So, anyway, so everyone knows about the ChatGPT moment 2022, right? This is actually a very interesting moment because for the first time we saw almost every industry change. Every industry now is re-understanding themselves as an AI effort.
And this is the first time we used ChatGPT. All of the research in the market evolved towards working on this problem. And the fundamental technology that enabled it was the intermingling of reinforcement learning for the first time with large language models.
And then we got our first really important buzzword. This is the first time everyone knows this term, right? Prompt engineering. It changes the way we build software. It changes the way we solve problems in our lives.
And then there was another important buzz term, and this is the buzz term of this year. It's agentic AI. So, if you look at what the key thesis of Andreessen Horowitz, a16z, if you look at what Anthropic is doubling and tripling down on with co-work, with Claude code, making that private, it's a bet on this type of technology defining how we work. It's starting with software engineering, but it's bleeding into other types of knowledge work for the first time in this year.
It's causing a massive disruption. There was a $280 billion sell-off of software stock. Companies like Salesforce, companies that do SaaS software are in panic right now. And it's because the thesis is changing. The thesis is becoming that everyone will be building their own software and not buying a SaaS solution. And it's because of this particular trend. So, that's where we are. But let's talk about these three trends that I promised is going to redefine next year and everything that we know.
Let's look at the first trend. Open Claw. So, right now in China, grandmothers are building their own Open Claws. Right now, everyone is calling each other, 'How do I create an agent to run autonomously for days on end solving problems and operating, taking actions on the behalf of the human while being able to be productive otherwise,' right? And so, this is the fastest growing repo in the history of GitHub. It has more stars than Linux. It has more stars than PyTorch.
And it was bought within 3 months by OpenAI. And now, Meta bought Open Claw, Claw Claw, what was it called? Multibook, which is a social media network where these agents are talking to each other. It's actually causing a massive disruption. But, what's really interesting is what the market is excited about isn't connecting to a cloud to run Open Claw. They're running it on local hardware, the cheapest hardware they can find. The most exciting thing is spending $20 to have a platform that you can run an agentic AI on today.
And at the same time, the Mac Mini sales went through the roof from Apple. Okay, that's interesting. We're now outsourcing our work in new and interesting ways. What's the other trend?
This one you don't hear much about. And I'm going to say an audacious thing. I think it's just a matter of time before Anthropic and OpenAI become irrelevant. Now, why don't you hear about these models? So, these are three models. So, there's Kimik 2.5, everyone knows about Kodaks, Claude, etc. There's Kimik 2.5, there's a model called GLM 5, and there's another model called MiniMax M2.5. Sounds very cryptic.
But, the interesting thing about these models is that they're open weight. That means that they're free, and you can download them and run them yourself. You don't need to connect to a cloud. And what these graphs show is that they're competing and beating Anthropic's Claude. And Anthropic's Opus 4.5, Opus 4.6. They're beating GPT 3 5.3 at many of the tasks, and they're free. The only problem is it costs about $100,000 of hardware, of GPU hardware, to run them at that level. But, the model is free.
There was a year ago I gave a similar talk, and what we saw was 'Oh, the open weight models are catching up.' But, there was a massive gap. That gap has closed in just a year. But, it's about $100,000 of GPUs to run it. The Nvidia thesis, right? But, then something else happened just a couple weeks ago. I don't know if you guys are paying attention to Apple's M5. Apple's key thesis and their mission is to take AI out of the cloud. That's what they're quoted saying.
Now, if you look at, if you want to predict what's going to happen, look at the economic interests. Apple has no cloud story, right? They have no big model. They sell boxes, and they sell laptops. And so, their play is to allow you to run these models at home. Right? And it's actually already here with the M5.
And there's a Mac Studio coming at the end of the year with an M5 that can run these models at 100 tokens a second. That's my prediction. Right? Today, if you spend $8,000, you can run it at 26 tokens a second on your desktop, you're going to get 100 tokens a second. At that moment, there's no reason to connect to any API, and you have the model locally at home, which everybody wants. They want control of the model. So, everything is going to change again. Right? So, I think that's an interesting question for us to look at, and I want to talk about our signal, what we're working on.
The problem is the popular languages Python and JavaScript, they were invented 30, 40 years ago. They haven't changed. They haven't evolved. But the way we write software and the software themselves are wildly different. Indeed, if you want to launch anything to scale, we have a polyglot problem. You have to front end in JavaScript, TypeScript, back end in Python. You might need some fast code to run using Rust. You have to have REST APIs, SQL for your database. Complexity is insane. You need a team of 30 people in order to launch a simple application.
And it's also a problem for AI, because if you want an AI to build a Twitter, it has to get all of that right, from the YAML files of deployment to your database organization to the code itself. And so, what we're working on is one language created for the era of AI. Created for us humans and also for the AI to perform.
And so, here we have on the left the most popular stacks. This is the savviest to-do list app that you or an AI would have to touch. This is the code that you or an AI would have to touch if you use React with TypeScript, FastAPI, SQL Alchemy for your ORM, etc. If you don't know any of those terms, that's okay because our language treats that as middle, as IR, as an intermediate representation. So, the same application on the left is written on the right with Jack and Jill Saki.
Additionally, if you want to create a cursor, you have to write hundreds of lines of prompts today. So, on the left it's the best you can do with today's technology and on the right we have a key piece of technology that gets better performance for your AI and you have more rails and less complexity and we published that work. And that work is in this Jack and Jill Saki stack.
I won't get into the details to how this works, but an AI was able to create the levels at the bottom without any prompts. Cutting-edge piece of compiler technology, so I encourage everyone to engage this technology as you build. We have folks here that can help you on the journey and we'll interact with you and engage with you. We have a number of special prizes including free subscriptions. We really believe we need to democratize access to models. It's one of the biggest barriers to entry right now is you have to have a key to build something cutting edge. And so, our prizes will be access to models. And with that, thank you guys so much for having me.
