The GCC Pod Ep 2 Part 1: Agentic AI, Ruliology, First Principles, and Micro Neural Networks

So Omer, today we're talking about this quiet but massive shift that's happening in AI. And I think we've all seen AI that can chat and can summarize and answer questions. But a different class of systems is sort of emerging and has already emerged, which is systems that don't just respond, but they actually act. And these systems are starting to sort of plan and decide and execute and even learn across real business workflows. So I know you've been working very closely with a lot of companies that are experimenting throughout this shift that's happening. But I wanted to start, I wanted to get your take and I wanted to start from the very first principles. Most people still think of AI as something that just answers questions, but what is fundamentally different and what is fundamentally changing with agentic AI? I think let's start there.

So when in 2022 ChatGPT happened, that was the first interaction of people choosing something that has AI in it. Like I'm just talking about a broader, I mean ChatGPT was in the lab, there was, I mean we know chargeability from ChatGPT3, but there was a chargeability two and then there was charge GPT1. Right. So we were having some, we, we were using some beta versions of Charge BT2, right. It had nothing like what chat GPT3 when GPT3 happened. So we know about AI through this chat GPT moment. Right now everyone is talking about charging. But really what's happening in the world of AI? I mean before we go to agent and all that, what's happening? The word affairs. We need to understand like how charge you works. What's this? What is the ruliology of charge? What's the simplest rule that that's, that's all this. AI chats and AI systems are falling really is a term that Stephen Wolfram is using a lot. And it's his idea that most of the world's complexity emerged from the very simple rules. That's what's called radiology. Right.

I was about to ask you.

That's a made up term from Stephen Wolfram. So he's a physicist who runs a consulting business and he's a big time inspiration. Anyways, so what is the ruliology of AI systems? Right. The rheology of AI systems is if you keep bombarding neural network with lots of data, it would start making sense. That's what I mean. ChatGPT and ChatGPT was some major algorithmic changes and updates. But then there was a lot of data at the time. You keep bombarding neural networks with data and it would start making sense. So that's what AI is all about. And it's being developed exactly the way a human brain works. So that's the core of AI. So we need to understand this before we go into agents and songs. That's how AI works. I mean it's exactly, it's, it's, it's its model exactly how human brain works. And if it is modeled the way human brain works, then human brain is very good in generalization, right? You don't need to write if and else. I mean, every time you see this, let's say bottle in your mind, you don't have this if and else statement going on to see if there is a blue color, there's something fluid going on, then probably it's water. You don't do that. Right? It's pattern recognition. It's that information compression engine that you have. It immediately, instantly tells that this is a bottle.

Yeah.

Now this is the foundation of AI and this is a thing on top of which now we are trying to build systems. Right. Now the first thing is chat. Of course, when we talk about chat, it's not about that chat existed chatbots. I mean, I remember we used to make lots of chatbots using dialogflow. There were lots of instances if someone would say this and then we would say this, then some predefined rules and rules over rules and rules and rules. The chat. You are experiencing of chatgpt an agent, right? There's a lot of agents. This agent doesn't have any predefined rule. It works on generalization, right? The way human generalize things.

Right?

So that, that, that's. First of all, that's a difference between the chat. We think chat is, I mean when someone says it's a chatbot, Chat GP is a chatbot. Sure, ChatGPT is a chatbot. But does Umar Khan is also a chatbot. Right? Because you chat too. You, your source of information connecting to connect. Word is chat, right? No, it's not a chat, it's not a chatbot. It's a very different kind of chatbot, but you like to call. Okay, it's more of like an agent that thinks like human brain.

Right?

So now, so that's where I establish what is, what is AI. How this chat, what we. What you just called that, it's. Is it, is there anything more than chat? So I just established it's not the chat. You think chat. It's not dialog flow under the hood that's powering chat. Based on rules. It's. It's the way human brain works that makes it very special. So just establish first this, that how important this, this work is. Right? I told you. Really? Ology, right? Yeah. Now, now the second thing is what we can do, okay. One is like we can do, we can interact, we can, we can word. We can interact with it and we can have the world's information from this and organizing and one of the things, there was a site in Lebanese, I think 17th century. Yes. The organization of human knowledge into one place has been the vision and vision for many scientists from 17th century. Google was the first company. I mean what we've seen that it has done incredibly well. Their experience was access towards information. Now with chat like experiences with agents now you have much better way to interact with the person knowledge. So yeah, one thing is like with chat like experiences one is chat but then beyond chat one is how can you make this chat like experience that is valuable to perform certain actions. Currently what's happening is AI is very good from middle to middle. It needs something to start. For example for chat, for interacting with ChatGPT, you need to give it something and at the end you need to fine tune as well. It's very good from middle to middle and it's not very good from the start and it's not very good at the end. Now what we are trying to do and what you asked is like how to make sense of the agent of this chat like agent, how to make sense of this. This middle is already solved. How can we build agents at the start that can just start doing certain tasks, use this AI like ChatGPT or any other AI to do processing of information, do so arrive at some decisions and at the end also an agent that take that thing and make it as one complete set of action or thoughts. That's what is called agentic AI. Right? So I just help you bridge the chat like experience using chat like experience first to establish that it's not just chat. Don't think of it like chat. It's lots of, lot of things are going on. Because when you interact with a human you don't think it's a chatbot, right? You give it up, you give it a complete structure. You have human. When you interact with human, you don't think of that, you're just another chat chatbot, right? You get human brings lots of things with the same things are being powered by chat, GPT or these AI or LLMs. Now we are building the first mile of that chat and the last mile of that and just to close the loop and make it agentic AI experience so we can outsource certain tasks to these agents. So that's what agentic AI is all about.

And I love the fact that you established AI in that lens as well. And I think what I understand from what you're saying is that essentially it's almost like with agent AI the system is kind of being trusted a little bit more with both initiative in terms of like the first mile and then also completion in terms of the last mile. Right. So it's kind of like giving it a little bit more agency and hence the name. But you know, I think I want to go a little bit deeper because you know, people hear the phrase oh, execution of actions and immediately, you know, they think automation. So how is this different then from. I mean we know it is, but I want to dive a little bit deeper into that difference. Right. How is this different than from automation or rpa which is, which is something that enterprises have been using for years. So if you could talk a little bit about that.

So when you talk about rpa, Robotic process automation. I mean again, the traditional RPA was based on rules. There were some predefined rules. You write lots of rules and you get certain outputs. Right. The same thing can be achieved with.

Now.

AI systems with agentic AI, but it would have a far more flexibility and generalizability. So eventually what I say is like you can do, you can do everything what an RPA can do for you with AI, but it's not the other way around. Okay, interesting, right. So what I think is again, that's why we need to establish. What we need to establish is that radiology principle that is being modeled based on the human brain. So RPA would be there. I still think RPA would have its certain need where it is already there. The real comparison of these systems, these digital beings is not with RPA and something because you want predictive rule based system anyways, right? You don't want. Because if you didn't want predictive rule based systems there, then probably it should have been done by the humans at the first place. Right. The real thing of thinking about AI is what parts of that currently humans doesn't want to do in our workplace, in factory floors, can we replace agents on that part? So for example, someone just hit, for example moving data from one Excel to another Excel sheet. Consolidating, consolidating data. Continuously doing lots of boring accounting work. In terms of software, we talk about people hate bootstrapping application. They just Want to focus on business logic, all those things where I mean we have this definition of what are the kind of what are the stuff that human doesn't want to do and when you talk about doesn't want to do. You give humans enough time for anything. They would develop flow in anything. But the question is like what, what is, what's harder for the humans to get started with, right? So because that compromises their flow state. I think wherever you find, you know, in the workplaces that the areas where human doesn't find, where you find that humans doesn't like that kind of work, that's the first kill, that should be the first kill of AI. That's how I like to look at it. When you come, when you talk about automation and all and getting ROI by deploying a. That's a, that's a great lens too. But I think currently our focus primarily is finding out areas where we can build copilots, where we can, where we can build agents, where humans probably human brain probably is not finding its flow state in the workplaces. So we just, that's, that's our first line of action. And then also coding has this vision of keeping human in the loop too. So that's currently how we think about AI and that's how I think about RPA versus this agentic AI systems. That agentic AI has to be compared with human brain, not with rule based system.

And you talk a little bit about replacing this grunt work with agentic systems and software in general. But if we move away from this theoretical discussion that we were having before and double down a little bit on this part of what you said, like, you know, where are companies actually deploying this? Like what are some of the spaces, the use cases? Not in terms of like demos, but like real uses of this. Where are you seeing, you know, some early deployments?

So because Codinger is a software business, so most of the companies that we interact with and internally most of our work is around internal tooling and workforce orchestration that makes software development easier. That's the first thing that we are seeing lots of value here. But some of the areas that I think which are very seeing and other than that, other than software development and testing, some of the areas that I think where it is adding a lot of value is operations monitoring and fraud and disk investigation. These are the areas where I find that currently there's lots of value currently being tapped by deploying agent TKS system. So for example, if you are a supply chain business or if you are a logistics business or if you are. So these, I mean, where you need a continuous surveillance, you can build very good video AI agents in those parts and deploy those agents, which means a camera. And this micro neural network, you can pretty stack up. And that would be a human, digital human for you would be continuous monitoring. Whenever something unusual happens, it spikes. So these are the areas where currently I see lots of value.