FDE Episode 8 : AI Is Turning Thought Into Reality | Weekly Tech Update

Hi everyone. Welcome to the Codinger spaces. It's our eighth episode. So our first news is about AGI. What happening is basically there are some bold claims. Elon says that 2026 will be the year of singularity. What does it mean? It means that artificial intelligence will be more than human intelligence. And you know, at the same time, leaders at anthropics are saying that the term AGI will become obsolete. These are two extremes. So, Umar, what are your take on it?

So, so AGI means that artificial general intelligence, this means like computer can do. A model can do a lot of tasks that humans can do. So it's more of like generalization aspect of a human brain. Like, for example, we think about like, I can talk to you, I can walk, I can be very good in math. I. I can create music, I can drive a car. This is one human who is doing a lot of different tasks, right? So this is, for example, initially, AGI means like we would improve the model performance in a way that it can perform like human. And when you talk about like human, I mean, already models perform in some areas far better than humans, right? Even computers, even before AI, computers used to perform far better than humans. It's about the generalization aspect of it. And one of the hardest thing to crack historically was language. So language, you see, models have crack language. So in some aspect it is AGI, right? So they have crack language. But yeah, that's one part. That's one part. Like, what does AGI mean? And then once we have established AGI is really that generalization aspect of model performance, you can generalize across lots of different things. If this is. So that's one thing that we've stuffed before. AGI is. Now there are two parts. One is like, have we already achieved AGI? Or it's just a very concorded term. It does not exist. It means nothing. So I think both views exist. So, for example, one view is like, yes, we have to achieve AGI and there has to be a North Star. So what does that actually mean to. Sam Elfman, while talking with David Dutch, said, like, if you. If AI can prove quantum mechanics, so would that mean that we have achieved AGI? He said, yes. So that's one way of looking at another reason. We have already achieved AGI because it's. We've already achieved it because it has cracked language. So there's a bit of a dichotomy, but the real thing, I think the question as an engineer, we need to ask is like, I mean, this thing by Jay Cooper, Jake Cooper is the CEO of Railway. It's a cloud infrastructure company, right? Railway is not like Railway, it's a cloud infrastructure company. And he had this. So he gave Claude a system specification which he personally theorized over five years, right? So imagine you've been working on some theory for five years and then model produced a complete distributed runtime in Golang in four hours. Not a prototype graph, a working code base. Think about it, just think about that. You have in five years you have built, you have, you have. Think about like imagine 2010 or 2008, you have imagined something like Kubernetes or Docker, right? For three, four years you were always thinking there should be something like this, there should be containerized environment and all that. And AI happened. You gave that theory to AI. AI build, Docker. Think about, I mean one person could, I mean essentially create a $900 million business. So think about it like that's what AI. So we need to think about this part more than what, what is AGI and what's not AGI? That's very exciting and I told you I have some very extensive props that I have theorized for the last five years, six years. Not like, I mean it's not like that. Now I've seen J. Cooper and now I'm going to give you. No, no, I mean I've actually had those prompts that's basically regarding hyper. So for example in the last six, seven years one of the things that I've been thinking about is like how to help developers pursue great ideas like technical technology ideas, how to solve. So I've had Pinterest for software ideas, right? So think about like. So I've had that theory, have theorized all that knowledge base and then I'm probably going to give you in next few weeks and then when I embed it into Hyper.

So our second use is also related to something we really talk about, self recursive software. So second news is how AI models are showing self confidence. What's happening? Basically the calibration researchers at RAND Corporation reports that plot models can increasingly estimate whether their answers are correct or accurate before executing code. And this capability is being described as self reflective confidence as well. So the main question here is how does that different from hallucination and why should people care about this?

So interesting. So basically this, where do you get these terms self confidence and all that anyways? So basically what Claude, the scientists Rank Corporation shows that like Claude can really predict the outcome before actually giving the result, right? It's calculating the probabilistic outcome before it executes the outcome that is very powerful because it means a lot less brute force retries. And when you talk about a lot less brute force retries, it means a lot less token usage. And when it means a lot less token usage means a lot less energy. Right. So energy is the innermost loop. So that I think is very interesting that if somehow you figure out how to somehow model has that internal innate sense whether what they're doing is right or not based on probabilistic way, not in terms of if else that would be a very powerful thing to execute response in the least token consumption way. So that, and by the way, by the way, this is just this very earlier version of the results. Think of it as like intuition emerging from statistics. So that's what for software teams. Rafael, for example, if you're asking in terms of hyper, this probably would collapse QA cycles. You don't need much qa, just imagine why do you need qa? Let's question it. Why do you need qa? QA exists this one single thing that there has to be that developers can't produce quality free software. That's what, that's why QA exists. Right. Bug free in terms of like in any way. It's like whether you were not able to capture the customer requirements right or you just captured it right. But there were flaws in the code. But think about AI. AI writing the software and with the ability to predict the software having that kind of self confidence as you like to call would essentially collapse the QA cycle. There won't be any QA anymore, I guess.

But don't you think events are the QAs for the agents?

Yeah, I mean you can think of. Yeah, eventually that cycle would collapse too, I think.

Let's see. But right now events are being greater and popped, so can't say anything.

So we need to integrate EVAL in Hyper two. You told me that you were going.

To do it then. Our third news is related to the operating layer agent. So AI is colonizing the operating systems agents as the new app we really discussed App for the cloud, app for the ChatGPT. So what's happening here is these are the news being shared by the Microsoft as well that the OS layer is the operating system layer is shipping from passive software to active agents. What does it mean? They are saying that AI agents are the new app. So what does that practically change now?

Nothing. We knew that already. Microsoft knew it now. So basically what they're saying is like nothing. Going to embed agentic AI into The operating system probably in the Windows or something. Right. So they're going to do lots of this agent, they're going to integrate lots of agent TK experiences at the OS level. So that's one thing. Already there are so many agents of BCG. For example, I was reading up somewhere, BCG has 36,000 GPT clones in the organization. Already there are so many agents in the organization. So for Microsoft, if you're releasing Windows then at the operating system you're going to use lots of agents natively that would be embedded into the operating system. Right.

So yeah, the fourth news of the day is more related to the hardware. We really talk about this compute intelligence. So hardware is becoming new bottleneck. It's just like what's happening basically AI demand is pressing the global supply chain as well. So like why memories are jumping 70% higher. Doesn't software supposed to mean to be more cheap?

So but it's very interesting topic when you talk about AI there are two core things. Compute, right. Raw compute like that's required to for the training run. But then inference, inference is all about memories. Inference is like when you ask a question to model and it answers you at the edge, right? It's inference. So inference is all memory, your context windows are memories long like I mean I normally talk. I mean for example if I'm driving for let's say 300 for 400km, I keep on talking in the same session with ChatGPT wearing these glasses and talking all sort of stuff that's going on in my mind in one session. Why is it possible? Because there's a lot of memory. I have ChatGPT Pro, it's a very long context window. So memory is the bottleneck. Compute is not, Memory is the bottleneck. So Samsung is investing a lot in this. So Samsung is one company who is investing a lot in inference chips. There is a startup called ferosa. It is also trying to create power efficient inference chips. So the opportunity is in system Optim JI frugal infrastructure innovation. But yeah, I think now we have lots of GPUs but we need a lot of innovation now at the inference level. So that's why I mean memory is becoming this bottle like not hardware memory.

Agreed. So the effect news of the day is more towards logistic energy. The earth matters are becoming programmable. So what's basically happening is that physical systems are getting and entering into the software feedback loops. So these are the things being shared by SpaceX Texla. So the main question here is why does AI suddenly include shapes, batteries and rare earths. Like when we are talking about AI, we are at the same time talking about batteries and these materials as well.

So really the question is of a software, it was always possible. Like you can build something, launch it, learn from the customer, feedback iterate, then again relaunch. It was never for the hardware. So because hardware engineering is hard, I mean it's the feedback cycle in hardware. Imagine you build a rocket and it's not like launch a rocket in two months and learn, right? So but with AI and when it comes to AI, it comes to for example when think about self driving systems, which essentially is a software play. So every time a self driving car does something wrong, it just learns from it. Right. So it. So now that physical world has become, that has become the same kind of iteration layer that software, the world of software had always been. So that's what you're talking about. Like it's far more easy. I won't say easy. It's far more, it's far more possible to iterate in the world of hardware. In some sense. In some sense. Not in every sense. I mean I remember one of my friend is building this mobile device and he talks a lot about production. Hell, it's not, I mean in hardware you can build a prototype, but going to scale is incredibly hard thing. But yeah, some parts of physical world, let's say self driving systems or autonomy systems, they are more like software because they are no software.

So our sickness is related to biology. It's becoming editable software. So what's basically happening here is that medical breakthroughs are increasingly more towards computationally driven things. These are the research and things being shared by us, Food and Drug Administration, Stanford University as well. So the question here is, and the idea here is that are we talking about reversing diseases like cancer as well? Is there any possibilities? Yeah.

I think perhaps not curing everything, but changing the slope. Yes. So I mean for example recently FDA granted this Brakto status to Daxonovis, it's a company for pancreatic cancer. Right. So basically giving a breakthrough stratus means you can put this thing into steroid and you can fast track the whole thing. And this is basically AI led drug discovery. So yes, I think if you really Rafa, if you really understand, I mean biology or computing, whatever, at the end of the day there's some basic rules that are powering all of these things. Right. We need to understand everything in the world is a computation. Your body, information processing all follows some very basic, very basic fourth grade standard laws and then when, when, when you just. And then complexity emerges on top of it, right? Historically, for example, it was very hard for, for us to understand human cellular structure, right? How cell division happens. I mean what goes wrong in case of cancer. I think with AI it has become the first time, I think the first time ever. I generally think treating cancer is within human reach now it's first time I think you would see in next 20, 30 years. I mean my mother died of cancer, right. I have seen how the whole journey of how it takes everything from someone. I mean first you say okay, it's. But it is killing cancer kills patient every single day a little bit, right? And eventually you just. And when you talk to doctors because you feel like he's just, he's just fine or she's just fine, she's just recovering herself. The doctors knows exactly what's going on. They know probably two years. But I've seen that, how painful that is and why. I think it's a civilizational challenge. But I think with AI and not a general purpose AI, it won't be like large language model would be feeding cancer. It's more like protein folding kind of things where we'd be just able to predict with a lot more accuracy that what's going on there. And then based on discover some drugs and all that. So yeah, I think this is the area that I'm. My mother died of it. This is the area I am very passionate about that what can be done. And the way I. The lens I use is, is. Is computation. I read and studied Stephen Wolfram a lot. I think, I mean, I don't know if you know Stephen Wolfram Founder of Mathematica Alpha Wolfram Consulting So I just, I, in the last a decade or so I built that lens of looking at the world, that computational lens and it makes physics lots more, a lot more easier. It gives you that generalized capability to look at the most complex problems that if you really see. He talks a lot about Rule 30 where it says in theory of automata, if you for example, it's some seven, eight rules. And if you just play with those rules, it just. Complexity just emerges, right? He talks a lot about second law of thermodynamics. But if you really, really see what's going on is very simple computational rules to understand those rules. I think AI is going to play a lot because human mind cannot process that level of complexity. So to, to. So currently we are living on top of complexity. My body is, is basically a reflection of complexity to transform this body into a very basic rule. I have to. I have something that can just dig deeper and come up with that basic rule. I think that's possible with this company we just talked about, Directional Risk. This name is very mouthful, by the way, but this is exactly what they're trying to do.

Exactly.