Claude Code source code LEAKED

All right, so this is absolutely wild news. So, Anthropic accidentally leaks Claude code's source code. No, this isn't the mythos leak from 5 days earlier. This is a whole new massive massive leak out of Anthropic. And this time it's not a blog post. It's everything that makes Claude tick. It's all the secret sauce just oozing out. So, an Anthropic engineer accidentally leaks the source code. The internet does what the does. It archives it within hours. It clones it, forks it all over GitHub within hours. We're talking about tens of thousands of copies. Anthropic did confirm the leak. So, some engineer accidentally includes the map file. It's like 60 megabytes. It's a dot map file. The source map can be sort of opened up, converted, transcribed into the entire source code. It's the simplified obfuscated version of the source code, but but you can convert it into readable source code. So, it's something like 200,000 type script files, 600,000 lines of code. One GitHub mirror got forked something like 42,000 times almost immediately. Anthropic, of course, pulls it back, but at this point it's too late. One very interesting thing about this release is there are tons of features in there that are being worked on, that are being developed, that are behind these flags that are set to false for public builds. So, basically, it's there. It's It's in Claude code. It's just It's not rolled out to the public. These are basically unshipped features, but features that are being worked on. These are real features that no one up until now knew about other than people at Anthropic. So, first and foremost, let's talk about these major major unshipped features because, wow, not only do they give us kind of a road map to kind of where things are heading, and some of them are pretty wild, it also kind of gives an idea of how big the open Claude release was, how much of an effect it had on these AI labs. Because I think it's very fair to say that all of these features are in some way, shape, or form a response to open claw. But, interestingly, this highlights a whole other can of worms that I think we're going to have to deal with and that AI introduces. Think about this for a second. Let's say there's a piece of software like a Photoshop or I mean just whatever you can think of that's a standalone piece of software. It's owned by a company. That company makes some money from it. Let's say I managed to fully clone or copy the entire code and just recreate Photoshop. And I call it West Photo or something, right? But it's the same code. It's the same source code. I just managed to get my hands on it. I start selling my version of it. What happens? What happens is I go to jail or there's some sort of a legal liability because of this. So, if I copy the code, that's bad. Go to jail. Do not pass go, etc. But, if I copy the functionality, that's totally okay. And I'm sure there's a million exclusions and variations and exceptions for this. But, the point is in general, this is how it works. You can't just steal code and use it as your own. But, you're welcome to create your own version of that product from scratch by yourself. Here's the thing. I literally was about to release a video talking about this exact thing about the potential of it happening. And literally, this happened this morning. So, Anthropic leaked the Claude code source code. We talked about that. Somebody forked it. We talked about that. 32,000 stars, 44,000 forks. All right, we're caught up. The person that forked it, well, needless to say, they're scared of getting sued. Cuz if you're caught, you copy the code, bad potentially right? You don't have permission. Could be issues. People have gone to jail for this in in certain situations. I assume it's mostly civil liability, but I I there there's some situations where it could even be criminal if I understand correctly. So, what does the person do? They convert the whole code base from TypeScript. If if you recall, the leak was like 2,000 TypeScript files. TypeScript is a coding language. You don't need to really know any details. It's just it's one language. And it got converted from this Anthropic wrote it in TypeScript, it got converted to Python. It's another coding language. So, in effect, what happened was they copied over the code and then they used the code to kind of copy over the functionality, right? Cuz now the functionality is the same, but the code is different. The code is no longer the code that Anthropic wrote. It's entirely different. Same functionality, completely different code. How did this person, you know, completely recreate Claude code in a different language? They did by sitting there and very rapidly, you know, retyping 600,000 lines of code. Keep in mind, this is happening probably within hours of of this leak, maybe within 12 hours, I'm not sure. This wasn't like days, this was hours. How did they manage to replicate it? Well, with Codex. Again, you know, OpenAI's AI coding tool, right? So, they used OpenAI's AI coding tool to completely, basically, recreate Claude code. By the way, Eugene Jin, great follow, highly recommended. As he's saying here, AI is quietly erasing copyright. Right now, it seems like AI developed this functionality for people gives the ability to do what sometimes in the past has been referred to as clean room engineering. It's like if you want to reverse engineer some other companies' products, we've had examples in the past of companies going to great lengths to reverse engineer in such a way that in the court of law, if they ever get sued, there's no connection between that product and and what they did. They they create this sort of a layers of separation to make sure that it's like, "No, we didn't copy anything." But, it was difficult, expensive. Here, it's easy, fast with the help of AI. I'll leave that there. We're not going to dive too deep into it, but mark my words, we're going to be hearing a lot about this. We don't have laws that clearly explain who's in the right here and who's in the wrong. I'm sure there'll be lawsuits and lawyers and judges will set precedents precedents, like precedents, but for right now, it's kind of a gray area. I'm not a lawyer, let me know if I'm wrong, but this seems like a legal battle in the future. If you're replicating some specific software with AI, does that mean that you just erased copyright? If I fork an open-source project that says, "You're welcome to use it however you see fit, we just have these stipulations like you have to use our licensing agreement moving forward." or whatever it is. Okay, but what if I do this completely reverse engineer, do that clean room engineering, completely copy, so it's more or less identical, just with different code, but the functionality is the same. So, now I've basically stripped it of whatever license that the original, the creators of it wanted it to be used under. If they had any copyright or anything like that, that's been effectively erased. Anyways, whatever else comes out of this thing, uh this is something that's going to have to be discussed and to figured out. But, let's dive into some of the huge features that are being built out in Cloud Code that are there, under development. They're just sort of a flag, so that the public, the general release, we don't get it, but it but it's there. One of the features, this is probably not the biggest one, but I thought it was interesting. It was kind of a Tamagotchi Easter egg. It's a little pet, kind of that that sits next to your input box and reacts to your coding. I'm not sure if it was going to like talk crap about what you're what you're coding, how you're going about it, or what, but apparently it was like a whole thing with Tamagotchi, different creatures that did stuff and grew and had various like legendary status, common status, whatever it is. It was like a whole thing. And I think it was meant to be an April Fools' Easter egg. First of all, can you imagine if the reason for this update being pushed live was because of they they they were hurrying to try to get this Easter egg in time for for tomorrow, for April Fools'? Can Can you imagine? All right, we'll get back to the Tamagotchis in a bit, but what are the big things that are going to be shipped pretty soon? So, first and foremost, there are mentions of the Mythos model. So, a few days ago we all saw the leak out of Anthropic. They were referencing a mythos model code named a capybara. So, first and foremost, that that is referenced in there. That model is there along with the next generation of models that are coming out. There's also something called a chiros, a background agent. So, it's an autonomous agent that runs constantly without human user input. It monitors GitHub repositories and sends updates. You can ping it from anywhere to get answers or to get it started on a task. So, you can think of it as a coding agent that runs while you sleep that updates you when it's done with stuff that you can ping from anywhere. Another one is, I think it's called auto dream. I'm not sure if that's what they reference it in in the code or that's what sort of the community, the name that we collectively gave it. But, it's this background agent that dream. If you think about how when we sleep, we have, you know, our REM sleep, etc. Or you say it REM sleep. REM is the is the band. REM sleep is the is the thing we do that kind of consolidates our memories. That's one of the functions for sleep is to consolidate the memories, kind of like compress them, you know, save what you need and discard the rest. This is what auto dream does. It's this agent that runs when nothing else is happening. Kind of consolidates all the memories, reviews your past context interactions, etc. There's a voice mode, a real-time voice chat with AI agents. So, similar to, you know, how you have it in Grok, OpenAI, etc. Now, something like this is coming to Claude. And there's something called ultra plan. So, this thing spawns a a separate remote in the cloud sort of a work session. It's a 30-minute work session that's remote. It's run by a very sort of kind of expensive deep planning model. And it fully plans out your entire task before starting. So, if you have some sort of a software thing that you're trying to run, it's complicated, you write everything out. I assume I'm guessing here based on the description, but I assume what happens is it kicks off this ultra plan before doing anything else. A remote session gets started and this specific model fully goes through the entire plan, fleshes everything out, you know, gets all the checklists, everything it needs, and then from that, the next step happens, whether that's like a coding agent that gets spun up or whatever. But this is a standalone 30-minute remote session for planning only. And kind of hand in hand along with that, you also have the coordinator mode. This is a multi-agent swarm. So we're getting to see these uh multi-agent swarms. It's funny because everybody used to call them in the past when we were kind of thinking about what the future's going to be like. We're like, "Yeah, AI agent swarms." That was the terminology that we used. Now that these companies are actually developing them and they're trying to use them for, you know, enterprises, you know, they're not going to call it swarms. That's, I don't know, not professional, I guess. So it's interesting that all the different companies they came up with, you know, coordinator or orchestrator, like words like that instead of the swarms. So they do a simple one, Claude agent orchestrates a bunch of other Claudes. Each worker has its own little scratchpad, its own set of tools that it can use. And what we see in this source code is a a whole sort of multi-agent architecture, how that happens to be, how that's all managed, etc. Basically describing how this orchestration happens. There are things like agent scheduling, cron jobs. So we've seen this before, but more functionality kind of built in with that. So running things on a set timer or on a certain schedule. There also seems to be a real browser control. So not some sort of a scripted browser interaction, actual full browser with a browsing agent. We also have persistent memory across sessions, so real memory doesn't get wiped between sessions, it accumulates. So again, that's another function that's already built in, it's just not exposed to users. It's there, we just don't see it. So I got to say a lot of this is things that we've seen with open Claude, all right? So like heartbeat.md accounts for some of this. I'm not exactly sure what they were working on beforehand or not, but a lot of this seems to I assume was triggered by open cloud how how popular it was, how powerful it was, how many developers jumped on board to start using it, you know, Anthropic and all the other companies they saw how incredibly successful it was. And so now they're trying to build in a lot of those features here. There's a tons of other stuff like commands like {slash} advisor where you get a second model to overview Claude's outputs. There's also {slash} good Claude, {slash} bug hunter, and {slash} teleport. We don't know what those do yet, but those are some intriguing commands that apparently are coming soon. Teleport seems like switching between sessions. Bug hunter seems like, you know, hunting bugs, but we we we don't have full information about what exactly those do. And apparently Reddit deep dived into the code. They found a complete virtual pet system hidden in the code. So {slash} buddy command creates this little ASCII companion based on your user ID. There's 18 species like duck, capybara, dragon, ghost, axolotl, and something called a chunk. Not sure exactly what a chunk would look like. I assume it's going to be a big boy if it's called chunk, but not sure exactly what sort of an animal that's going to be. We have a full gacha rarity system common to common to legendary with a 1% legendary drop rate. I hope these aren't like spoilers. Um I guess spoiler alert after I already said it, but yeah, 1% is the legendary drop rate. And there'll be various shiny variants and hats for your favorite pets to wear. Crowns, propeller hats, a wizard hat, maybe even a tiny duck hat. And the pets will have different stats. Can you guess what the stats are going to be? No, it's not like strength and agility. No, no, no. It'll be things like debugging. That's stat number one. If you're thinking, "Oh, I got it. They're going to be like software development themed." No, stat number two is chaos. And in case you're not fully confused, the third one is snark. They're going to be snarky to you, I assume is what they mean. So debugging, chaos, and and snark are the three main pet stats. One other interesting thing that jumped out at me is that apparently Claude as you're talking to Have you ever got frustrated cuz it didn't do something right or it's failing to do something? I had one moment where I really was getting frustrated. It was with open Claude specifically using a Claude model, but a new open Claude update broke some functionality. So it kept saying like, "Oh, I'm going to do it." and then didn't do it. And I'm like, "What are you doing? Stop. Just fix it. Go." What I didn't realize, I think most of us didn't realize, had no idea this was happening. But apparently, based on this leak, Claude is sort of watching your language to try to figure out if you're getting frustrated, if you're maybe losing patience with Claude. So there's some sort of a process there which like, "Is this user about to lose it?" There's also something called an undercover mode. An AI researcher flagged it and there's something interesting happening here. So people are still investigating. So keep an eye on that. I'll I'll do a follow-up once we know exactly There's something weird happening there. We just we just don't know what yet. And if you've ever wondered, you know, that that spinning text when Claude is thinking, like how many different variations of that text is there? Cuz it does seem like almost every time it's like a new one. So in case you wanted to know, 187. 187 different verbs for the loading spinner text. You're welcome. And just to clear up any misunderstandings, misinformation that may be floating out there. So model weights were were not leaked. No model weights were leaked. No proprietary AI training secrets, no customer data, no API credentials, nothing like that. This was just a lot of the harness around Claude code that made it go and that made it go so well. Also, apparently there's a lot of X402 references. So this is kind of like the protocol for crypto coins. So if you were building an infrastructure to allow agentic crypto payments, there's a lot of references to that. So it seems like Anthropic is building that functionality. So, there's no coin to buy, there's no S coin to to dump your life savings into. Don't worry. I'm sure some scam artists will soon spin one up. So, if you're if you wanted to lose your money, you will soon have the opportunity, I'm sure. But, nowhere in there is there any coin. Don't buy any coins. Just stop it. Seek help. Don't buy any coins. But, it does seem that Anthropic is building some protocols for agentic crypto payment. The recently leaked capybara model is mentioned. So, that's the mythos model that was mentioned. Now, some of the more eagle-eyed watchers or listeners will say, "Wait, but isn't that one of the Tamagotchi pets? Was the capybara one of them?" So, capybara is one of the little Tamagotchi pets along with chunk and duck. But, apparently their own build scanner, like when it scans the code, it sometimes flags certain words. So, like if you want to call a model capybara and you want to then create a little Tamagotchi pet called capybara, it's not going to allow for that. It'll flag that conflict. So, what they did is they just encoded all the 18 species of the Tamagotchi pets in hexadecimal. If you're not sure what that is, it's like if you have a dog and every time you say walk, it loses its freaking mind cuz it knows what walk means and it like starts like running in place rapidly and running up to you and whatnot. So, you can't say walk, right? So, what do you do? You spell it out. Maybe if you need to talk to somebody else in the household and you're like, "W A L K." Hoping the dog won't soon also figure out that you spelling the word also means the same thing. That's basically what they did here. They just spelled the duck in in spelled it out with hexadecimal. So, what does this all mean? What's the big picture? It means we all can now see what Anthropic is effectively building, what the road map is. The entire road map effectively became public. All the competitors are looking at it right now as we speak, guaranteed. And there's going to be a lot more coming because everybody's digging through this stuff. We haven't seen some of the deeper hidden stuff like the undercover mode. There's a lot of stuff there that that we have to kind of a parse through. This is just the tip of the iceberg. So, stay tuned. Make sure you subscribe. More coming soon. I'm sure of it. If you made this far, thank you so much for watching. My name is Wes Roth. I'll see you in the next one.