Anthopic did a thing...

All right. Hello. What's up, Greg? Good to see you again. Hello. Hello. Yeah. So, they dropped a they dropped a new thing and it is very much anthropic coded [laughter] and that's what we're going to be talking about. I I already read through it, already highlighted it. What's up, Luke? Ahi, Mark Anthony, what's up? Industry shocking. I haven't said that in a long time. We need to bring that back. This shocked the entire industry. I missed that. Ah, good times. Heck, how do I feel about doing a guide on how to generate genuine income using AI? I don't feel good about it. That's not something that I like talking about because the second I put something out there about making money with AI, everybody sees it and all the alpha disappears. Um, and also it's just not something I'm super interested in teaching people how to make money. I'll teach you how to use the tools and then you could decide how to make the money. Patrick, what's up, Patrick? All right, we're going to get into this pretty quickly. Just give it another minute or two. Let folks stream in. All right, let's see. Can everybody see this? When AI builds itself. Perfect. It's not cursor 3. No, it's not composer. No, this is anthropic. Composer is cursor. Anthropic. Yeah. Uh, Osman Anthropic just crushed the AI industry. Absolutely. Yeah, this is live. CAD Mural, this is live. Yeah, buddy. Welcome. What's the RSI note? Never heard of that. Are all startups killed again, Greg? possibly. Let's find out. All right. All right. Let's start. Um, so Anthropic dropped this paper. And by the way, I'm going to be recording a video, so I'll be saying things multiple times if I need to. But yeah, Anthropic dropped this new paper. When AI builds itself, it is all about recursive self-improvement. And that is a very, very big and important concept. And that is what we're going to be going through today. This is their thoughts, Anthropic's thoughts on how they're seeing recursive self-improvement already inside of Anthropic and how they build their tooling and al also how they research and develop their next generation of models. you know, one model, Opus 47, helping research and develop Opus 48, 48 to 49 and so on. And what that means for society, what that means for safety, and of course, all along the way, they are doing their anthropic thing. They are doing their fear-based marketing. They're saying, "We can do it. You can't do it." And it's just so interesting. It is just so anthropic coded through and through. And we're going to review it all. So it starts out just straight to the point. We are delegating a growing share of AI development to AI systems. And there is so much to break down there. There really is because like when when you start offloading the development to AI systems, you become you as a human kind of get abstracted away further and further from the details, further and further from the problems. So keep that in mind and we we'll be talking about that later in the video. Roman, what's up? and taken to its logical conclusion, given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor. And immediately what comes to mind, at least for me, um, the situational awareness paper, the famous graph that I have referenced so many different times. That's not a good one. There we go. shrink it down. Uh, where is it? All right, this is it. But you all can't see that. Let me try to open this up. I'm looking for the Let's just I'll just find it on the actual site. Come on. I cannot believe I'm having this much trouble finding this thing in this moment. Come on. Base scale up of compute. Where is it? This is nuts. I can't find it. Okay. Um, it's this one. Intelligence explosion. There we go. Got it. Okay. And so when this paper came out, situational awareness, this was the graph that jumped out at everybody. And my goodness is Leopold Ashen Brener the ultimate predictor. He kind of foretold what was coming very early on. Very early on. And he like every single thing that he has said so far has come true. Especially look at the date. So what we're seeing here is as compute ramps up, that's what we're seeing over here on the Y-axis. And as time progresses, what we're going to get to is what he calls an automated Alec Radford. basically automated AI researcher and once that happens we get the intelligence explosion the or super oh aka super intelligence basically just an absolute explosion in the capabilities of these models because they are improving themselves. They are no longer bottlenecked by humans. They are no longer they are no longer bottlenecked but they are no longer bottlenecked by how much code a human can write, what research decisions a human makes. It is just off to the races and they don't actually know what that means. They being anthropic. So, let's keep talking about it. God, the universe king, thank you for the super chat. I appreciate that. Okay. And I I think that's also like they're again they're doing all the kind of very anthropic coded things in this essay like safety number one. We don't really know what's going to happen. Uh we'll get into all of it. Now, interestingly enough, they say recursive self-improvement is not inevitable. It definitely seems inevitable at this point, but there is one missing ingredient that we don't have today. And so, stick around because I'm going to explain what that is later in this video. And I want you to decide whether you think that it's inevitable that we're going to get that ingredient or not. And of course, they have to mention safety because they're enanthropic. Full recursive self-improvement also might increase the risk of humans losing control over AI systems. That's putting it lightly. I definitely agree this is a major alignment problem that needs to be solved. All right. Now, what they're going to show in these little graphics that they put together, which by the way, beautiful just top tier graphics on the anthropic team. Now they are going to show the progression of humans being abstracted further and further away from the actual development of artificial intelligence from the actual code being written the research being done everything and and it's wild to watch. So by the way 2021 to 2023 I see. Okay. Uh oh. Yeah. Uh let's see. I'm going to move. How's that, Alex? Okay. Yeah, the my camera is over. I know. Let's see. Let's see if I'll resize the fill canvas. That helped a little bit. More padding on the edge of screen. Scroll. Okay. No, but there's going to be a little graphics. I'm just going to go through this and I'll I'll move my camera as needed. What did Enthropic do? We're We're finding out together. Oh, RSI is recursive self-improvement. Of course. Yeah, thank you for clarifying that. All right, so first what we're going to see building the first claude. And look, this is just a few years ago. This is not that long ago. In the early days, work at Enthropic looked like work at any other tech company. You know, you're an engineer, you write code, and that code gets shipped to the end user. That's what it was like. So there it is. Okay, so you have the human developer writing directly into a computer and the computer is building cloud. Great. Looks like good old-fashioned development work. But things changed and they changed quickly. In the subsequent couple years, we moved to chat bots. Okay, this is chat GPT. This was the chat GPT moment. This is when humans started talking to a computer. The computer was really just a chatbot and that was building claude. Okay. Then then that is the point at which we really started seeing something different happen. The 2025 to 2026 era of coding agents. No longer was the human writing code directly. We had the person talking to a computer which was really a chatbot which delegated out to an agent which wrote code to build Claude. Now check this out. It's pretty seriously the graphics department in Anthropic is top tier. What we see here and if you look at kind of the density of the pixels and the claude logo it's um very very big. Then they start to get more dense, more pixels and it's exactly telling as to what's going on. As the human becomes or as the human becomes more and more abstracted away from the actual problem, it's able to we are able to actually produce a lot more code, a lot more research and that's what we're seeing here. So then we had autonomous agents. Okay. Um, let me just make sure the camera's in a good place. Okay. So, then we had autonomous agents. The human spoke to a computer, typed into a computer. That computer delegated out to a chatbot, delegated to an agent. That agent delegated to sub agents, to workers. Now all of a sudden we had major parallelization happening. That means a human typing one little thing, a little prompt could result in a tremendous amount of code. Now lines of code is not the measure. That is not the measure of quality. It is one signal to look at. We're going to get to that in a little bit. Okay. And then look at this. Once again, the Claude logo becomes even more dense with pixels. Now, here is where we close the loop. Okay, let me see if I can see this. All right, closing the loop. In the future, agents could become capable enough to build and train models themselves. If this happens, future versions of Claude could be continuously improved by Claude itself. And there is only one bottleneck at that point. Can you guess what it is? Because if you look at this little graphic right here, there is something missing. The human. In every previous iteration of this graphic and of the evolution of how humans work with artificial intelligence, there's a human involved. And then finally when we close the loop when we develop u when we develop recursive self-improvement that is the point at which humans are no longer involved and as you can see here the only bottleneck is compute. How much compute can you throw at the problem? How much parallelization can you have? How many how many agents are working on that problem? Okay. Super interesting stuff. Shout out by the way. Yeah. Shout out again to the graphics team at uh Anthropic. And if it's Claude, shout out to Claude. All right. All right. So, next what we're going to talk about is the progression. What they are actually seeing inside Anthropic as they're building these tools. They're developing Claude code. They're developing Claude AI. They're developing the actual model of Claude. What are they seeing? They actually shared a lot of really interesting information. Let's look. Of course, they use the word accelerating a bunch of times. The length of tasks that can be Sorry. The length of tasks that Oh my god. One more time. The length of tasks that they can reliably complete on their own that is AI agents has been doubling roughly every four months up from an earlier trend of doubling every seven months. That is the definition of accelerating. The acceleration is accelerating. That is kind of the strongest signal of recursive self-improvement that I've ever seen. Now, they are very clear in this article that recursive self-improvement is not here yet, but we are seeing hints of it and I've been talking about this for a while. You know, whether you're talking about the auto research project from Carpathy, which we're going to talk about later in this video, whether you're talking about meta harness, which was an incredible paper that I covered, which actually allowed the model and the harness to improve itself, right? These are all self-improving signals that we're seeing. Give me a sec. Augmentotos, what's up, man? Good to see you again. I literally use an RSI framework daily. So, yeah, we have it. Yeah, we do have it. Okay. Jonah just passed me this uh Ethan Mollik, famed AI researcher, professor at Wharton. I really think it is worth reading this piece on RSI, recursive self-improvement at Anthropic. There is a bit of naval gazing, some marketing and a lot of very sincere beliefs about what anthropic thinks is likely the future of AI that you will probably be aware of, you want to be aware of. that is. Now, this is exactly my point like kind of I kind of touched on all of these things. We're going to get into it. Thank you, Jonah. Um, okay. So, let's talk about the progression. March 2024, which is, you know, a little over a year from when Chat GPT was first launched. March 2024, Claude Opus 3 could complete software tasks that take humans about four minutes to complete. Okay, that doesn't dict that doesn't describe how long it takes the model, the AI model to actually complete that task, but basically what it's trying to do is set a standard of a long horizon task and how much it can complete. And what it's saying what they are saying is that if h if a human had a task that took them four minutes let's say you know writing a function um now opus 3 could do it or back then opus 3 could do it. Um and four minutes is not that long but again just like a year previous all we had was tab complete. So like uh a model being able to complete something that takes a human four minutes was actually a big deal. But we've obviously gone a huge amount, but we've obviously come quite far since then. A year after that, Claude Opus 4.6, which is crazy to think about. What? Oh, sorry, sorry, sorry. A year later. Yeah, that that would be crazy acceleration. A year later, Sonnet 3.7, which is interesting that they didn't mention Opus again, but Sonnet, their workhorse model, their slightly less capable, but much cheaper model, managed tasks that took about an hour and a half. So, it is now able to complete a much longer Horizon task that humans took about 90 minutes to complete. Then one year later, which is pretty recent, Opus 46 managed 12hour tasks. And let me tell you something, I've been finally uh coding and developing new projects and boy, I'll give some of the like codecs or cloud code. I'll give them a task and they just go off. I've had tasks run for 40 hours successfully. Like I can't wait to show you some of the stuff I've been seeing lately. more. I'm gonna create some videos on that soon. Okay, again, let's keep going down the progression. If this trend holds, tasks that take a skilled person days could come into range this year, 2026, this year. In 2027, AI systems could be capable of tasks that take a person weeks. I think we're closer than I think we are closer to to that. Sorry. I think we are closer to that than most people realize. I really do. Every single time we hit some milestone, the next milestone comes faster than even the absolute experts predicted. And that is what we continue to see. Now, they start to talk about something called corebench, which I've reviewed on this channel. But corebench tests the ability of a model to reproduce existing research. Why is that important? Okay. So if a researcher, a human researcher does all this work and writes a research paper, a white paper and describe a method in which to reproduce the results, the the ability of a model to be able to read the white paper and successfully re and successfully reproduce the results is an incredible skill to have. Now, it is missing one thing and they talk about this later. I'm going to give a hint at and I'm going to hint at what it is now, which is the actual task of discovering that novel research in the first place. The thing that the human did to write the white paper, that is something that the models are not really doing today. That is the missing ingredient. But if the model can read and understand novel research that likely is not in its training set, reproduce the results, that is a very valuable signal as to what's coming. So AI systems went from succeeding at reproducing the results roughly 20% of the time in 2024, two years ago, to saturating the benchmarks 15 months later. It can AI models can now successfully reproduce AI research paper novel ideas successfully nearly 100% of the time. Crazy. By the way, if you're watching this stream, I would very much appreciate if you liked the stream, liked the video, subscribed to the channel. Much appreciated. It helps get the word out there. So, thank you. Uh, Jean, you uh built a Golden Eye clone in 3JS. Very cool. Very cool. Okay, let's keep going. Thank you for liking the video. Uh, I just saw a bunch come through. Thank you for doing that. appreciate it. Okay. So then they get to what's actually happening within anthropic and there's again two pieces now and I've already kind of described the difference. There's the engineering piece which is how much code how much of the total code and how much new code is anthropic or sorry is Claude actually writing. That's the engineering bit writing code standing up the infrastructure and overseeing model training. That is likely what most of us, you, me, are are using these models for. Then there's research. And this is the part where models are kind of only scratching the surface of capability right now. Deciding what experiments to run, interpreting what comes back, and figuring out which ideas to try next. This is very important to keep in mind. This is the difference between having a new novel idea and knowing how to verify it. having taste in research, taste in a project that you're building versus actually going to execute that project, going to actually build it. We are at the point in AI right now in which the actual development is now the easy part. The ideas is becoming the hard part. I'm going to touch more on that in a moment. Uh yeah, we we've now gotten to the we've now gotten to the point we've now gotten to the point where Claude can be handed an underspecified problem and figure out how to solve it. Now they are still being we are still handing the problem to Claude. The problem is still the realm of humans to determine. But even a very unspecified, excuse me, even a very unspecified problem can be determined and developed by Claude. And I know firsthand that that is the case because I basically just say like my prompts are so underspecified. It's crazy. Uh let me know if yours are too. Like I literally just take a screenshot of an error, put it in cloud code, put it in codeex, and just say fix it. Okay? Like these are highly underspecified prompts and they do just fine. Never prompt like Matt. Thank you, Brian. [laughter] No, that's the way, man. Just screenshot, fix it. That's it. Yeah, Sakana AI. Um, they've been talking about self-improve. Sakana AI is a research lab out of Japan. Um, super impressive work. They're they're kind of doing their own thing over there. They have some really cool stuff, but they they've talked a lot about self-improving AI, automated research. They've done a lot of projects like that and and they've been way ahead of the game, too. Uh, Robbie, what are you reading? Is it a website? Yes, it's the Anthropic website. This is a paper that they just dropped. Okay. Yeah. Okay. Actually, I want to talk about this. Uh Whoops. Garrett, um computer use look at this error and fix it. You know what I've been noticing lately and what I've been doing especially with codeex is I'll I'll like give it a problem or or tell it to set up some infrastructure for me that requires going to let's say Digital Ocean or Verscell or something like that and it will pop open the browser and use browser use and click through and actually set it all up for me. I don't have to do anything. It's so lovely. Okay, let's keep going. Um, but large performance gaps persist when it comes to claude exercising judgment in choosing goals in both engineering and research. And I found this as well. If you've built any kind of project with any of the modern mar with any of the modern marble, oh my god, I can't say that. Modern models. Modern modern. Wow, that's a try saying that three times. Um, with any modern mortal, I can't say it. I literally cannot say it. Modern marbles. Mar. [laughter] That's right. Oh my god. I know this is going to end up as a clip from Brian. Okay, great. Um [laughter] uh if you've used any of these modern AI models, okay, that helps. Uh you've probably noticed that you give it a big problem, but you still have to tell it the problem and it goes out and builds it. But actually knowing what to build next, it's not really doing that today. Challenge. Say it five times fast. All right, I'm going to do it. Modern models. Modern model I can't do it. Modern models. Everybody, please try. Let me know if I'm just I have a lame if I'm lame. I just I like cannot do it. Uh okay, let's keep going. Um right. So like this is the type of uh prompt that I give. the export button isn't working. Please fix it. Now, this is, you know, pretty specified because you're saying this is the problem. The export button isn't working. Although it's short, like the please fix it, you're not telling it exactly what's wrong. Okay. But then the kind of the next step, the way you can also do it is investigate why the network slows down under heavy load. Here's the problem, and it's it's just a bigger problem. Um, but the thing that these uh but the thing that these models can't do quite yet is what should the team build next quarter? And I found that I said, "Okay, what do we build next?" And it kind of gives me some decent answers. But it's not great. And I I think that's probably a reflection of the fact that truly novel ideas are not coming out of these models yet. And that is the missing ingredient that I talked about earlier. Truly novel ideas by definition are not possible when everything coming out of the model is a derivative of the data that was put into the model. Right? It's such an important concept to note. Now will we get to a point in which they can come up with novel ideas with current architecture? Possibly. Maybe it does require some kind of new architecture. Maybe We're uh maybe LLMs just won't get there. We'll see. Um Okay. All right. So, here's an interesting thing. We probably already knew this, and I think at least if you're coding regularly, you're probably already feeling this yourself. But listen to this. As of May 2026, more than 80% of the code we merged into Anthropics codebase was authored by Claude. Now, this is super interesting. I just spoke to a CEO of a tech company, 300 people, and he tends to be a little bit more pessimistic about AI and its ability. And I told him and I was talking to him about it and I was like, "Hey, the vast majority of code being written by Anthropic is written by Claude." And he looked at me and he goes, "Yeah, that's why it's so buggy." And so like he he might have a point, but nonetheless, Claude is shipping the vast majority of lines of code for Anthropic, and they are moving faster than any other company on the planet, bar none. So before Claude Code launched in research preview in February 2025, this number was in low single digits. So low single digits up to 80% in basically a year. That is crazy. That is crazy to think about. That shift also shows up in the amount of output per engineer. This is where you guys, especially if you're traditional engineers, traditional developers, might start to cringe hard. Let's talk about it. Code contributed per person by quarter. Lines of code. Okay. The ultimate uh the the ultimate annoyance, the ultimate metric that really pisses engineers off. This is like the old school way where especially like 90s early 2000s you had non-technical managers managing developers and the non-technical managers would measure many engineers based on how many lines of code written and it is such a bad measure because it doesn't matter how many lines of code you write. It's about how high quality is the code, how readable is the code. Um, how efficient is it? Right? Does it actually get the job done? In fact, there's an argument that less lines of code that can achieve the same outcome, achieve the same output is actually much more valuable. And so again, they do say this, they do anthropic does caveat. Okay, so caveat bienthropic lines of code is an imperfect measure. That is an understatement to say the least as it measures quantity over quality. But it is one signal. Okay. And it's interesting to see this acceleration. Look right around Q3 Q4 of last year. That is when their lines of code per engineer absolutely exploded. And that is when the vibe shift in agentic engineering happened. And if you've been following AI for at least, you know, 6 months, nine months, you you felt it. You felt something happening in, you know, October, November, December of last year. Something changed. And there were really two things that changed. Opus 4.5 came out and GPT5 came out. Both of these models were better, like significantly better at coding than all previous models. There it was a step change. It really was. And so we saw it from Q4 to Q1, we had a massive increase, two and a half times the amount of code to 5.8. Now, what's interesting is they're using Mythos internally and and I think about why they might be doing that. I think about this is like there's so I have so many thoughts going on right now. I'm going to try to break it all down for you. They have been using Mythos since Q1. They are using this frontier model and they did not release it publicly. Now, this is the most anthropic coded thing I've ever heard because first of all, there was Let me see if I can find it. Anthropic cut access to XAI. Boom. Uh, shout out Kylie, uh, who broke this story, I believe. Um, but this was January 9th, okay, two, 2026. XAI staff had been using Anthropics models internally through cursor until Anthropic cut off the startup's access this week. They said, "No, no, you can't use our models to code your models." Which is insane to me. That is insane. That is massive platform risk. Okay. So they they uh it was reported Anthropic basically banned their competitors from using their own models and at the same time oops and at the same time they were starting to use Mythos internally. Now what happened again? Rather than doing the same thing rather than releasing Mythos and then having to like turn it off at XAI, turn it off at you know whatever other competitors there were out there instead of doing that they simply didn't release it. They're using it internally. They're accelerating their code development. They're accelerating their AI research internally. They're basically trying to and I know everybody loves this, but win the race. They're trying to do this and they did it without kind of looking bad and and it's like they it's so self-erving. This, you know, Mythos is so scary. We can't release it. But we're going to be using it. By the way, don't worry. We got you. And we're going to be using it safely. We're going to be we're going to be using it to accelerate our own AI models. But huh, also that lets us get further ahead. Hm. Well, we didn't intend that. Don't worry about that. Don't don't think about that too much. It's crazy. Um, and so I really think that's what's happening. And so they got the fear-based marketing going. They got to accelerate their own development while all while not sharing any of that with uh the broader market. Cool. Thanks anthropic. Now uh here here is important and I agree with this. Eight times the number of lines per code per engineer per day in the second quarter of 2026 is almost certainly an overstatement of the true productivity gains. that statement is doing a lot of work and I'm not sure everybody realized it because in fact when myself and my team were going over this I don't think I realized it and I I think it was Jonah who pointed it out. He was like that actually means that the code being written by Claude is worse. It's worse than human written code. And maybe that's true. there's actually a strong likelihood that's true. There's also an argument that they simply don't know what to do with all of that code. And this is something that I talked about in a previous video, I think actually my last video, where it's like you can write as many lines of code, you can develop as many features as you want. And and if all of this stuff happens at an accelerating rate, the bottleneck becomes everything else involved in releasing a new feature. the marketing, the sales, the adoption, the documentation. How does all of that? You have to like in in any machine, in any system, when you unbottleneck one part of it, you expose the next biggest bottleneck. And that's what we're seeing here. That's probably also in conjunction with Claude not writing as good as good as code as a human can write. But it's probably also they're not able to actually deploy that code as successfully. Think about this. Imagine you have a car with a 1000 horsepower, but you're only able to get 500 horsepower to the tires and then the tires are basically smooth and they can't get traction to the road. That's basically what we're seeing here. All of that horsepower only goes so far if you can't actually deliver traction to the road. So in March 2026, a poll of 130 employees from acrossanthropic research teams, let me just make sure the recording is looking good. Yep. Okay. And you guys can see all the text on screen. Okay, cool. Um so uh in in March 2026 in a poll 130 employees from across anthropic research teams the median respondent estimated that they produced around four times as much output with mythos preview as they would have without access to any AI model. That is a significant improvement. Four times your productivity. But remember, they're producing eight times as much code. So that code is half as valuable as purely human written code. Now overall, they're still much more productive as they would be otherwise. And that actually just shows me that with a little bit of tweaking and finding other areas to remove bottlenecks from, they're going to be so much more productive. They even if the model capability stops today, they still have room not only to uh deploy the model, increase adoption, improve the scaffolding, the harness around the model, but actually uh unblock other areas of the business to deploy the code effectively. So this code mythos preview is 50% less productive than human code. Very interesting. A significant fraction of anthropic technical staff is accomplishing their core work multiple times faster than they could without AI assistance. And actually we didn't even talk about this with my team, but what does this tell me? If all of a sudden the development team, the technical staff is able to produce so much more code and develop features at such a higher clip, then they need more people to market it. They need more salespeople to go sell it. They need more customer support people to service their customers. So that's kind of an argument against this kind of AI is going to have this huge job apocalypse, right? There is a strong argument that as productivity increases, so does the need for humans. And we're going to continue on that thread throughout the rest of this paper. You're going to see multiple examples in which this is just shown with data. All right, let's keep going. So um ah here's another important part on that thread. We are we also see evidence that people at anthropic are using claude to do work s to do work that simply wouldn't have happened otherwise. It's not just that they're able to automate things that they were already doing. It's that they are doing new work, net new work. That is very exciting. That's the frontier. A lot of people are just thinking about how do I automate the stuff I'm doing. So, you know, like kind of unfortunately there are some companies that are thinking, let me automate this thing so that I don't need as many people. They're going to lose. Those companies that have that point of view, that have that posture towards their employees are going to lose. So, let me see where I was going with that. Um, Aaron Levy has talked about this. Shout out to Box, uh, sponsor of my channel. Just a great partner, so shout out to them. Um, net new work. We use them like crazy, by the way. Um, let's see if I can find one of his tweets. He talks about this all the time. I don't know if I'm going to be able to find the tweet right now. Maybe uh somebody from my team can help me. But he I I'll just describe it. He basically talks about net new work. He's like the old way of thinking about it. the old as it being like two years ago is like okay what work do we already do that we can automate the right way to think about it is what work would we not have done because we don't have these tools what net new value can we be driving that's the exciting part to me uh here's a an anonymous anthropic employee now been five months since I last wrote any code myself five months not a single line of code. Now, the part that they're not describing is the actual reviewing of code, and they're going to talk about that later. Are humans reviewing code? Is it even possible to review AI written code? Because if AI is writing eight times as much code, you need technically eight times as many people to review that code. And I guarantee they're not doing that. So, how do they do it? Well, of course, they use AI. We're going to get to that in a moment. Uh Brian, there was um a tweet and you know Aaron Levy is extremely prolific, so it might be difficult to find, but he basically talks about how AI I think traditionally is being seen of like how do I automate existing work that we're already doing versus what new work can we be doing because of AI. I think Jonah just posted it. Let's see. Let's see if this is it. Oh my god. Yeah, he's basically saying when was this from Oh, this is re Oh, this was yesterday. Oh, we're on the same page, Aaron. All right. Um, okay. We're going to go over a couple of Aaron's tweets right now. Uh, by the way, great follow if you want to learn about how AI is affecting the enterprise in the greater economy. Um, so he's talking about jobs data. Uh, I actually covered a lot of this in my recent video, but what we're actually seeing is that no, jobs are not being automated away. There are new things to be done. And we've been talking about that on this channel. And if you want to have known that a long time ago, follow now. Follow the channel. subscribe to the channel because we've been talking about this for a while. So, the jobs data coming out uh the jobs data coming out continues to suggest the opposite of what a lot of people had thought would happen. Just take engineering as the prime example of the area with the greatest AI impact. Most companies now have far more software projects than ever before. Okay. um uh you can get by. So a lot of nontechnical people building software but eventually someone has to understand what the thing is that got built has to maintain it has to fix security issues that come upgrade the system beneath it and so on. That's all jobs agreed. Then you need to hire more in sales because agents can let them process more leads. You need more customer you need more customer success. You need more customer service, new marketing roles because you are now launching 10 times as many products. So, yeah, this is literally what I just said. Let's see. We got one more uh tweet right here. Okay, I think it was a while ago. That's okay. I don't need to find it anymore. Basically, that's what he said is like net new work is where the value is. Okay, so here's Claude Code Sessions success rate. By the way, uh I just saw somebody give a thumbs down on this stream, so let's counteract that. If you can give a thumbs up, I would very much appreciate it. Um okay, so Claude code session success rate. Um, thank you for the thumbs up. Um, okay. So, trivial task about the same. Here's where it's super imp uh here is where we've seen the biggest improvement. [laughter] Um, okay. Open-ended problems and substantial tasks. I love it. Yeah. Yeah. Yeah. Yeah. Okay. Uh, now I know. Don't ask for thumbs up, thumbs down. Yep. Got it. Um, yep. That backfired for me. Brian, you gave the thumbs down. Thank you. Okay. Um, so what we're seeing here, and again, this is all around the last quarter of 2025. This is where we really saw the improvement. This is really where Agentic Engineering became crazy valuable. Okay, so right there we're seeing it. Now, here's something super interesting. Okay, we we talked about how do you actually review AI code? How do you actually do it if it's producing so much more code? How do you do it? So session success is determined by a claude judge. This is crazy to think about. Okay, as I said earlier in this video, the the human in the loop is becoming more and more abstracted away from the core problem. And so we have AI writing code and now we have AI reviewing code and we're just getting further and further away from understanding the details. And that actually reminds me of another tweet which was like I I think about this daily this tweet. Um no I think it was what was his name? Gosh what is his name? Oh right yeah Carpathy even tweeted it. So uh cash you can outsource your thinking but you cannot outsource your understanding. So as we become kind of more disconnected from the systems, more disconnected from the details, how do we maintain our understanding? You can offload your thinking, meaning you can have AI build the systems, uh determine what research projects to take on, actually go and execute the research projects, but ultimately a human needs to understand it. If they don't, that is the recipe for AI misalignment. And if they do, humans become and may are continue to be the bottleneck in the entire system. And as long as humans are the bottleneck, then it is the kind of recursive self-improvement is rate limited by human cognition. All right. Um, cool. By the way, uh, we're streaming to Twitch now. We basically have no following there, but if you enjoy Twitch, if you prefer Twitch, you can check us out there as well. Um, okay. So, I I found this to be very interesting that Claude is the judge. Very very interesting. So, we we talked about that discrepancy between the increase in the number of lines of code and their perceived output increase, the value, the actual productivity as the like perceived by anthropic employees. And we noted 8x as much code but 4x as much productivity which of course leads us to believe the code being written is not as valuable. It's not as good. So they're writing this they're writing more code to accomplish the same amount of value. And listen to this. There isn't full consensus among staff at Enthropic, but many believe that the clawed written code was still worse in quality than human written code at Enthropic in late 2025 and is still and is roughly sorry and is roughly at par. We expect it to be better within the year. I sure hope so. So that's the number. As there is 4x as much output or sorry as there is 8x as much output in code there should be 8x as much output in value. Now that number is never going to match because every other part of the system from documentation to graphics, marketing, sales, customer success, customer support, all of that needs to also keep up. And the the the vanguard, the tip of the spear is the code being written. All right. Um, okay. So, this is really interesting. Now the next thing that they talk about is how Claude and AI models in general are accelerating research. And we already talked about that research decisions, the novel ideas of what direction, the taste, what direction should we head in, what should we experiment with? That is still the realm of humans. So they developed a test, a miniature version of an experimental research loop in May 2025. Oh, sorry. Uh the the job given is to find speedups by rewriting the code, running it, timing it, and repeating. So they're basically looking for ways to improve the latency of code. In May 2025, Opus 4 averaged 3x speed up over the starting code. In April 2026, brace yourselves, Mythos preview was achieving 52x 52x speed up from one year prior, less than one year prior of an average of 3x speed up. Now, how does that compare to humans? Well, they told us a skilled human researcher would need 4 to eight hours to reach 4x. So although AI today is uh there it aren't developing they aren't coming up with new and novel ideas for areas of research or at least not successfully. They are accelerating humans. They are allowing humans to be much more productive in deciding, designing and executing experiments. Right. Ah, and interestingly enough, Andre Carpathy released a project that went mega viral just a couple months ago called Auto Research. So here March 9th he talks about auto uh wait yeah so he talks about auto research and this is a project to train a small model I think it was a GPT3 class model but to do so as quickly as possible and he basically just gave AI an overarching goal kind of directionally let it come up with its own experiment ments execute the experiments and then self-improve and he was incredibly successful with this and it was a very very interesting project and if you're wondering oh Andre Karpathy I've heard that name before where have I heard that name before well he just joined anthropic that was major is okay. Major major news. 27.4 million views worth of major news. Personal update. I've joined Enthropic. I made a whole video about this because it was mindblowing to hear. Reminder, Karpathy was an OpenAI co-founder decided to go to Anthropic. I think he saw something. All right, let's keep going. Right now, they do anthropic does say Claude is getting better at proposing its own experiments. And when you have unlimited compute, when you have the ability to parallelize to incredible numbers, it's kind of like that uh that notion that if you give uh you know a million monkeys with typewriters would be able to come up with uh the the best you know works of writing in human history. It's kind of like that. If you give them enough compute, if you let them come up with enough experiments, it doesn't really matter if they have good taste or not. Taste is important when you have limited resources. Now, obviously, companies still have limited resources, but still, if we have unlimited resources, taste does not matter anymore because they will just figure out, they being the AI, will figure out everything to test, test it, and just find the best. But it is getting better. It is getting better at proposing its own experiments. But for now, the taste of deciding which experiments to run is the realm of humans. Now, it took me a little while to understand what this chart meant, so I'm going to try to break it down. Basically what they're showing is they took past uh experiments to try to let's say improve AI or you know let's say make it faster make it more make it better improve the development of it and they looked and these are experiments that had already uh occurred already happened and they basically applied AI to it to try to see okay where the human failed when the human made a poor decision about the direction of an experiment. If we applied AI, would AI have gotten it right? Now, Claude Haiku 3 back in March 2024, 22% of the time it would have done better. Fast forward to Claude Mythos preview, it is now at 64% of the time. So there is this notion that claude or AI in general is going to be able to decide directions or decisions later in the experimentation process better than humans can. So we view this result as an early signal that AI systems are getting better at making the kinds of judgment calls that AI research depends on. Now, if you've been using artificial intelligence at all, you probably know this. You probably see it. The comparative advantage of humans as of right now is still in seeing the bigger picture and thinking beyond the confines of the immediate task. This is prompting and verifying. I also want to bring up Balagi uh you know one of the smartest big thinkers on the planet talked about this a few months ago and this is another one of those posts that I just think about all the time. The new bottleneck on AI is prompting and verifying since AI does tasks middle to middle not end to end. Okay, so prompting and verifying that middle to middle section is what AI handles really well and accelerates really well. And that middle to middle is expanding outwards. But the ends are still necessary. The ends are still the realm of humans. All right. Let's see. Let's see. Let's see. So, what does this actually mean? What does it mean for humans, especially humans at anthropic, which they're going to talk about, but humans in general, because you can extrapolate what they're learning from how humans and AI are interacting to develop software and develop AI systems to broader knowledge work. So, let's see. The evidence suggests that the human role is narrowing at each step in the AI development process. Once human and AI authored code quality reaches par reach par humans will stop writing code entirely and shift to only reviewing it. Now here's the important part that I mentioned earlier. But if they can't review code as quickly as cloud can generate it, human review will become the bottleneck to AI development. Let me read that one more time. Uh human review will become the bottleneck to AI development. I want to click on this. So this is something that we talked about as a team on my team. There was this notion for a long time in Silicon Valley that ideas are cheap, right? Anybody could have their the next billion dollar idea. The hard part is actually going out and building it. And not just the code, but recruiting, marketing, sales, fundraising, the the kind of the grind, the years and years of grind a founder needs to go do to be successful. the idea changes over time. The idea, you know, it's called a pivot. It's kind of a a very normal thing to do in Silicon Valley, a pivot. You have one idea, didn't really work out how you thought, and thus, I'm going to pivot to this other direction. But here's the thing, execution is becoming easier now. Execution, which was the hard part, can now be done by artificial intelligence. at least so far the development of code. And so does that make ideas even more valuable or actually valuable? Because if you have a great idea and you could just push a button and let the AI go and execute it all, sounds like ideas are the important part. interesting to think about. But they still say research, taste, and judgment, including choosing which problems matter, which results to trust, and when to approach and when an approach is a dead end, is still the realm of humans. All right. So, next they talk about what if they're wrong. What if all the things they're talking about self-improving AI is wrong or it just doesn't happen. So, it is unclear today whether the today's training methods and architectures could unlock that capacity. That means kind of endtoend recursive self-improvement. And this is kind of the idea versus execution trade-off that I mentioned earlier. Edison said that genius is 1% in 1% inspiration. That's the idea. And 99% perspiration. That's the execution. But we see perspiration becoming increasingly automated. What a what a wild time. every notion that I had kind of let's say quoteunquote growing up in Silicon Valley is is like just being completely like flipped right in front of my eyes. It's it's so wild to think about. So even if we suppose that Claude never achieves good taste, good research taste, a conservative reading of our evidence still implies compounding acceleration. So it just means if humans are forever prompting and verifying that prompting and verifying is going to have uh kind of this massive multiplication effect. Humans spend most of their time on the singledigit fraction of work that is direction setting. And I believe that's going to be the case for a long time now. in domains that are fully verifiable. That's when it's like, okay, infinite compute kind of removes the necessity of taste, but we're never going to have infinite compute. It's it's like at least not for the like a very long time because we're still bottlenecked by first of all just the production of silicon and data centers and most of all energy. You know, unless we somehow discover unlimited energy, we will never have unlimited compute. Okay. So then they talk about the three possible futures. Number one, the trend stalls, but today's AI capabilities are widely diffused. Okay, that is assuming we have this S-curve curves up, gets really exciting, and then all of a sudden just flattens out and we do not get any more progress. That is one potential outcome. But they say that is unlikely in their opinion. And something I've been saying for a while, even if model capabilities were frozen at today's level, we would expect major changes to occur in the world that is the capability overhang. That means the models are so capable everything around them has not caught up yet. And that includes the scaffolding, the harness, that includes the other area, sorry, that includes the other areas of business and society that can't that haven't even caught up enough to leverage all of that additional code being written. Okay. Yeah. So they say right here, we include this scenario for completeness, but we don't believe it is likely. Second potential future AI labs continue to see compounding efficiency gains. AI development becomes substantially automated but humans continue to set research directions and judge results. Okay. So this is the outcome in which endto end does not happen. Humans taste judgment that is still critically important for the foreseeable future according to this outcome. But it does mean that a 100 person company can do the work of a 10,000 or even a 100,000 person organization. That is major productivity gains for the world. Very exciting future. But and this is what I talked about speeding up one part of the process often just shifts the bottleneck elsewhere. Overall pace is capped by the parts that haven't sped up. Exactly what I said. Okay. Now, the third outcome, and this is both the most exciting and also the most scary. AI systems themselves become capable of full recursive self-improvement and begin building their successors. I'm going to pause for a second. Let me know what you think. Which of these three outcomes is most likely in your mind? One, two, or three? One, the trend stalls and whatever we have today is what we'll have for the foreseeable future and we'll get more out of it, but basically the innovation stops. Two, we have massive productivity gains, but humans are still in the loop. And three, it is full endtoend self-improvement. Recursive self-improvement. Tell me what you think. One, two, or three. Let me know in chat. And why? I want to know why. If technical trends in advancing capabilities continue and AI systems are able to develop the capabilities inherent to transformative human ingenuity, then it is plausible that AI systems could design and refine themselves. That is recursive self-improvement. That is the intelligence explosion. Super intelligence. Whatever word you want to apply to it, that's what they're that that is what they're talking about here. I'm seeing a lot of threes in chat. Some twos, some threes. But why? Tell me which number, which outcome do you think is most likely and why? I'm genuinely curious. CG flows two, ratty whale three, effects three because it already is starting. Yeah, totally one. We have a one out there. Okay. And this continues in this world pace of progress and AI development becomes determined entirely by the availability of compute or the speed or of the speed well actually not even that for AI systems. Now boy does that sound like something who has heard the term permanent underclass. This is the concept that once we hit recursive self-improvement of models or even AGI, wherever you are in the societal classes, the the class structure in society, wherever you are, that is where you are forever. If you've made it above the permanent underclass, you're good. If you're below it, sorry. Now, there's a lot to unpack there and I think this is probably pretty reasonable because if the po if we get to the point at which AI can do endtoend recursive self-improvement, humans are no longer needed. The only bottleneck left is compute. And if we say compute, really what we mean is energy. And thus, how do you acquire compute? Well, you need capital. You need to buy it. And so at the moment in which recursive self-improvement happens, whoever has capital at that moment will just buy up as much compute as they can. They will buy up as much intelligence as they can. And if you have capital, you're you're in a good place. But if you don't, you're not. And that is the whole notion, the whole concept of the permanent underclass. That is a scary future. It really is. It might even be at the point at which that happens. It's not about capital. It's whoever had the compute, had the energy production mechanisms to begin with. Whoever had it at that moment, it's just all of society freezes. Now, here's the scary part and very anthropic coded. We do not have good intuitions for what this world would look like because our economy is currently driven by humans and human-built tools. How would they know? What does it look like if you have infinite intelligence? What about if you had embodied intelligence aka robots and infinite robots, infinite labor? What happens in that world? Anthropic does not claim to know. And this is, you know, they might be genuine in their admission that they don't know, but that is also very much fear-based marketing in my mind. They talk about embodied intelligence robots right here and they think they expect that robots might quickly follow recursive intelligence and follow a similar path of increasing returns at decreasing cost. Now they say something here that I don't necessarily agree with. Achieving recursive self-improvement alone does not suggest an immediate change in how industrial production occurs. societies organize or markets function. More intelligence can't learn what a drug does over decades of use. Is that true? What about if we could just run simulations, unlimited simulations? Can't we predict with pretty darn high accuracy what the what a drug will do over decades of use? I don't really understand that. All right. Now, here is the part in which, brace yourselves, anthropic does anthropic stuff. Okay. This is Yeah. Someone Wall Wall. Uh, the bitter lesson. Yep. Yeah. Uh, Guy Smiley, the AI is on drugs now. Yes. uh just walked in. Give me the headline. Um we're talking about a new anthropic paper that just dropped. They're talking about recursive self-improving AI. Okay. Um All right. So, here's where I'm going to get a little angry because this is just so anthropic coded it hurts. If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing. They are saying, "We think we should slow down. We think we should slow down." Now, they're going to go on to say, "We'll only slow down if everybody else slows down," which is completely understandable, logical even. But that's a nice thing to say when you're in the literal lead, when you are winning the race. If you're an Olympian and you have 10 other competitors that you're racing the 800 meter dash on and you're in first place halfway through and you say, "Hey guys, why don't we all slow down? Equally slow down. You're always going to be in first place. Of course, you would want to do that." But I guarantee XAI doesn't want to slow down anymore. I guarantee other countries don't want to slow down anymore. But Anthropic, so nice of you to say that you think we should slow down because you're in first place. Great. And they go on to talk about it pretty pretty accurately. But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe. So everybody's got to slow down to the pace that Anthropic dictates. And they're right though. If Anthropic being in the lead slows down, I guarantee the third place, fourth place, fifth place, they're not going to slow down. Other countries, they're not going to slow down. That is not how human nature works. Without a global coordination, excuse me, without a global coordination mechanism, companies and governments will have to make difficult decisions about safety while under competitive and geopolitical pressures. We believe it would be good for the world to have the option to slow or temporarily temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology fearbased marketing. I just don't get it. It it it hurts to read this stuff because it's like it's so self-serving from anthropic. They can they can be the good guys by saying, "Oh, well, we we said, you know, we said we should slow down. We said it. Look, we can point back to that essay that we wrote about recursive self-improvement. We said we should slow down. We should give society a chance, but nobody said but but nobody agreed with us. You know, we were in the right. We were the good guys. We had all the morals, you know. No, it was everybody else. Boy, that's a nice position to be in. I guarantee you they would not have said that if they did not have the absolute frontier of AI right now. But they do say it's not impossible for the entire world to slow down. It would be extremely difficult, but not impossible. It would require multiple well-resourced labs at near or uh sorry at or near the frontier in multiple countries agreeing to stop under the same conditions. It would also require that each can verify that the others have actually stopped. How do you even do that? They go on to say it's actually much easier to determine if nuclear proliferation, if nuclear development has stopped or continues than it is to determine whether AI systems are continuing to be developed or not. The detectability element of this arms control problem is much more challenging than with other technologies. Training runs are far easier to conceal than missile silos. None of this is impossible in principle. They here again, here's some more fear for you if you didn't have enough fear from reading all of this. Here's some more. Those regimes those regimes like uh kind of um detecting whether other countries were building new missiles, new nuclear systems. They took decades to build both the infrastructure and the trust. We don't have that long. AI is developing and evolving faster than any of those other technologies. Great. And yeah, they're saying a unilateral pause by one lab would be basically useless. That lab would pause, everybody else would catch up and excel. So, that's it. That's the essay. Uh, what do you think? I would love to read some of your comments right now. If you want to drop some comments, let me know what you think about this because I thought this was fascinating. A lot of what we knew already um a lot of new things, a lot of kind of signals from inside Anthropic that show us how they're thinking about it, how they're progressing with, you know, just code code written code development research on AI models. Let me know what you think. Let me know what you all think. Okay, I'm going to stop my recording and start uploading that. Yeah, I'm going to give you all a minute uh if you want to ask any questions. Um if you What is going on here? Yeah, I want to know your thoughts on this. I'm just going to upload that video uh very quickly. So, just give me a moment. I'll give you a sec. All right, uploading. Let's see some of your comments. Are we doomed? Um, no. I mean, I don't know. I like to be optimistic because otherwise I'd just be depressed all the time. So, I'm just going to say no. We're not doomed. Self like fully endto-end self-improving artificial intelligence does seem very possible. It it seems also like in the kind of the mid the the short and midterm we're going to see humans in the loop which is not only a good thing for uh kind of AI developments but also a good thing for the economy. Are they saying we should all stop using claw to slow them down? Exactly. Do you see any viable routes other than LLMs actually showing promise? No. Um there's like some early work in world models. Fe Lee just put out a great essay. Um I haven't read it in full, but just describing how they think about world models, you know. Um oh my god, I'm forgetting his name. the previous chief AI scientist at Meta um Yan Lun he you know he talks about like Jeepa and needing some new architectures new technology new innovation um Ilia Sutzkver same thing yeah thank you Jonah Lun Ilia Sutzkver there developing safe super intelligence another open AI co-founder um he [snorts] believes we need some new innovations new research thank Pro a productive dude. Yan lakun. Yes. So very very interesting. Um couple other things I want to talk about real quick. Um oh my god. Look. Look. Be Jazos right here. Literally what I called. This is why they didn't release mythos. incentivized to compound their lead rather than make models available to their potential competitors. Almost word for word what I said completely accurate. Um there was something Oh, I made a website I want to show you all. It's just completely unrelated. Um, this is a new website I built. I had this idea Theo um like over the last like day or two said, "Hey, I'm starting to see some issues with codecs." So that leads me to believe they're going to reset the quotas. And I was like, "Oh man, that's really cool. What if we could predict when they were going to reset quotas?" So, I built this website using codeex last night and it's a prediction of will codeex get reset? Will the quota get reset? Codeex never resets. It literally reset yesterday. Productive dude. Literally yesterday it reset. So, it takes in a number of different factors. Let's see if I see if I list it. stuff like vague posting uh upcoming parties or potential releases from OpenAI issues. So, OpenAI status. So, it looks at this and it and it tries to put together and or it did it put together an algorithm to try to predict when codecs will be reset. So, go check this out. Willc Oh, sorry. Willcodarreset.com. I'll drop it in the chat right now. It's It was just for fun. Uh, that's all. Um, MLE Grand, it'd be nice if we can get an alert of codeex resets as well. Uh, totally, totally. I think it's it's like it's cool to get the notification of the reset. I think I also want to add in notification of um when you know certain thresholds happen, like above a 70% chance of reset, then you get a notification. I think I'm going to do both of those. Um, I'm just taking a note to do that. Great, great call. What does it mean? reset live streamer that just so you you get a a quota with your subscription to open AAI you get a quota and as you're using that quota sometimes quite often actually Tibo the guy who leads Codeex uh reset your quota which means you just get a lot more usage you get a lot more tokens out of it um Garrett totally or Gareth sorry Gareth if I make money from Koshi using your website I'll let you know cool uh 10% cut. Um I've been using a lot of codecs lately actually. I've been been having a lot of fun with it. Yeah. And so like the funniest thing was when Anthropic was cutting all of their quotas, reducing their quotas, at the same time Open AAI was increasing and resetting the quota. I love it. Um anyway, so go check this out. Um, willcodexquotarreset.com. Go check it out. Um, cool. I think that's going to be it for today. Um, thank you guys for watching. Thank you for joining. I really appreciate it. This is always fun. I love seeing your comments. I love seeing you guys show up. Uh, and when I start recognizing you guys and your like your names, the profiles, everything, I really it's it's a lot of fun for me. So, thanks for hanging out with me while I'm recording this video. Um, one last promotion, one last thing. Uh, go to forwardfuture.ai. That is our website. It is our newsletter. It is something that we're putting a tremendous amount of investment into, both time and money. We're developing our own I I'll actually show you right now. We're we're developing our own original content right now. We're working with other people. Here's one from Ahmad Osman, who is kind of the local inference god. Uh, we have one about UBI from Scott Santins, an absolute expert at universal basic income. We have one from friend of the show, Dave Shapiro. Uh, so go to fordfuture.ai. I'm so proud of the work we're doing there. I would love your feedback. Drop comments, subscribe to the newsletter. Thank you so much. Um, productive dad, thank you for the $10. Really appreciate that. Thank you very much. And um, I'm going to drop it uh for future.ai in the chat one more time. Thank you very much and I'll see you on the next stream. Bye everybody.