The Karpathy loop: learn an AI concept by rebuilding it and teaching it

The Karpathy loop is simple: pick one idea everyone uses but few understand, rebuild it from scratch as small as possible, explain it out loud step by step, ship the code so others can run it, name what you learned, and repeat. Andrej Karpathy has run it for more than a decade, from Stanford's CS231n to micrograd, nanoGPT and nanochat. You can run it in four weekends. A coding agent can write the toy version in minutes, so the work that is left is understanding it, and that is the point. In MAO, you put a Builder agent, a Student agent and an Explainer agent on one canvas, wired to the same notes, so the explanation gets tested before anyone else hears it.

What is the Karpathy loop?#

Rebuild a famous system small, explain it in public, ship the code, repeat.

Look at what Karpathy does with his time rather than at his job titles. He takes something complicated, rebuilds it in as little code as possible, and explains it so clearly that a stranger can follow. Put together, his public work reads as six steps:

1. 1Pick one idea everyone uses but few understand.
2. 2Rebuild it from scratch, as small as possible.
3. 3Explain it in public, out loud, step by step.
4. 4Ship the code so others can run it.
5. 5Name what you learned.
6. 6Repeat with the next idea.

None of these steps needs a famous employer, and each one compounds: the code you ship is the next person's starting point, and the explanation is what people remember you for.

Where does the loop show up in his work?#

At every stage, before and after the big jobs. The titles changed; the loop did not.

The repositories are the clearest example of step 2:

• micrograd: backpropagation you can read in one sitting.
• makemore: character-level language models.
• nanoGPT: a small, readable GPT trainer.
• llm.c: GPT-2 training in plain C and CUDA.
• nanochat: "the simplest experimental harness for training LLMs", from tokenizer to chat.

The lesson is not "read his code". It is "build the small version yourself".

Why does rebuilding beat reading?#

Because each level of contact shows you something the one before hid.

The last row is the one people skip. Explaining something out loud, to someone who stops you at every word you have not defined, is where gaps you did not know you had come to the surface. When you explain something well, you also get to name it, which is how "Software 2.0" and "vibe coding" entered the vocabulary.

How do I run the loop in four weekends?#

One concept, one weekend per step. The first one will be rough; the fifth will be better than many paid courses.

1. 1Week 1: pick. Choose one concept you use every day but cannot explain: attention, embeddings, RAG, an agent loop, backpropagation. Write down what you can already say about it and three questions you cannot answer yet.
2. 2Week 2: rebuild. Write a toy version in the smallest code you can. No library that hides the part you are learning: for attention, write the softmax and the matrix products yourself. Ugly is fine; working is the point. Every step should print or show something.
3. 3Week 3: explain. Record yourself explaining it from a blank screen. No slides.
4. 4Week 4: ship. Publish the repository and the recording together, then one post about the part that surprised you.

Doesn't an AI agent make the rebuild pointless?#

No. It moves the work from typing the code to understanding it, which was always the part that mattered.

A coding agent can write a working toy transformer in minutes. If you stop there, you have code you did not write and do not understand, which is worse than reading the paper. Used well, an agent changes what the four weekends are spent on:

• It writes the scaffolding, so your time goes to the one idea you picked, not to data loading.
• It explains on demand: you can ask why this shape, why this constant, what breaks if this line goes.
• It can play the newcomer. An agent told to know nothing about the concept, run every step and ask about every undefined word is a patient first audience for your explanation.

What it cannot do is decide what deserves to be understood, rebuild it with taste, and make it click for a room. That part stays yours.

How do I set the loop up with coding agents in MAO?#

Put three agents on one canvas, give each a role, and wire them to each other and to four notes.

1. 1Open a canvas on an empty project folder and add three agent terminals, for example Claude Code or Codex. Name them Builder, Student and Explainer.
2. 2Give each one a role from its agent sheet. A role is a reusable block of instructions; save each one with Save as role so the next concept starts from the same three.
3. 3Add the four notes and write the concept and the rules in the Concept brief.
4. 4Connect the Builder to the Student and the Explainer, and the Student to the Explainer, so they can send each other tasks. Connect each agent to the notes it needs: a wire grants access only to the node at its other end.
5. 5Tell the Builder to start. Each agent reads and appends to the notes with the mao note commands and asks the others with mao ask, for example:

mao ask "Builder" "Step 3 changed the output shape and the log does not say why. Explain, then fix the log."

1. 1When the canvas works, select it and choose Save arrangement…, so the next concept starts from the same layout.

You stay in the loop the whole time: you pick the concept, you read every step, and you record the explanation yourself. The agents make the rebuild fast and test the explanation before anyone else hears it.

The Builder, Student and Explainer instructions are also written up as a pack, Learn by rebuilding, with the four notes and two arrangements (the full four weekends, and a weekend sprint with only the Builder and the Student). It asks for the concept and the allowed libraries when you add it to a canvas. It will be listed for the MAO release that adds packs.

What are the ways to run the loop?#

You do not need MAO for the loop. Choose by how much help you want with the explaining step.

Frequently asked questions#

Sources#

• karpathy.ai: education, CS231n enrolment (150 in 2015, 330 in 2016, 750 in 2017), OpenAI and Tesla roles and dates.
• Neural Networks: Zero to Hero: the course, from micrograd to GPT and its tokenizer.
• The Unreasonable Effectiveness of Recurrent Neural Networks: char-rnn, its datasets and its MIT-licensed code.
• Software 2.0: the 2017 essay.
• nanochat, nanoGPT, micrograd, makemore, llm.c: the repositories.
• Eureka Labs: "a new kind of school that is AI native".
• Save and assign agent roles, Connect agents, notes, and portals, Work with connected notes from the CLI, Send a task and collect its result and Save a canvas arrangement: how MAO's roles, wires, notes, mao ask and arrangements work.

Checked on 2026-10-29.