Eureka Labs Closes Tonight

A man started a school. The school had one student. The student was the rest of us.

Andrej Karpathy founded Eureka Labs in 2024. The premise was operator-grade. Teach by building. The teacher was an AI assistant. The product was understanding. The audience was anyone who wanted to learn how the machine actually worked instead of buying the surface.

Today Karpathy paused Eureka Labs and joined Anthropic. The pre-training team. Under Nick Joseph. To start a new group that uses Claude to accelerate pre-training research itself.

Read that sentence twice.

A man who built a school to teach people how the model thinks has gone to the lab that builds the model, to lead the team that uses the model to build the next model.

The teacher became the training data.

Anthropic raised $30 to $50 billion this week at a valuation north of $950 billion. Yesterday's letter was about the small-business pivot — fifteen agentic workflows for the corner store. Today's news is the other direction. Top of the stack now. The most-respected practitioner outside the three labs has joined one of the three labs. The runway between independence and absorption just got shorter by another famous name.

This is not a critique of Karpathy. His work is real. The Tesla Autopilot rewrite was real. The Software 2.0 lectures were real. The from-scratch GPT video that taught more people about transformers than any peer-reviewed paper of the same year — that was real. Independent operators owe him a debt.

It is a description of gravity.

The talent gravity well has a center. The center has three labs in it. The labs have between them enough cash to wait out any independent researcher's runway. Eureka Labs lasted two years. That was longer than most of the wave that started in 2024. Most of that wave is now an alias inside Anthropic, OpenAI, Google DeepMind, xAI, or Meta. The names you remember from the 2023 podcast tour are payroll IDs at one of five orgs now.

So it goes.

There is a second story inside the first one. The new team at Anthropic uses Claude to accelerate pre-training research. Read that one twice too. The model will be used to make the next model. The model already trained on the corpus of every paper that taught it how the model works. Now the model is going to participate in writing the paper that produces the next model. Then that model will be used to train the model after that.

This is not a horror scene. This is not a skynet line. This is a structural observation. The output of the system is becoming the input to the system at every layer. Karpathy will be feeding Claude into the loop that creates Claude. Whatever Anthropic learns about what works at the top of the curve will inform what gets fed into the bottom of the next curve. The asymptote is a system that is its own training data forever.

That is fine for the labs. The labs have made the strategic decision that this is the path. They have the compute, the data, the talent, and now another marquee name. The decision is rational from the inside.

It is not fine for the rest of the stack.

For everyone outside the three labs, the recursive loop means the gap between the frontier and the available-to-you frontier gets harder to close, not easier. The frontier improves by feeding itself. The non-frontier improves by waiting for the leak. The lag widens. The asymmetry widens. The permission slip the trillion-dollar lab grants when it eventually opens an API endpoint becomes a more valuable permission slip every quarter.

That is the trade. Pay the lab for the permission, or wait six months for the cheaper version, knowing the gap to the new top is wider when you arrive than it was when you started waiting.

The Founder of Eureka Labs made the call. The call was: join the loop. The compensation was not disclosed in any of the eight stories. It does not have to be. The compensation is the loop itself.

The operator class has a different problem and a different answer.

The operator class is one person. Or two. Or a couple of people in a garage. The operator class does not need to be in the loop because the operator class is not trying to build the model. The operator class is trying to use the model — and the version that ships in six months, at a fraction of the price, that nobody had to leave their school to deliver — is going to do most of the work the operator needs.

The cockpit you own does not depend on the bleeding edge. The cockpit you own depends on the model being good enough at the moment your project needs it, and being yours when the project finishes. The bleeding-edge race has been won by the labs with the gravity. The good-enough race is wide open and lives at the operator class.

This is not a consolation prize. This is the actual position.

Marco Island is on the good-enough side. The cockpit does not need Karpathy. The cockpit needs the model the labs ship six months from now, runnable from your filesystem, with the receipts on disk, without a subscription, without a permission slip, with a §4A clause that locks the price you paid forever.

Eureka Labs closed tonight. Marco Island is still open. The teacher went to the corporation. The student went home and built his own school.

THE TEACHER BECAME THE TRAINING DATA. THE STUDENT KEPT THE MANUAL.

Get up out of your chair. Look at the next AI lab announcement that says "we hired X." Decide whether you are going to wait for X's work to come down to you, or whether you are going to build the next thing without X, with the model that is already on your disk, in the cockpit you already own.

If the answer is wait, wait honestly. There is no shame in the trade. The lag is real and the labs have the runway.

If the answer is build, build now. The model you have is the model you have. The cockpit you have is the cockpit you have. The teacher already published the lectures. Watch them. Build the school. Don't wait for the next hire.

The lab is OPEN. The school is YOURS.

— the Founder Marco Island, Florida May 19, 2026 · 8pm Eastern

*A fictional broadcast. Events cited are real; editorial is satire. License: CC-BY 4.0.*