Karpathy + Anthropic is INSANE! 🤯

Andrej Karpathy just joined Anthropic. I want to tell you why this is one of the most important things to happen in AI in 2026 and why most people are completely missing what it actually means. So everyone's covering this as a talent story. Big name joins big lab, that's it, that's where they stop.

But there's a deeper story underneath this and it has massive implications for how you use AI in your business right now. So stick with me on this one because by the end of this video you're going to see something most people simply aren't seeing yet. First of all, who is Andrej Karpathy? So because if you're not like from a technical background you might not know his name and you should.

So Karpathy is one of the founding team members at OpenAI back in 2015. He ran AI at Tesla for five years, came back to OpenAI, left again, then started his own AI education company called Eureka Labs where he made free courses teaching people how AI actually works from the ground up. He coined the phrase vibe coding which is basically what most of us are already doing. You describe what you want, AI builds it, and you steer the direction.

This is someone who spent years shaping how people think about and understand AI. He's not just a researcher, he's actually a translator, he takes the hardest technical ideas and makes them feel obvious. And on May the 19th, 2026, he announced he's joining Anthropic. Now here's the thing that jumped out at me the moment I read his announcement.

His specific role. He's not joining to be a spokesperson, he's not doing marketing or educational content for them, he's going into the pre-training team and his job is to use Claude, Anthropic's own AI, to make further versions of Claude smarter. So if you think about that, he's using AI to improve AI. That's the story and it's a much bigger deal than most people realize.

Now let me explain why. Before Kapafi joined Anthropic, he spent the last few months building a project in public called Autoresearch. And this is where things get really interesting. Autoresearch is basically an AI agent research loop.

His artwork's in plain English. So the AI proposes a change, it tests that change, it checks the results. If the change made things better, it keeps it and runs the loop again. If the change made things worse, it actually undoes it and tries something else instead, over and over again, whilst you're asleep, whilst you're making coffee, or whilst you're doing anything else.

Kapafi ran this about for two days, I think it was two days straight. During that time, the agent ran 700 experiments on its own. It found roughly 20 improvements that could all be layered on top of each other. The result was roughly an 11% improvement on speed on a standard AI training benchmark.

Now, 11% doesn't sound crazy, but here's the thing, this was running on a home computer with a tiny model, built in about 30 lines of code. That was the test version. Now imagine that same approach running at scale inside one of the most well-funded AI labs on the planet, aka Anthropic, with access to massive compute, with Claude already one of the most capable models in the world running the experiments. That's what Anthropic just bought in-house.

And here's the open loop I want to leave you sitting with, which is, if AI can now run experiments to improve AI faster than any human team, what does the next version of Claude actually look like? We're going to come back to this. But first, I want to talk about the wrapper, because this is the part that matters for you and your business right now. Most people still think about AI the wrong way.

They think the model is the whole game. Claude 4 versus, you know, GPT-5 versus Gemini, which one is best, who won the benchmark, which one should I use? And yes, the model matters. But the longer I use these tools, the more I realize the model is actually only one layer.

The thing that actually changes your results day-to-day is what wraps around that model, right? And Kapafi has been talking about this for months. He's the person who started calling it context engineering instead of prompt engineering. And the difference is huge.

Prompt engineering is writing a better question. Context engineering is building the whole environment that AI lives in, right? So, for example, your files, your examples, your workflows, your histories, instructions, your style guides. Here's a simple way to think about it.

So you can open a fresh chat with Claude, he knows nothing about you, your business, your customers, or what good looks like for work, right? It can still help you, but it's guessing. You end up explaining things over and over, and the results are pretty generic. But now, imagine Claude has your meeting notes, your past client conversations, your SOPs, your brand voice, your best performing content, your offer details.

Now it's not just guessing, it's working with real context. Same model, but a completely different output. And that gap between Claude with no context and Claude with full context is the biggest opportunity in AI right now. And it's mostly being ignored because everyone's focused on which model to use.

When the real question is, what context have you built around the model you're already using? Kapathi has been obsessed with this. His LLM Wiki project from April 2026 was basically a system where an agent takes all your raw documents and builds a living, connected knowledge base out of them. Not just folder files, but a structured, searchable, relational knowledge base that the agent can actually navigate and use intelligently.

Think about what that means, for example, for your business. You've got your client notes, your past proposals, your email threads, your training documents. All of that gets turned into something Claude can actually work with, instead of just sitting in a Google Drive that no one looks at. And this is the direction Claude code is already moving in, right?

And with Kapathi now inside Anthropic, this is going to accelerate fast. Now, Kapathi has been teaching this publicly, and it's what the best AI users are already doing, right? They're not prompting Claude. They're building a context layer around it.

A real system, their business information, their client data, their workflows, all fed into one place, so Claude actually knows what it's working with. And that's exactly what the agent operating system inside the AI profit boardroom is built for. It's a full operating system for your AI, and Claude plugs straight into it. You feed it your business context, your SAPs, your password, your goals, and Claude then knows your business.

All your agents know your business. It generates content in your voice, follows up with leads the way you would want to, writes proposals with your actual offer details, not generic outputs. And because the context layer is already built in, every new Claude feature that ships, including everything Kapathi is now building, makes your whole system more powerful automatically. You get the full setup, the prompts, coaching calls, where we walk you through the whole thing, and 3,000 business owners in there already running it.

Link in the comments description or go to the AIprofitboard.com to get access. Now, back to Kapathi. And the thing in this announcement tweet that most people completely skipped past. He wrote, I remain deeply passionate about education.

That one sentence is a clue because Eureka Labs, his education company, wasn't just about teaching AI. It wasn't just about creating courses. It was about making the hardest technical concepts feel accessible to normal people. He has a rare ability to explain something incredibly complex in a way where you go, ah, that's obvious.

Why didn't I see that before? And here's why that matters for Anthropic. Specifically, we are entering a phase of AI where the bottleneck is no longer the model. The models are already capable, right?

Very powerful. The bottleneck is now adoption, education, people not knowing how to build the context, set up the workflows, or use the tools in a way that actually produces results for their business. And Anthropic has been building the model. They've been building Claude code.

They've been building the infrastructure. But there's a gap between what Claude can do and what most people know how to use it for. Karpathy is the person who closes that gap. And that brings me to the three things I think are coming that you need to be watching out for.

The first is a context marketplace. I think right now Anthropic has skills and plugins inside Claude code, but I think this gets much bigger. Not just prompts you can copy and paste, but actual domain specific context packages like workflows, memory files, evaluation systems, examples of good outputs for specific jobs, an accountant's monthly close process, right? A recruiter's intake workflow, a marketing agency's brief to content pipeline, all of it packaged up and pluggable into Claude code so anyone can instantly level up their results in their specific area.

And the model doesn't matter as much when you have the right context. The person who builds the best context packages for the most valuable use cases owns a huge piece of this. The second thing coming is more slash goal style commands. Claude code already has a forward slash goal feature.

You give an objective, it keeps working until the objective is met, right? Not just one step, but a sustained loop. So you can set the direction and you come back when it's done. But I think that's just the beginning.

I think you're going to see much more of this, right? Specialized loops for research, for debugging, for content creation, for lead generation. Commands that say keep going until this specific condition is true in this specific vertical. The interface stops being basically about individual tasks and it becomes about setting outcomes and letting the system run.

That's totally different. And the third thing is, and I think this is probably the thing that has the biggest impact for people watching this, is an education layer for building your own workflows. Right now, most business owners are using Claude like very reactively. They have an idea, they open a chat, they ask a question, they get an answer, and that's it.

But it's kind of like one-tenth of what's actually possible. The real power is in building systems. Setting Claude up with your context, your memory, your instructions, so it can run parts of your business consistently, not just answer questions on demand. Carpathia has been building and teaching exactly this.

How to set up agentic systems that run loops, build knowledge bases, and work towards goals without you having to babysit every step. And now that thinking is going to be built directly into how Anthropic designs Claude, which means the gap between people who have built these systems and people who haven't is about to get much, much wider. Here's what I think I want you to think about right now. Most business owners right now are still at the sort of ask Claude a question and copy the answer stage.

And that's fine, that's where everyone starts, but the people who are like six months ahead of you are already running Claude in the background on the business, right? He knows the clients, he's running on a schedule, he knows their offers, he knows the processes, he generates the content, follow-up emails, their proposals, their reports, all with their specific voice and context built in. And the people who are 12 months ahead are doing what Carpathia described, right, which is agent swarm systems that run whilst they're still loops to test, improve, and optimize on their own, right? And this isn't hypothetical anymore, it's already running.

We have systems for this inside the Anthropic boardroom. And the Carpathia hire tells you Anthropic is betting everything on this direction. Now let me come back to what I mentioned before, which was the open loop. What does the next Claude look like?

Well, if auto research can find 20 improvements in 48 hours running on a home computer, and Carpathia now has access to Claude's full capabilities and Anthropic's compute budget to run these loops at scale, well, the next version of Claude isn't being improved the old way, right? It's been improved by Claude itself, running experiments, testing hypotheses, stacking improvements automatically. That's recursive self-improvement. And Anthropic isn't hiding the fact that they think this is the path forward too.

Jack Clark, Anthropic's co-founder, publicly said there's a 60% chance we hit fully automated AI research by the end of 2028. Not partially automated, but fully, right? AI running its own research cycle. So Carpathia is the execution plan for that bet.

And here's the part that should make you sit up. You don't have to understand the technical details of recursive self-improvement to know what this means for your business. It means the models are just going to get better, faster, and they're going to improve much quicker than anybody expects. And the context, workflows, and systems you build now around today's Claude are going to become even more powerful as the model underneath them improves.

Now, you're not starting over every time a new model drops because everything is compounding. Every SAP you gave Claude, every past project it learns from, every output you correct and save as an example, every workflow you build, all of it gets better results as the model gets smarter. And that's a flywheel. And the people who don't build that context layer, they're stuck back at the reactive stage, right?

Rewriting the same prompts, getting the same generic outputs, whilst everyone else and their Claudes are getting better and more and more specialized and useful, right? And so the belief that AI is just a tool that you use when you need it, right? Like Googling something, it's the thing that is going to cost people the most over the next 12 to 18 months. AI is not a search engine.

It's closer to a team member and team members need context training and systems to work inside of. Carpathia understood this before almost anyone and that's why he spent the last year building in public, for example, LLM wiki, auto research, context engineering frameworks, right? He was showing people what the future of AI-powered work looks like and now he's inside the company building the future of Claude. The last thing I'll say about this is something Carpathia mentioned before he joined.

He talked about staying outside a frontier AI lab for too long and kind of losing your edge. Your judgment kind of starts to drift. You lose touch with what's actually happening at the bleeding edge. So he chose this moment to go back in.

He looked at Anthropic's trajectory, what they've been shipping, where they're headed, and decided this was the most important place to be right now in 2026. For a guy who could be doing literally anything, who already has more credibility than almost anyone in AI, that's a signal, right? Now if you want to build the context layer in your business before this gets away from you, before the gap between early adopters and everyone else becomes impossible to close, the AI profit board is where to start. We actually have systems built around this.

The agent operating system. We have an infinite context engine that is built for specifically this sort of stuff. We run four coaching calls every week specifically about how to implement Claude and AI automation in your business. And we have actual walkthroughs of how to set up Claude with your business context, how to build the memory files, how to create automation workflows that generate leads and save hours every single week.

We've got a 30-day roadmap, daily tutorials, and a community of 2,800 business owners, sorry, it's 3,000 now, who are building these systems right now. People you can connect with, ask questions, and learn from directly. So there's always someone online 24-7. Link in the comments description or go to the AI profit board to get access to that.

One last thing, the question I keep coming back to is, after all of this, is simple. If Kapafi, who could be teaching, who could be building his own company, could be doing anything, if he decided to join the most, you know, one of the biggest AI labs in the world, which is Anthropic, and he decided that's the most important thing to do, you know, working on Claude to improve Claude, what does that tell you about where AI is actually heading? More importantly, are you building the systems today that put you on the right side of what's coming? That's the question we're sitting with.

Thanks for watching, I'll see you in the next one. Cheers, bye-bye.