I Turned Karpathy’s Second Brain Into an AI Operating System

• Channel: AI Impact
• Video: Watch on YouTube
• Tags: karpathy second brain ai os system

Overview

AI Impact covers Andrej Karpathy’s second brain / auto-research system — how it combines knowledge graphs, vibe coding, and agentic architectures into an AI operating system for your business.

Key Takeaways

• So this week, the big news was that andre carpathi, the ai godfather, the guy who created the term vibe coding, who was pushing forward knowledge graphs, and also created auto research system. and all these things have gone super
• And claude are starting to realize is that a lot of the power of ai is actually not hidden in the models. it’s actually hidden in the layer on top. so what all these ai companies have realized is that
• On top of the ai. and the goal thing is this can be free. you know, this can be something that maybe you work with a claude max subscription, maybe you work with other tools. but this isn’t something that
• Everything is changing. every week there’s new news, there’s new tools, there’s new models, everything else. but i do think that the ai os operating system is likely it is starting to solidify. and i think that that’s something that’s
• This is where you want to focus. you don’t want to try and learn some random tool, even though i spend a lot of time with claude code and codex and things like that. i think they’re super valid. but
• You can break the data that you need to give to llms into two big categories. one of which is quantitative data. and this is like sql stuff, business intelligence, such as numbers. and so that’s pretty straightforward. there are
• Strategies that you’ve written or brand guidelines or even like visuals or images or different things like that. and that i think is becoming more and more clear exactly how qualitative data needs to be stored for ai. and there
• Go into what is a knowledge graph. so it’s been a lot of times trying to explain knowledge graphs to people. and i think i’ve found the best way to explain it. so when you’re on your computer, most people
• The seasonal promotion library. and then you’re having to decide where to store like this file. maybe. but it’s a draft. it’s a draft that you have. and so what most people are used to is you have to just
• May tie to multiple different things, but you just have to find some hierarchy that makes sense. and it just forces that situation. it’s what we would call like in data like a one-to-many relationship. like there’s one parent folder
• Tool email. it ties to the workspace, marketing department. it’s a part of our seasonal offer. and also this ties to our company voice. so you then can just have the file. and then it can have six references. and
• And see connections that we can’t even see as humans. but on the traditional file path, they get very limited. and so if you’re trying, even if you give it access to the files, it just spends way more time

Transcript

So this week, the big news was that Andre Carpathi, the AI Godfather, the guy who created the term vibe coding, who was pushing forward knowledge graphs, and also created auto research system. And all these things have gone super viral and have really been like the trend setting for a lot of the industry. He’s now joined Claude. So what does this mean and why does it matter for you? I think what this means is that what AI and Claude are starting to realize is that a lot of the power of AI is actually not hidden in the models. It’s actually hidden in the layer on top. So what all these AI companies have realized is that the model is just not enough to really bring true value to people’s lives. And what Andre Carpathi and what everybody else that I’m seeing is found is that you really need like an operating system that’s customized to you on top of the AI. And the goal thing is this can be free. You know, this can be something that maybe you work with a Claude Max subscription, maybe you work with other tools. But this isn’t something that you have to pay a whole monthly subscription for. It’s really something that you use to really define and structure the way that AI is getting information to you. And the goal thing too is that I think that there everything is changing. Every week there’s new news, there’s new tools, there’s new models, everything else. But I do think that the AI OS operating system is likely it is starting to solidify. And I think that that’s something that’s very important when you’re working in like a fast-moving field like this. If you can figure out what is not going to change, that’s really what you want to invest in and focus on. So I think that with AI, this is where you want to focus. You don’t want to try and learn some random tool, even though I spend a lot of time with Claude Code and Codex and things like that. I think they’re super valid. But you really want to understand how you can build your own operating system on top of that. And something that I’ve sort of noticed working as like an AI architect for the clients that we work with, I feel like you can break the data that you need to give to LLMs into two big categories. One of which is quantitative data. And this is like SQL stuff, business intelligence, such as numbers. And so that’s pretty straightforward. There are some unique things that with my analytics company we do differently for AI. And I’ll show that in the future video. But what I want to talk about today is qualitative data. So this is like written content or like strategies that you’ve written or brand guidelines or even like visuals or images or different things like that. And that I think is becoming more and more clear exactly how qualitative data needs to be stored for AI. And there are a couple other sections as well. Being able to store the runtime states and also having canonical knowledge, but really the quantitative and the qualitative is probably what you should spend most of your time on. So let me go into what is a knowledge graph. So it’s been a lot of times trying to explain knowledge graphs to people. And I think I’ve found the best way to explain it. So when you’re on your computer, most people are used to a traditional file system with a hierarchy. And so let’s say you have a Memorial Day, Clavio email that needs to be sent on May of 2026. And it’s tied to the marketing department. It’s tied to the seasonal promotion library. And then you’re having to decide where to store like this file. Maybe. But it’s a draft. It’s a draft that you have. And so what most people are used to is you have to just pick something. Like you might have a file system that’s like all right, we have the company folder and then the marketing department. And then maybe there’s a a Clavio folder. And you just have to pick, I mean it may tie to multiple different things, but you just have to find some hierarchy that makes sense. And it just forces that situation. It’s what we would call like in data like a one-to-many relationship. Like there’s one parent folder to many subfolders. But when you have a knowledge graph, all that breaks. You can actually just say here is my Clavio draft. And by the way, this ties to the project of a Memorial Day. It ties to the tool email. It ties to the workspace, marketing department. It’s a part of our seasonal offer. And also this ties to our company voice. So you then can just have the file. And then it can have six references. And it’s more like a much more like flattened, broad, like known brain. Instead of like a traditional file hierarchy. And I’ll just tell you this, AI’s love this kind of data. They can like fly through and analyze them all and see connections that we can’t even see as humans. But on the traditional file path, they get very limited. And so if you’re trying, even if you give it access to the files, it just spends way more time like going through finding things, it’s harder for define connections and do its own self-expiration across files. So these knowledge graphs are going to be really, really key for AI. And to be honest, it doesn’t matter if you use OpenAI, it doesn’t matter if you use Cloud, it doesn’t matter if three years from now they create a super AI, like the AI is going to need knowledge. And it’s going to need your company’s knowledge, your personal knowledge. That’s going to be something. So I think this is something that you definitely want to spend some time focusing on starting to build these knowledge graphs now. But a recent video I did is I actually have not liked a lot of the second brain type content. I went super into that a few years ago. But too much of the second brain type content was focused for human readers. But I’m never going to open my knowledge graph. I’m going to have an AI open it and explain things to me and write details for me. To be honest, it really sucks copying all this and doing manually. So it’s like one of the perfect things for AI to do for you. You can actually build the brain, it can track things you can ask with make changes. But because it’s going to be AI, I’ve actually found that there’s several things that you should change with this. And I released a video if you’d like to see on my infinite brain system. And I’m planning to open source it in the next few weeks. So subscribe to the channel or look at our other videos to learn more about the infinite brain system. But the key things you want to do is you definitely want to like the most important part I would say is indexing. And with indexing, you want to be able to have like maybe there’s 5,000 nodes you have on a specific topic. You want to be able to have like an index that’s like, hey, here’s all of our marketing images tied to this product. And here’s all of our marketing and products tied to this. And have like a short summary of each piece. So the AI can quickly read the indexes and know like, oh, here’s where I can go to find other details because the AI could read maybe 500 notes before it would have to actually come back to you with an answer. So you ask a question, it goes to the knowledge graph. It’s too much sport to just fit all into one prop. So it reads across and finds relevant things to inject its knowledge into its knowledge. Maybe it will read like your brand pillars, your brand strategy, different things like that, or your images, your products, or previous marketing campaigns that work. Maybe even tie it into some data to show what works well. So you could basically have all of that to where the AI could research all that in a few seconds and then come back to you with the perfect answer. But that’s the dream and the vision of knowledge graphs. One thing that I added to my infinite brain system that I’m seeing a lot of real missing is that I am adding these edge types and few other things. So definitely if you’re interested, go check out the video and I’ll walk you and it walks you through more about the infinite brain system. And what I used to have is I used to have a system where I would have like a folder, honestly like the whole hierarchical folders that have like agents that have like skills, I’d have workflows. And within that I would store each of those pieces. But now what I’ve done is instead of having it in that hierarchy, I’ve started just making it to where everything is an infinite brain. Like the more I thought about it, the more the infinite brain just ate everything except for quantitative data, which I’ll talk about in the future video with to do with that. But pretty much the whole system has just become an infinite brain type system. And so now I have my own I’m finalizing an operating system that I’m going to open source for free and we’re going to talk about in our school community as well. But it basically uses mainly just free tools to free or cheap tools to interact with this operating system. So you can continue to build out this AR architecture for you. And then at that point you’re like agnostic. It doesn’t matter if clog becomes way better or if rock comes into the race or google suddenly pulls ahead. You can just take your operating system and move it between the models. And if there’s other really cool tools, they’re probably going to integrate with systems very similar to this kind of architecture anyway. And so now in my infinite brain instead of my hierarchical folders I used to have, I now have entity types that could be like an agent. So you might define this as, you know, with a group. Persona capability, I am making it where all my agency with paperclip. You also can have skills, you could have workflows, you could have rules, tools, knowledge. So this is like just knowledge with a knowledge. Like if you want a whole knowledge base just tied to one particular topic, you could have that. A different project, output. So anytime it creates something, my AI, I always save it. But what’s cool about this system is that you can store all this in Markdown. And Markdown is basically just English. Like maybe it’s like a little bit overwhelming to read, but you could read all this. But you don’t need to learn any crazy coding languages or anything else like that. You just need to know English and how all these components can work together. So all of this can then be stored in a way that’s compatible with infinite brain. And what’s cool to me is that you can store all of this in a folder. And what I’m starting to do with companies that I work with is we actually will give every single one of their team members one of these types of folders. We’ll train them on it because it does take a little bit of training on how

(Transcript truncated — full length available on YouTube)