How Andrej Karpathy ACTUALLY Uses AI (10x Faster)

0:00Andres Karpathy is one of the best AI researchers humanity has ever seen. He co-founded OpenAI and worked as the AI director at Tesla. Yet, he still uses the same chatbots and tools as non-technical people, which raises an interesting question. If he has access to the same technology as everyone else, what actually makes him so good at it? So, in this video, I want to show you the truth about that and reveal his seven little-known tricks. And I'm also going to help you work out how to properly use the last one to get compounding results 10 times faster than you normally would. So, the first trick is the one nobody expects from a guy at

0:30his level because it's almost embarrassingly simple. You see, most people sit on their computer trying to write the perfect prompt word by word, but Karpathy does the complete opposite. He barely types any words, and the few he does write are nowhere near a perfect prompt. So, what he does instead is use a speech-to-text [music] tool, which basically gives him a messy dump of everything in his head. This might look lazy, but it's actually smart for one simple reason. You talk about three times faster than you type. Now, you might be wondering if that hurts the quality of the AI results you get compared to perfect prompting. Well, it

0:59actually doesn't because AI has gotten so good nowadays that it can structure and pull all the details it needs from your whole description. That's why Karpathy says all of his engineer co-workers are doing the same thing. So, basically, everyone is whispering at their agents all day. Now, the easiest way to do the same is to use the built-in dictation features inside the chatbots. For example, here in Claude, we have this small microphone button, but I find it quite annoying for two reasons. First, it misspells some words or doesn't pick them up properly. And second, it usually stops listening after about 30 seconds, which isn't great when

1:30you've got a long description to get through. That's why I'm using Super Whisper, a third-party app that runs on your computer so you can use it everywhere and for as long as you want. And you can actually use it completely for free, even if the paid subscription is just a couple of bucks. So, with this first trick, you'll save at [music] least 10 hours every week. And to make things even easier for you, I made an entire document with all the tricks from

1:50this video so you can save it for later. It's completely free, and you can find it in the description down below. Now, using this speech-to-text trick only gets you so far because every answer you get from AI comes from its memory, and that's exactly where most people go wrong. You see, AI has become extremely powerful over the last 3 years, but it can still make mistakes. And the worst part is that it can't tell when it's making them. So, instead of warning you, it just tries to convince you that whatever it's saying is true. But even though the fix for this is almost

2:16obvious, almost nobody actually does it. Karpathy calls this approach the LLM Wiki, and the whole idea is to create a small library of your own material that you want the AI to answer from. Because if you don't, it just goes across the whole internet to find an answer. And depending on how complex your question is, it can hand you some pretty generic stuff. So, to solve this, Karpathy always uploads PDFs, documents, or whatever data the AI needs to give a proper answer. Inside Claude, all you have to do is press the plus button, then upload from computer, and it opens up all your files. So, let's drop in

2:46this report and give it this prompt. As you can see, it starts answering based on the document I gave it. And this is still a pretty basic task, but imagine what you can do with it on a bigger project, like a full report or a big piece of research. And that's exactly what Karpathy does for literally anything. But when it comes to research especially, he revealed in one podcast that there's a small trick that completely changes how he works with documents. Now, how many times have you spent hours reading a study or a summary that made you feel like you'd actually learn something? But then the moment you

3:13go to apply it, you realize you don't really know that much. Well, it happens to me all the time. So, the trick Karpathy uses is to completely change the way we interact with documents with the help of AI. Instead of just reading them, we can start interrogating them, which forces us to actually think. And in my opinion, the best tool for this is Notebook LM. To create a new project,

3:31you just press create new notebook. Straight away, you get a pop-up where you can drop your files to build that same little LLM Wiki we saw in the previous trick. Once they load, you can simply ask questions inside the tab, and it answers based on your documents. And studies have shown that asking questions and getting answers is one of the most

3:48efficient ways to learn in the the run. And what's really cool inside Notebook LM is that you get some genuinely useful features, too. There's the mind map, which builds a fully structured plan from everything you imported, and the flash cards, which help you learn even better, especially if you're a student. But all of this has been about words, and that's only half of how he uses AI, because the other half has nothing to do with text at all, which brings us to the fourth trick Karpathy uses. And honestly, this is his favorite way to use AI, because it gives him a big advantage over normal people. But first,

4:19let me ask you something. When you want to find a specific answer to a problem, how do you usually do it? I used to just ask AI about it, but I don't do that anymore, since I learned from Karpathy that the best way to get answers is to think like a data analyst. For example, he revealed that he used this exact trick to understand which jobs are most

4:36exposed to AI and what's coming next. But instead of relying on hyped-up news, he simply told the AI to go and pull the jobs data straight from the Bureau of Labor Statistics. Then he waited for it to come back with a full spreadsheet of data, and he used that to get extremely accurate information on what's actually [music] happening in the world, instead of listening to other people's biased opinions. So the secret here is to think like a data analyst and turn anything

4:58into numbers and charts that don't lie. But even with all this, there's still one really annoying thing about AI, and it's quietly wasting a huge amount of your time. You might not consciously realize it, but you've probably had this happen loads of times. You open a new chat and you feel annoyed that you have to re-explain the entire context to the AI all over again. But the simple solution for this is to use the same LLM Wiki feature Karpathy made to give the AI a permanent memory. But this one's a bit different from the trick we saw earlier. This time, it's one single system where the AI adds files and

5:28documents based on the conversations you have with it, and it updates live. Now, the setup is honestly way simpler than it sounds, because it's just a folder full of normal text files on your computer, which is basically how Claude's projects work. [music] So let's get back to Claude, and on the left side, you can see two options. Let's press on code, and And get the normal prompt box, but besides that, you also have this feature where you can choose the folder to work into. If you press on it, you can either create a new one or choose one from your computer. If this is your first time doing this, just

5:55create a new one, and then you can start chatting with Claude and tell it how to organize your own LLM Wiki. So, I'll paste this prompt, and now let's press generate. As you can see, it also creates the claude.md file, which is basically a set of global instructions the AI reads every time you open a new chat. After that, all you do is tell Claude to read everything inside the folder, and anytime you want to add something new to its memory, you just tell it like this. And the best part is, you don't need any complicated or expensive software, just a normal Claude subscription. Now, once the AI knows

6:22this much about your work, using it just for tiny little tasks starts to feel like a real waste of time. A big mistake I see almost everyone make with AI is treating it like a search engine that can only solve small problems. They ask it to generate one email, one YouTube title, or maybe quality check a piece of code, which helps, of course, but it's like using only 5% of what AI can actually do. That's why Karpathy uses AI in a completely different way, especially in his coding work. He said he went from writing most of his code by hand to letting AI agents handle the big

6:51chunks for him. So, instead of asking for one tiny thing, he gives it big tasks like creating entire features or doing full research on a topic. And you can do the same thing outside of coding, too. For example, instead of asking it to write a title for this video, I'll give it the whole task at once using this prompt. Now, the difference is actually huge because the AI is thinking through the entire project instead of just filling in one small blank. So, the trick is to delegate bigger and bigger tasks to [music] it, which also brings us to the seventh and most powerful trick, the one I promised you right at

7:21the start. So, the idea is simple. If AI can handle one big task on its own, why not run several of them at the same time? Karpathy actually talked about creating an entire team of Claude agents at once, each one working on something different. [music] And there are two ways you can do this. The first one is for beginners who don't want to overcomplicate things. All you do is open a few separate chats and give each one its own job. And if you've used the previous tricks from this video, they'll all already have the context of your project, which means you don't waste any time explaining [music] to each one what

7:49you're working on. Then, for the more advanced option, you can use Claude's skills to basically create an orchestration. [music] One AI agent manages multiple sub-agents, and each one has its own task. Like I said, this one is a bit more advanced, but the results you get with it are way, way better. So, I might even make another video on this topic alone. But, until then, I'll post it on our free school community, where I share everything I learn about AI. You can join using the link from the description below. Thanks for watching, and I'll see you in the

8:14next one.