← Back
Nate Herk October 5, 2026 10m

I Built Another Andrej Karpathy Using Claude

Read full transcript 10 segments
  1. So, this is not the Claude So, this is not the Claude you know. I you know. I you know. I just turned just turned just turned one of the one of the one of the greatest minds in greatest minds in greatest minds in artificial artificial artificial intelligence into Agent intelligence into Agent intelligence into Agent Claude. Andriy Claude. Andriy Claude. Andriy Karpatiy is one of the Karpatiy is one of the Karpatiy is one of the lead engineers at lead engineers at lead engineers at Anthropic, a co-founder of Anthropic, a co-founder of Anthropic, a co-founder of OpenAI, and the man who, as OpenAI, and the man who, as OpenAI, and the man who, as everyone says, explains everyone says, explains everyone says, explains AI better than anyone AI better than anyone AI better than anyone in the world. And today I'm going to in the world. And today I'm going to in the world. And today I'm going to show you the exact show you the exact show you the exact four steps to four steps to four steps to turn his turn his turn his thoughts into your thoughts into your thoughts into your own Agent own Agent own Agent Claude. So let's Claude. So let's Claude. So let's look at this. Okay look at this. Okay , quick, who is this , quick, who is this , quick, who is this guy? He was guy? He was guy? He was one of the founders of one of the founders of one of the founders of OpenAI back in 2015. OpenAI back in 2015. OpenAI back in 2015. He then spent about He then spent about He then spent about five years leading five years leading five years leading AI at Tesla, leading the AI at Tesla, leading the computer vision team behind computer vision team behind Autopilot. Autopilot. Autopilot. He then left to He then left to He then left to pursue education pursue education pursue education full-time full-time full-time and founded Eureka Labs. And and founded Eureka Labs. And and founded Eureka Labs. And as of May of as of May of as of May of this year, he works this year, he works this year, he works at Anthropic on the at Anthropic on the at Anthropic on the pre- pre- pre- training team, which trains training team, which trains training team, which trains Claude's models before they are Claude's models before they are Claude's models before they are released to the public. released to the public. released to the public. And besides, he's And besides, he's And besides, he's released so many released so many released so many different YouTube videos different YouTube videos different YouTube videos and courses on and courses on and courses on creating creating creating with AI. So what with AI. So what with AI. So what we are building we are building we are building today is not a today is not a today is not a voice clone of voice clone of voice clone of Andriy Karpaty. This is Andriy Karpaty. This is Andriy Karpaty. This is not his impression. This is an not his impression. This is an not his impression. This is an agent that compresses the way agent that compresses the way agent that compresses the way the best the best the best AI teacher AI teacher AI teacher , Andriy , Andriy , Andriy Karpatiy, explains Karpatiy, explains Karpatiy, explains things, and it teaches you the same way things, and it teaches you the same way things, and it teaches you the same way . So, he . So, he . So, he follows seven follows seven follows seven rules, all of which rules, all of which rules, all of which come from his come from his come from his own writings, blogs, own writings, blogs, own writings, blogs, and lectures, such as and lectures, such as and lectures, such as building the building the building the smallest version first, smallest version first, smallest version first, predicting what predicting what predicting what will happen before you will happen before you will happen before you run, run, run, showing a broken showing a broken showing a broken version, and never version, and never version, and never handing over code that handing over code that handing over code that you haven't run yourself.

  2. you haven't run yourself. you haven't run yourself. So, it's more than So, it's more than So, it's more than just a human clone. just a human clone. just a human clone. It's more about It's more about It's more about cloning the way cloning the way cloning the way her brain thinks. her brain thinks. her brain thinks. So, I guess you guys are wondering So, I guess you guys are wondering So, I guess you guys are wondering now now now , why , why , why even even even create this? Let's create this? Let's create this? Let's think about what think about what think about what happens when you happens when you happens when you ask Claude ask Claude ask Claude to explain something to explain something to explain something you don't understand. This you don't understand. This you don't understand. This may be over- may be over- may be over- explaining. It might explaining. It might explaining. It might make something up along the make something up along the make something up along the way. And you way. And you way. And you actually just actually just actually just feel like you have feel like you have feel like you have n't learned anything and n't learned anything and n't learned anything and feel even feel even feel even more overwhelmed more overwhelmed . And if you're building . And if you're building . And if you're building something for a client, that's how something for a client, that's how something for a client, that's how you end up with you end up with you end up with code that you can't code that you can't code that you can't fix or fix or fix or explain when it explain when it explain when it breaks. And Andriy breaks. And Andriy breaks. And Andriy Karpatiy faced Karpatiy faced Karpatiy faced exactly the same exactly the same exactly the same problem. Back in problem. Back in problem. Back in August of last year, August of last year, August of last year, he publicly he publicly he publicly posted that he posted that he posted that he was trying to get was trying to get was trying to get Claude Code to teach Claude Code to teach Claude Code to teach him, at the same time as the him, at the same time as the him, at the same time as the code he was writing; code he was writing; code he was writing; and according to him, it didn't and according to him, it didn't work at all because work at all because he actually he actually he actually wants to write code more than wants to write code more than wants to write code more than explain things as he goes along explain things as he goes along explain things as he goes along . So, . So, . So, creating this creating this creating this agent here will fix agent here will fix agent here will fix AI's worst habit, AI's worst habit, AI's worst habit, which is that which is that which is that it sounds it sounds it sounds right but isn't right but isn't right but isn't right. And the bigger right. And the bigger right. And the bigger reason, frankly, reason, frankly, is is is this: the quality of what this: the quality of what this: the quality of what you create you create you create with AI with AI with AI depends on the expertise depends on the expertise depends on the expertise that's in the system, and you're that's in the system, and you're that's in the system, and you're not going to be an expert at not going to be an expert at not going to be an expert at everything. You know, I'm not a everything. You know, I'm not a everything. You know, I'm not a machine machine machine learning researcher. I'm not a learning researcher. I'm not a learning researcher. I'm not a machine machine machine learning expert, but I learning expert, but I learning expert, but I can certainly can certainly can certainly ask agents to ask agents to ask agents to pull out everything that a pull out everything that a pull out everything that a real real real expert said, compile it into expert said, compile it into expert said, compile it into something that Claude something that Claude something that Claude can actually can actually can actually use, and use, and use, and then I can just then I can just then I can just talk to it until I talk to it until I talk to it until I understand. And then, understand. And then, understand. And then, when I understand a little better when I understand a little better when I understand a little better , I already have , I already have , I already have all this knowledge in my all this knowledge in my all this knowledge in my project, and I can project, and I can project, and I can integrate this integrate this integrate this knowledge into my own knowledge into my own knowledge into my own skills, my own skills, my own skills, my own agents, and my own agents, and my own agents, and my own workflows, which is workflows, which is workflows, which is exactly what we are doing exactly what we are doing exactly what we are doing today with Karpathy. And what's today with Karpathy. And what's today with Karpathy. And what's really cool is that really cool is that really cool is that once you once you once you do it with Karpathy, you do it with Karpathy, you do it with Karpathy, you can do it can do it can do it with practically with practically with practically anyone. You can anyone. You can anyone. You can do this with your do this with your do this with your favorite, you know, favorite, you know, favorite, you know, sales teacher, sales teacher, sales teacher, or an author you or an author you or an author you really like, or a really like, or a really like, or a coach you're coach you're coach you're already paying for. And it's already paying for. And it's already paying for. And it's essentially the same four essentially the same four essentially the same four steps, just like how we steps, just like how we steps, just like how we create this kind of create this kind of create this kind of second brain.

  3. second brain. second brain. Now really quickly, Now really quickly, Now really quickly, before we get before we get before we get to the build, I have to the build, I have to the build, I have this completely this completely this completely free free free AI OS kit for AI OS kit for AI OS kit for you guys that will you guys that will you guys that will help you help you help you build and build and build and scale your scale your scale your operating systems. operating systems. operating systems. If you combine that If you combine that If you combine that setup with setup with setup with things like Karpathy's brain, things like Karpathy's brain, things like Karpathy's brain, it's just a it's just a it's just a huge huge huge discovery, and that's how discovery, and that's how discovery, and that's how I can move I can move I can move so fast. So so fast. So so fast. So the link to that will be at the link to that will be at the link to that will be at the bottom in the description, but the bottom in the description, but the bottom in the description, but let's get to the let's get to the let's get to the build. Okay, so I'm build. Okay, so I'm build. Okay, so I'm going to do this going to do this going to do this inside of Claude Code, and there inside of Claude Code, and there inside of Claude Code, and there will be six will be six will be six prompts here to run, prompts here to run, prompts here to run, and I'll show them all on the and I'll show them all on the and I'll show them all on the screen so you can screen so you can screen so you can just take a just take a just take a screenshot and paste screenshot and paste screenshot and paste them into your own Claude them into your own Claude them into your own Claude Code. So, first Code. So, first Code. So, first Claude needs everything Claude needs everything Claude needs everything that's in Andre that's in Andre that's in Andre Karpati's head. Things like Karpati's head. Things like Karpati's head. Things like his blogs, his blogs, his blogs, lecture transcripts, lecture transcripts, lecture transcripts, GitHub repositories, his GitHub repositories, his GitHub repositories, his posts on X–all of these posts on X–all of these posts on X–all of these will be practically will be practically will be practically free and free and free and public. But Claude public. But Claude public. But Claude Kod may not be Kod may not be Kod may not be able to just take able to just take able to just take all of this for free, all of this for free, all of this for free, as it may be as it may be as it may be behind some kind of paid behind some kind of paid behind some kind of paid access or access or access or software barrier. software barrier. software barrier. So here's what I did for YouTube So here's what I did for YouTube So here's what I did for YouTube : I asked it : I asked it : I asked it to extract subtitles to extract subtitles to extract subtitles using two using two using two free free free Python packages: YouTube Transcript API and Python packages: YouTube Transcript API and Python packages: YouTube Transcript API and YTDLP. And for X, I YTDLP. And for X, I YTDLP. And for X, I use an API use an API use an API called twitterapi.io, which is a called twitterapi.io, which is a called twitterapi.io, which is a paid API, but I paid API, but I paid API, but I pulled almost pulled almost pulled almost every post from it since every post from it since every post from it since 2023, and it cost 2023, and it cost 2023, and it cost about a dollar. This is about a dollar. This is about a dollar. This is not a very not a very not a very expensive API at all. Here is expensive API at all. Here is expensive API at all. Here is a tip for a tip for a tip for scanning. It's scanning. It's scanning. It's basically basically basically about where to get everything, about where to get everything, about where to get everything, what tool what tool what tool to use for to use for to use for each source, and I each source, and I each source, and I told it told it told it to run one agent to run one agent to run one agent per source so per source so per source so they all run at the they all run at the they all run at the same time and in same time and in same time and in parallel so it doesn't parallel so it doesn't parallel so it doesn't take forever. I take forever. I take forever. I ordered him ordered him ordered him to write down the to write down the to write down the material he received so you material he received so you material he received so you could check it.

  4. could check it. could check it. And it took me And it took me And it took me 45 minutes to an 45 minutes to an 45 minutes to an hour to get all these hour to get all these hour to get all these agents up and running agents up and running agents up and running and everything put together. As and everything put together. As and everything put together. As a result, we ended up with a result, we ended up with a result, we ended up with one one one raw folder with raw folder with raw folder with one subfolder for one subfolder for one subfolder for each source and each source and each source and over 700,000 words over 700,000 words over 700,000 words directly from directly from directly from André Karpati himself. André Karpati himself. André Karpati himself. You could just You could just You could just stop here and stop here and stop here and point Claude to this point Claude to this point Claude to this folder, but all these, folder, but all these, folder, but all these, you know, jumbled you know, jumbled you know, jumbled words, they don't quite words, they don't quite words, they don't quite fit into fit into fit into Claude's head in a way that they Claude's head in a way that they Claude's head in a way that they can filter them well can filter them well can filter them well , do , do , do they? It will be too they? It will be too they? It will be too messy. It's like messy. It's like messy. It's like looking for a needle in a haystack looking for a needle in a haystack looking for a needle in a haystack . So, we will do what . So, we will do what . So, we will do what Karpati advised Karpati advised Karpati advised everyone to do back in everyone to do back in everyone to do back in April. He wrote about April. He wrote about April. He wrote about how his how his how his law degree compiles his law degree compiles his law degree compiles his raw sources raw sources raw sources into a wiki. The raw into a wiki. The raw into a wiki. The raw files don't change, files don't change, files don't change, but LLM actually but LLM actually but LLM actually reads them all and then reads them all and then reads them all and then links them to links them to links them to pages, keeps pages, keeps pages, keeps an index, keeps a log, and an index, keeps a log, and an index, keeps a log, and now you can now you can now you can actually actually actually query them in a query them in a query them in a way that makes way that makes way that makes more sense more sense more sense because there are those because there are those because there are those links. And you links. And you links. And you put that whole idea put that whole idea put that whole idea into a gist, and that's what I'm into a gist, and that's what I'm into a gist, and that's what I'm going to put into the going to put into the going to put into the coding agent. coding agent. coding agent. The link to the gist is in The link to the gist is in The link to the gist is in the description of this video. the description of this video. the description of this video. So, what we do is So, what we do is So, what we do is pretty funny. We pretty funny. We pretty funny. We store store store Karpathy's brain in our own Karpathy's brain in our own LLM Karpathy wiki memory system. And if you LLM Karpathy wiki memory system. And if you guys have been following guys have been following guys have been following me for a while now, me for a while now, me for a while now, you know that I've you know that I've you know that I've created quite a few videos created quite a few videos created quite a few videos on this topic, and I have a on this topic, and I have a on this topic, and I have a bunch of different LLM wikis bunch of different LLM wikis bunch of different LLM wikis set up in my set up in my set up in my own AI OS own AI OS own AI OS . Okay, here's the . Okay, here's the . Okay, here's the second command to second command to second command to configure the wiki. You're configure the wiki. You're configure the wiki. You're going to insert going to insert going to insert the gist and then tell the gist and then tell the gist and then tell Claude Claude Claude to use it to use it to use it for the raw folder, which is for the raw folder, which is for the raw folder, which is everything Claude just everything Claude just everything Claude just unpacked for you. And what unpacked for you. And what unpacked for you. And what comes out is a comes out is a comes out is a wiki inside Obsidian.

  5. wiki inside Obsidian. wiki inside Obsidian. Obsidian is just a Obsidian is just a Obsidian is just a visual layer visual layer visual layer on top of what I'm on top of what I'm on top of what I'm showing you here. And you showing you here. And you showing you here. And you don't need Obsidian for don't need Obsidian for don't need Obsidian for this to work, but this to work, but this to work, but if you want if you want if you want to look at it to look at it to look at it visually, Obsidian visually, Obsidian visually, Obsidian works. You can works. You can works. You can see here: we have see here: we have see here: we have sources, we have a sources, we have a sources, we have a page for page for page for every thing he every thing he every thing he wrote or said, wrote or said, wrote or said, we have topics about what we have topics about what we have topics about what he knows, we have he knows, we have he knows, we have principles, we have rules, we have principles, we have rules, we have methods, we have how methods, we have how he explains, we have how he explains, we have how he explains, we have how he debugs, we have, he debugs, we have, he debugs, we have, essentially, as I essentially, as I essentially, as I said, how said, how said, how his brain works. And his brain works. And his brain works. And every time you every time you every time you see one of these see one of these see one of these purple links, purple links, purple links, it's a wiki that connects it's a wiki that connects it's a wiki that connects one page to another one page to another one page to another or one method to or one method to or one method to another. Each rule another. Each rule another. Each rule refers to refers to refers to the source from which it the source from which it the source from which it comes, or to the comes, or to the comes, or to the sources from which it is sources from which it is sources from which it is derived. Each derived. Each derived. Each source cites the source cites the source cites the rules it rules it rules it supports. So supports. So supports. So instead of instead of instead of just having a dump of just having a dump of just having a dump of notes like we had notes like we had notes like we had before, we now have before, we now have before, we now have this interconnected this interconnected this interconnected network or something network or something network or something like a relationship map like a relationship map . And every time . And every time . And every time something new comes up something new comes up , you add a new , you add a new , you add a new blog or a new idea, LLM blog or a new idea, LLM blog or a new idea, LLM adds it back in and adds it back in and adds it back in and connects it to a bunch of connects it to a bunch of connects it to a bunch of other concepts. The other concepts. The other concepts. The next step is next step is next step is getting him to think like getting him to think like getting him to think like him, not just knowing him, not just knowing what he said in the what he said in the what he said in the past. So, here's past. So, here's past. So, here's a tip for a tip for a tip for setting up rules. setting up rules. setting up rules. She essentially pulls She essentially pulls She essentially pulls his rules from the wiki, and his rules from the wiki, and his rules from the wiki, and each rule has to each rule has to each rule has to be linked to an be linked to an be linked to an exact quote from exact quote from exact quote from it and where it it and where it it and where it came from, because we don't came from, because we don't came from, because we don't want Claude want Claude want Claude just making it up. just making it up. just making it up. So, we have seven So, we have seven So, we have seven rules. Number one - rules. Number one - build it, or you build it, or you build it, or you won't understand it.

  6. won't understand it. won't understand it. Number two - Number two - Number two - first term first. first term first. first term first. Essentially, find one Essentially, find one Essentially, find one important element, important element, important element, show how it show how it show how it works, and then works, and then works, and then add only one add only one add only one element at a time. And element at a time. And element at a time. And almost his entire course is almost his entire course is almost his entire course is structured this structured this structured this way. Number three - way. Number three - predict, then predict, then predict, then run, then run, then run, then compare. He will say compare. He will say compare. He will say the number he expects the number he expects the number he expects before running before running before running the cell because, the cell because, the cell because, he says, most of the he says, most of the he says, most of the time it time it time it will be learning, but will will be learning, but will will be learning, but will imperceptibly imperceptibly imperceptibly perform a little perform a little perform a little worse. Number four - worse. Number four - show the show the show the wrong version first. wrong version first. wrong version first. He intentionally leaves He intentionally leaves He intentionally leaves his own mistakes in the his own mistakes in the his own mistakes in the record. Number five— record. Number five— prove it, don't prove it, don't prove it, don't claim it. Number claim it. Number claim it. Number six—tell me what you six—tell me what you six—tell me what you guessed. His guessed. His guessed. His main complaint about main complaint about main complaint about Claude's code is Claude's code is Claude's code is that the models make that the models make that the models make incorrect incorrect incorrect assumptions on assumptions on assumptions on your behalf and then your behalf and then your behalf and then simply execute them simply execute them simply execute them without checking. And without checking. And without checking. And number seven—that number seven—that number seven—that winning is easier. And winning is easier. And winning is easier. And what's cool is, you can what's cool is, you can what's cool is, you can verify any of verify any of verify any of these claims. You these claims. You these claims. You can click on can click on can click on a quote and it a quote and it a quote and it will open the page will open the page will open the page it was taken from, with a it was taken from, with a it was taken from, with a timestamp or with the timestamp or with the timestamp or with the exact source. And exact source. And exact source. And if there is no source, if there is no source, if there is no source, the agent should clearly the agent should clearly the agent should clearly state that he is state that he is state that he is drawing a conclusion based on what drawing a conclusion based on what Karpati said. Now Karpati said. Now these rules need to these rules need to these rules need to be somewhere where the be somewhere where the be somewhere where the agent can agent can agent can actually read them. actually read them. actually read them. So, here is the So, here is the So, here is the fourth query. And he fourth query. And he fourth query. And he asks Claude asks Claude asks Claude to convert them to convert them to convert them into two files for us. This into two files for us. This into two files for us. This first file is a first file is a first file is a subagent. And if you don't subagent. And if you don't subagent. And if you don't know what that is, it's a know what that is, it's a know what that is, it's a separate Claude with separate Claude with separate Claude with its own its own its own instructions, its instructions, its instructions, its own memory, and a own memory, and a own memory, and a context window so that context window so that context window so that if you want to if you want to if you want to pass a task to a pass a task to a pass a task to a subagent, it doesn't subagent, it doesn't subagent, it doesn't drag down drag down drag down your entire conversation. And you your entire conversation. And you your entire conversation. And you can see here, can see here, can see here, inside this inside this inside this subagent file, seven subagent file, seven subagent file, seven rules. We rules. We rules. We can also see how it can also see how it can also see how it communicates, and we communicates, and we communicates, and we can see the loop can see the loop can see the loop it executes for it executes for it executes for each individual each individual each individual task. So it's task. So it's task. So it's essentially like a little essentially like a little essentially like a little mini-Agent Karpathy.

  7. mini-Agent Karpathy. mini-Agent Karpathy. Now, the second one is a Now, the second one is a Now, the second one is a skill. So you skill. So you skill. So you can see here that can see here that can see here that when I type {slash} when I type {slash} when I type {slash} Karpathy-teach and then what Karpathy-teach and then what Karpathy-teach and then what I'm stuck on, I'm stuck on, I'm stuck on, it goes to the it goes to the it goes to the agent with the rules and agent with the rules and agent with the rules and then evaluates then evaluates then evaluates the response against the the response against the the response against the checklist, checklist, checklist, one line per rule one line per rule , and then it actually , and then it actually , and then it actually shows me everything. shows me everything. shows me everything. So these are the two things So these are the two things So these are the two things we have: the Karpathy agent and the we have: the Karpathy agent and the we have: the Karpathy agent and the Karpathy skill. Good. Karpathy skill. Good. Karpathy skill. Good. Now here's another thing Now here's another thing Now here's another thing that's really important that's really important that's really important when we're when we're when we're setting this up. We setting this up. We setting this up. We want the agent to want the agent to want the agent to make sure they are make sure they are make sure they are executing their code executing their code executing their code before responding to before responding to before responding to you. And that's why you. And that's why you. And that's why building this building this building this inside Claude's code inside Claude's code inside Claude's code works so well, works so well, works so well, because he can already because he can already because he can already write and run write and run write and run code. And this is a step that we code. And this is a step that we code. And this is a step that we need to make sure the need to make sure the agent cannot agent cannot agent cannot skip based on skip based on skip based on Karpathy's own rules. Karpathy's own rules. Karpathy's own rules. So, here is the fifth query So, here is the fifth query . This is essentially . This is essentially . This is essentially adding a adding a adding a launch gateway. It adds a launch gateway. It adds a launch gateway. It adds a hook, which is a hook, which is a hook, which is a small script small script small script that runs that runs that runs immediately after the immediately after the immediately after the agent attempts agent attempts agent attempts to complete its turn. to complete its turn. to complete its turn. So if he wrote So if he wrote So if he wrote the code and never the code and never the code and never ran it, the hook would ran it, the hook would ran it, the hook would actually block actually block actually block the response and send the response and send the response and send it back with a single it back with a single it back with a single message: "Hey, message: "Hey, message: "Hey, you have to run this you have to run this you have to run this before you can say it before you can say it before you can say it works." So we're works." So we're works." So we're just implementing a just implementing a just implementing a verification cycle. And verification cycle. And verification cycle. And you know, obviously you you know, obviously you you know, obviously you don't have to don't have to don't have to write write write this hook yourself. You simply this hook yourself. You simply this hook yourself. You simply request it request it request it using this using this using this request. So, now request. So, now request. So, now it will run, it it will run, it it will run, it will test, it will test, it will test, it will fix, and only will fix, and only will fix, and only after that will you after that will you after that will you see the final see the final see the final result. And then the result. And then the result. And then the last step, last step, last step, of course, is just of course, is just of course, is just testing. So, testing. So, testing. So, a few different a few different a few different queries that we queries that we queries that we used. I used. I used. I asked him here to asked him here to asked him here to build something build something build something real and teach real and teach real and teach me how it works, which me how it works, which me how it works, which was a byte pair tokenizer was a byte pair tokenizer was a byte pair tokenizer that that that converts text into converts text into converts text into numbers before a numbers before a numbers before a big language model even big language model even big language model even sees it. And sees it. And sees it. And look what he's look what he's look what he's doing here. First, doing here. First, doing here. First, it writes down what it writes down what it writes down what "done" means "done" means "done" means before it touches before it touches before it touches any code. And any code. And any code. And then he creates the then he creates the then he creates the smallest version that smallest version that smallest version that works with tiny works with tiny works with tiny input data, and input data, and input data, and before he before he before he runs it, he runs it, he runs it, he tells me what he tells me what he tells me what he expects to see.

  8. expects to see. expects to see. Then he runs it Then he runs it Then he runs it and shows me the and shows me the and shows me the actual result. And actual result. And actual result. And then he shows then he shows then he shows me the version that doesn't me the version that doesn't me the version that doesn't work and why, and work and why, and work and why, and then he fixes it. And then he fixes it. And then he fixes it. And then at the bottom he then at the bottom he then at the bottom he lists lists lists practically everything that practically everything that practically everything that was run, what was run, what was run, what came out, what changed, came out, what changed, came out, what changed, and what rule and what rule and what rule was followed for was followed for was followed for each of these steps. each of these steps. each of these steps. So, instead of So, instead of So, instead of just getting a block of just getting a block of just getting a block of code and then the agent code and then the agent code and then the agent saying, "Hey, this works," saying, "Hey, this works," I essentially walked I essentially walked I essentially walked through it one through it one through it one piece at a time, and piece at a time, and piece at a time, and now I can now I can now I can understand much better what understand much better what understand much better what just happened. Now just happened. Now just happened. Now another thing you another thing you another thing you can test can test can test is its review. So is its review. So is its review. So before I before I before I hand something over to a client, I hand something over to a client, I hand something over to a client, I might ask them might ask them might ask them one question: "Would one question: "Would one question: "Would the client accept this, and what would the client accept this, and what would the client accept this, and what would they like to remove they like to remove , or, you know, , or, you know, , or, you know, questions like that?" And here is a questions like that?" And here is a questions like that?" And here is a script that script that script that Claude wrote for me Claude wrote for me that I can that I can that I can use as an use as an use as an example. This script example. This script example. This script gets comments from gets comments from gets comments from YouTube, and it worked YouTube, and it worked YouTube, and it worked fine when Claude fine when Claude fine when Claude tested it. The agent tested it. The agent tested it. The agent reads it, reads it, reads it, predicts it will predicts it will predicts it will crash on an emoji in a crash on an emoji in a crash on an emoji in a regular regular regular Windows terminal, runs it Windows terminal, runs it Windows terminal, runs it that way, and then that way, and then that way, and then it crashes it crashes it crashes before writing any before writing any before writing any files. So the claim files. So the claim that it worked was that it worked was that it worked was only true in one only true in one only true in one specific specific specific setup or setup or setup or terminal. So here terminal. So here terminal. So here you can see that you can see that you can see that Claude was able to go through and Claude was able to go through and Claude was able to go through and cut out a bunch of things cut out a bunch of things cut out a bunch of things that didn't deserve that didn't deserve that didn't deserve their place. He their place. He their place. He was able to give me an analysis was able to give me an analysis was able to give me an analysis of how we would actually of how we would actually of how we would actually fix it, fix it, fix it, clean it up, and clean it up, and clean it up, and revise it before revise it before revise it before we could actually we could actually we could actually ship. And then ship. And then ship. And then he proves that the he proves that the he proves that the trimmed version trimmed version trimmed version answers the same answers the same answers the same questions by running questions by running questions by running them both side by side and them both side by side and them both side by side and actually showing actually showing actually showing me. And then another me. And then another me. And then another thing I would like to thing I would like to thing I would like to bring up is that this bring up is that this bring up is that this thing thing thing will just keep will just keep will just keep learning. Back in July, learning. Back in July, learning. Back in July, Karpati wrote about Karpati wrote about Karpati wrote about how he would how he would how he would chat with a model chat with a model chat with a model for 10 minutes and for 10 minutes and for 10 minutes and then let then let then let the model clear all the model clear all the model clear all that text. So, I that text. So, I that text. So, I can run {slash} can run {slash} can run {slash} Karpathy {dash} ingest with Karpathy {dash} ingest with any link.

  9. any link. any link. It will then check the It will then check the It will then check the raw raw raw data folder, write the source data folder, write the source data folder, write the source page for page for page for the post, and then the post, and then the post, and then update every update every update every page that page that the post touches. A new behavior has been added to the sixth rule new behavior has been added to the sixth rule . So now . So now . So now the agent knows to the agent knows to the agent knows to ask ask ask me a couple of questions me a couple of questions me a couple of questions when my request is when my request is when my request is small, rather than small, rather than small, rather than just guessing just guessing . And the rule that . And the rule that . And the rule that was lying on the bench with was lying on the bench with was lying on the bench with one source behind it one source behind it , received its second , received its second , received its second source and became a source and became a source and became a real rule. And real rule. And real rule. And then, because it then, because it then, because it understands how this understands how this understands how this wiki works, it will update wiki works, it will update wiki works, it will update the index, it will update the the index, it will update the the index, it will update the hot page, it will hot page, it will hot page, it will update the log. So, my update the log. So, my update the log. So, my brain just brain just brain just got better after got better after got better after I added one I added one I added one link. And I link. And I link. And I never had to never had to never had to touch the wiki touch the wiki touch the wiki manually or manually manually or manually manually or manually set up those set up those set up those connections or links. connections or links. connections or links. And that, of course, was And that, of course, was And that, of course, was his second rule. his second rule. his second rule. Okay, so Okay, so Okay, so moving away, we moving away, we moving away, we had four steps. We had four steps. We had four steps. We gave Claude everything that gave Claude everything that gave Claude everything that was in someone's head, the was in someone's head, the was in someone's head, the head of Carpathia, and we head of Carpathia, and we head of Carpathia, and we put it together in a wiki. We put it together in a wiki. We put it together in a wiki. We turned this into turned this into turned this into his rules with a his rules with a his rules with a quote for each one. quote for each one. quote for each one. We run We run We run it before it it before it it before it speaks, and we speaks, and we speaks, and we test it on test it on test it on real things. And real things. And real things. And the phrase I'd like to the phrase I'd like to leave you with today is one of leave you with today is one of my favorite quotes, my favorite quotes, my favorite quotes, and it's something he and it's something he and it's something he said himself this year, which is: said himself this year, which is: said himself this year, which is: you can you can outsource your outsource your thinking, but you can't thinking, but you can't outsource your outsource your understanding. You won't understanding. You won't understanding. You won't be an expert at be an expert at be an expert at everything, and that's everything, and that's everything, and that's okay. But you okay. But you okay. But you can definitely can definitely can definitely bring in an expert and bring in an expert and bring in an expert and develop them. develop them. develop them. Anyway, Anyway, Anyway, that will be enough that will be enough that will be enough for this video. So, for this video. So, for this video. So, if you enjoyed if you enjoyed if you enjoyed the video or the video or the video or learned something new, learned something new, learned something new, please give it a please give it a please give it a like. This helps me a lot like. This helps me a lot like. This helps me a lot . And, as . And, as . And, as always, I appreciate you always, I appreciate you always, I appreciate you watching watching watching the video to the end.

  10. the video to the end. the video to the end. See you See you See you next time. Thank you next time. Thank you next time. Thank you all.

Summary

This transcript discusses the creation of a specialized AI agent designed to replicate the teaching methodology and cognitive style of prominent AI engineer Andriy Karpathy. It references Karpathy's background with OpenAI, Tesla, and Anthropic, along with his core educational philosophies such as building the smallest version first and avoiding unverified code. The practical takeaway is a four-step method to transform standard AI assistants like Claude into personalized pedagogical agents that teach and explain code effectively rather than just generating it.

View original episode ↗