FDE Episode 3 : Agents creating businesses on their own | Weekly Tech Update

So hi everyone, welcome to the code Ninja spaces. It's our third episode and a lot of news to share with you guys. So AI is beginning to recursively self improve itself. Even those 4x product lead recently shared that advanced models are now monitoring their own training runs, creating their performance and correcting failures automatically as well. So is this the first example of recursive agents in production?

Recursive self improvement? You know we talk about this a lot. Recursive self improvement is the holy grail of AI where AI starts to for example train its own data, starts to fix its own bugs. So intelligence for the intelligence and then making it in the loop is something every AI lab is trying to build. Actually this is an incredible capability. So and if you see, so for example Claude CEO says like 80% of the software is written by Claude itself. And basically what they want to have some level of step function where they can just demonstrate to the world that we have achieved recursive self improvement. So yes, I think what we are seeing here with this codex is like at some level AI is now fixing its own bugs. So basically what we are trying to achieve here, I mean these AI labs are trying to achieve here is like where you, where inputs doesn't match the output. So you probably doesn't need hundreds of engineers and thousands of more complex training data sets and compute if somehow we achieve this recursive self improvement. So lots of exciting stuff when it comes to recursive self improvement. And we also need to see when we are building our own products, we need to at least aim for that. Can we achieve some level of recursive self improvement in our own products or services that for example I asked you to build this in hyper for example and you never listen to this. But yeah, that's what.

Architecture. So you know, on another news, AI doesn't stop there. One of my favorite news is for today is that the value creation with near zero human labor, an experimental autonomous AI agent reportedly created around 1,000 Etsy listings in 72 hours. And guess what, it generated its first revenue around $1,000 without any human effort ongoing. So do you think we are entering into an era where economic value is being created with almost, I mean no marginal human effort at all.

So that's quite a news. So I read this like a couple of days back that I mean we have this AI agent that can just set up its shop and then browse the Internet, come up with these 1000 listings and actually sell some of the products and generate hundred thousand dollar value, $1,000 value all the way up. So that's, that's. I mean this is the one such example of creating economic value with AI agents. But there, there's a catch here that we need to make sure that in well structured domains this is possible. So AI, I mean one of the challenge with agents is the humans are very good in tool use. So if, if I am using, if so imagine if so if I have to perform certain work, we just take it for granted that for example I've opened Hyper, I have opened lots of other platforms, I've opened Vs Code Cursor, Open Outlook and I've opened, I'm creating posts and I'm creating content. I'm just switching with so many different tools instantaneously. So, so we are very good in tool use, right? With agents, if you want to for example achieve this level of performance, as you just mentioned, that we need to make sure that AIs are very good in tool use and context, switching from one tool to another. And to make that happen, what's really required at this stage is that we need to make sure that we give AI is very structured data, very structured domain. And if we give these kind of things, if we give these kind of scaffoldings to these agents, they perform much better than for example, we don't. So in this news, if you go deep into this, how they achieve this, one of the core thing was that in very well structured domains, for example, in this E commerce reasoning chain of thoughts and tool use, AIs can perform, agents can perform much better. So that's one takeaway from this is like if you are trying to build agentic AI systems where you have a very specific domain, where you have a very structured data and where you can basically enable these AIs to switch the context from one context to another, when it comes to tools, AI would perform much better than for example, if you don't have these things.

Exactly. So on another front, JP Morgan forecast that memory chip companies are reaching around 1.5 trillion valuation by 2027. So it's not unfair to say that how much? 1.5 trillion by 2027. And it's not unfair to say that silicon is becoming strategic infrastructure as well. So there is a real question is that why is memory becoming the real bottleneck for the AI era, not softwares?

So I mean I have this different, what we call software. I mean, I mean I think in, in our last part we discussed software is not just these workflows. I told you that software is three things. Energy, what we call Software like lines of code and knowledge work. When you combine all these three things, that's a new way of thinking about software and that's new software. But to the point, one of the things that we need to think about is when you talk about memory is in terms of AI is like context windows, long context windows. So currently I think we grok has somewhat around 10 million tokens for context window. What it means is something like in this one context you can just Give the all 12 books of Harry Potter and it can maintain the context. But if you really see how it's doing, it is like how those new how these information is being stored. It's basically how this is being stored. It's basically the memory. So so and intelligence, when we talk about intelligence, what an intelligence is information compression and correlation. So when you talk about correlation, generating different ideas from let's say from last two years and then mix it up in a real time conversation with something that's happening in real time. If you see the very important component of that is memory in even in humans. So memory I think is fundamental to generating intelligence. It's part of intelligence. It's intelligence is actually is part of memory. It's not two different things. Intelligence is different and memory is different. I mean so memory, when we talk about it, it's like why they have so many high valuation, they're just like for example, I also know there's a company called Skywater Technologies, they have also 3D carbon nanotube transistor chips in the US so when we see so many investments and so much technologies happening at the memory space, that's why, that's the reason because we need long context windows. We need to have these AI systems have really long memories to make sure they can correlate different set of ideas to generate that level of intelligence, what we need. So every time when we talk about, let's say for example if I'm talking to large language model and you see that in some cases they're able to generate good ideas, but in certain cases they're not. Where they're not is they don't have that level of that much level of context from last so many years. But you have in your mind. So that's the reason why you are very good in generating those sort of ideas and AI are not. So one of the core is memory. So yes, that's a space to watch out.

I do agree.

I just want to end with like historically also AI was constrained because of the data movement and not because just math. So when we solve, we've. In terms of map. If you talk about. So even if you know about Stephen Wolfram from Wolfram. So they've done a lot of computational work in terms of generating what we call now intelligence. So computation has been. Intelligence is basically part of computation. That has been. The groundwork has been laid out by Stephen Wolframer W. Of Alpha, if you use those languages of Mathematica, for example. But there the challenge was we didn't have that infrastructure, the data, the memory. And once we had that, now we can think of like, okay, now got the real intelligence. So that's also one of the things why memory all of a sudden has become such an important part of generating intelligence.

Agreed. Intelligence need hardware. So there's one more thing then. Because of that we are burning physical molecules just to produce digital intelligence as well. So do you think this is fair?

So again I'll refer back to the core principle. What is intelligence or how to look about. AI is like it's intelligence produced per gigawatts of energy. I think Elon's. I saw one of the tweet of Elon, that's how he thinks about. AI is like intelligence produced per gigawatts of energy. And what is energy? So energy is lots of data centers. Energy means like we need. I mean if you know a small nuclear reactor is required to power a big data center, just think about like. So data center and energy are highly correlated. And when you talk about these centers, when you, for example, if you, if you, if you have read the news, the copper prices have gone up to $12,000 per ton. So and why they've gone up? I mean copper, I mean how do you connect these racks? It's all copper wiring. So data centers. So. And why do you need copper wiring? Because you want these systems to be energy efficient. So when you talk about intelligence and physical infrastructure, it's, it's, it's. We need to think about that. Intelligence is actually a physical infrastructure. Intelligence is energy, intelligence is data centers. That's why I said like copper prices have gone up to $12,000 a ton, I think. And because it's, it's. And all of these price price hikes are because of data center buildups.

Yeah, true. On another front, we have a newsletter, Tesla is operating unsupervised in Austin. So do you think is software is officially living the physical, the digital world and entering into the physical reality now? It's not.

So I think, I think I know what news you're referring to, but I mean I think it has. I mean, it's not the new thing. I mean physical, I mean AI for the physical world. I mean, it's been for more than a decade now. Uh, so. But yeah, I went with Tesla. The real thing here is when Vemo did it, it was lidars. And you can't think of it as a pure AI because there are a lot of sensors and stuff. But the Tesla yx, the big news is it's essentially cameras and lots of software that just processes in real time all those feeds that's going, that's coming in and then makes decisions. So, yes, one of my passion, you know, in Code Ninja, in Koninja is. And a lot of people disagree with me. I remember people come back to me and say, what? I mean, what is it? I mean, actually, people ridiculed me about this. Like when we, I used to talk about software for physical infrastructure and we need to think about how to build software for the physical world rather than just building agents and stuff like that. But now you see, FSD is one such example. There's so much, so much untapped potential in this, in this area. Whole embodied intelligence of robotics is basically AI for physical infrastructure. People find it hard to understand. People think probably the LLMs that are powering the intelligence that is powering our neural networks that are powering our, our chat, that's what a different kind of intelligence and intelligence that's required in, let's say, robots is different. It's essentially the fundamentals are the same. It's just like in robots, you need, of course, the dexterity to achieve in robots is hard. For example, my hand is, I mean, you can't just take it for granted this kind of dexterity is required in robots. But then at the end of the day, robots are being powered by the same level of same neural network, same core idea of intelligence that's powering our chat, that we call it large language model. So I think imported intelligence is an incredibly exciting area. And for engineers, for entrepreneurs like us, we also need to think about that. It's time probably to move away from just giving complex data sets to train these large models for these businesses. We've been working with lots of labs to give them these complex data sets, but we also need to think of how can we train robots now. So that's what I call, that's, I mean, that's what we call the body intelligence. So yes, it's not just Tesla Robotech. This AF infrastructure is something that we need to think on. That's where now the real Gains are.

I think, yeah, I do agree there's a new state. Biology is becoming programmable. If we look towards the patron as well, just you talk like right now as well that. Do you think we are turning in biology into our engineering discipline as well?

I mean again, this is not a new news. I mean we've done it in 2016 when Google DeepMind launched AlphaFold. So there's nothing new, I mean it's nothing newer. But for example, if you read the news of ILI Lily, they've come up with a medicine that has demonstrated to play its part to reduce weight by let's say 30%. So drug discovery using AI is something that's one of the, I mean if you know all these AI CEOs for example, if you watch Sam Altman and everyone says that they're gonna cure cancer, right? So that's what they talk about because I think the level of job displacement and disruption that's gonna AI gonna cause. The only thing that can justify this level of disruption is if we achieve something as big as cancer. But yes, AlphaFold, I mean structure biology, that is already, that's being done. So there's nothing new here. I think it's just like, I mean we're seeing all these things now more widely accessible for different entrepreneurs and startups and they can now build these things on structured biology, computer science and math. But AlphaFold has done it in 2016. Alpha fold 1, 2, 3. So yeah, I mean biology has already, I mean we've already made biology programmable as you call it. I don't know where did you get this line?

But yeah, so there's one more thing as well. Meanwhile researcher used around I guess 185nmuv lights to convert methane plus carbon dioxide into fuel. So it's not unfair to say that energy is approaching zero marginal cost. So what do you think? What happens when energy becomes cheaper than coordination? That's huge.

But still where do you get these, I mean get these lines. Energy becomes cheaper than. I think you do a lot of strategy anyways. Okay, got it. So I liked it. So I think yeah, energy is already, I think energy. The true lens in, in the last part is look soft in terms of energy. If you don't do that, we're missing the whole point. So. And the cost of energy is going to zero. Just imagine one of the biggest case study in the world is of Pakistan itself. I mean it's fourth, I think largest in the world in terms of mass urbanization of solar energy. Just imagine Pakistan, right? And Most of that is decentralized. I mean it's not wabda that's powering these homes and all that. So think about it. So energy cost of energy. And it's not just I'm talking about there are solar farms, some very large textile winds are being powered, some very large plazas, commercial landscapes are being powered by these solar farms. And solar energy in Pakistan. So energy is, the cost of energy is going to zero. The more and more in the next 10 years, we would somehow find a way to tap into the solar energy. It's energy. There are two forms of energy that we think we have, especially in Pakistan. One is that solar, we've seen it, but that is nuclear. And both forms of energy, energy would be widely accessible, especially for example if the competition goes into orbit and we use start using orbital compute, then in orbits energy is far more dense and intense than for example here on earth. So y cost of energy would be down to near zero. If not zero, it would be near zero. And the real question is what do we make out of it? What can we do out of it? Can we really solve some interesting challenges? Can we make our climate cleaner? Can we solve cancer? For example, if everyone stocks work, can we make our workplaces better? Fractive flowers work better. Can we bring purpose back to our workplaces? I mean, what do we do out of it other than just writing a very good email?

Got it.