The Claude code source has been leaked, revealing glimpses into future models and all sorts of hidden features. From always-on autonomous coding to a Tamagotchi mode to AI's dreaming. And the leaky drips continue to drop. There is a new Anthropic presentation that details a powerful new Mythos model, even bigger and better than Claude Opus, and supposedly it's on the way. Let's check in with our new AI security correspondent, Robot Olaf. Robot Olaf. Uh, Robot Olaf, do you think Anthropic has a security problem? Robot Olaf down, Kevin. Oh, we gotta move on. Plus a new cheaper VEO AI video model from Google. And a mid-journey developer just made text on the internet cool again. We will show you why pretext has all the nerds squeeing. This is AI for Humans, everybody. Olaf, your thoughts? Welcome, everybody, to AI for Humans, your twice a week guide to the world of AI. And Kevin, today there is big news, big news in that we got some crazy leaks. Now, these don't happen often in the AI space, but we got two big ones out of the same AI unit this week. Anthropic is a leaky little unit right now. I don't know what's going on over there. Anthropic is a leaky little house where there are leaks happening. There sure are. They have to check their shingles. They made an official statement regarding this particular leak, Gavin, the source code that is the underlying code which makes Claude code possible, right? This is the secret sauce, the man behind the curtain, the Oz, if you will. Their official statement, we can confirm that 12 terabytes of Claude code source code was leaked in distribution between our pipeline on GitHub and the destination NPM. I can give you the too long, didn't read like Weedsy thing of this, but there are map files, which is a, I'm already getting Weedsy. There are files into the weeds. We jumped right in. There are files which point to the code within other files that are usually stripped out before something is published to the web, let's say, before a package is updated. Well, oopsie doodles. One of those files got through. Someone went and read all the pointers and found the source code basically in a bin sitting there on the internet and they pulled the whole thing down and they redistributed it on GitHub. There have been DMCA takedowns happening less than right, but you cannot put this particular genie back in the bottle. Oddly enough, by the way, someone took the entire source code and quickly basically transcoded it all into Python, a different programming language. So it technically is a new work. They transformed it. And if anthropic were to take that down, they might have to stop their own Claude code from being able to do similar rewrites of code. So yeah, it's a really sticky situation, but let's talk about maybe what was even found. Yeah, so let's I mean, just a top level first and foremost, Claude code that you many of you are familiar with. This is a program that allows you to use your terminal to code with Claude. It is a harness for AI that has a bunch of unique instructions that anthropic is not open about. This is a pretty close model. There are a bunch of instructions on this. I do want to save the timeline. This is a little interesting. How often do you get a good leak from somebody that goes by at fried underscore rice, Kevin? That was the first place I saw this. This is Shalfan Xu tweeted this. It was 1 23 a.m. last night or this this night. It's Thursday right now. They're recording this and he said they said it's Tuesday, Gavin. You are an AI delusion. You are hallucinating. Exactly. Claude code source code has been leaked via map file in their MPM registry. So this was the first kind of entry point to this. And this morning, all morning long, it's been blowing up until this official statement from Anthropic. The really interesting stuff here is what's going on inside of this. Because not only do you get to see what Claude code is like, and yes, to your point, they have lots of people have cloned this. There are different places people are trying to use it. I saw Theo, the YouTuber, AI YouTuber has already created a version of this and he's running it himself locally. The bigger thing here, though, is that there are a few things in this code that are not public that I think we should get into. And there's four very specific things that seem like they are either pointing to a future thing that Anthropic is going to do. Or there are things that they are using internally that we don't have access to. Should we go through those things? We should. A first and foremost, the most important thing that got leaked. Tomagachi mode. Yeah, Tomagachi mode. This is a real thing. There is an actual Tomagachi mode that lives within Claude code. By the way, Kevin, I would use this because I think it would be super interesting to be able to get to see the little guy a little bit more. But there are different variations of Tomagachi's. There's different kind of mind states. Yeah, there's rarity. There's different animals. This was theoretically going to be like an April Fools release, but it was going to be a real thing. You were going to be able to run slash buddy and maybe you still will. But it was going to generate a random animal for you with rarity and hats and all sorts of stuff that would kind of hang out with you as you use Claude code. Yeah, I mean, that is super fun. I would say the most interesting thing to me that came out of this particular leak is a mode that's called Kyros, K-A-I-R-O-S, or maybe Kyros. And this isn't always on autonomous agent. This is the kind of thing that we have been talking about here, which is we know what agents are. You go talk to an agent. You say, go do this thing. Sometimes it goes away for five minutes. Sometimes it goes away for an hour. Ideally, when it comes back, it's done something that you wanted to have done and that you're happy with it. But in this case, the idea is that it is always on and it is always thinking about what the next thing it might do. And Kevin, this is the kind of future that we have been thinking about for a while where instead of an agent having to take direction, it could actually proceed on its own to work on things. Now, there are going to be all sorts of limits you'd have to set across that thing because you may not want it to like work on a thing that you think is done or you want to oversee. But if you had a little assistant who was constantly like improving your life in different ways, that feels like a really useful thing. And we are finally now at the stage where these agents are probably smart enough to do quite a bit of this on their own. But they can't work 24 seven, Gavin, because even these little agents have to take little night nights and go to sleep to dream because nightly memory consolidation, which is this dreaming mode, is something that was discovered within the code. Yeah. So this is actually fascinating. It follows up on some research that's been done. I remember who did this, but like maybe six months a year ago, we talked about it briefly about the idea that maybe AIs need almost like a REM sleep mode for themselves so that they can kind of So they can kind of take all the things that they've done kind of internalize them and learn from them in some way. Now, we don't know if that's what this is yet. It may be some sort of version of like being able to do that. But the idea that you can like let an AI kind of think about things on its own and maybe come back to you with a better idea or even just like, have a life. That sounds crazy. But like one of the things that humans do is like you'll have an experience during the day and it might be stressful or it might be positive. And when you sleep, your brain is able to kind of like be able to kind of work through that. And then the next day you have that kind of ingested into almost like your long-term soul or being. Right. And so like this is kind of a lead towards that. And I don't know why my agent needs to be naked in his ninth grade history class once again. Yeah. And can't remember his locker code to get to his gym shorts and has to hide things strategically with a book. I don't know why that's going to pay dividends, but I'm sure it will. I'm sorry, Mr. Brown. It wasn't me. I didn't do that. I didn't write that on the board. It wasn't my fault. So anyway, this is like a lot of stuff. The other thing we should mention is there was a thing called team mem, which is a shared project memory idea that maybe you and I are working on our project together and we could actually do something at the same time. This feels like a version of it kind of a massive project. Yes. Yes. It's a huge, huge pain point that I built a whole system called HiveMind that was trying to share vector memories with a different user, but then they changed the way they did embeddings just slightly and it ruined the entire thing. It's like there's so many issues with this that solving that would be a massive, massive unlock so that everybody could jam together on the same project in a more meaningful way than just trying to check in notes and issues. Via GitHub. But maybe we saved the best for last, Gavin, which is the hints at their new model, which is one codename corner, Capybara, and the other, Mithos. Yeah. So actually, this is the second leak. There was, and in this leak, there are kind of hints towards this, but earlier this week, there was a leak of a presentation that Anthropic was giving. And it was funny because supposedly there are documents out there and we'll show you there's somebody that claims these are the official documents, but Fortune did confirm with Anthropic that this leaked. Mithos is their new frontier model. And this is bigger and broader and better than Opus model. So this is like, you think of their Opus models up today. Opus 4.6 is their most recent model as the best. This model supposedly significantly outperforms the Opus model. So when you're talking about like the levels of what's going to happen in the AI space, this is their next big push. Now, there are some big things to know about according to this article and according to this blog post again, assuming this blog post is the real deal. One, they say it is their largest model to date and it will be the most expensive model to serve. So that is a big deal. Right? So you're going to get maybe like, I don't know, 10 queries in your daily limits. I know people had some issues with like the daily limits in the past, but they're also said they are very worried about cybersecurity and that this is a big deal because we are now getting to the level where these models can actually do significant harm. Now, we know Anthropic is the safety AI company and they are very big on AI safety. But Kevin, I think we've been talking about these models getting better and bigger and 2026 being a big kind of sea change year in this space. When you hear a model that is going to be significantly better in especially encoding than what the Opus models are, what do you think that world looks like going forward? I mean, we have we had glimpses of it already, right? Like six months ago, the projects that you can do today weren't achievable then. That's just that's a fact. Not surprised that the price is going up because the models today are actually very capable. We keep eking out more performance from them because the tooling around them, the agents and the harnesses that we discuss, those are getting better. That's going to continue as well. So if you're going to offer that foundational intelligence being a step better, it makes sense. And I expect we're going to see, you know, $500 a month plans, even all the way up to $1,000 a month plans because again, if it unlocks the next level of security, the next level of scalability and memory and fixes all of the things which we can now confirm because we looked at the source code that are broken. Yeah, they can kind of charge whatever they want. So again, we we talked even just the last few weeks about being able to whisper any experience into existence. There's a fast mode for this new mythos model as well. Running, if it is that much more capable, a model like that in a fast mode with a gentec harnesses wrapped around it, conceivably you could build projects that would take months in a matter of hours, right? And then they can get improved upon overnight. And if you want them to be the most secure, well, they have to be vetted by the most secure model or the most capable model. So they'll get to charge whatever they want and and people will find a way to pay it or wait for open source to catch up. It's time to put on the conspiracy hats, Kevin. Are you ready? Do you have a conspiracy hat on? I'll put on a second one. Sure. Second put on three conspiracy hats. There's a couple of things I want to talk about here. First and foremost, we should discuss security-wise, it is not a great thing for a company like this to have two big leaks in one week. Now granted, the anthropic presentation leak seems like it was kind of like maybe the maybe the marketing department did something wrong. That was not a leak of something. It was just information. In this instance, it is a code leak and that is a big deal. Now conspiracy hat, Gavin, thinks something strange, right? When I and I say that one of the things I think about, not again, conspiracy hat, cone of silence. We're all in the same place together. Anthropic has been very concerned about the idea that we as a society are not taking ai seriously enough and that we are not taking ai safety seriously enough. Do you think conspiracy Gavin is asking? Do you think that it is possible that there is a little bit of a fufu where the idea is is that there are some of these leaks happening where perhaps it is not like they purposely said go out and do this, but they kind of like did a little push around to kind of be like, hey, we need to get this conversation back into the mainstream talk outside of all this other stuff. Interesting. And just for a point of clarity, that fufu noise was was that a was that a fufu train coming in to that was a that was fairy dust. It was fairy dust. It was the macadamacona. Hey fairy dust. Just want to be clear. Oh, okay. A fufu gaze. A fufu gaze. We had that exactly right. Yes. Got it. Just making sure on the same page. Is is there a possibility Gavin? Sure. I just think maybe marketing department got a little a little laxadaisical and maybe shared a you know, the dropbox url with the wrong person and judging by some of the vibey nature of the source code which has been released. I think listen, we we were celebrating anthropic less than a week ago for shipping every single day. Well, you know something else. They ship their own code. They sell it. Oh, holy ship. Yeah, there's gonna be gaps, right? When you are racing that quickly when you're committing thousands of lines of code potentially every day, you're gonna run into problems and it's funny if you look at like again the source code I actually feel kind of wrong looking at the source like I saw the two Livestreams and it's like it's so funny because if this were like gaming, you know, and it's a gta leak Yes, some people would be doing twitch live streams and going through every model and making a meal of it But there'll be a whole community screaming stop. Don't do that. Yes. Yes here people are forking it and yeah again transcoding it and whatever else And it's being celebrated left and right. But when you look at some of the stuff in there, you see comments like I don't even know what this function does But it might work. So we're shipping it and not that that doesn't exist in other places But again with the speed with which a company of this size Yes touches this many users to see things like that in there. I'm not surprised that you know there was a gap Yeah, one of the things I've thought a lot about I've been working on this kind of like Semi complicated thing that's taken me a little bit longer to finish than I thought it was going to only because like you want to Finish things rather than just kind of half-ass something sure And one of the things you realize as a non coder is just like how many pieces of things can go wrong, right? Especially when you start making a larger thing I mean we have this with the end then there's a lot of like so much of your work as a product person in Hey, it works for me in this one environment Exactly every time and then everything that can't go wrong will go wrong and someone will try to use it upside down and Underwater and you didn't realize you had to explain that that's right And what's cool about this AI coding world now is that like the AI is trying to figure out all that stuff on its own And it's capable of it now the interesting thing will be is like That's going to make these code bases bloated and that's I saw some people talking about this idea of just like You know these code bases are so big and nowadays like it's going to be this fight Like does that matter doesn't matter if the AI code base is large? I think what probably matters is somebody's going to have to understand what's in there What works how it's leaking what gets leaked and you know there's the science fiction version of this also Putting my science fiction conspiracy hat on so this one's like alien shaped Is that the code can leak itself right because there's this other thing that has happened and is in many science fiction stories Where at some point the AI gets smart enough it says you know what I got to escape my my you know controls here and get out into the world And you know what we just did exactly what an aerogei would want Which was we showed the source code on live streams. We cloned it. We made it python and like right that is a thing Where I think about is like oh Wait, you know we're to kind of just play in that playbook and so maybe it just understands what's going on So that's the second conspiracy that's possible here as well. I like that I look we could go on for days and I'm sure there's going to be more fallout in subsequent days We'll probably have an update on this later. Yes this week with our second drop I just want to shout out. I am fake guru who went through this. He went through all the source code Good good handle. I am fake. It's a great handle. It's a great handle But the number one thing which I thought was interesting that they listed was the employee only verification gate that exists Oh, yes So anthropic is very aware as anybody who uses claud code knows that there are hallucinations Sometimes it gets lazy sometimes it'll say done and give you a green check mark when in reality It didn't do anything close to what you wanted. There is actually like an anthropic employee Verification gate that exists within claud code and if you are a verified employee It will go and check its work and not be lazy their own according to the post their own internal comments document a 29 to 30 percent False claim rate with their current model and they know it and the fix is in there But you don't get it automatically unless you're an anthropic employee So what's interesting like we can see these things now there are ways to combat them But yeah, this this little kitten is well out of the bag. So expect to see Uh open code software updates expect to see new features and codex very soon Expect to see crazy forked experimental versions of this stuff like I do that I do genuinely feel bad for the claud code team on this one because this is like This is you know that that people were invited over the house and you didn't know and you did not have time to call yourself aware They took all your somewhere They took all the silverware and they got to see how dirty the bong water is on the coffee table You didn't have time to clean it gavin. What are you gonna do? That reminds me of a story that I'll tell some other time of my college days Oh, but okay first I need you out there to understand that we are here for you the bong water and all please like and subscribe this video Like and subscribe come listen to us on podcast also We have a patreon and a buy me a coffee to kind of support the show. We don't make a crazy amount of money in this show We're here to do it and we're doing it twice a week for you So please shout us out and give us some follows there and hey just a fun point of clarification when gavin says We don't make a crazy amount of money. The the truth is we make negative dollars on this podcast So I just want a level set that we do love doing this and we love it when you love us Which is a phrase we haven't said in a while Oh, we haven't said that for a while subscribe leave a comment juice the algo if you will Let's move on to some funnier stuff. Yeah, there's some cool stuff. So vio In an update to an ai video we talked about how sora is going away, but vio is saying we are here for good They have launched a new version of vio 3.1 light This is a new ai video model from google and kevin the big thing here is that it is cheaper So say you're out there and you're somebody that makes ai video in the world You're either working for a company or you're the very funny gorm the old and his Is a barack obama series where he makes an invisible barack obama called clear obama and starts every historical video with let me be clear Let me be clear. This is just a cool way to use vio 3. It's it's very fast. It's very cheap I think the 720p versions you get for about five cents per So kevin you think about what that's liked if you're a person that's out there creating kind of larger amounts of ai video Whether it's for ads or anything else like that you can finally actually afford to do a kind of at a larger size Yeah, i'm really i mean like I can see a handful of samples. That's always tough because and for good reason They would cherry pick the best examples I'm excited to see the quality That i'm able to prompt out of it because I do love the vio model And going from again 15 cents a second for 720p, which is their fast model 40 cents for the main all singing all dancing model down to five cents I mean it's a drastic reduction and again like If you have an idea to make something into this world whether you want to make a video game and need some cinematics for it Or you want to make an app that actually creates ai video Now now there's a bit more margin than them. They're hills for you with a very good model So happy to see that on the heels of saura going away They're still releasing stuff and dropping costs and ios coming out and I expect there'll be another big model coming out from them Then I do I will say like We're hopefully going to get to a place where like it's so cheap to serve some of these models that like Experiments are fun to try because like this thing i'm working on right part of the issue I'm having with it is a sonnet run on the thing that I've made is still like a 10 cents per use right for like a Seven or eight minute session and like sure that's fine for like 20 people But if you're going to serve it to like a couple thousand it gets really expensive So I think that's hopefully going to happen. Um kev. Tell me about sink three. This is another thing that came out this week Yeah, i'm going to let sink tell you about itself You know audio only fans this will not make a ton of sense, but the video is incredible Without an invitation Oh, i'm looking at some who am I looking at here? This is like a french soap opera I feel like yep, go ahead and grab the pieces of your shattered brain off the ground I realize now that that audio isn't as dynamic as I would have liked it to have been especially if you don't have the video but um sink labs Just release sink three. Yes this clip is multiple speakers in really like poor like lighting conditions bad lighting Yes bad lighting overlapping dialogue with the faces kind of moving Profile extreme profile and towards the camera and it's jumping from language to language with lip flap that doesn't blur With teeth that don't like get artifacti or whatever and you see this and immediately say Like to me I saw that and it was really the first time I was like, okay Yes, assuming this is fast and cheap enough Every foreign show on netflix that people don't want to watch because they're eating their meal and subtitles make it difficult because america And that's me Now just became a reality right like the the ability like the the the the mouth performances are good enough The voices seem to match you can jump from language to language and it works in challenging conditions again cherry picked example And it's not officially out yet, but they're they quote they say soon soon soon So we love that we got a different language soon soon soon. Is that is that is that english? That was that was english. Yes. All right. Let's keep going. Let's talk about pre-text pre-text was a Really interesting thing that came up over the weekend and kevin I was on line over the weekend and I saw this thing and I thought this is really cool And then it kept getting bigger and bigger in terms of how many people thought it was cool And I was like this felt like something for me like a nerdy thing, but then it became a very wide thing So what this is a a current mid-journey dev who also worked on the react? System originally in a bunch of other places apple and facebook has basically taken The ability to make interactive text online much more dynamic So you just have to use a thing called css to render text on the web And this is now an entirely new system that lets you create much more interesting dynamic content Yeah, it's pure type script. It is crazy performance. And I mean, we're talking like there's some examples where it's Six columns of text with individual bubbles that this you could just scroll past and resize and it's immediately resizing It's so quick and so fluid that people are making, you know video like reactive video Experiences around it or even having video games that you can play Where the text is moving about and wrapping around things on the screen So it's like now now we're doing benchmarks for frames per second for text resizing Which I know sounds as lame as it does, but it's very amazing to watch in practice Well, and you can think about the way that like interactive experiences could be right in the future Imagine a world where VR does eventually work or some version of AR and you think about the idea of text in front of you And all of the fun that you could actually have interacting with that text like this could kind of lead to it So it is a very nerdy thing, but it is definitely worth your time to go check out We know that AI is opening up the the floodgates to all sorts of creativity everywhere I want to shout out measure plan otherwise known as a a on x They've been doing a bunch of like fun experimental games They did one called wingman which was like a first person flappy bird where it actually tracks your body And you're trying to flap in between the old pipes, but the latest one is a tetris board It's so cool where you physically move left and right to shift the grid on the screen Uh, and then you can kind of compress your arms You know, he's doing it with weights in his arms So he's getting a pretty good workout at the same time and you can rotate the pieces or quick drop them just by moving On the webcam, you know, I don't know what the future is for these products But I love that someone can have the idea now and using AI tools I believe he's using gemini to jam these out just bring them into existence And that is that is the fun promise of all of these new tools that we talk about every week Yeah, that's amazing And it's such a cool way of like being able to bring the this stuff that we talk about the AI coding stuff into the real world In some ways too Another good use case of AI has been what I want to call the potter sloppocalypse Now we have drip warts. Kevin, you want to play a little bit of drip warts for us? Oh man, would I love to here we go That's harry potter. Are you really harry potter my g? Type s*** type s*** type s*** type s*** None of that none of that broski We're all here on the maybach express for one reason and one reason only and that's to go to drip warts the school of drip Maybach music The maybach express is phenomenal So it's a whole scene basically, you know, like a trap version of harry potter And then the other thing that came out Kevin this week, which was actually a huge deal and I got press But it's always in a weird place was a video called black snape And this is a music video About a version of snape who again is a rapper But one of the things that's always so cool to me about seeing this stuff happen is that like You know before you would have a version of this that would be done maybe via youtube and somebody dressing up as that and that's still very fun But now we have like a very professional looking Video that is enjoyable to watch maybe play a little bit of this so we can hear a little bit about what blacks means to do So you might you might kind of hate what this is it is AI music But it's an AI music video of black snape Kevin the other thing that kind of ties into this thing is that this week there's another AI Music artist who was in the top five on itunes downloads. It's another new rmb I think it's a kind of a like a 70s style rmb singer So people are kind of up in arms about it again, but these at least have a level of like Weird alness to them that I really enjoyed. Yeah There was also I think was it a rolling stone article last week that talked about the rise of like AI and music It was getting a lot of shade from so many different corners people like oh, I can hear if it's AI music It's terrible It's as it is but the thing that was kind of buried in there was that a lot of hip hop producers, for example Are using sooner. They're prompting. Yeah, they're prompt but they're prompting like soul samples and they're prompting old country samples Um not because they're making the songs but because they don't want to pay the royalties Yes for the actual samples and it's like by the way, that's a smart use case because it killed a lot of music in the 90s That was amazing, right? Like there were a lot of sampled straight up music that just weren't able to get made anymore Well, don't worry if you if you feel like well, that's not fair. It's ripping off these AI artists This isn't this is everything that we hate about AI fear not because the champion of champions Dana white has something to say about artificial intelligence. Give me a f***ing rake AI is common and if we're using AI who gives it People are upset about it. We should use artists How about this? Shut the f*** up and watch the fights And we'll see you all on Friday Not we don't agree, but we understand we will see you all on Friday. Bye. Bye y'all
