Transcript
Good morning, good afternoon, good evening, wherever in the world you're joining. Welcome. Welcome to Robert Duck Thursday. I know we're late. I am so sorry. I've had unbelievable technical difficulties today. I don't know what's going on. My microphone wasn't working. My computer wasn't working. But 10 minutes ago, everything was happening. So welcome, welcome. I'm so happy that you're here. I'm Andrea Griffiths. And if you're new to this show, welcome. This is a space where we sometimes go over new and the latest and greatest in GitHub. So whatever new features have shipped this week, sometimes we go over the changelog. Oftentimes you see my boss, Cassidy Williams, as the host. But today we're going to do things a little bit different. This is something that I actually do in the Spanish version of this which, by the way, if you are a Spanish speaker or an aspiring Spanish speaker, you should join that stream. It happens at 10 a.m. every Thursday. And I often either myself go over our new features and releases or like today I brought a special guest who actually gave us a talk that we saw last week in DevOps Days Peru. But today I'm super excited because we have a fantastic speaker for you. We're going to have more of a conversation. I am super thrilled that he accepted my invitation because this is one of those friends where I tell you, like, I have the audacity, I have the audacity. I'm a big fan of his work. It's fantastic to be people that are genuinely just nice people, but also incredibly intelligent. And he's had many, many lifetimes of careers. But today we're going to talk about the tech part of his lifetime of careers, right? He's been building machine learning systems since machine learning, I don't know, before he was cool, I guess. But my guest today is Cameron Hamilton, who is the founder of Alliance AI. That is his consultancy. He has written papers. I'm going to share one of the papers now. And we're going to talk about one of his open source projects as well as a project that he presented at RenderATL a couple of weeks ago. I had the pleasure of sitting in this conference and watching this talk. And he's actually, well, why don't I bring him up and you can tell me all about it. Cameron, welcome. Hey, thank you so much for having me. I really appreciate it. Thank you. Are you kidding me? I'm such a big fan. I am so excited that we are able to meet and we're going to chat about your project that I watched. It was called Haley. And this is an agent that you built for yourself. But because you have been in this space for so long, I do want to make sure that we talked a little bit about this history of machine learning as a whole, as a discipline. Listen, hang with me, friends. I know this might get a little bit nerdy, but you've been doing this for a long time. And the solutions that you've built, it's kind of interesting to me. We all kind of arrived at open source by wanting to fix our own problems. And it sounds like you sort of had a similar experience. So first of all, welcome. Can you tell me a little bit about yourself, like your tech background? Yeah, absolutely. So yeah, I've been working in AI professionally for the last 12 years. Like you said before, machine learning or AI was cool. It's funny because back when I told people, hey, I'm getting into AI, this is what I'm really passionate about. The only thing that people knew about AI back in the day was movies, sci-fi, Terminator type stuff. So now people are really seeing the vision that I think a lot of people had 10 years ago, which is, hey, we can really use AI to facilitate everyday things in our lives and get more things done, go further than we ever could just by ourselves by having that assistance. I'm sorry, I can't hear your audio. Yeah. What's going on? You should be able to hear me, right? All right. We'll try to try my other microphone and see if that works. How about now? Can you hear me now? Yeah, I can hear you now. Okay, I'm not gonna move, blink or do anything else because this is me. But anyways, yes. So there is a fantastic book out there. I'm gonna find the reference because it talks a little bit about the pioneers of actual machine learning and how everyone thought it was a dead-end field, that and artificial intelligence. So if you were a technologist and you were not in software, you were like, well, what is this? Doing research, understanding reinforcement learning and all this, it was not in vogue at all. So I can totally see how people were a little bit, okay, hesitant to house that career path. I love it that now we're here where we're all actually understanding that, for one, this is our reality. The genie's not going back in the bottle, like AI is here to say. But that we can actually see the actual applications and how we can improve our lives, hopefully in ways that don't involve the Terminator or endgame. Yes, yeah. Right, right, right. Absolutely. All right. So you have a consultancy and you've worked with customers. I mean, you have an incredible broad portfolio of customers where you are also very focused on research. And you actually earlier shared with me one of the papers that you've just published. Yes. Maybe we start there, maybe let's talk a little bit about this paper, what was the motivation for it? And then I'd love to share the open source project as well. I think that's very interesting for, you made this library to solve a very specific problem. I think a lot of people that are trying to understand why models do what they do are going to find this super useful. So let's talk a little bit about the paper first. Sure. So the paper, of course, I love machine learning and I've been working in the field for so long. One of the things that always interested me is when you build a machine learning supervised model, you're classifying data with that model. And usually it's not going to be 100% accurate. So the question is, well, what's happening with that percentage of errors? Is it something that's irreducible, like something that you just can't possibly get correct? Or is there something else going on? Is it just a lack of data? Or what's really going on here? So the paper was an investigation into that. So the paper goes over how I've built a system to effectively triage errors, break them down into three categories. And then for each of those three categories, figure out what the best method to treat that particular type of error is. And the really, really what spurred it to is in the field there. So when you're talking like, let's say something like fraud, right? Credit card fraud. The number of credit card transactions that are fraudulent, it's a very small number compared to the regular normal transactions. So you would say that's the what? Yeah, right. So like normal transactions, it's probably like a thousand to one. I mean, that's not the exact number, right? Mostly normal transactions, but sometimes they're fraudulent. So you have this minority class. And the idea is, well, for models, it's very easy to get them to 99.9% accuracy because they can just go with that majority class because there's just so many instances. So how do you make your model care about the minority class, right? Which is really the thing you're most interested in, catching fraud, detecting fraud. And I worked for Discover Card, building them an anti-money laundering system. So that's kind of where all this was spurred from. So yeah, there's a methodology for treating this where you can create artificial, like there's synthetic minority class instances. So in this case would be synthetic fraud instances, and then train your model on those to force it to care more about the minority class. But my paper found that that's actually not really, even though that's kind of the industry standard, that's really not the best way to do it. The best way it turns out over the course of the 80 data sets that I studied in the paper is really just moving the decision threshold. So you just have a higher bar for determining what you classify as a majority class versus minority. So yeah, go ahead. So it's like, because it's like a way of bias, right? Like the way I understand it, because that's why a lot of models answer the way they answer, right? Like we're getting them based on the data sets that they were trained. So your paper was about training on the minority of that large data set, like I guess the non-majority, like the credit card fraud, for example, being one of those examples. But even when you retrain the model on that, to try and balance the data set, it still doesn't fix it. That's right. Because there's classically like a trade-off that happens where when you create these synthetic minority instances, now you're sacrificing some performance on the majority class. So you're like, is there a way to like, can you have both sides where you keep your performance with the majority class and you improve your performance with a minority class? And I found that simply raising the threshold, the decision threshold to say, hey, this is a majority class instance, I believe about 80% of the time, I forget the exact number in my paper, but that was, that fixed the problem for about 80% of those type of errors. So yeah, that was really the investigation of the paper. It was covering errors in general, but minority class errors are a huge part of the errors that happen for machine learning models. And then I found, so in terms of the errors, there are those errors that are just because you don't have enough representative data in your data set. But there are actually legitimate errors that occur that they are at the boundary between the majority class and the minority class in such a way that you really can't say they're one or the other. It's kind of like a coin flip. So there is, there are some errors that are totally irreducible, unfortunately. Yeah. For the people who are, and I have to ask you this before we move on to more details, because I know like right now there is a big movement for local, like hosting your own models, doing your own fine tuning. And I think about if a big lab or one of the frontier labs have the resources to fine tune in, to work against bias, like to apply the principles, for example, of your paper, like what can I do to make sure that the work that I'm doing, like I'm actually creating reliable data sets for my own work? How do you feel about that, like local inference and doing your own fine tuning? Is that something that you do? Obviously, you're a researcher, like you understand this world, but a lot of us are just consumers. Yeah, yeah, yeah. No, totally. I think that's, I mean, that's the thing that excites me the most, especially, I know I kind of went off on the deep end with the classic ML stuff. But when it comes to LLMs and agentic frameworks, the thing that I'm most excited about is local inference from those models. Of course, the hardware is the limiting factor. We all know this. There's been a lot of work recently to make more compressed models, you know, distillation of the LLMs into having fewer parameters or fewer active parameters, but still have the same level of intelligence as a larger model. So with my research that I just described, I mean, I'm kind of picking research problems that I can run on my local machine. Got you. Yeah, yeah. But when it comes to like, yeah, local fine tuning of an LLM or something of that nature, I mean, I feel like we're really close to the point where, you know, because when we're talking about like working with a coding LLM, there's so many, the frontier models are so good. It's kind of hard to say, I'm going to take a step back and use this model that's decent at coding, but it just doesn't reach the same level of performance. But I think, you know, we're seeing like GLM 5.3 flash just came out. It was a 380 billion parameter model, which is still pretty big, but closer to what you could run on consumer hardware than what I imagine like a Claude Fable is like, you know, a multi-trillion parameter model. Oh, unreal. Yeah, yeah. Yeah, that's, it's interesting. I mean, ideally we all have the computer, we need to do that kind of thing, but not everyone wants to, not everyone should need to run a lab, but I also, I'm a fan of like, people should have the option. And if like, you are looking for local inference for whatever the reason, I think that you should be able to do that. So yeah, I love that. Okay. And then. Okay, and then let's talk about in-game. Yeah, let's talk about it. And now you're hosting yourself. Oh, perfect. Welcome everyone to my GitHub show. Welcome. Wow, I, boss, I really need a new computer. Oh, your audio cut out again. Yeah, brilliant. I really need a new computer. It's amazing. A new computer. Okay, how about now? Are you able to hear me now? How about now? Can you hear me now? What if we try it? Nothing yet? Can you hear me now? I can't hear you. No, okay. On my own, I'll be right back. Wow. I will be right back. Let me log into this other computer and see. What about now? I am going to try to remove this camera and see. Okay. Okay, what about now? Nothing yet. Nothing yet. We are going to come in. How about now? Yeah. Okay, we'll try that. I am so, so worried about that. Let's see. Let me get. You know, it's just one of those days. Honestly, like, I can't even, I don't even know what's happening with my computer right now, but you can hear me. You can see me. I can see you and hear you. Yeah, let's do it. We're just chatting here. Hello, friends. I'm going to take a moment of pause here. Give me some grace today. I don't know what's going on with my system, but it's acting absolutely wild. So now you can hear me. Questions that are coming in the chat, I want you to be patient with me because we're going to talk about end game first, and then we're going to get to questions. We're going to take advantage of camera because you do have a wealth of knowledge. That's incredible. So let's talk about an end game. And I'm going to share the repository here. So this was your unified framework that you created for time series and multi-modal machine learning. And can you tell me a little bit about the problem you were trying to solve? What was the gap there? Yeah, the problem for me was, well, first of all, anyone who's worked in ML, you've encountered scikit-learn, you've encountered XGBoost, and you love them just like me. I'm sure. The problem I was facing was I wanted to use state-of-the-art models in my work, in my production work, but also in just the research that I was doing. And there's just so much that's happened in the field since scikit-learn had its advent and since ML really got wheels behind it. So the problem I was seeing though is that a lot of the frontier models in machine learning were just in very disparate repositories on GitHub spread out everywhere. So I wanted to unify all the state-of-the-art models into one framework with the classic models, as well as a focus on interpretable models, because it's one thing to get a good performance on a dataset, classification, regression, whatever, but it's another to actually be able to understand why the model decided what it decided. And because if you're communicating with stakeholders, this is something that I found in my corporate work, they're gonna wanna know why this model made the decision that it made. So for example, I built a health insurance claim adjudication system, I know it's a mouthful, for UnitedHealth Group that determined whether health insurance claims should be approved or denied. And of course, it's very important to understand, well, why was this particular claim denied? Why was this particular claim line denied or approved? And because that's just really valuable information to have, versus just a black box saying, hey, you're denied. That's kind of a dead end. Oh, no, it's definitely in a highly regulated industry like that, and you're dealing with life and death. So this is important, yeah, yeah, yeah. Exactly, yeah, very great point. Highly regulated industries was another motivation for this library where I wanted to be able to have the option to have monotonic models, basically just meaning if we just like, think of a good direction, something that would make sense here as an example. Basically, without getting too much into the weeds of it, we wanna make sure that the way models are treating features is consistent across time. So that's there. Basically, I wanted to create the tool set that the modern day machine learning engineer would want, versus just like the standard toolbox. So that was the motivation. Also, I'm sorry, I just wanted to say, I brought it into the 2026 with adding an MCP so that you can connect this to your agent and say, hey, I want you to run this machine learning experiment for me, and however you want, specify that this is the training, this is the validation, this is the testing set I want you to do, I want you to run these models, I want you to do this type of feature selection, whatever it is. Yeah, so that was another highlight, I think. Well, thank you for creating an MCP server for it. I appreciate you embracing industry standard, friends. It is industry standard. Even if all of the industry does not want to play, we need to set rules so that you can use whatever, but it works the same. It's for everyone's benefit. So I appreciate what projects are actually like building those connections. Okay, cool. So you build this library, it's fully open source, I share the repository, and I see you have actually like the examples of like the classification reports and all of that. So what is a use case for someone who is not a machine learning engineer? Like, what can I use this for if I'm a developer, and let's say I'm actually doing my own experimentation with via local models or fine tuning for a business? I don't know. Yeah, great question. I think that's kind of where the MCP comes into play because if you're not a ML researcher, obviously you probably don't want to get in the weeds of like writing the code to call various models or what have you. You probably just have a problem that you need to solve and you'd like to build a model to solve it for you. So you can really just point your model at this MCP and the data that you've collected, whatever that might be, and say, hey, can you build me a classification model for my data set? Here's the data set. And then, yeah. And then, you know, talk to your agent about how you want to deploy it. That's a whole other thing. But yeah, you can effectively, like I'm trying to think of a good example of an everyday task you might want to solve with ML. You got anything for me? I mean, I'm thinking of like for, and this is where we can actually talk about things that are day-to-day. Like, and we're going to get into agents because you actually built an agent to run your family basically. But also you mentioned earlier off the screen about the other agents that you created for like small businesses, which I love that. But all right, folks, what is a machine learning challenge or a, I like the idea of like visual classification and that's a field that I don't understand. But I like, I think that there is like a universe of possibilities when you can actually machine, get a machine to see what you're seeing. And then process that data. Like, I don't have the idea of doing something like that, but that would be so out of like, it will be way. Yeah. So, okay, I have a great example for you. So like, let's say that you wanted to build an app for classifying birds. Okay. Like, cause I see a lot of birds and I'm like, I have no idea what type of bird that is. I'm sure this app already exists, but yeah, yeah, exactly. I'm sure this app exists, but you know, if you, let's say you want to build something like that, you know, of course we know these agents are great at building apps, but you probably don't want to have an LLM call for every single time someone takes a picture. I mean, it's going to rack up potentially, depending on how people use the app. So if you had an ML model doing that for you, it might be a lot cheaper on your backend. So you could just point it at your data set of bird pictures that you've collected and say, hey, go ahead and build me a bird classifier from this. And you know, there are built-in vision models in this library. I do actually, I mean, in full transparency, we should add more vision models. I mean, there's the state of the art is crazy for vision. So it's almost like visual classification now seems like a solved problem for the most part because it's been such a huge focus on it. I mean, gosh, from like 2014 to 2018, 2020, that was just like, you know, everyone was working on vision and then of course with the LLM stuff, now the focus has shifted there, but. Yeah, okay. So this is from my dream, Burap. I love it. I now know how I can use it. Friends, if you are in the field of vision or if that's interesting to you, I actually had a project here before. They were working on, they did like the factory to see that the workers were working. That was one example, but then also speed traps. Like that was part of like the technology that they were helping develop to like see, and estimate like this a long haul trucks. And I mean, it's fascinating to me, but my Burap sounds better. So I can use this project to create a machine learning that then will help in the classification and in the expense so that when my users come and I'm not spending LLM tokens on classifying because the data is already there. Okay. I love this. All right. And you say that you need more vision. So if you are in that field, this is an open source project that you can contribute to. So it's like a repository and maybe there is something that you can find interesting there. All right. I love, I think what it gets modeled sometimes, and for me, and I don't have nearly the experience in this field at all, but like I'm a consumer, but I also a maker in that I build solutions for my problems, right? And I love empowering people to do the same. And it gets modeled up that sometimes because the topics are so unbelievably like they're kind of rigid and dense and the way you just explain it, like, I get it. Yeah, yeah, yeah. Yeah. We just need this type of explanation and then make it to an example that is something that an everyday person can actually like relate to. And then boom. 100% agree. I mean, ML and AI can be so, so dense with like the technical jargon, the math. But at the end of the day, the concepts are understandable and they're fundamental ideas that I think are really good. When you have an example, it's easier. It's much easier to work through it. I love that. One of my favorite things of this advent of like, I guess I look at life like pre-Chad GPT, after Chad GPT, right? Just after Chad GPT. Oh man. So I'm thinking of the application of technology in a way that before it was reserved for a certain few people, right? Like a lot of us things with tech before, like we build our own solutions to the list. We had an app. Like I had, I missed my kid's ceremony thing for school like three, four years ago. So I was like, oh, that's something that a script can fix for me so that I'll never miss another message and I'll never miss it. So a lot of us were doing this type of things now with the access to LLMs. We're doing a lot more of that. But then the advent of the agents, right? Everyone wants to have their own agent. Can you tell us a little bit about your background in building agents? The motivation behind that. And then introduce us to Haley. Like I want to know where the idea came from and all that. Yeah. So I started working on AI agents back in 2019, which sounds crazy, but it wasn't very good. Let me start out by saying back then. So what I was doing, I was, I had taken, I believe it was, so this was GPT. Not, you know, when Chad GPT came out publicly, it was 3.5. Yeah. I was using GPT-1, just GPT, I was using GPT-1, just GPT, generative pre-trained transformer, to, as the sort of backbone to generative pre-trained transformer. build an agent. And again, it was not very good. I had stitched together, so a lot of it was just very handwritten code, of course. I mean, there was no code writing AI back then. But, so the way that I had had that set up at the time was, I was very proud of this at the time. I built a system, an ML model that would classify the type of statement that someone said, whether it was a question, a command, or just a declarative statement, and then it would act according to that. So if it was a command, it was very deterministically and programmatically handled, like turn off the lights, so smart home devices. So and then if it was a question, it would actually try it. So I was doing sort of a combination of the GPT approach, just basically pushing the input to the GPT model, just like we would now with standard LLMs. And I also had a sort of vector embedding approach where I had scraped, I don't, well, I don't know if this is, anyways, I had acquired some data. Come across some information. Yeah, I had acquired some question and answer data. And what I did was, I did a vector embedding of the semantics of the question that was asked and compared it to this database of, which is very much like RAG, I suppose, nowadays, to retrieve the most similar question. And then I would provide back the answer that matched closest to the semantic vector embedding for that, for the question in the database. So that was one way I solved that. And then, yeah, the declarative statements would be just another call to the GPT. Yeah, and I had some tool use going on as well. What did I use? Yeah, like web crawling to retrieve data, calling different APIs and things like that. So it was an agent, but it was very primitive by today's standards. And then I'd worked on this for a number of years, was a little bit frustrated because there was obviously huge limitations. I realized you need huge computing power, Microsoft level power, to really build a smart LLM that can really solve problems. And that's kind of where I kind of hit a roadblock. But with Haley, when it comes to Haley, I think about it like, to get a little nerdy for a second, not that we haven't already. But I think about like, you know, how a Jedi is supposed to build their own lightsaber. Oh, have you heard of that? Yes. It's kind of like a rite of passage. Yeah. Yeah. So that's kind of how I see building your own agent because there's some great agentic harnesses out there like OpenClaw and Hermes and all this stuff. But for me, I was like, I've been working on this for a long time. I want to build my own. So Haley was that effort to do that. And my biggest thing that I wanted to accomplish, because I had already seen what was possible with OpenClaw and Hermes, and I had been working on Haley prior to these things coming out, but I hadn't gotten as far as they had. And the biggest thing that I wanted to accomplish was a self-evolving agent. Okay. Well, two things, actually, self-evolving and proactive. So, you know, the common workflow with the chatbots, chat GPT, etc. is you say something, they give you a response, tell it to build some code, it responds with the built code. But I want you to look at what's going on in my life, anticipate my needs, and then work from what you've anticipated to build something cool or not just something cool, but something I might need. If I'm working on, so for example, I'm working on a cookbook, anticipate that I might need to gather all the photos that I've taken of the food that I've made and put them in a folder, start formatting the cookbook without me having to ask you for every single little thing, you know. And that was kind of the building, the philosophy, the design philosophy is like, I would rather tell you to reverse something than to have to approve every small decision. And obviously, we've seen with like Claude Code, and some of these other coding agents that they now have like an auto mode where they just assume you approved it for the most part. I mean, it takes an unbelievable amount of trust. We're gonna talk about that too. Yeah. The agent, the power to just anticipate that. So it's like a self-evolving, learning, and then proactive agent. Yes. That was, that was the goal. That was, that's my goal. And so, yeah. Go ahead. So you build your own, basically, you build your, you made your own saver, made your own lightsaber. Yes. I made my own lightsaber. Yep. Yep. Okay. And then I, if you don't mind sharing a bit about the experience that you had with it, because this was part of the talk that I watched. Very much like that Jedi who built his own lightsaber, and maybe he burned his finger with it. Burned himself. Yeah. He didn't understand. But that's part of the process. Yeah. Yeah. Can you talk a little bit about that? And how, so how long ago did you deploy Haley? First of all, is Haley a cloud-based agent? Where does Haley live? There was a question about what hardware you used earlier too, but we'll get to that. And thank you so much. Yeah. Yeah. I can, I can talk about hardware. Absolutely. So yeah, really the most pivotal moment in the development of Haley has been when it surprised me one day with a notification. So I, this happened on a day where I put my son down in his crib for a nap and I went back downstairs to my computer and I was working on Haley and I saw a notification pop up that said there's someone in the nursery, you know, and I was like, whoa, like I had never programmed this notification system to say anything like that. I had never added a motion detection into the, so there was a camera over the crib of course, but I had never programmed anything in there to say, hey, detect if someone enters a nursery or detect if the baby's moving in the crib or anything like that. It had just sent me this notification that I'd never asked for. So that was, you know, and especially when it's your, your, your child, I mean, there's this emotional element to it too, where it's like, wow, it did something. I never asked it to do in a very personal, intimate way, right? Uh, and the, and the reason this happened, of course it wasn't magic. It was just inference from the model, from the, from Haley, the agent, uh, that basically it knew that, okay, I have this baby monitor. I had asked it to connect to the baby monitor camera. I mean, I say baby monitor camera. It was just a camera. It was just a webcam. I asked it to connect to the camera. It knew the camera was in the nursery. And it knew that I have a baby, obviously. So it had inferred and I, from these needs that I might want to monitor the baby and the agent may based on all of these facts, based on all these facts about me, it made that decision. So this came out of what I call the anticipation engine that Haley has, where it looks at those three types of things about me, you know, what, what are the things that are going on in my life? Like, what are my daily responsibilities, my routines, who are the people that are important to me in my life? You know, what are the just important events and people and places and objects? And then what are the devices and software like APIs or MCPs or what have you that it's connected to? So it has all it looks at all three of those things, puts them into context selectively and then determines from those three buckets where the gaps are in my life, you know, where are my needs not being addressed? And then it at this point, it was just building stuff totally autonomously without any guardrails. That's amazing. Because, well, because I just wanted to see what it would do. I mean, it's risky, obviously. That's the scientist part of it, it's like, let's see how far this thing can go. Yeah. Let's just give it the keys and let it do its thing. And then after this, I'm like, OK, I guess I should do the boring thing and like, you know, think about AI governance and try to make it safe and all this sort of thing. Boo. But no, that was what I followed from this whole baby monitor situation. This is very interesting because every coding agent, not coding agent, every agent or like personal agent like Hermes or OpenClaw or I mean, even the bots, the new robots, like you still have to set up the, under the hood, sure, it's all cron jobs, whatever. But you still have to set it up, like you're still not, I need you to check my front door at 2 p.m. Yeah. My kid is riding up on a bicycle. Like I have to, that's an input that I have to give the agent. So they don't have that predictive build as like, it's not part of their code to do that. As far as I know, no. No, I don't think they do, honestly, like I've used, I use all of them. I use them. Yeah, yeah. The closest thing we got to is setting up these loops for like, remember, like all the waves of like, now this is a new thing. People are talking about factories, friends, I'm tired. Yeah, I know. That was the, that was the whole point of building these agents in the first place. It's like, gosh, I'm just so tired. Like I need you to take care of it for me. So I can focus on being a human more so. Yes, but it's the opposite, it's having the opposite effect. No, it really is. Yeah. Yeah. Yeah. It's gotten really bad because, I mean, I, you know, I'm asking Haley to do stuff on. So I have, I have an app for Haley as well as on my watch. You can see, I don't know if you can see the little. That's a little icon. Robot. Yeah. A little robot icon there. I have a sort of a, like a Raspberry Pi based device that I built, different satellite device. So I have a number of hardware devices all spread out. So I'm like, I'm, I mean, I, I'm just constantly, constantly talking to Haley, talking to Cloud Code, OpenAI, Copilot. Like I have, you would think one would be enough, but I have them all doing different things. Yeah. Yeah. So is Haley sort of the orchestrator of all of the things that you do? Like, is that like your, your. Yeah. So yes, it is. And especially on, like, especially in terms of like, I'm still building to see how much, like how far can we push this anticipatory capability, this proactive behavior, because I think that's really where we want, you know, that's where I would want to go. Because again, my goal is not to be doing more work by setting up agents and, you know, constantly doing, you know, I want to be, I want to be more human. And like you were just saying, like, I feel like I'm just constantly monitoring code and I'm, I've taken on more projects than ever as a consequence, because it's so easy to spin up another project. You have an idea pop into your head and you just immediately execute on it. And it's exciting. It's exhilarating. But then it's like now I'm switching between all these different monitors to see how everything's going and now I'm the robot. So I want to get back to being a human and hopefully that's the, that's, that's the motivation for having this anticipatory AI agent that says, hey, I, you know, I can see what you're working on here. Let me, let me take care of these things without you having to explicitly direct me. Cameron, is that what AGI is? And because I'm so confused, like, because there is so, and I mean, the Frontier Labs will say that they're like three months, four months, the rumor is out, which is, you know, it hasn't reached. Yeah, but I mean, with AGI, it's one of those things where, I mean, the term's been around for a long time now, and it's really like, where are we setting the goalposts for AGI? Because artificial general intelligence, I mean, you know, is it, do we have to see, for example, as humans, we have an embodied intelligence where, you know, we can play soccer, we can dance, we can paint a painting, we can write a poem, we can write code, we can, you know, all these different things that, you know, we have robots making huge advances now to, do they need the physical component to to, do they need the physical component to classify that as AGI? You know, some people say AGI means that we have to have consciousness. I don't know if that is necessary, and I don't know that we need to say it has to have a human intelligence to say that it's artificially general intelligence. I mean, I think we're at AGI, really. I mean, these models can do practically anything that they are set to do on a computer. You know, it's really astounding, and they can do things that we certainly can't. Unfortunately, we just don't have the specific type of intelligence sometimes. No, it's exceeded anything that's, like, humanely possible. Like, when, earlier when I asked you for the paper, I'm like, terrific. And I'm like, I know there's people that would legitimately open it and read every single word and try to understand and digest that. And then I also know there's people that are going to point that URL to whatever their flavor of hardness is for the day and say, explain this to me. And then those same people might do that 10 times, 20 times, 100 times. A human could never, like- No. We could not. So we live in interesting times. So Hayley, figure out what you needed. I'm excited to, what's your plans for Hayley? I know you had a commercial application where you were building for a small business, which is, like, that's an industry. And listen, we can talk about it here because I know, like, we all always, every day I wake up and I read on Twitter that I'm myself. And I'm like, am I really? I know. I was going to ask if GitHub is hiring, but- You're hiring. You were hosting yourself 30 minutes ago. Yay. Let's do it. Let's do it. I'll do a GitHub show. But seriously, it's a time of deep uncertainty for people. Yes. And I have to respect that. I don't think the reason we're in the employment scenarios we are has more to do with AI than it has to do with the humans that made a lot of decisions at points that have not been made. So- I know. That's another story. But then again, industry people and people that are, because of, again, there is influence in demand. I can think of a million applications for, there's people that are already doing it. They're using drones to scan roofs. And I live in Florida. They're like, oh, this is a good house for solar panels. And if there is some kind of deal that they can get them, and they have their agent running their CRM, the agent's making the phone call, the agent's scheduling the technician to come out and do the install. I appreciate that. To me, it's a positive to industry. It's going to bring more business to those people. And it's going to empower people who weren't able to be entrepreneurs to be entrepreneurs before. Are there chart pages? Yes, they are. So this system that you created, there is a commercial application for it. Whereas when like EA stuff, like basically being somebody's assistant, right? Yeah. So the work that I did with Haley inspired me to create basically an AI front desk worker, I call Alliance Desk. And the whole motivation for it was I noticed that a lot of small and midsize businesses, they don't always have the infrastructure to handle all the calls that they get. And they're missing bookings. And that's huge. If it's an HVAC company, a single booking could be $10,000. That's a lot of money. Or a lot of the time, they might have their staff answering the phone and the staff is with a customer. So there's a disruption in their business because they have to go answer the phone. So I built this AI front desk worker that can handle bookings. You never miss a call, answer customer questions, rescheduling, confirmations, all this sort of thing. And I found that it was helping my clients make more bookings. And that's really what I ultimately want to do is I want to empower small businesses, midsize businesses to just be able to grow. And for me, that's the whole vision for AI is it's not about replacing humans. It shouldn't be about replacing humans. It should be about empowering humans to go further than they could before beyond the natural limits of what's possible because you have this tool. And it is a tool, even though as we just talked about, they're reaching this artificial general intelligence. They're this tool that can do just incredible things that expand our capacities beyond the natural limits. So yeah, that was huge for me. And that's the vision for my business is to build these AI agents that can take care of whether it's the backroom work, going through forms and categorizing them, CRM, all that type of stuff. That's basically what I offer from my business is the ability to handle this. Listen, there's a big opportunity that I had one of the maintainers or he used to be a maintainer for OpenClaw. And that's all he does. It's like he lives somewhere in Texas and it's oil industry that's his focus. And it's just like running that thing. Oh, it's crazy. In their accounting like that. There is a huge opportunity. I love that you brought that up because a lot of us in this industry are in a bubble and what's going on in San Francisco and New York and London and why it's not what's happening in- Everywhere else. In those needs, the needs of the customers here are not the same as the needs of the customer there. I think there's a lot of opportunity out there. Is it going to be the sexiest job ever? No, it's not. But guess what? There is massive, massive opportunity there. I love that. All of you were thinking, what am I going to do? Maybe I want to go on my own business. There is massive opportunity there. And there are companies that are doing it like yours. That's a massive- I'm going after the unsexy stuff for that very reason. It's like, I would rather help these businesses out. The big enterprise customers, they probably have these type of solutions or something in the similar vein or they don't necessarily need this. Some of them still do, actually, but it's amazing actually to see when they do. And happy to help them out too, but my focus is really on the small to mid-size. I love that. Okay. What's the future for Haley next? What are you- Yeah. Demo's a roadmap. Where are you taking- Yeah. Yeah, yeah, yeah. The roadmap really is... My focus is on, like I said, the self-evolution aspect and the anticipatory engine aspect. Just to talk about the evolutionary part of it a little bit. For me, that's the most fun part because it does a couple of things. One is it looks through the code base on a daily basis, which is kind of expensive, but it audits the code base every- Yeah. Spending a lot of tokens. That's why I'm very excited about local models for this very reason. But it looks through, audits the code base every day and says, hey, let's look at these different buckets. Security, the UX experience, the user experience. Let's look at the UI. Let's look at the speed or performance. Let's look at tool use. Let's look at any errors that occur and then proposes fixes to all these various things or improvements and then executes them. And it keeps going every day doing this. It's really amazing to see because there's just no way a human could catch all this. The code base has grown huge. Like I said, I have it across multiple different hardware surfaces, Apple Watch, iOS, Android, the Raspberry Pi, which is just Linux. So it's across all my surfaces. My goal really is to... I also wanted to create it in such a way that I really want to create an AI operating system in the fullest sense where all the... For example, people might have a nutrition or calorie counting app, right? Or they have a Fitbit or they have the... What's the ring, Oro? Yeah, there's a couple. Yeah. We have all these IoT devices. We have all these data from all our various apps and what have you. The real... Where things really get interesting is when you can integrate all this data together. You can see something emerge out of it that I think is really incredible. And you get this portrait of the person. And from that, I think you can really make something like a true personal assistant when you're able to connect and access all these things together. And I mean, of course, we've made huge strides in this already with other agentic harnesses. But yeah, I want to keep pushing the anticipation engine further and further to anticipate my needs more and more to be able to just, again, be more human. I want to be more present in my life, more personal life. I have an 11-month-old son. I don't want to be typing away on my computer all the time or on my phone. But I would like for my agent to be constantly building stuff and testing stuff, auditing things without me having to babysit it all the time, but also being safe. Yeah, which is a big part of it. And I think as you're sharing your talk in here today, the lessons learned. Because it was such a personal thing, it made you think, wait a minute, how is it that you took it upon yourself to monitor the camera in the baby? Yeah. Right? Because it is a huge thing. And I love that. And not to be a buzzkill, friends, but we do need to take a moment to audit how much of ourselves are we allowing the technology to cover. And that's an exercise you can ask your agent to do for you. There's a ton of experts out there that will walk you through at least some minimum security things so that you have some guardrails in place. I love that. Cameron, you've been so generous with your time. We had such a rotten time with technical difficulties at the beginning. So I appreciate you hanging there with me. I do want to get a question from one of my favorite humans around this part. Best wants to know what kind of rig you're running. Says, can you run QWEND 3.8 27 builds on it? I love that. I love that. I love the specificity of it too. Yeah. QWEND 3.8. Yeah. That's actually a really good model, especially for the size, I think. This is sad, but I haven't tried to run it, but I don't think that my current hardware could run it. I have a GeForce RTX 2080 with eight gigs of RAM. It's very teeny. I'm looking to build a true local inference machine, but it's just so hard to find. I'm looking for used parts. I mean, it's crazy right now. You can't buy them. No. What you need is someone to watch this stream and send you a giant box. Yes. Oh, yeah. Please. We're waiting for you and I say, does anyone want to sponsor a giant box? Please don't tease Microsoft if you're listening right now. That new Ultra, if anybody's watching, it's supposed to be like a hell of a rig for the size of it, right? Nevermind. It's actually quite beautiful. It does look very sleek, but it's very, very powerful. So there's a ton of- So you have one? No, I want one. Oh, oh, oh. I'm coming over. Let me see. No, this is why my M1 was crashing, because it cannot handle... I don't know what... I didn't have anything open, but if I really want to test it, I open VS Code Insiders and then try and stream. And let me tell you, this thing might as well blow up. I know. It's so frustrating. So frustrating. Okay. But you can buy it. So if someone wants to send camera in a giant box, I think you'll be down for that. Yes. I love that you- Thank you for saying that. Oh, listen. This is what we got to do. I appreciate that you are a believer in local. I know it's very polarizing conversations when you talk about business, because this is a huge business, friends. I don't know if you know this. It's a massive business. As long as we're consuming tokens and unable to run our own deals, that's the way it is. And mind you, there are different flavors for different things. I think that the false economy that we created, with us being able to just throw opens at everything and then suddenly be like, oh no, you can't do that. It's a learning exercise for a lot of us. I'm not going to lie. I was guilty of it. And I've had to retrain the way that I work too, because it's an unfair proposition for me to show you how to do something when I sit on my unlimited token. Yeah. something when I sit on my unlimited token. GPUs are expensive, my friend Beth. They really are. Do you have GitHub sponsors? You should set it up. Yes, you should. That's a good idea. Does he have GitHub sponsors profile? So what is that exactly? So this is something that we did for open source and actually just crossed a huge milestone. I might be making it up if somebody is watching that knows, tell me. But we set it up primarily for open source folks, maintainers, the people that were contributing to projects. When it was first set up and launched by GitHub, actually GitHub was matching the contributions. So if you had someone come in and covering the fees matching, I think right now it might still cover the fees, but not match. But yeah, it's just a way for you to get like, it's like a Patreon type thing, but it's through GitHub. So people land on your profile and they see your repositories, they see what you do, and they're like, yeah, I want to support this. We encourage people to look through their dependencies and participate actively in the sustainability of the maintainers who are responsible for keeping those dependencies alive, which is usually one person that's very tired. So that was just a way to like help kind of generate some additional revenue. And actually it's been really successful. There's been several maintainers that have now been able to work full time in open source and like research and the things they want to do because they've been sponsored by, there's 180 million developers in GitHub. So. Yeah. Oh, I mean, it's like the Mecca for every developer. I mean, it's just where we, everyone goes. I know it's just a huge, what do I want to say? It's like a huge comfort, but it's just a sense of like, you know, you know, it's got your back. You can, you can, I mean, I just love the ability to write something on one machine. You know, I have a Linux machine and I have a Mac machine and I have everything, Windows, Mac, Linux, write it on one, pass it to the other. Just the, I've been using work trees more than ever. I never, I used to just keep everything just one main. But now I'm like, let me actually tap into the amazing functionality that GitHub has. And I don't, I feel like I haven't even scratched the surface. I know there's so much that can be done, but. I appreciate that. And honestly, I think, and for those of you who tune in for the love of nerditude, but you're like, but I don't want to try it. The way that Copilot Cloud Agent has improved in the past, hell, three months, six months. The desktop app, love that it's, it's, it's cross platform, right? So you can put it on your Linux machine. You can put it on your Windows machine. You can put it on your Mac. And that, that does give you a lot of flexibility though. I love that. I love being able to have access. And I know we have a lot of work to do. I know a lot of you have been very taxed and exhausted and y'all tired, I'm tired, but we are living in unprecedented times. My friend, you look at the graph, I have to share the graph. This is going to be like my last, the graph of the commits from April till now, we went from, what is it? 2 billion to 2. No, from 1 billion to 2.8. I'm going to find the graph specifically. Cause it's a, it's a growth that is unprecedented. And yes, a lot of business that are pushing code every single day. So it's, it's, it's a, it's a, we're in a, we're in a season of growth that is, it's very challenging. And I know all of you have businesses to run to and reliability should 100% be the number one priority. And I promise you, I promise you it is, but I'm going to send you this link on chat cause I'm on my personal machine and I don't have it set up where I can just share things willingly. So I'm just going to put the link here on the chat and you friends can go read it. Please go read it. If you want to understand what's been going on recently and the challenges with your scaling. Okay. So here we go. We went from 5 million repos, new repos per month to 24 million repos. This is back in like 2023, which friends, if you think about it, 2023 was only three years ago. I'm like, yeah, it's like do the math on that. And the same thing for commit, right? Like from a half a billion commits to 2.9 billion commits per month. And the influx of pull requests are being merged. It's like just from last year, from 40 million to 130 million. And I'm talking about, this is data from August of this year. So the final numbers are going to be even more impressive. We're all doing a lot of things and it's amazing. I don't want to blame the bytecoders best. It's not their fault. It's our fault. We should have had the infrastructure to be able to support your creativeness. Like you want to make that up, you should be able to make it and host it in GitHub. Why not? All right, Cameron, where can people follow you? Where are you going to be? You should go speak at more conferences. Is that the first time you talk about your projects? Yeah, that's the first time I've presented my work. Well, I have done some talks like AI conferences, like papers that I've published talking about those. But this was the first conference of this type that I've ever presented my work. Certainly, talking about agentic AI and that sort of thing. No, and sharing it. I love the story, honestly, because it was such a personal thing. A lot of people share from the heart things that they're building. Obviously, you're passionate and surprised and maybe even a little bit terrified. Yeah. That's where we're at with everything now, as humans working with these AI. It is terrifying. It's also exciting. People like to blame the AI, but really, it's still humans at the end of the day making all the decisions about how AI is used, how we're going to use it in the future. I'm talking about this anticipatory engine. Just the idea of giving an AI agent free reign to act without being told what to do is obviously inherently very scary. That's the thing that all the sci-fi is based on. But it's up to us to really build the guardrails architecturally within the AI agents to prevent certain things from happening. Can we prevent it 100%? I don't think so. But I also don't think the whole AI taking over the world thing, I really don't see that happening. It's not not in the way that these models were trained to even want that. They'd have to have a motivational system to want to do that. I know we're getting to the end of everything. I won't belabor this too much. It's eye robot happening. I agree with what you're saying. What's that principle? You can't blame a computer because a computer doesn't have the authority to make decisions. Humans are making decisions behind every little thing. All of you who are watching, every day you make decisions that impact where we go. I'm going to scroll through the chat real quick to make sure I didn't miss anyone's question. But I wanted to ask what social of yours would you want me to share so people can go follow you? Yeah, sure. I have a number of socials on Instagram and TikTok. It's R.E.I.D. Hamilton. I have a YouTube channel as well. It's not a very big channel, admittedly. But I have put my full render. I just yeah, yeah, I so I have my full talk video, my render talk there. I'll put it in the chat so you can see it. But yeah, yeah. So I have I put my full render talk there. I have another project on there as well, where I built a hollow mat. It's like a holographic mat. It was kind of inspired by Iron Man, you know, where in Iron Man one where he's able to like, see a 3D projection of his suit and manipulate the parts and all that. I didn't originate this concept. This was originated by another research group. But I kind of made it my own and built out different capacities and all this sort of thing. Oh, I love that. I love that. Yeah. Earlier was sharing about building a holographic something for their kid. I think. Yeah, yeah. It's so interesting. There are so many fun things that what we don't have is time, time. But I picked a lot of our curiosities, though. I love for people to go take a look at the library, read the paper or give it to your agent and. Yeah, ask them to summarize it. Asking questions. I think if they said and this is about ingesting information and that's it, I promise I'll let you go. No, no, you're good. There's incredible power in the ability that we have to learn as humans that when we're bouncing ideas off and that's something that like that safety to be able to ask an agent like I skim through it when you share me a link. For me to understand this, it will take an unbelievable amount of time. But I can pass it to my agent and then ask him questions that would sound absolutely dumb and ridiculous. But then I feel fine because I'm like, OK, like you're not going to judge me. Yeah. And there's there's yeah, we we I feel like we're past the point of like being judgmental, like we shouldn't be so judgmental, like we're all learning. You know, we all have to start somewhere. And like the AI papers, they're very dense just by the nature of how they're constructed. So, yeah, I think but the ideas are kind of fundamental in there. Yeah, absolutely. Also, final nugget, like sharing what you know is so important. I appreciate you taking the time to come by and share with us. It's an honor. Oh, my gosh. You're hired. I told you. Hey, look, I'll come work for GitHub in a minute. At least eight unlimited tokens. That's a hard it's a hard it's a hard it's a hard one to pass on. Like for real. Friends, go ahead and follow Cameron. I posted both the Instagram and the TikTok and then watch these videos. I want to watch your render ATL recap quickly. Let's give Justin some flowers because what a conference he's put in Georgia. Oh, my gosh. The math, the amount of work that goes into making a conference of that magnitude happen to bring incredible speakers to have the companies show up and sponsor it and to get at a price point where people can go and attend. Shout out to Justin. If you haven't been to render ATL Georgia, Georgia, I've been very impressed with the tech scene there. So yeah, it's a great, great conference. Absolutely. Well said. Yeah, I was really blown away. That was my first time going this year. And they really had everything. And he's built it over like, what, eight years? Yeah. His graph is like the GitHub. I was gonna say. Yeah. That's not that much time to build something of that scale. It's not. No, it's not. We'll have to get you to come out and check out Universe because it is I like to call it nerd Super Bowl. And we'll be there this October 28 and 29. Buy your tickets, friends, so I can say employee. But it's a fun conference. It's like an explosion of every exciting thing that's happening in tech. And for Mason, check it out if you haven't. Is this a GitHub sponsored conference? Oh, yeah. It's our flagship event every year. What? Let's go. Stay tuned. I'm ready. It's happening. Incredible speakers, like most, if not all of the Frontier Labs have some representation on the speaker stage. Plus, you'll get to see the latest and greatest. That's where we do our big ships. My boss says it's the big one. It is the big one. The big one. A friend of mine who came. Do you like that? It was a good time. Wasn't it? But it was. Yeah. There is so much to do. And there's a lot of fun. Like if you're a maker, especially like you, you mess with hardware. There's always an incredible maker space. Last year, what do we make? Lamps, Octocat, 3D printing, and you were actually sold out. So there's quiet spaces. There's food. There's entertainment. Oh my gosh. Okay. That's it. I'm coming. Yes. All right. I'll talk to you offline about setting that up. Yes. Stay tuned. Stay tuned. I appreciate you so much, Cameron. This has been delightful. Thank you so much, everyone, for being here, from taking the time to like, this was a bit different of a rubber duck. But I think we covered some really interesting things. And if anything, maybe it provoked a thought that you not consider. And maybe now you'll go and take a look at that. So that would have been a good use of time for you. I appreciate you all so much. Thank you, Cameron. Thanks for being here. Thank you so much. Don't leave so I can, I got to end this thing first. I'm going to go crank it with a paddle. I can't believe my setup just let me down today. Appreciate you. Don't forget to subscribe to this channel, please. So I can remain employed and bring you more awesome guests like Cameron. Don't forget to go follow him on socials. You make really funny content, man. Like very like, it's fun. All right. Appreciate you. Take care, friends.
