Quick context if the name is new. Andrej Karpathy was on the founding team at OpenAI, ran AI at Tesla, and taught the free course “Neural Networks: Zero to Hero” that trained a large chunk of the people now building production AI systems. When someone with that much signal chooses where to spend the next few years, that is not a job change. It is a forecast.

This week he joined Anthropic. Not as an executive. Not to run a division. He joined as a researcher to build a small team that uses Claude to accelerate Anthropic’s own pre-training research. In plain English, a few AI workers running thousands of autonomous experiments to find better architectures, data mixtures, and training recipes for the next generation of Claude. Anthropic is betting that the frontier moves through AI doing the AI research, not through bigger compute alone.

I’ve been running a multi-agent AI team in production for months. Not as a thought experiment. As the actual operating structure of NexHam. Three things in this Karpathy announcement matter for anyone running AI in their own business — or thinking about it.

The operational model just got validated at the frontier

What Karpathy’s team is going to build at Anthropic looks a lot like what I run at NexHam. A handful of AI workers, each with a defined scope, running autonomously, surfacing problems, with their outputs reviewed by a human manager. The scale is different. Anthropic runs experiments on H100 clusters; I run mine on a couple of Mac minis. The operational shape is the same.

If you have been treating AI as a tool — open a tab, type a prompt, copy the output — this is the trend that argues against staying there. The frontier is moving toward AI running AI. The downstream version of that for a business is AI running parts of a business. Same shape, smaller stakes, same skill set.

The compounding skill is not prompting

If the frontier looks like teams of AI running experiments autonomously, the people downstream of that frontier are not the ones who got good at one-shot prompts. They are the ones who learned to manage workers they cannot see. Precision in written rules. Review at the right cadence. Expecting problems to be surfaced, not hidden. Evolving the system as the work runs.

That is the skill that compounds. Prompt engineering is one summer of relevance. Managing autonomous workers is the rest of your career. Karpathy moving to Anthropic is one more data point that this is where the line is moving.

The best teachers in AI are still the ones building

In his own announcement, Karpathy noted he plans to resume his education work in time, but is going back to R&D first. That sequencing is not incidental. The people who teach this stuff well are the ones who are also shipping it. You do not trust a coach who has not played the game.

The same applies here. If you are learning AI workflow from someone who has never actually run an AI team for paying clients, you are reading the rulebook from someone who has only watched the sport on TV. The frontier moves too fast for that gap to close from the stands.

This is the part I have been quietly betting on for the last six months at NexHam. The AI team I run is not a side project. It is the operating model. Everything I have learned about how to build, manage, and scale this kind of team has come from doing it for real, on real client work, with real failures along the way.

Your move this week

Pick one piece of work you currently do alone. Write down what “good” looks like in two short sentences. Hand the work to Claude. Review the first output and tell it specifically what is wrong. Run it again. Then run it a third time.

The first cycle is always wrong. The third cycle is where the output starts to come back better than you would have done it yourself. Most operators give up after the first wrong attempt. The ones who stay through three cycles end up with the start of an AI team.

Karpathy is betting his next few years on this direction. Worth one week of yours.

If you want the full curriculum on building an AI team that actually runs your business, the AI Deployed community is where I teach it end to end. skool.com/ai-deployed-1960