FDE Epi : 16 | AI Agents, Billion-Dollar Infrastructure & the Future of Intelligence

Hi everyone, welcome to the Code Ninja Spaces, episode 16. This week we are talking about how the intelligence is breaking down the institutional boundary. It's governments versus AI labs, and the question is who controls the intelligence. The US Department of Defense recently demanded the unrestricted access to AI models from Anthropic, citing the national security and supply chain risk. Yep, Anthropic refused to do the cooperation related to the mass surveillance an autonomous weapon. But something was more interesting. At the same time, OpenAI signed a mega deal deploying their AI models to the Pentagon classified network. In result, 591— around 591 Google employees and 93, I guess, OpenAI employees signed the protest letters against it. Even though some of the professors from the Wharton, Edson Mollick, was saying that when AI becomes this strategically important, this is obvious things to happen. So the question here is from you, that why does governments are trying to control AI labs for the defense perspectives?

So, uh, specifically about this cloud thing what you talked about, uh, historically US government, uh, was working with a company called Palantir, right? And, uh, regarding intelligence and things like that. But when you talk about cloud, so the whole architecture works is in a way that Palantir has this incredible capability to process very large amounts of data, right? And then when you bring cloud on top of it, you can take actions at the speed of thought. At the speed of thought, for example, if you're in a high-pressure environment, especially in defense and things like that, you need to take decisions extraordinarily faster in what's called speed of thought way. So with Claude, with AI, intelligent compression, Claude, even very large chunks of data, GBs of data, video data, text data, intelligence data, it processes at the speed of thought. As you're talking, it's giving you the real-time information. This is incredible, incredible value for defense, for the government organizations. So that's what I've seen. What we've seen is US government wanted to use Claude for certain, uh, defense ops, uh, and then Claude said, uh, you can't use it because of these two clauses. And then there was— you've seen, uh, on X how it has unfolded, uh, and now OpenAI has said, okay, we would do it. But clearly OpenAI is not even close to what Claude, especially in these— in those— in those, uh, scenarios. So I think at one end, I'm not sure how to think about it, but AI is— is— it's It's incredibly powerful. This is gone. It's mathematics on steroids, right? Or it's for governments. I think they've realized it's important for them to control this. It's just like they're controlling other critical assets. So this is what's happening in this cloud versus the government and all. But for us, for people who are listening to this, I think they try to understand What you can make of it, right?

AI is becoming an economic actor. Multiple signals this week show that how AI agents are moving from tool use to actors. So even reported that their agent usage is now more as compared to their autocomplete tab feature. Law firms are branding themselves as cloud-native companies. Some of the researchers at UCSD are also creating agents that are living in simulated societies. That is something huge and interesting as well. An agent named as— moreover, an agent named as Einstein is attending university lectures and completing the assignments autonomously. So what does it mean when AI agents go from tools to actors?

So it's interesting, uh, if you see, uh, software currently is 1% of the US total GDP. 8%. White collar work is somewhere around 21-22% of US total GDP. Either 21-22% software, that's, that's— I mean, it's just a matter of time. Or this is a major— at one end, it's a major threat as well. Long-term job displacement is on the horizon. Uh, that's one part. But when you talk about, uh, uh, economic labor and all that. So if you see OpenAI, it's started to publish evals like GDP evals, percent of GDP, uh, the AI can do how much, what percent of the GDP and stuff like that. So now that's a benchmark. When we talk to the customers, CIOs and CTOs, so for example, I won't name one of the major health insurance business in America, and that their cost currently per dollar is somewhere around 90 cents, and they're saying how can they do it in 20 cents, right, per $1 spent. So AI is going to take a lot of labor, but it means it's going to replace lots of jobs. So at one end, it's exciting. There's so much work that previously— there's so many latent ideas you previously— now I am doing it, right? It is helping me a lot. Like, but the thing was like, you need a lot of energy to do boring work. Right? So just scaffolding of work, you're doing it yourself. But on the other hand, it also means there's lots, lots of, uh, job displacement. There's a risk of long-term job displacement. For example, this is one part, but Meta is building its, uh, a new organization around AI and the rule of that organization is there should be one manager and 50 individual contributors reporting to him. So, so flat structure, possible. So there should not be manager over manager over manager, right? Now largely the organizations will be flat. One guy would be able to manage lots of different things, and, and that's— and I think that's a good thing. So there are some ups, some downs, but I think largely there would be more— there are lots of latent ideas. So now the context, so I can execute lots of work, I like it. So it is taking a lot of lots of, lots of labor.

Is it possible because of co-work majorly as well?

Agents. So, I mean, historically, as David Shew used to say, like, AI is good from the start. I mean, but it is good from middle to middle. Real shift started to happen when co-work, especially co-work execution, when they started So when it started to do work on my behalf and it gave me a lot of time to think. So currently, for example, if I am— so I outsource my work to cloud cowork, let's say, agent architecture. And then I am, while I'm— when it's outsourced, then I'm thinking about new ideas along the way, like what else, like what else needs to be done, what else needs to be done. So For me, it's very exciting. Uh, there's a long-term job displacement up here too.

So third news is related to the massive consolidation of AI infrastructure. There are some major infrastructure development going on, uh, which includes even those— OpenAI has recently, uh, raised $110 billion at the valuation of $730 billion. With that, OpenAI weekly users are 900 million. That's huge. And with even those Amazon and OpenAI announced $50 billion partnership using 2 gigawatt, uh, 2 gigawatt Trainium compute. And just to give the context, 2 gigawatt Trainium is roughly, uh, the power needed for a small city. That's a huge amount of power. Or in, in the parallel, NVIDIA is developing new infant-specific processors that's gonna change a lot of things. So why are companies investing hundreds of billions into an AI infrastructure?

I mean, 900 million users, or so the whole— so Claude is entirely focusing on enterprise primarily, or OpenAI, it's focusing on retail end users. But it's all computation, it's all math. Or— but it requires lots of compute, and safe computing, lots of energy too. So mostly what ChatGPT and Claude is trying to do is they're building the entire compute infrastructure, they're building energy infrastructure, and then they are also building the— on top of it, software stacks. I think I mean, it's neither cloud, not— it's Google. From top to the bottom, infrastructure already consolidated. Like, they have their own TPUs, they have their own data centers already here. So when they launched cloud, so where GCP services, AI/ML solution, this is when you— and tagline be like run your processes and infrastructure on the same infrastructure that powers YouTube, right? So, so true infrastructure power consolidation. So, but yeah, I mean, I think again, we are too much focused on— here in San Francisco, everyone is focusing on one way or another, just another LLM. And 5 million Super Hugging Spot models. So we're not short of models now. It's stable sticks, which we need, right? We— what we— innermost loop is energy. So I think we need to think about, um, I mean, we are an AI company, but I think what I am pushing internally is how to think about energy, how to think about— so today I had a meeting with someone and we discussed at length what to think about Pakistan in terms of infrastructure. So Pakistan Pre-training infrastructure, it requires a lot, lots of real estate and lots of energy, but energy to access, but inference infrastructure, Pakistan, so for example, inference, Pakistan, I think for all the inferencing can be routed, and that's one of my goal is how to convince important businesses to set up their inference infrastructure within Pakistan or within emerging markets. So yeah, that's what it is.

So AI is like everywhere. The port news is related to the AI accelerating the scientific discovery. Drug discovery historically used to take years and years. Now there are recently scientific breakthroughs that include some of the major things, that is FDA-approved drug of lung cancer, and next year something in 44 days as in period. That's huge. Even those cortical labs demonstrated biological neurons, neurons learning to play What does it mean that scientists view living neurons, brain cells, in a lab? And even though there are things when eventually in a world neurons gonna be replaced with these silicon chips, how does the world gonna look? These are some very, very fascinating things because the brain uses around 20 watts of energy and data centers uses megawatts of energy. So how you look towards the AI accelerating this scientific service?

Where you want to accelerate the innovation, especially in sciences. So we talk about protein folding a lot. Protein folding to 1970s. So at least we had all those amino acids, right? It was a very well-structured problem, uh, ripe for, uh, this age of AI. Problem framing. So, uh, scientific discovery, time compression. Like, for example, given— so can we cure cancer in the next 500 years? The answer is probably yes, but it means truly what AI brings to the table is 500 years. Can we compress those 500 years into 5 years? Right, so, uh, I think probably number one thing for scientific discovery is the problem has to be well structured, well framed. Uh, for example, in terms of math, can AI discover new maths? Maybe, uh, like for example, I am from a school of thought of Robert Penrose, like for example, AI is mathematics, it's not cautious. Or, uh, it— so if you take that, then before you update, okay, what is, uh, Gödel's incomplete theorems, right? Uh, what it means is like there's a lot of emergence, there's a lot of computational irreducibility, uh, that you cannot, uh, prove everything in the world. You cannot run— if you can't prove everything in the world, it means you cannot— if you can't come up with proof for every math, uh, uh, for every math statement or mathematical rule, it means you can't run that on the computer. Right? But if you can't run that on a computer, it means not by design. It's not a mathematical process. If it's not a mathematical process, then clearly AI views speech as garbage. So this is one thing. But yes, same thing, like where you, for example, think about protein folding. We had these amino acids, but we didn't have that insane amount of compute infrastructure, insane processing. It can process huge chunks of data and bring that, those patterns. Uh, it was going to have a long-term impact. Impact, but still I would say scientifically, probabilistically, what is the probability AI would be able to find a drug that would cure cancer in next 30 years? Yeah, probability is probably more than 20% or 40%. In next 100 years, yes, probably 70%. So that's how you want to think about it. But by the way, it's in terms of scientific discovery, you talk about cancer, if we cure every person who is living, who is alive and is struggling with cancer. So average age won't increase more than 2.5 years, average of the planet. So real challenge the world is facing is probably not cancer, though it's a very important problem. It's probably longevity if you are in that group that thinks that humans should live longer, or these war-like situations. Yes, problems.

So moving on to the next news. Governments is losing AI control as we're doing, looking that defense AI is moving so rapidly. Government institutions are struggling to regulate AI. What's happening basically, California requires now complete operating systems to collect birth date information. At the same time, AI-generated public comments are influencing the climate-related regulations. That's a huge thing. These all are AI-generated comments and public comments, but that are affecting the climate decisions directly. So what do you think, why is government struggling to keep up with AI?

Because of the slow process or governments, the social media struggling. I think AI is far more pervasive. I think the fundamental difference is it's a technology not designed to be boxed, right? Like if they just say no one can use AI and stuff like that, it is, I think requires new kind of laws. It requires new kind of governments even, right? Uh, so currently, I mean, governments— imagine governments are organized, government departments are organized to think about how to, uh, do policymaking, how to manage. We just talked about that all major private organizations are bringing this flat hierarchy, right? So, uh, it's a challenging thing. I think again, I would say I would prefer Imad, Imad Mustafa, like how to think about— he talks a lot about city-state intelligences, provincial intelligences, and then national intelligences. So sovereign AI stack deployments, or government at some point, how they would start thinking of AI as just like they thought of thinking about other agents, other intelligent species. It's not, it's not designed to be governed as just a government process.

Exactly. So you mentioned blockchain with AI. What's that combo gonna look like?

So blockchain is the other method, for example. Think about currency, UCST, uh, currency, for example, think about money, money programmable, uh, people think about if I'm using card, that probably is like I'm like, for example, credit card use, debit card use, that's what, uh, it's not programmable. Programmability— when you think about, uh, let's say after executing this podcast autonomously, we disperse some payments. Podcast complete, there are certain amounts that will be dispersed. Uh, just think about it. For example, if someone wants to send some money from US to, let's say, Dubai, or let's say Pakistan, it takes 3 days, 5 days, and all that. Why not instant? So cryptocurrencies, and they can be sent instantly. The reason is because it would make money a lot more programmable, like decisions at the edge. So there's so many new business models, like for example, that can, uh, that can, um, that can— you see, you would see startups would be built around those. And I think, uh, programmability, government issues, government's heavily relying on taxes, heavily relying on I don't know how to think about it. How to think about it exactly.

So the last news is related to that. Intelligence is becoming cheap as we're looking towards the trend that ChatGPT weekly users are around reach 900 million, and cost per unit of intelligence is rapidly falling as well. One of the primary reason is that a lot of investments in the AI compute infrastructures as well. So what happens when intelligence eventually will become so cheap?

So for example, I spend a lot of money in running my workloads. It depends what you're doing. But if you want to run heavy workloads, like for example, if you are doing real labor work from Cloud, and the model is called Intelligence by Drink. Like, you use it for— there's a base price, but then you use it, uh, for a certain window. So it's not cheap, uh, it's very expensive. And that's the reason historically SaaS pricing, the SaaS pricing, $20, $30. But currently most of us are paying $200, $300, $500, in some cases thousands of dollars a month, for doing their workloads. So it depends what kind of work you are doing, uh, on this. But, uh, so it depends. So, uh, if you don't, uh, if it's— if your operating work operating system is not agentic, like you're not outsourcing a lot of work to agents, but if agents, uh, uh, they maybe So 1800 tool calls. So for me, it's 3 hours, 3.5 hours, but then, uh, it took, it took most of my day to limit the tokens. So it depends, but down the line it will become easier, but I think heavily dependent on What kind of questions we ask, uh, capability has, uh, one would ask God to write a better email, another probably would say predict the molecular structure. So the cost, it depends also, I mean, discovering a new molecule, curing someone, curing a disease, and then, uh, uh, A very famous quote that it never depends on the model, it depends on the user, the quality of output.

So yeah, I did. So that's it for the today.

Thank you. Hyper is a composable white coding platform, or as we like to call it, CVC platform, to help knowledge workers build internal tools faster and better. So rather than scaffolding the entire solution with AI, Hyper helps build blocks of code through automation and then uses NTPi orchestration to help stitch all the those blocks of code to give more deterministic internal AI tooling you can trust. So essentially it's a better way to build custom tooling or custom software. The idea is to have developers use the AI tools maximally in a one unified workflow. We've built agents like business analysts to help developers groom the requirements, and then we've built data modeling that helps you transfer those requirements into a complete data model. Then you bring all those things into HyperAI Studio where you build tables, hook the databases, build MCP tools to hook with a bot code and build custom workflows all using HyperAI. The idea of HyperAI is to land up at a point where helping organizations self-build their internal software systems rather than outsourcing to services vendors and coming up with a platform where the entire solution can fit into a mind of one developer. Both seemingly different ideas come together as one holistic outcome. So it has a dual use. And I think it's a much better way to build software where you're not outsourcing everything to the vibes and we're not only using quad code to generate all the things. It's basically a spectrum between vibing and architecting software. It's a new platform that we call Composable Vibe Code.