011. What Andrej Karpathy’s Role Is and Why It Matters
Andrej Karpathy is one of the most important AI researchers in the world. In May 2026, he announced he is joining Anthropic, the company behind Claude. His specific job is to work on the pre-training team.
He will use Claude, Anthropic's own AI, to make future versions of Claude smarter. This is not a marketing or spokesperson role. He is using AI to improve AI.
This is called recursive self-improvement, and it is a much bigger deal than most people realize.
Before joining, Karpathy founded Eureka Labs, an education company that teaches people how AI works from the ground up. He also coined the term vibe coding, where you describe what you want and AI builds it. He is a translator of hard technical ideas.
His move to Anthropic signals that the company is betting everything on making Claude not just smarter, but also easier for regular business owners to use.
022. Karpathy’s Autoresearch Project: AI That Improves Itself
Before joining Anthropic, Karpathy built a project called Autoresearch. He showed it in public for everyone to see. Autoresearch is an AI agent research loop.
The AI proposes a change, tests it, and checks the results. If the change makes things better, it keeps it. If it makes things worse, it undoes the change and tries something else.
The loop runs over and over, even while you sleep.
Karpathy ran Autoresearch for about two days straight. During that time, the agent ran 700 experiments on its own. It found roughly 20 improvements that could all be stacked together.
The result was an 11% improvement in speed on a standard AI training benchmark. And this was running on a home computer with a tiny model built in about 30 lines of code. Now imagine the same approach running at Anthropic with massive computing power and Claude itself running the experiments.
033. How to Set Up Context Engineering for Your Business
Karpathy has been teaching that context engineering is more important than prompt engineering. Prompt engineering is writing a better question. Context engineering is building the whole environment the AI works in.
Your files, examples, workflows, histories, instructions, and style guides all matter. Here is how to get started building your own context layer:
- Collect all your raw business documents. This includes meeting notes, past client conversations, standard operating procedures (SOPs), your brand voice guide, your best-performing content, and your offer details.
- Use a tool like Karpathy's LLM Wiki project. This system takes all your raw documents and builds a living, connected knowledge base. Not just folder files, but a structured, searchable, relational knowledge base that the AI can navigate.
- Feed that knowledge base into Claude. Instead of opening a fresh chat where Claude knows nothing about you, give it your full context. The same model will produce completely different, much better outputs.
- Test and refine your context. Add examples of good outputs, correct mistakes, and save them as references. The context layer gets stronger over time.
044. How to Use Claude with Full Context: Examples from the Episode
Most business owners use Claude reactively. They open a chat, ask a question, get an answer, and stop. That is only using about one-tenth of what is possible.
Karpathy shows that the real power is in building systems. For example, you can set up Claude with your meeting notes, past client conversations, SOPs, brand voice, and offer details. Now Claude is not guessing.
It knows your clients, your offers, and your processes.
You can then set up automation workflows that run on a schedule. Claude generates content in your voice, follows up with leads the way you would want, and writes proposals with your actual offer details. The AI runs in the background of your business while you focus on other things.
The episode showed that people who are six months ahead are already running Claude this way. People who are 12 months ahead are using agent swarm systems that run loops to test, improve, and optimize on their own.
055. Results and Comparisons: What Autoresearch Achieved
The Autoresearch project produced clear, measurable results. In two days, the AI agent ran 700 experiments and found roughly 20 improvements that could be stacked on top of each other. The overall result was an 11% improvement in speed on a standard AI training benchmark.
This is significant because it ran on a home computer with a tiny model and only 30 lines of code. The episode contrasts this with the old way of improving AI, which requires human teams running experiments manually for weeks or months.
The comparison is stark. If a small-scale version can produce 20 improvements in 48 hours, what will happen when Karpathy has access to Anthropic's full computing budget and Claude's capabilities? The episode also compares business owners who build a context layer vs. those who do not.
The ones with context get specialized, accurate outputs. The ones without get generic outputs and have to rewrite prompts every time.
066. Tips, Limits, and Mistakes to Avoid
Karpathy warns that most people are thinking about AI the wrong way. They focus on which model is best — Claude 4 vs. GPT-5 vs.
Gemini. They ask who won the benchmark. But the model is only one layer.
The bigger opportunity is the context layer you build around the model. The biggest mistake is treating AI like a search engine. You do not open Google and keep it running in the background.
But AI is closer to a team member. Team members need context, training, and systems to work inside of.
Another mistake is not saving and reusing your context. If you open a fresh chat every time, you lose all the history and examples. Build memory files, keep a knowledge base, and correct outputs so the AI learns.
Also, do not try to build everything at once. Start with one workflow, like generating follow-up emails, and expand from there. The episode emphasizes that the gap between people who build these systems and people who do not is about to get much wider.
077. Who This Is For and Why It Matters
This episode is for business owners who want to use AI to save time, generate leads, and automate parts of their business. It is also for anyone who wants to understand where AI is heading in 2026 and beyond. If you are still at the stage of asking Claude a question and copying the answer, you are falling behind.
The people who are six months ahead are already running Claude in the background with full business context. The people who are 12 months ahead are using AI agent loops that improve on their own.
Karpathy's move to Anthropic matters because it shows the company is betting on recursive self-improvement and making Claude accessible to non-technical users. Anthropic co-founder Jack Clark said there is a 60% chance of fully automated AI research by the end of 2028. The context, workflows, and systems you build now around today's Claude will become more powerful as the model improves underneath.
That is a compounding flywheel. The episode ends with a question: Are you building the systems today that put you on the right side of what is coming?
