I'm the luckiest guy on the planet because I didn't have to lift a finger and it's come together beautifully from the beginning. And so I have to give a round of applause to the organizing team led by Ling Ja Tang. Too many names to list because it's like 10. So yeah, Batel Panita of course, David, it goes on and on. Janica, etc., etc. Okay, if I missed anybody, oh my god, I'm thinking about you in a minute. Anyway, let's give them a round of applause. This is awesome. This is phenomenal. If I could get a clicker, that would be super cool. But I'm going to kick off the talk while, you know, hopefully we can get a clicker. So this has been a crazy journey. You know, I've been working on this with phenomenal people for six years now. And so six years of grind got us to where we are today and there will be more years of grind into the future.
And so it started with an observation, right? So there's an observation that what the world needs is better AI, right? We actually need... so 2019, I said in this TEDx talk that we need to aspire for human-level understanding. And then the very next year I started the Jeki project and the Jack language, and it was really about understanding how we can code when we reason about the code that we write having to have agentic properties like little robots or ants. Indeed, the very next year while I was working on the Jack language, of course, it started as a slow painful grind with many, many eye rolls, right? Like this is like the thing that you're not supposed to work on. You know what I mean? I wrote a paper, this was the very first paper I wrote on Jack and Joki, and I'm talking about agent-oriented programming and this was before ChatGPT was invented and it's also the last piece of me before ChatGPT was invented because there's a very embarrassing spelling mistake on this paper, data spatial, I spelled it wrong. And this paper is incredibly valuable and I recommend to everyone that you find a piece of writing that's in your voice before the AI came along because now I use this paper all the time to teach the AI how I sound. And so you guys should try to find that even if it's like a school report from high school.
Anyway, so shortly after, later that year came and that changed the world, right? It's very hard to even imagine the world before the first ChatGPT where every single word written was written by a human, right? Every single line of code written was written by a human. It's kind of unfathomable and it's only four years ago. And then just a couple years later agentic AI became a thing, right? And so what now? So that's where we are. We've got AI. What now? Why work on some of these things? Well, as I've been saying, Nvidia gave us the hardware to make AI possible. And then the researchers gave us the model. But now it's time for us, the builders, to invent how to build the software. And so that's the journey we're on today. And in this talk I'm going to talk about three things: where we are now, why the way that we build software is already dead. And you guys, you guys are in the midst of it. You may not know it yet, but it's already dead. And we'll talk about the next era of how we start building starts today, right?
Okay, so let's talk about some trends. And this is a very important trend. So, I've used this slide before to do a few things. Let me move this drink so you guys can see. Very important slide. We have agentic systems. This is what everybody's doing. Not everybody's training models. Training models is very expensive. It's unreasonably hard. It requires incredible expertise, lots of pain. We're not all building training models, but the big guys, the billion-dollar organizations are building agentic systems, but the best agentic systems are built by us, Legion. We are the legion of builders. And then as soon as one of us builds something interesting, they pluck us out of the market, right? And so it's up to us to build. And we have two generations of agentic systems, two classes: the in-loop AI, which is what we're really using a lot, and then we have out-of-loop AI. And we're building these systems ourselves because we want to control what the model's able to do.
Another important trend is the model is becoming a commodity. You're not special if you have a general AI anymore, right? It's kind of embarrassing in a way. I'm going to say this in a bombastic way. It's a little embarrassing that these billion-dollar organizations who are special because of Fable 5 and Mythos are being out-competed by smaller labs, right? The only differentiator they have is the money to train. And so Xi Jinping, he can solve the money disparity for a few labs in China. And these labs are crushing it. So, of course, we have Kimi K3 and everybody knows Kimi K3 is challenging the Fable 5s of the world and the Mythos and the GBT 5.6 Souls of the world, etc. That's an interesting trend.
And another important trend that's defining the era that we're in today is the trend towards hardware designed not for the data center but for the tokens that you want, right? Remember the data center thesis? Oh my gosh, stock market grows. Nvidia stock triple, double, quadruple because we have to build data centers. Oracle, oh my, x billion dollars to build the next data center because that's the future. But the semiconductors that are being manufactured today is to run outside of the data center. Apple really spearheaded this with their M5. That's really when this conversation got really key. But it's also fueled by the market. Everybody's buying Mac minis. The market is listening. There is an ASIC that can run a 120 billion parameter model at around 40 tokens a second that plugs into your computer like an external hard drive. So you don't even have to power it. Even though they have it in the picture, they have the power. You don't even have to power it. You plug it in like a hard drive. 120 billion parameter model, 40 tokens a second. Whenever you use Opus or Fable, let's say Opus, you're getting about 40, 50 tokens a second in practice just from the network latency. That's interesting.
What does this all say? All what this says is the market has made its call. We do not want to tap into the token drain of our wallets. We want control of the model and we want to build. There's another really important message here, too. When you look at these trends, it doesn't make you special if you have an AGI. But what does make you special is if you could build the box, right? I think we're going to move in the era of hardware dominating our top market cap companies in the world. We're already seeing this trend and Nvidia and Apple, TSMC is coming up, right? This was before SpaceX. The reason why I think the reason why SpaceX could command that valuation is because they have a vision for hardware, right? And they're buying companies to help build the hardware.
But okay, good. So, we're in this world. We've got Fable 5. Awesome. Is there a problem even to solve? Like, why work on a programming language? Okay, cool. So everybody knows this massive drama. How many people heard about this drama? All right, there was massive drama, guys. So the creator of Bun, how many people have used Bun? Knows what Bun is. All right, cool. You're going to use it today. The creator of Bun, he took Bun and he rewrote it in Rust. It was a Zig project. He rewrote it in Rust. Anthropic acquired Bun. He was sitting on Anthropic. Anthropic loves Rust. He's moving from Zig to Rust. It caused a lot of drama because he wrote a big blog article and then the creator of Zig like flamed him, called him like a beginner, what was it? Beginner energy. I was on the internet, it was like massive, like, you know, drama. But okay, so he did that and he used Mythos because he's Anthropic. So what he did is he rewrote a compiler for which there's like a million from one language Zig to a similar language Rust. So you've got super smart people. We're talking about Anthropic people helping him do this. Doing it with his colleagues at Anthropic with their model that's so amazing. It's so amazing that they have to ban it. It's like too smart AI model. Very well-defined problem. What he was doing was going file by file to rewrite each file from Zig to Rust. And he didn't even do the type safety, the memory safety stuff. Everything was unsafe. So it was like literal code one-to-one translation.
Do we have a problem? That's how much it cost. $165,000 to do almost a literal translation from one language to the next. And that's our game-changing Mythos technology. I think we do have a problem and that problem has a name. I call it learned helplessness and the polyglot, right? For 40 years, 70 years depending on how you count, our languages haven't changed. We're still JavaScript and Python. That's the code. The cycles that are executing are JavaScript and Python cycles. And then you got C++ which is kind of underrated, but I'm not going to do that fight. You got C++ and then we have variations. Rust, Zig, etc., right? Mojo, all of it. So, okay. So, we have this stagnation and it produces the polyglot. That means, what is the polyglot problem? In order to build a real application, your application has to span multiple languages with multiple boundaries, interrupts that you have to create, right? Twitter is not one language in one program. Twitter is many different programs talking to each other to give you a Twitter experience, right? It's across these languages. This is kind of the tooling that you're used to. This is the space you have to navigate when you want to build a thing that's real. That's a problem. The model has to know all of the details about all of these things and get it right to build you a thing, right? And so that's a problem for humans and it's a problem for AI. If it costs $165,000 to convert Bun to Zig with a super smart model, we've got a problem.
So that's how we've aligned the strategy and ingenuity in the creation of Jack to exist in this space where we might want to lift it a level above the rest. And so in order to have a to-do list app that's full stack and decent and average, your model has to produce all of these tokens, right? This is using SQLAlchemy, FastAPI, React, etc. The same application is realizable with Jack with just those lines of code. And that's the mission. And one interesting thing, these companies don't know it yet, but ecosystem code is what they want. These acquisitions of Bun and Astral, who does Ruff and UV, etc. They're acquiring these companies. They're slowly eating into the tool chain. They're eventually going to want to build a language and they're building models. You want to co-design these models. You want to co-design the model, the language, the tool chain. So this is a hint as to what's coming next. Imagine something like Llama where you can fine-tune your ninja model. It only has to know one language to build anything. And you can put it in a model hub that your tool chain knows about, right? That's a to be continued. It's getting cheaper to train these models. Kimi K2.5 was trained for our estimate, this could be totally wrong, 100 to 150K.
And now what I want to do, follow me on Twitter. I'm supposed to be on Twitter. I'm supposed to be doing more social stuff is what they tell me. And I'm just like, no, I just want to code. It's a distraction. But anyway, I'm going to try to do the Twitter thing so you can watch to see what's going on. I'll be doing blogs and stuff. At this point, we're kind of feature complete with Jack now. Of course, there's a lot of road map, a lot of hardening, right? It's a new language, a lot of hardening, a lot of work still left to do, but we do have this feature completeness checkbox, which is very exciting. So now I'm going to give you guys an ultimate tool for your hackathon today. With this, with what I'm about to tell you, you're going to be able to build hopefully some of the best software you've ever built in your life. Letting me know after if I'm totally wrong, but I think I'm right. So, I'm going to show you some of that and then we'll kick it off.
Right. Okay. So, I created this the last like two, three days. I created... Oh, how do I... It's not coming up on the screen. Change. Oh, I got to exit from... Got to exit from that and do this. Okay, cool. I created this thing. This is a site. It's the jackland.org site, right? But it's a little bit more than a site and let me explain. So Jack has a property we call sneakic. It needed a new name because there's nothing else like it in the world. So this application is one program that is running in the cloud. It's running in your browser and it's generating borrow-checked native code. Okay, one application. So it's native binary plus Python bytecode plus JavaScript in your browser and it's one program. So what we did is we actually took this part which is supposed to be where you like have some graphic or something that looks cool. This is actually a code viewer of the codebase. So this is actually the codebase of this site, right? And when you scroll down, it kind of scrolls with you. Of course, we made it look cool and you can do all this stuff. Oh, and then you can look at the code. Of course, you could say see the GitHub and then you'll get the actual GitHub and this is the GitHub that can unlock what you can build.
Okay, cool. So, you've seen a lot of JavaScript looking stuff, right? Well, let's see this native capability. So, here we have a video game, right? And so I'm playing this game, right? This is actually natively compiled. So one thing you can do is you can actually take this codebase and compile this to a native binary. It's under 100 kilobytes and then you can play the game in native on any platform. It's actually going through the same backend. Oh yeah, it's because I was holding it. It's going through the same backend that Rust, Zig, C++, and C uses. So when you write a program and you compile it, it's the same instructions completely borrow-checked, right? Okay, cool. And that's actually part of the same one program.
All right. And so we talk about the features. You can check this out on your own. I'm going to show you just a couple more things so you can navigate. I have this wait what section. This wait what section really shows you what's going on in this codebase by using the actual files. It pulls it from its own backend. It's pulling its own codebase, of course. You don't have to write REST APIs when you have a language like this. It's like wait what? And this is how we got that game. We just did an import and it's importing Jack. It's a Jack module. It's importing and then we just await call it and that's type-checked across that boundary. That's never been done before in any other language. Type checking across the polyglot. It's never been done. It also means you don't have to write validation code. It also means the compiler can generate the optimizations where you cache non-writing backend APIs so your front end doesn't have to do the network traffic to get it. That's a compiler problem. That's one dev's work that needs to own it. And that's one AI's work that needs to make sure it isn't messed up and break your code. Anyway, there's tons of this, right? There's like tons and tons and tons of this stuff. But you want to know what this is? I didn't create it for you to read. I created for your models to read so it can generate projects like this, right?
We also have this ninja scores section. This is fun. You can put a code base in here that has Jack code in it. So, let me just do this. Hopefully, it works. And it'll score your code. What? Okay. It'll score your code. That's the code base for this site. So, I mean, I guess... Oh. Oh, it has no community. Yeah. All right. Well, you know what I'm saying. So, you can put code in there. And it'll score it. It does. I'm going to make that like really fun. And then this doc section is kind of crazy. Because this actually uses some of the data spatial features, object spatial features, right? So, like there's walkers. The way this execution works is a little different than any other language in existence because it has this way of having your compute, your functions can go to your data on a graph instead of your functions there and you have to pass parameters to it. So it inverts the model. It's an option. You don't have to use it, but it's one of the cool language features. And so the documentation here is essentially this graph.
And I'll kind of... So this is actually the docs when it loads. So this is the documentation. So each node, by the way, you guys all have this. This is open source, right? It's the code base is 100% Jack. And so you can see it's 94% Jack. Correction, because you know you have to have CSS. And so you guys can see how that was built. Now, another thing that you guys are going to get an update on is we have this jackhammer.ai. Of course, if you have a new language, people are going to want to host what they build. And even though hosting is super automated, if you host just from the CLI, you're going to have to know a little bit about getting a cluster up. So, we have this platform that'll help you host. It also has a lovable capability where you can build. Everyone gets free tokens when you sign up. So you could use the free tokens, build UIs, whatever, build full stack apps and enjoy it. But really the deployment is really clutch and this is going to be relevant to what you guys build. So there's a deploy button. You guys are going to have sandbox. This is free forever for anyone on the platform. So you can deploy and you can share it with the world, right? Anything you build, you can share it with the world, deploy it instantly available to everyone. And so we want you guys as part of your submissions to deploy and share it with us, right? So that'll be the instructions in your sheet. And also tinker with the building capability, have some fun. It's a good way to start, get a little prototype together, and then you can move into building in the way that you feel most comfortable.
Okay, last thing before I wrap. I promise I'm going to wrap here. Last thing before I wrap, I want to give you the top secret unlock. There's like one prompt you need after you install. So this is how you can install the site, right? So that you can run it locally and see it. Check out this book. It's got all the details. Free book, dude. All the details. But this is your key prompt. That's all you need to get started with Jack. Install the one binary. Everything's one binary. You can delete anything else on your system if you wanted to. You don't have to. And then you can build all of this with one binary. You can run all of it with one binary and that's the only prompt you need to get started, right? And a pro tip is you could also point it to the repo of this actual open source project and then your AI is going to be able to help you and teach you how to use it. And with that, I want you guys to have fun. I want you guys to build and I want you guys to do your best work today. And with that, I'm happy to take a couple questions.
