Speaker 0: In this Date, using AI AI Graphs agents with A.I web interface. So that means, let's see if any of these illustrations are helpful. Otherwise, I'll just verbalize this. Yeah, AI mean, AI this. AI, I mean, I can say that it's accurate at least. Speaker 0: So, this is complicated. You know, the Riggen that my previous example made sense is because of how complex this example is. Because there are a lot of ways that it falls short. We can go into those if you'd AI. But for example, if someone, someone's computer runs out of battery in the middle of generating an ad, all of the other systems are probably going to be blocked thinking that they're still working on the ad. Speaker 0: You know, you could get around that by starting to like get really sophisticated with your signaling and your statuses but at that point you're really over complicating something that is not too dependable. This in this case, these are cloud containerized agents running through kubernetes and docker containers with a dispatch broker that is just a simple API and a web interface so that non technical users can go in and run these agents similar to how you would on chatcpt.com, but have them spawn out the the container that they need to do what they're asking to do. So for example, if they're in this case, like running through millions of records inside of, what's it called, Snowflake, it's going to need to have a lot of memory in there in order to bring those down and make sense of them, create the report for the user A.I then give it back to them. But if they just wanna have a chat about, hey, like what Notes our schema look like for, you know, customers who signed up in the last 30 Date? Then you need a small container, small memory. Speaker 0: The dispatch broker can control that stuff so that your costs are under control because something like this can spin out of control pretty quickly when it comes to costs. So you have a browser workbench controlling a dispatch broker that's a simple API connected to the Date, to artifacts. Artifacts are stored within the standard. I think on Amazon AI EFS or elastic disk storage or whatever it's called. So that when they're written and the broker goes down or that agent goes down, it's still persisted. Speaker 0: Still persisted within our cloud environment in a way that, like, the IT and security team are comfortable with and used to. You can make a system like this run through all of the permissions that the org is already using. In this case, it was through, what is it called? The oh, man. It's Okta. Speaker 0: Right? So you can do that by doing, JWT handshakes. So you have an enterprise system using Okta for me to log in as a data engineer. It knows that I have access to x, y, and z through the token. It's just a json file that we can trust inside AI the broker. Speaker 0: The broker now knows what I have access to. So if I say, hey, I wanna check out the schema for HR and I want all of the data on everyone's salary. It's actually gonna know and not provide me, not even try to provide me with A.I container to do that stuff. Because the Okta JWT tells it what I actually have access to. And if it made a mistake and tried to give me access to that stuff, then the systems downstream would stop it anyway. Speaker 0: So, yeah, it's interesting AI these 2 examples are actually really far on the spectrum. Right? 1, you're like, hey, we're gonna use cloud code on your computer, a simple database, and you're gonna be generating ads with your teammates without causing problems for each other. And then the other case is like, no, we're going to go through Okta A.I web interface. We're going to control the containerized environments. Speaker 0: It's all going to be sandboxed. It's all going to be auditable. AI guess, o c 2 compliant auditable. And, you know, one's pretty easy to set up and quick and messy, and the other 1 takes a lot of time, months of effort probably 2026 to actually get it deployed in a real org. So after all of this, there was something released by Lead called routines AI Anthropic. Speaker 0: So now you can use something called cloud code routines that lets you deploy your agent via a GitHub repository into the lead. And then it has access to an API so you can use Loops from it A.I you can host to it in order for it to start up and then use that API to hit your systems. So now the the 1st example that I gave of a local running agent in tandem with others is probably best represented in that way because you still have the simplicity of this just being a GitHub repository that is an agent AI any other cloud code designed agent, you know, Agents. AI or cloud. Md skills, tools. Speaker 0: But instead of having to sync up all these devices, you can know that it's running in the cloud and has access to your API. And you can set up a button on a website and someone clicks the button, it runs that routine. And it can take input parameters too. So you can say, hey, when when someone clicks this button, it looks at our database, it finds the brands that need ads, it sends that info, it sends 1 of the Graphs info and like their previous ads to this routine, it generates ads, it hits our API to get more data and more context A.I then it writes a file on the end, right? So that's what I would recommend if you're trying to scale agents in the most simple way possible. Speaker 0: And AI API, so the, I would say like the middle ground here, something called Pydantic AI. There are similar competing sorts of solutions across different languages. Pydantic A.I is the Python library. That is a more like deterministic way of writing agents using code, using classes in Python that still are agentic, but it's running through a deterministic workflow. And you could take that really far if you want 2026, the point where you know, they have something called AI AI Graphs, where you can Date any point in the Workflows. Speaker 0: So let's say we're in the middle of generating an ad for a brand and we're checking out their competitors and what they've recently been showing on Instagram and the systems go down, you have you have something in that graph that is saved to your Date. Right? It's the state right there. When it comes back up, you can just run it right from that point deterministically A.I it'll behave the same way every time. It'll go through the nodes in the same way. Speaker 0: Any questions? What is the name of that? Routines. All routines. Yeah. Speaker 0: Yeah. It honestly, I AI think so I think that might be a game changer. Like, you don't you never know how much support they're going to give to this. For example, I don't know how many parallel calls you can make to a single routine. But if it's stuck in the dozens, then I think it's a game changer for a lot of people. Speaker 0: Because the alternative would be to use something like Vercel and transform your cloud code native agent into some web application that is agentic using AI AI or whatever, lang Graphs or something like that. And that is just so much more work for a non engineer. Right? So so for someone who's comfortable using cloud code or cloud desktop or co work even to automate things, I think routines are going to be a game changer for them because they don't have to go sign up for Vercel and learn this and that and that. They can just use the routine. Speaker 0: Now there will be some struggle with how do you save the outputs of that routine. But I think your agent can help you with that. And I'm really curious is Anthropic gonna lean into that? Are they gonna offer databases or AI, you know, some clogged code native database that works author teams or something? I don't know. Speaker 0: It's very new. I think they released that file 3 weeks ago or 4 weeks ago or something. But, yeah, I have I have an experiment running on it. It works well. It's slow. Speaker 0: It's a lot slower at the moment than what was running on, a computer, but AI, you know, we'll see, but I think that I have ways to optimize that too. And even if it's slower, if you can run multiple requests at once, then it's probably faster in the end if you have that parallelization. Any other questions on, agents scaling them different patterns? Yeah. Hey, Kevin. Speaker 0: Yeah. What type of tasks would these routines be best used for? I think that they would be I wouldn't use them for coding or, like, for generating code, I don't think. AI use them for generating static assets, HTML pages, text, images, videos. I think for anything you might use a cron job for, anything you might put into a Lambda or AI a microservice traditionally, would they'd be really probably good for. Speaker 0: It's like, if you've if you've had A.I a CLI agent file Lead Code do something for you more than once, and, like, you you go back to it. Like, for me, it might be scraping some data. Right? I'm like, I don't really wanna write a scraper. I don't wanna use a 2026. Speaker 0: I'm just lazy. Like, hey. Like, what's happening on even for, like, the the I'll I'll give you a real example of of something I might do. Speaker 1: So so, Speaker 0: actually, the the brand information example is good. If if you're some ecommerce brand and you have 5 competitors A.I you wanna keep up with what they've been posting and then you wanna make sure that, like, your messaging is up to date with that, you could use cloud code or the cloud desktop app to do that for you every day. You can say, okay, check it again. Next day, okay, check it again. AI day, okay, check it again. Speaker 0: Anything like that, you can host up into the routine and then set a timer to just hit the button Event day. So it's not on your Autocomplete. And then you can have that routine text you at the end or send you a WhatsApp message or a Telegram message. Maybe, like, a, like, a daily instant dump or something like that? For sure. Speaker 3: Yeah. Okay. Yeah. I have right now, like, some automation still through skills, like, to do regular work for me just like that. And I'm wondering, should this be your Speaker 0: routine instead of You should you should consider it, try it. Speaker 3: Yeah. What would be the benefit of Speaker 0: doing So versus If you you can use the routine in order to run that workflow that you figured out without your computer being on. I don't know if that's helpful to you or not. Does that not happen with skills? There no. Not no. Speaker 0: So Lead does have something called I don't remember what the other thing's called. It's AI where to run your stuff in the cloud. And you can download the mobile app and talk to it. That's another thing I haven't experienced with, but probably works fine too. I think the routine is good when you have a code Date that an agent works really well with that you are finding yourself using repeatedly. Speaker 0: And if you want, honestly, probably, if you're solo, you probably don't need it. But if you're trying to scale this to multiple people, it's a very good thing. So, like, if you have somebody Workflows your laptop, but you wanna invite a colleague or a friend or whatever to use it, that's going to be difficult to keep tidy, but the cloud routine makes it very easy to keep tidy. Yeah. Yeah. Speaker 0: Okay. I think we can get to knowledge graphs now. A.I you have any questions, concerns, thoughts before I start talking about this? Yeah. Speaker 2: Sure. So you've been here for a while. There's, you know, a couple months ago, the Kona low storm. I was just curious what thoughts or ideas you had for you, like, this, this map here. But do you have anything that kinda came to AI, file, any epiphanies or ideas that came up after the storms? Speaker 0: Yeah. I mean, I'll just speak off the cuff. I don't know if you all are familiar with that flooding. It was crazy. There was 1 day where we got 13 inches of rain in, like, 2 hours at, like, 2AM or 3AM. Speaker 0: Nuts. In Waialua, there were people who would have died without the local guys with heavy machinery driving around Date saving people. The police were telling those guys to stop helping because the next day A.I least, stop helping because the state government wanted evidence in order to get funding from the federal government for relief. Right? So so those guys author they were done saving people were cleaning up neighborhoods. Speaker 0: Like I have photos and videos that are are nuts, you know, just on their bulldozers, big trucks, you name it. Hundreds of people with shovels. These guys just kind of coordinating with 1 another. The fire department didn't have the tools or training to help people during that initial night of flooding. Turns out the fire trucks can't go in water that is greater than 18 inches deep A.I they don't have fast water training. Speaker 0: They were instructed A.I, you know, probably understood that they couldn't really help people that 1st night of flooding in the dark especially. So the neighborhood really was left to itself and honestly the guys who don't really follow the rules a lot of the time, right? These were guys who they're very rough, you know, the cowboys and and contractors, construction guys, people who can't vote, you know, like they have a felony. They were out there helping people. And the police and fire department didn't have an answer that they had an answer for. Speaker 0: So that's really interesting. I think other than that, there were a lot of people with a lot of knowledge about why that happened. They had not been listened to A.I that was captured. There's there's 1 meeting in particular in these in this data that that was captured really well in. So people that knew, like, why those areas flood. Speaker 0: You know, historically, it makes perfect sense they're flooding in this way. Also, like, there are, you know, 13 houses on Waialua Beach Road that have illegal AI. And also the waterways have not been cleaned out A.I also like there's this litigation where it's not the state's responsibility to clean them out, it's the private landowner but no 1 even knows who the landowner is A.I now there are squatters. So like all of these issues came up. And there were some people who really understood what was going on but they had no Agents. Speaker 0: Because even though they would voice those things Notes like there's no real solution, you know. Another thing is AI the military doing armament testing on A.I mountain range that changes the mountain, that changes the way that the water comes down, causes issues. So there was a lot of, I guess, knowledge that I saw was being just Graphs compartmentalized, etcetera. And then on the emergency response side, it was just people who seemed totally handcuffed by, bureaucracy that didn't really make sense anymore for these situations. I didn't see anything that was AI silver bullet. Speaker 0: Let's use AI to solve this, unfortunately. But but I I did come away with it AI, oh, AI I me personally, I care about this stuff AI I live there and my kids are there, it's my community. You know, when you see those things firsthand, you see these guys at, like, 4AM, 5AM nonstop Engineering the highway that I can use to go get to my kids, you really, like, start to feel strongly about it. Like, okay. Who are these guys? Speaker 0: How do I help them? How do I make sure that AI they're empowered instead of whoever is telling the police to stop them, for example. So that was the inspiration for me to build what this thing is. You know, I don't know if it's going to do much. I'm open to it doing a bunch, I'm open to it Notes. Speaker 0: But in the meantime, I'm I'm happy with it technically for sure. So and a lot of this is predicated on work that other people have Date. And a lot of that, by the way, is is not seen either. So we have people within the Hawaii government who have been doing a very good job getting a lot of data to be publicly accessible. So there's something called data.honolulu.gov. Speaker 0: There's great data here. So public safety, finance, transportation, there's a whole catalog here, business. This is very dense data and you can quickly have an agent write an API to, consume it. So AI for example, I'll run through the demo of this thing and then I'll actually show you the database and and explain why I'm saying knowledge graphs give you a ton of power right now. So let's just look at this map, right? Speaker 0: The, the top are different categories. So voices, these are community meetings A.I the topics that happen at them, sourced by A.I YouTube channel channel that Honolulu County keeps up to date. So they film these community meetings A.I they upload the videos. Ahupua'a. So this is something that used to be very relevant on this island. Speaker 0: These were agricultural systems, food systems typically spanning from the ocean to the top of a mountain. So you have an entire food system for some Community, you have some people who own it A.I then you trade between each other. This is really important stuff if you care about agriculture on the island. Probably more for historical posterity AI also for inspiration more than any like literal usage. Unfortunately, it'd be cool if we could do this again, but it's just not feasible now. Speaker 0: Right? Legislature. So I've spent very little time on this, but there's indexes of legislature in here. Agents is interesting. So these are arrest records. Speaker 0: You can learn a lot through these. You know, like I I organizer there are just not that many arrests up in North Shore compared to AI what I see actually happening there. And there are reasons for that. You can start to to look at this and be AI, why is that the Date? And then learn interesting things. Speaker 0: You could talk to the police and they're like, yeah, there are 3 of us that span from Mokolilla to Kahuku. There are more now, but you know, when I talked to them a year ago, there were 3 of them A.I they don't typically arrest people unless they really need to because it takes them 6 hours to go down to the station and book them and fill out all the paperwork A.I then they're not up there to like help. So you can really quickly find, really big problems here that AI Notes one's going to disagree with you about by looking at some of the data, which is cool. A.I then building permits, you know, this it can be interesting depending on what you're into and etcetera. But for me the voices part is most interesting. Speaker 0: North Shore flooding, Waianae Coast, Whitmore Knowledge. Okay, so the reason that this says flooding 196 is because on the left side I clicked on flooding. AI I go to all topics, I have 359. I think this is from like the last 4 or 5 months of those YouTube videos. So the way that that system works is I I think I have a reasonable AI thought I had a reasonable slide here for it. Speaker 0: I don't. No worries. Actually it might be like no that's not it either. Okay. No worries. Speaker 0: So the way that this system Workflows, this started with an agent like Claude Code and me in a continuous loop is go look at the YouTube channel, channel, find the videos, what date was it, where was it, map that to our database of places on Oahu. You know, the LLMs already know all the places in Oahu. It's a small island. Cool. Transcribe it. Speaker 0: Chunk that. So you need chunks of the transcription. You cannot take 2 hours of a transcription and send all of that at once to something for analysis. You can, but in this case, I want more fine grained, data. I want more rich data. Speaker 0: So I chunk that. That means that I get chunks of it. Analyze that into our database schema. Do a sanity check on it, and then repeat. Now when we repeat, we are going to be considering the data that was already written. Speaker 0: Because, for example, we have these categories like flooding, community, emergency, and governance. And when someone says, hey, the military keeps shooting the mountain A.I that's a problem for flooding in minute 1 A.I then they say in minute 60, hey, the military should be here to talk about this with us. It needs to know that those are related. Right? Now it can know that those are related by generating labels for them like military, flooding, concern A.I they're going to have the same labels. Speaker 0: But it can also, sometimes if you wanted Date, have summaries of those previous things and say, hey, these things are related. So LLMs are smart enough to do like a lot of different ways of, joining categories. So AI would like to find the actual statements here. Now the board chair said the parks committee head would work with the speaker to see what can be done to make the proposed community project happen. Happen. Speaker 0: Okay. We have all these things that people are saying across the island. As far as flooding goes, pretty much everywhere across the island, people were talking about this. Right? Because we had a Kona Loops storm, we've had other storms, flooding is just a problem here, it happens. Speaker 0: Emergency services. Okay. You can tell AI the North Shore 1 way more than the rest of the island. Can I trust that? No. Speaker 0: Because my system's biased to the North Shore meetings. So now I I need to, like, have a way of knowing what has been processed here. I don't think I've solved that in this system, but here for HawaiiTechWeek, if I go to sources, see how AI see which videos I've actually consumed. So I know that there's a bias here. What I do is AI don't wanna consume all of this data right now because I'm refining the system still with Asia and I'm AI sanity checking it A.I these are passion projects, so I can't spend all day on them. Speaker 0: So over AI, I just do a little bit at a time and make sure, yeah, the data looks good. If the data looks bad, AI worked with KK agent in order to improve that. And that can look like it changing code maybe, but probably changing the prompts that are being used to analyze this stuff A.I sometimes AI data schemas. So knowledge graphs, what are they? They enable you to do something like this where you don't just have a bunch of text that gives you a good amount of information if you can read all of it and understand it all in your lead, but it gives you a way to move this data in different shapes where you can say, hey, what parts of the island are talking about, the number of police officers that are there? Speaker 0: Like, how many parts of this island are complaining about the lack of police or the lack of ambulances, for example? How many of them are complaining about, like, military stuff? You know? This is interesting thing to things to know and the reason that I can pause it certain questions that I couldn't predict yesterday, and 1 of you could throw a question out to me right right now, and we could figure it out with this data is because of the way Speaker 3: the data is organized. So I have a question that might be dumb. But No. No. No. Speaker 3: Were you sourcing the I thought the knowledge Graphs were from, like, government websites. Speaker 0: Are they Oh, okay. So it's a combination. So, actually, to that question, AI I thought I had this prepared. Yeah. Okay. Speaker 0: So in my database, I'm I'm storing the ingestion runs. So every time we go out and fetch data. At the moment, I have 3 types of runs in here. So 1 of them is crime incidents. 1 of them oh, you can't really see that. Speaker 0: It'd be better if you did. Yeah. Down here. HPD crime incidents, that's coming from Date, and that's it. N c o YouTube, that's coming from YouTube, but the the Honolulu YouTube channel. Speaker 0: Right? I can open that up actually because I have I have the provenance for these in the UI now. So so this statement, the speaker said Community members helped perform rescues. That happened right here. Speaker 2: That's the mayor. Speaker 0: That's that's mayor Rick. Speaker 3: Yeah. Is he summarizing this video and taking data from it? Speaker 0: Or It is downloading the video, isolating the audio file from it. So so I get a video. I have tools running on the computer to grab the audio from it. Now that's an audio file. I cut that into pieces. Speaker 0: I send it to this opening AI service using whisper. It sends back transcriptions. I save all of the data AI the timestamp, for when that was said in the video. So that when I Date some logic about that text, I know that it's tied to minute 10 on this video now. So I know that when the mayor said a lot of those rescues happen as a result of the community and people helping people, I know where the source evidence is AI I can link straight to City, which is linking me here. Speaker 0: So the reason that that's even possible is because of just a few database tables that are designed in a way that knowledge graphs are, that I worked with an agent to design A.I that I worked with an agent to hydrate from those YouTube videos, which literally could be as simple as saying, hey, here's the YouTube channel. I want you to create a knowledge graph from it that's super simple in SQL. And it'll basically do this if you have a smart enough model. Other considerations will be like, well, what about automating it, keeping track of the ones you've already ingested, keeping track of partial ingestions and failures, etcetera. Those will require other database tables and more code. Speaker 0: But the the thing that's beautiful about 1 of the things that's beautiful about AI is that I can Graphs. So not just that YouTube channel, but Honolulu Date, they also have information about these meetings because they they have a website that they post 2026 also. So I can cross reference these things. Sometimes if I don't have speaker information, because of some technical difficulty here. I don't know who's talking right now. Speaker 0: Sometimes the description on the Honolulu data source that is textual will mention that the mayor was there. It's like, oh, maybe it's the mayor. Just a little example. So you can cross reference exam, data really well. You can throw totally, Loops say AI Notes AI data at l l m's and have them normalize it for you. Speaker 0: So back in the day, we would have to normalize all this stuff in like really heavy duty systems. Because L. L. M. S can kind of make sense of this stuff for us. Speaker 0: We can say, hey, take all these disparate sources of data, videos, images, forums, databases A.I throw it into this knowledge graph. And also AI, tell me what schema we need so that we are representing it in the most accurate way that is actually, versatile. Right? So accurate way that is actually, versatile. Right? Speaker 0: So that we can use all of the data from all these things in the best way. So those are the types of injections that we have. So crime incidents, NCO, YouTube, and then building permits. I would expect more. I don't know why there are only those 3 right now because for example, you know, we have, legislature too. Speaker 0: So, like, why isn't that there? I'm not sure. It'd be interesting to to check that out. As far as ingestion runs Notes, so so these are the actual runs of ingesting, for example, this YouTube URL. This was the model that was used, GPT 4 0 Transcribe. Speaker 0: In a year, if there's a way better transcription model, I can tell an Agents, run through these, re Riggen the transcription using this better model and save a new version in the database A.I then look at the results and tell me if it's worth upgrading to this model. Sometimes it is, sometimes it isn't. 1 of the things I didn't mention so far is this concept of evals or evaluations. As a solo dev, it's not too important but it's worth thinking about when you're changing a model, when you're changing your prompt, what's coming back and how the hell you're gonna know if it's better or worse over AI? That's hard. Speaker 0: You can build a workbench to do that. You can build a workbench to for example, I could say, right now, I wanna try Google's transcription service. Build me a workbench to show on the left AI, the video. In the middle, the GPT 4 0 transcribed result, and on the right, the Google AI result. And I'm gonna go through A.I I'm just gonna eyeball which ones are better and I'm gonna tell you or I'm gonna click a button, to to tell you which ones are better A.I then we're gonna see statistically, you know, what's better. Speaker 0: If you could even say like this model is better for shorter videos and this one's better for longer videos that way. There's a lot of considerations here since we have models that are parameters, right? Like like we could use clause, we can use GPTs, we we can use this 1 or that 1. Okay. The media assets part of this, This keeps track of AI the actual videos. Speaker 0: Right? So now I have history of all the videos that this system has consumed. I have a this is basically City fingerprint. So even if someone went and uploaded a new video at that URL, I could find out that this isn't really the same video. The codec used m p 3. Speaker 0: Right? So so that's the audio file that was generated from the video file. Local status deleted. So part of this process is downloading these videos on my computer, doing that processing for them, and then uploading the outputs to my database. And I could get clever and do other things like lead upload the video to my, bucket so that I have a backup of all these videos. Speaker 0: Or maybe I wanna do something weird AI transform these videos into, you know, what Hawaii looks like in 100 years. AI I'm gonna do that by building another agent that consumes these videos that I've saved from YouTube. There there's a lot of once you have a system like this in place, there are a lot of options at your disposal. Because now I know, okay, AI, go tell the agent before you delete these videos Applied them here, you know. And then I can have another status that's AI uploaded. Speaker 0: Now transcript chunks. Yeah. So the you know, the gold in AI, I think, are these transformations. From a video, you get statements that people are are making and saying. 2 or 3 years ago, you'd have to hire someone to watch these videos and like write down the statements. Speaker 0: But now I have this thing that transforms these community meeting videos into these statements. This right here is that data. So AI, it's really worth thinking about what that goal is for you before you start on something like this. I, you know, sometimes I do that, sometimes I don't. It's A.I art and not a science. Speaker 0: Sometimes you don't know what you're even looking for, but you know in your gut that there's something there, right? When you write a database like this, this is a simple Lead database any coding agent can help you with. Then you have a permanent place to review these things and start to like really think about them and build off of them. So these are these are transcript chunks, not statements. I I take that back. Speaker 0: But within these chunks, I know that gold exists. So these are things that were set at the meeting, chunked up without any, reason behind their chunking except that they are a certain length A.I 1 comes before the other within the time of of what was said. Right? So so I know number 1, happened before 2. Probably. Speaker 0: I mean, I have a date time in here that is more, accurate about that. But once I have this transcript, I could copy paste this into ChantCPT and say, what were all like the, you know, the statements made that matter for people who live in AI? You all know that City would come back with something interesting. So now that I have this transcript and I have it in the database, I know I can programmatically get that meaning and then I can assign it into categories. So we have a statements table and I don't know, yeah, here we go. Speaker 0: Here are our categories. So if I do select distinct, statement Speaker 2: kind Speaker 0: from this table. These are such important decisions, right? File, if I if I change these these types of statements to, like, joke, insult, like, it changes the way that the LLMs thinks about analyzing those those chunks of transcript. So because this is a civic concern thing, these make sense. Commitment, update, question, request, decision, concern, claim. Speaker 0: Okay. If this was a business setting, they might be different. If City was an entertainment setting, they would be different. It's you know, I I can't say that I I have a process of, like, defining these super important things in a checklist before doing anything else, but but I know I know that when I'm talking with the agent, this is the sort of thing I have to be very careful about. I have to, I have to articulate myself reasonably well A.I usually they get it right. Speaker 0: So this statement kind is very important. Think about it as the way that you're categorizing the data that you're collecting. And let's back up for a 2nd. So we have these transcripts, this is the raw data, okay? Now the video is the raw data but I know that AI can transform it easily. Speaker 0: And so I know when I see a video now that's, you know, if someone's speaking, that also has raw data of text. Now in my mind, I just go there and it could also have raw data of images, of objects. You can take this these videos and instead of concerning yourself with this transcript raw data, you could concern yourself with the frames of the video A.I you can build a system that tells you all of the objects in the frame. You know, why? I I don't know. Speaker 0: I've thought of different reasons you might do lead. Security, business, retail, etcetera. And there are probably like more economical ways of doing City, but regardless that's a crazy thing to be able to do now. So this is the raw Date. If, if this knowledge graph was for something different, this raw data might look different. Speaker 0: It might not be text like this, but images or code or Date signature or audio Event. A.I then the categorizations of them would similarly be different. So if this knowledge graph were about music, these would probably not be statements but be like beats or maybe melodies. And then this statement kind, if you're trying to categorize melodies, you might have it depending on what you're doing as like an emotional response, you know, like file that that sounds, I don't know, pleasant or that's like creating tension that makes you anticipate what's coming next, that give makes you calm, that makes you excited. And then tags. Speaker 0: Yeah. So this is more like assigning metadata and labels to this is an ability or AI in the knowledge graph of assignments to make of these statements. So now this is very important too because this is going to bias everything also. I can tell right now these tags I think I did a limit of 20. Let's let's see, though. Speaker 0: Yeah. There are only 19 of these. So this is very shallow, I think. You know, like, the the reason that I keep this as something that in a CLI agent AI Cloud Code, for example, does for me, and then I don't have it as a programmatic loop yet, is because I still need help adding to this AI. And I want the agent to know that over over its analysis. Speaker 0: And I wanted to notice that, for example, there's no, like, noise here, but people are talking about noise complaints. Complaints. And now there might be something more in-depth that I'm not looking at that these are on top of, but, you know, this looks shallow to me. Maybe fireworks would be added to this at some point. Maybe, fires. Speaker 0: Oh, there is fire. But you get the idea. This this might be too shallow. And so if I were to run this for all these videos all the time A.I I have this shallow, number of tags, then it's going to bias it into these subjects and and others might be missed. And that's what's so beautiful about using the the agents in a manual loop. Speaker 0: Are you adding the tags lead tags or is the system? The system is almost entirely adding these things for me. But if I see that something's missing, actually, let's let's do a little bit of this right now. So so that's this go to a new session. I'm using GPT 5 6 max with Sol. Speaker 0: I think I'll, make this a little faster. So I'm just gonna switch the model. I haven't actually tried Luna. Loops fast and affordable. Yeah. Speaker 0: Let's just see what this thing says. K? This is I really like using the smartest models on the highest thinking, but and this is on extra high Luna. Let's just go to medium. That way, it'll be quicker. Speaker 0: Apply to plan mode override. Okay. No. I shouldn't have clicked that. Okay. Speaker 0: It doesn't matter. Okay. I just copied from here and I'm pasting it. I'm going to say, hey, this looks shallow. And I'm going to tell it what I ran so there's no question with results. Speaker 0: I guess my question is, like, 1, has this been making our statement generations biased towards these tags? And, like, 2026, do we have any, regular habit of updating these as we see gaps slash analyze transcripts chunks. So, you know, this is an approach that is more holistic, that that is going to be sustainable over time. This approach I'm taking where I'm like, hey. Let's think about this. Speaker 0: Do we have any methodology for this at the moment? The other approach might be, hey, let's make this 5 times denser. Look at the transcripts in our database to make sure that it's on point and then we're not just hallucinating A.I then add them. Right? So so it's saying yes, like it is biasing tag generation. Speaker 0: Boom, Folder this up in our instructions. Prefer these canonical style concepts when they City. So even right there, I'm like, okay, if I update these, does the agent even know about them or is it just looking at the hard coded instructions? Right? There's a lot to think about. Speaker 0: However, the table itself is not directly fed in into statement extraction. The flow is transcript to LLMs generates statements and raw tags. Raw tags are preserved. Code maps them A.I canonical tags using the table and aliases A.I then canonical tags are used for filtering AI. So the strongest bias comes from the prompt. Speaker 0: AI the database taxonomy reinforces it during normalization. Yeah. So AI, if we do fix the prompt and it's not limited to these things, once it is ready to hydrate our database, then it might be constrained to these things. So I don't know if I answered your question or AI even asked it. 2026 we're getting to the end of the presentation. Speaker 0: Yeah, so it says AI found no recurring job, scheduled review, dashboard, etcetera. So this is something that I would like, you know, if I was working on this at home, I would be using the strongest mode here. A.I I would think about it for a bit. Like, what I really want here, I AI probably have 2026 review everything 1st. Like, I would say, it has access to my database. Speaker 0: Okay? I would say, I want you to review, like, all of our data so far and determine if this truly is shallow. Are there gaps? Look at the statements. Actually, don't even look at the statements. Speaker 0: Look at the transcript chunks. Compare them to the statements we generated from it A.I their tags. 1, are we missing any tags in those statements? And then 2, are the statements gapped? Are there things that we didn't capture as statements because this prompt has biased the statements into these things? Speaker 0: Which I don't know, it's not really saying that. It's just saying that the concepts that I think are the labels to the statements. Is there anything that can block the filter, like, that thing that's seeking from the sources, is there any codes or something that would, like, don't let anyone pull from this page or something like that to block in? I don't know. To block, say, you from using that data in your situation. Speaker 0: Oh, so, are you asking if there are systems out there in the world that try to stop us from scraping data? Yeah. I think so. Yeah. Yeah. Speaker 0: But at this point, if you can get to it with your browser, you can pretty much have an agent scrape it. There's pretty much nothing that's really gonna stop you. Even, you know, even for example, if you had to, you could have an agent open up the browser, use your credentials, and take screenshots, and then extract that into text and into your data. If and then that's becoming more, feasible now. Computer control, screen recording, screenshots. Speaker 0: Those are going to bypass a lot of things. That's going to be a little bit disruptive and really interesting. At the moment, it's pretty expensive to do that, but it's becoming cheaper. Yeah. So not all data is public, but all of this data is. Speaker 0: So there's really nothing that stops me in this case. I I have a friend, you know, like, he's a data engineer. There are people who ask him to automate the collection of data from their vendors. Their vendors make it very difficult to write scripts to do that because they want that company to keep paying for their front end system. So he figured out a way using some open source software to automate the browser using the person's credentials A.I then analyze all the network requests that are happening, and AI be able to replay those in order to get that data. Speaker 0: And he said that they actively try to stop that from happening, but he's like, yeah, there's tech now that it just does it. Notes it's kind of file very sophisticated on the scale of things, but it's also becoming commoditized. And if you just happen to find it with your agent, AI, your agent will probably be able to use it and do what you did, honestly. AI feel like I buried the lead a little bit with, like, how you might design a knowledge graph. This is a little bit irregular and a bit sloppy in maybe you could say. Speaker 0: Like, you could normalize this in a way that scales to a bunch of other subjects. And what you would really want and need are nodes and edges. Like, So this is kinda all you need to know and think about. There are nodes, a category, and then there are edges. So you see this edge of type of, placed order Details, has address. Speaker 0: The edges can be almost anything that is a relation between Notes, And the nodes can be anything that you care about. So in this case, we have fresh Loops, we have fish, fruit, snacks. So this is about AI foodstuffs. So Speaker 3: Can I also ask, like, about you keep mentioning all the agents you have? How do you, like, just quickly, I mean, how do you organize all your agents? Speaker 0: So there are some parts of software engineering that you should all just lead. And they've been gatekept from you and whatnot. They seem more complicated than they are. 1 of them is called git, g I t. You can talk to your agent about keeping it simple for you, but all of mine are in projects that are tracked in Git. Speaker 0: I use GitHub, and I separate them, you know, by how you would probably separate projects. If it's for a particular company, a particular passion project, the this book generator is its own project. When I say agents, I'm really meaning this, which is codex, which is similar to cloud code, which is running inside of that project. And then, you know, I might mean, AI AI agents that are, more formalized versions of an agent like this that can run-in the cloud, AI. Agents is a very vague AI for all of these things. Speaker 0: For actions. That are happening. Yeah, it's AI, yeah, it's something that is driving LLMs. You know, an agent is something that is driving AI requests and responses And using tools, using skills, using memory, using data A.I databases, maybe collaborating with other agents. And this thing is an agent right here on the screen, but so is the cloud routine that I mentioned. Speaker 0: And so are, AI AI graphs defined in Python running on a Kubernetes cluster in the lead. And and maybe something on your phone too should be considered an agent. If you have to leave, feel free to Is it possible? Yeah. You can unlock it. Speaker 0: You can unlock it with your hand, and then you don't need to worry about locking. I'm sorry if you are worried. It's a little knob. Like, right No alarm. No. Speaker 0: Gotcha. Just wanna make sure Thanks for coming. Awesome. Next time I got it, it's really cool. Yeah. Speaker 0: Take care. Out of curiosity, Speaker 3: was it was the question more about, like, the mechanics of how to set them up or how to define them and organize them or Well, you have a lot of agents that are all talking to each other, and they're all doing their independent tasks all the time. So I guess I'm just gonna be, like, a project manager more. Like, how do you make sure they're all running well together and Yeah. It seems like it would be, like, a map, like, the knowledge graph, you know, or it just be a ton of moving parts. Yeah. Speaker 3: I just in folders, and that would be kind of hard for my like, to stay on top of. Speaker 0: Right. There's just so much to say about that. So there are different techniques. This is a very new practice. The term agent, you know, it it jumped into the culture maybe a year and a half ago or something. Speaker 0: You know, it's a it's a very new concept. So I would say organize them in the same way you would organize a file system. A.I that might be something that you need to learn in the same way that you would learn in like a computer science degree. Okay? So there in in computer science, if you're learning that in university, they're going to at some point talk to you about file systems. Speaker 0: That's gonna be based on UNIX and Linux. I would learn those concepts because the fact is these agents sitting on my terminal, the reason that they're so freaking useful and powerful is because they just use the file system and all of the open source tools that we've accumulated for 30 years that also basically use the file system. Every database, everything that messes with images, code, everything that runs code, all that stuff, AI, it comes down to the UNIX processing, the kernels and stuff. So I would organize your agents in folders with skills and agents dot MD and different documents for yourself. And then I would save them in Git and host them on GitHub so that if your computer disappears, they don't disappear. Speaker 0: But, you know, it's possible depending on what you're trying to do. It makes more sense to have 1 agent. It it just depends on what what you're doing and what's going on. The idea of a bunch of Agents working in tandem is a bit overblown. Just heads up. Speaker 0: Feel free to DM me or anything. You can email me any questions. I would like to, I think I'm going to organize all of this within a a new system called AI tinkerers with a forum in it that, you know, people can use to talk to each author, not just here, but globally. But in the meantime, Details fine. I'm happy to answer questions about stuff like that or point you in the right direction. Speaker 0: Yeah. But the way I organize agents, it depends what you mean by agents A.I typically like in folders that are code bases. Yeah. And then sometimes there are different agent processes within those individual code bases. Organizing agents within a knowledge graph is very interesting and AI people are doing that. Speaker 0: There are there are so many different ways to approach these things. It's yeah. Yeah. Oh. Yeah. Speaker 0: And, but I think that'll be the idea of what they're doing. Speaker 3: Could you also, like, just roll scraper like that? Like, scrape for all the agents that you have saved Speaker 0: on your file system. Yeah. You created all the trash. Yeah. And now you see it. Speaker 0: 100%. Just saying, yeah, you make an agent to organize. So yeah. A 100%. So I kind of did that when I was organizing that book. Speaker 0: I I started, you know, AI up high in the file system, and I said, these are the projects I'm really interested in, but, like, look at all the projects on my computer. What am I doing? What should I talk about? What should I write about? And it it did that really well. Speaker 0: Yeah. I'm just gonna pack up, and we can kinda chat and talk story if you guys want. You can turn on the lights if you want. Does anyone want lights on? I know it's kinda dark up here. Speaker 0: Here. I did that for the glare. Speaker 3: I guess that was a strange weird odd question Speaker 0: that the light state Yeah. Scraping, and they're pulling from the sources. And in video talk about in response Yes. Good catch. So there is a technique in chunking called, like, overlaps or windows. Speaker 0: There's there's a better term for it, but I can't remember. But but it's AI when you end 1 chunk, you want to have some overlap on this side for this chunk. A.I that way you have the context, both of them have the context to understand if the LLMs is not sure what this means. Okay. What is going on over here? Speaker 0: And if it really has to, it can back up into that analysis of the previous 1 to understand what this so if you if you wanted to program that in a way that wasn't dependent on a smart agent AI Lead Code or whatever, you would have to do a lot of engineering. Yeah. I understand. But Cloud Code is smart enough to if you give it that data and it has access to your database to know, like, hey. There's some vagueness here I need to clear up. Speaker 0: I'm gonna go back up in the rows and see what the hell we're talking about. It's smart enough to do that. You have to give it the right instructions, but but Yeah. Yeah. So Cool. Speaker 0: That's a bit adjacent to something I was wondering about if you're showing the database here. You know, you mentioned Yes. I've played with that a City, you know. I Field it Yeah. A couple weeks. Speaker 0: Takes a lot of effort to design the way you embed things in a way that is Yes. I'm gonna take a couple of weeks. More useful than using, traditional AI SQL knowledge graph querying system. Such a big topic. Yeah. Speaker 0: You can kinda do AI a it's own email harness for that. Yes. Exactly. And it's like at that point You can get something to it really. Yeah. Speaker 0: You have you can go really deep. Speaker 3: Right? Speaker 0: I've considered it. I haven't done that for any of these knowledge graphs. Classes. And, there's probably a reason to do it at some point. If you go on speaker, there's lots of noise there. Speaker 0: Oh, yes. Have you? Okay. I wanna know more. Let me join you guys. Speaker 0: I'm gonna turn on the lights so I wake up and got the cold Yes.