An unofficial microGPT course for learning GPTs from first principles.
Most people use GPTs.
Very few can open a tiny GPT file and clearly explain what every part is doing.
The Karpathfinder is a guided path through the code — from tokens to logits — so you can explain it, modify it, and build your own thing.
Founding: $695 Core / $1,095 Verified
One-time payment · Lifetime access · 14-day refund policy
The Karpathfinder is an independent educational product and is not affiliated with or endorsed by Andrej Karpathy or OpenAI.
You've watched explanations, but the code still feels slippery.
You can call an API, but not reason from first principles.
You know the terms, but not the mechanics.
You hesitate to modify model code because you don't trust your understanding.
You'll follow the path from:
By the end, you won't just "kind of get transformers."
You'll be able to explain what the code is doing and build your own variant.
Seven modules. One codebase. Each checkpoint builds on the last until you can explain, modify, and build your own tiny GPT.
Orientation, prerequisites, and what a GPT is actually trying to do. Set up your environment and understand the landscape before the hike begins.
Learner understands the project scope, has a working environment, and can articulate what next-token prediction means at a high level.
Documents, BOS, vocabulary, tokenization, and next-token prediction. Understand what the model sees and what it's trying to predict.
Learner can explain what the model is predicting and why, trace from raw text to token IDs, and describe the vocabulary.
Bigram intuition, loss functions, autograd, backpropagation, and learning dynamics. Where the actual learning happens.
Learner can explain where learning happens, how parameters update, and what the loss function is measuring.
Embeddings, positional information, self-attention, residual connections, MLP blocks, and layer normalization. The transformer core.
Learner can trace a complete forward pass and explain every major component of the transformer architecture.
Adam optimizer, training loops, logits interpretation, sampling strategies, temperature, and inference. From training to generation.
Learner can configure training, interpret logits, and generate text with different sampling strategies.
Modify the model, swap the dataset or architecture, and build your own tiny GPT variant. Ship something real.
Learner ships a working variant and can defend every modification they made.
What scales from tiny GPTs to real-world systems, and what changes in production. The bridge from learning to building.
Learner understands the gap between microGPT and production systems, and knows where to go next.
who want to understand model internals
building first-principles intuition
shipping products on top of LLMs
who teach or write about AI
You want only high-level fluff
You refuse to read code
You want passive entertainment instead of mastery
You want to pretend understanding is the same as understanding
[Instructor bio — add your background, credentials, and what drives you to teach this material. Keep it honest and specific.]
[email protected]If the first two modules don't create genuine clarity about how a tiny GPT works, email us and we'll refund your purchase. No questions, no hoops.
One-time payment. Lifetime access. No subscriptions, no upsells, no surprises.
Try the first lesson and quiz before you commit.
founding price
The complete path from tokens to logits, with lifetime access.
founding price
Everything in Core, plus a graded assessment and verified credential.
MicroGPT foundations for technical teams.
Stop treating GPTs like magic. Start understanding them from first principles.