FDE Epi : 15 | AI Agents Are Replacing Users - The New Intelligence Economy

Everyone, welcome to the Code Ninja Spaces episode 15. Every week we track global tech signals but this week it's a little bit different. What's happening is basically the news. First news of the day is that AI that earns its own money. What happened? Basically Conway Research launched the automation that's an AI agent that is capable of deploying products, entering into the trade market prediction, doing sort of cold calling businesses, running marketing campaigns as well. And you know what, it operates autonomously as long as it remains to be financial solvent and simple terms that it needs to generate its own revenue to pay for its own compute. Interesting. So in parallel N archives, that's a non profit digital library. Asked AI agent can you donate to the library if you are capable of financial actions? So that was very interesting as well. And on the other side the capabilities of models are getting increased day by day. What happened? Basically Claude Sonet 4.6 code highly around 16, 1633 score eloquent on the GPP val of artificial analysis and their financial agents are improving as well, around 63.3%. That's huge. So what's your take on it? That intelligence is getting stronger and affordable day by day as well.

So this agent, I mean if you, it's Lobsters 101. If you have read Accelerando book, there's a thing, I mean it starts from lobsters. So this whole agents doing all that work is basically after cloudbot I'm like now, now we've seen like if you, if, if you manage the memory architecture better and if you do tool calling better, if you help agents, if you, if you make agents access your file system better, then you can do a lot of work. A lot of work. And if you can do a lot of work then probably agents can basically not only build software, they can actually sell that software, market that software and earn money right on top of it. This is, I mean it's unheard of, I mean donating to the libraries and all that. I think this is, by the way, this is, I mean these AI agents are performing economically productive activities. Like you talked about automaton, like registering domains, launching services, you know, 97% of databases are being spun up by agents. I mean tasks from software engineering to tool use, most of them are now, I mean, I mean agents are now quite capable to do all those, all sort of tasks. So I think this is agents 101 I'm telling you. I mean, I mean we've been using a lot of agents in Code Ninja, you know and the real challenge is Context after very long conversation after. I mean I perform series of tasks and Sometimes it takes 300 tool calls or 400 tool calls and then it says. I mean then it starts hallucinating and then you figure out okay, that's probably. That's from this point onwards you need to do it manually. But imagine if you have tool calling if you have the context long enough. Currently Claude is 1 million tokens. So probably if it becomes 100 million tokens then literally you can do any task. I mean you don't need human for that.

Exactly.

I don't know what is it.

So moving on to the next signal of the day is related to the intelligence is improving the cursive. What's basically happening we discussed before as well did Sonet anthropic model of Plot Sonnet Fourpence has achieved the state of the art performance and and even those at lower cost at the same time Elon tweeted as well and announced that Grok Open 2 features are continuously post training learning. It means the models are recursively learning of the deployment as well. We have been seeing years and years software are always fixed and widgets no improvement. But now we are moving towards the recursive improvements on the parallel. If I got recently from AI Summit of India, Sam says 2028 is going to be the year we're going to achieve ati. That's something related to recursive intelligence as well. So what's your take on recursive intelligence?

Recursive intelligence is extremely scary. I mean not everything is good about it. It's procreation. Right. Recursive intelligence means that a software would, I mean self improve itself and then for example I'm not sure when Elon says like now they would do post training learning what does it mean? Like every time user interacts with the model with that interaction. That's not allowed currently. I think that's not. I mean it's. This is not the path. I think probably what he's trying to say is maybe from the user feedback from. From X or something like that. Maybe that's what he's saying because if you start training user on top of user data then probably even 2028, I mean you can do AGI in 2027 because that is that. That would. You would give models so much power. Right. So. But recursive self improvement in. If you see sci fi books, if you read most of the sci fi books there is some dedicated chapters on recursive software improvements and that if you see where it ends I Mean currently our finite minds can only comprehend I mean what probably. I mean we have our limits regressive self improvement. I mean you are operating at 110, 120, let's say IQ and a recursive self improvement. An agent that who's continuously recursively self improving itself. That's it's Maltese or lobsters. They would probably be operating at 700 IQs. So we, I mean they would produce things out of thin air, right? So I don't know what does it mean? And I think I've heard Eric Schmidt a couple of months back that probably we should not do two things. One of which was this that you should not post train on user behaviors, user interaction.

That's it. But we are into the AI war. So the next news of the day is related to same AI compute. The signal is that $80 billion infrastructure war for AI compute. You know what's happening? Anthropic expects to pay Amazon, Google and Microsoft $80 billion by 2029. Like up to three years. And Meta investing billions in Nvidia Blackwell Vera open chips. And these chips are capable to deliver 50% 50x more throughput per megabyte megawatt and 35x percent more cheaper, more lower cost per token that these things are huge. And you know what the companies like Google signed the contract with their company Ormet signed a 150. Ormet is a geothermal company. So they signed a 150 megawatt geothermal agreement that going to power Google power centers. These things like companies are spending tens of billions on compute infrastructure, just all for the intelligence. So what's your take on this?

So you asked two, three different questions. I mean you touched two, three different points. One is you talked about Dario and Claude. I think Dario recently on the podcast said like if they don't achieve AGI they would become bankrupt, right? They have to. Because somehow we have put by the way, I mean we shouldn't have gone that path. That OpenAI nonprofit thing was I think good enough. But now the genie is out of the bottle and now we are in war. So now this all what you're seeing is a competitive streak, right? That it's just about that now in a war. And then in that war we have to just. Everyone is thinking about how to win. So just think about it like, like clearly Anthropic is not going to be bankrupt. So they would do everything, whatever it takes to push for AGI, irrespective of what kind of disruption it would be and all that probably slow it down and put some moratorium on certain things and stuff like that. But that's one thing. Second thing is like electricity and tokens are the same thing. The only difference is in. In between there's a thing called gpu, right? So these data centers require massive energy and we don't here in the US we don't have energy. Right? We. As compared to China, we. The good thing about China is China has transitioned hundreds of gigawatts into solar. But here in the US it's not even close. And when you think about solar, you can probably power your homes and probably commercial offices and stuff like that, which is solar. But Honeywell CEO said, I think you can't cut the steel using solar. It's against the physics. So you need, I mean if you are manufacturing a lot of steel, if you are manufacturing a lot and America has this goal to bring manufacturing back to America, it means you need a lot of energy just for, for example, steel production, heavy equipment manufacturing and stuff like that. Right? And then if energy is required there and your homes and your offices and all that, they are also also not being powered by solar. They. Then there's little available for data centers. So energy is the real. I mean most of, I mean there are lots of GPUs. GPUs are already in access. But the real challenge now is energy. We don't have energy here in the US but. And everyone is trying to. So OpenAI is very close to closing their $100 billion round. Entropic has I think closed $350 billion round. I mean think about billions. Think about these billions of dollars of rounds. And most of that money is going

towards energy acquisition being in US have you seen any strategically move they're taking for the energy for compute? Because being in Pakistan there are a lot of solar energy, but is there any US is.

Pakistan is world's fourth largest urban solar adopt. I mean in Pakistan, Pakistan world's number four in urban solar adoption, right? So Pakistan is very high in terms of solar adoption because at the time there wasn't any other energy resource. So it was more of like we had to leapfrog to this. Like in US it's not, I mean, no, in. Not even close to what's happening in China or in Pakistan in terms of solar. But what US is doing is currently they're powering. I mean the plan I think probably is to power their data centers in the orbit. And probably if that plan gets successful, Elon is doing it. If that plan gets successful, then probably they have a Very good chance of finding a way out in like for example to meet these energy demands. I mean they would probably innovate their way out into this. Otherwise I'm not sure how this is going to happen. So lots of data centers are being built in the Middle east with Saudi Arabia. There's some very good proposal to set up some of the data centers in Pakistan as well. Right. So.

So yeah, perfect. So the fourth news of the day is related to robotics is entering consumer market. What's basically happening that Tesla manufacture their first cybercap robotaxi at Giga Texas and their plan to launch it for consumer availability of the price going to be around $30,000. And on the parallel unit we demonstrated coordinated human robots deployment. We have seen some hype of them in India AI summit as well. So moving on. For decades we have seen robots exist only in labs. And now all of a sudden we are looking robots everywhere around us as well moving into more consumer markets. But the shift indicates us.

So robots hardware was available from last 30, 40 years. It was like we didn't have those, that those sort of reasoning capabilities that we have now. So now robots can reason. I mean robots are not different than Claude cowork. If you see, you see how Claude uses tool calling and all that. That's if you really see the one is tool calling but on the other hand it's basically the model reasoning capabilities. Now we have those sort of capabilities and now we can now power that robotic hardware with that. The real challenge with robots is that like for example, can you outsource a plumbing task to a robot or can you outsource some mining task to a robot? Because mining by design is extremely dangerous for humans. Currently we're not there yet because the dexterity of the hand is not fully achieved. Now hand is Elon says is one of the most hardest part of the body to engineer. But maybe next two, three years here in the U.S. what we're seeing is I've seen some of the rewards in some of the exhibitions figure have. I mean it's not just Tesla. Optimus figure has some incredible robots. So I think we would see next few years that would have at least by 2030 it's expected that we would have same number of robots as we have cars.

Oh nice. As in human. I guess we are always really fascinated get fascinated by looking towards human robots doing work.

Humanoid robots are now pretty common now. I mean if you go in any exhibition and you will find out at least 10, 20 booths where they're demonstrating optimus look alike or something like that.

Exactly. So on front test we maybe we no longer need humans Robots going to be much more flexible.

Today I was in a conference and they demonstrate it. I mean robot, the first Robot was built 300 years BC to mix wine with the water. Oh, so it's not something. I mean human curiosity and quest to build things like robots being 2000 or 2500 years old.

Interesting. So moving on to the next signal, human machine interface is becoming physical and biological as well. What's happening? Basically just like meta, Apple is developing AI powered smart glasses, camera, AirPods and more towards wearable AI devices. On the other hand, Neurox beginning human brain computer interface trials as well. We can see BCI trials. So the question here is that AI is no longer just limited to the software, it's emerging into our bodies and environment as well. What does this transition mean?

I mean I've used these are meta glasses with my lenses. But the real challenge with wearable devices like this is like security in so for example in airports, there are certain airports I travel to, I travel from they have explicitly written you can't via these variable devices in certain public places. Even west is by default public spaces in west means nothing is private. Right. But however I've seen places where these are banned. So one of the things is security. What does it mean? But on the other hand I can tell you that it just powers your productivity 100x it's unlike smartphone. I mean it seamlessly integrates with your voice. Then you can currently the model is meta. I mean llama model is crap. But imagine if OpenAI built something like this, works on variable devices and then connects their models or Claude connects their models with them. It will be extremely powerful. This idea of for example opening this thing and doing this, this would look like pager. I telling you, I mean 10 years from now this would be a pager, smartphone would be a pager. No one is going to use it.

Exactly. So I guess that's great. So moving on to the next news, that's something very close to us as well. We already discussed it. Software business models are collapsing and getting rebuilt Even though Sathya said as well that SaaS is dead now what's basically happening is that companies like Snowflake and Databricks are moving away from seed per pricing because the trend of consumption based pricing is becoming more dominant. At the same time, Microsoft predicts even that computer based jobs are going to be automated within 18 months. Claude CEO said that as well. And the trends are moving towards more figma and Anthropic integrated AI generated code directly into design tools as well. These all are indicating a shift that for decades software companies charged per user. But this model is breaking day by day. So what's your take on this?

Yeah, I mean SaaS, I mean in some SaaS products the margins were 70 to 80% margins. I mean that era is gone. That's one thing. Now this is my favorite topic as well, the consumption bill pricing. So it's also called intelligence by drink. Like for example, you use the intelligence and you pay. Just like for example, you had a drink and you paid for it. Right. This model is called consumption model. It's called intelligence by drinking. I think there are three different tiers. Currently the top tier as we talk about, there's a guy called Ramanujam, right. He's an authority on coming up with pricing for products. Right. He's an ultimate, I mean he's written a book. Two books. He's written two books. One is monetizing innovation and second, there's another book, I don't know, both are around how to price your products. So he talks about this, about these tiers. At the top of the letter is for example, pricing outcome based pricing when it comes to AI, which is like for example, if there's a product called Intercom, Intercom is a customer service product. It only charges you if the conflict is resolved entirely by AI, right? So that's called outcome based pricing. If human resolves the conflict, then you're not charged. But it's hard to see. It's challenging to deploy that kind of border into other, other parts. However, that's the holy grail, that's everyone wants to go to, to this part. Like outcome based pricing. In the middle is this consumption based pricing. Like you would have base pricing. Like for example, if you're using Claude, so you would have $20 and you can use Cloud. But then if you want power doses of Intelligent, then you would take max package and $100 or $200 and stuff like that. And then if you end up consuming all that, then probably you would have new, more tokens. So this consumption bear pricing and the whole idea of this is like you don't want to be operating at 1000 IQ at 24 hours. You probably need some doses of that intelligence and the rest is okay. Right. So but this is, and I think in this part a lot of innovation can happen. If lots of companies talk about replacing SaaS. We also talk about replacing SaaS. It's not about building software and building, I mean if you can bootstrap something like Salesforce, that's less of a thing than probably. If you can somehow challenge your business model in an unpredictable asymmetrical way, then probably this is far bigger innovation than for example building the same software, a lot less pricing structure.

Interesting. Do you think just because self built softwares are getting more affordable. So that's why. That's maybe the primary reason SAS is getting dead as well.

The real challenge, I mean I want, I would Talk about software SaaS is for the business model that is over. I mean it's just a matter of time. It's over. I mean the real challenge with SaaS or with software now is software is very fixed. Even if you. For example, if I am building, let's say I have to built an internal tool for finance and I do, I do it with replit. Once I build it, it's there. I think what we need to think about is how to think about software computationally. The less you would build, the less software would be pre built, the better it would be, right? The whole pre built idea, for example bootstrapping database, performing certain functions is about giving context to the models in let the models compute in real time. The more computational it would be, the better your chances are to get closer to the reality you want to achieve. So for example, in software you want two things. Number one, you want to for example implement certain kind of workflows. And the second thing is you want to predict, you want to see the future. Elon tweeted a couple of days back, the best form of intelligence is the best, like the best way you can find intelligence, whether someone is intelligent or not is like to judge his ability to predict the future, right? So if you can tell how the next one month is going to look like, that's your intelligence. Now models can do that, but that's only possible if you design models in a way that. Sorry, software in a way. I mean if you have already written everything down and all you're doing is just probably doing some conversational UI experiences on top of it, on top of your software. That's not, for example, computational software. Computational software means like in real time, you're computing the future, right? For that to happen, you need to give a lot of context to the model. That's what we are trying to do with Hyper, right? So this was my rant on current software. Hyper is a composable wipe 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. Then use Agent API orchestration to help stitch all 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 help developers use the AI tools maximally in one unified workflow. We've built agents like business analysts to help developers groom their requirements. And then we build data model helps you transfer those requirements into a concrete data model. Then you bring all those things into Hyper AI Studio where you build tables, hook the databases, build MCP tools to hook with a platform and build custom workflows all using Hyper AI. The idea of Hyper AI 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 much better way to build software where you're not outsourcing everything to the vibes. And we're not only using card code to generate all the things. It's basically a spectrum between wibing and architecting software. It's a new plot of work that we call composable wipe code for computing the future. Right. For that to happen, you need to give a lot of context to the model. That's what we are trying to do with Hyper. Right. So this was my rant on current software.

Okay, so that's why moving on to the last news of the day is related to the financial infrastructure is adapting AI as well. What's basically happening that Stripe received approval to launch stablecoin from Trust Bank. Microsoft is investing $50 billion in South Global south infrastructure. And the companies like Palantir is relocating their headquarters strategically as well. All these moves and what do you say? Why are financial system adopting to AI so repeatedly so far?

AI blockchain. When you combine these two three things, it becomes very exciting. The current financial structure, I think that's that I'm not sure whether current banking system would adapt to this new reality or not. The chances are not very high. There are two things that's happening in terms of financial infrastructure. Number one is stablecoin or ucsd. So UCSD is basically a coin, but it's tied to a fiat currency. It's tied to a US dollar that gives its ability to, for example, to buy coffee from Starbucks and stuff like that. I mean it's not unlike other coins. So UCSD is tied to a fiat currency. That's you need to understand now people ask like if it's tied to fiat currency, what does I mean then what's the advantage of it? So yes, it's tied to a fiat currency, but real advantage of UCSD is its programmability. You can program everything in the world currently except money, right? Now when you bring money into this digital infrastructure either I mean, I mean it's okay that it's tied to a fiat currency, but now you can program the money. If I want to send you money from us to let's say Pakistan, I can program it in ways the way you want to receive it. The way the instantaneous nature of the money and all that currently takes a week or so. Right? So this is coming now for this sort of thing. I'm not sure current banking system is ready or not. I think they're not. So this, this so 20, 26 and 7. We need to watch out stablecoin anyways, Y Combinator has started to give finance and investment to startups in stablecoins as well. They can take in stablecoins so UCSD is a thing that we all need to watch out for the next two years.

Perfect. So that's it for today. Thank you everyone for watching.