Explains how language models change software development, from his original description of vibe coding to experiments where an agent changes training code and measures the result.
Why read Andrej Karpathy?
Start with Karpathy to understand what building through conversation means. His original vibe coding post describes asking a model for changes while letting the code recede from view. It gives you a reference point for a term now used for many different AI workflows.
His autoresearch repository makes an agent's work concrete: change one training file, run an experiment, measure it, then keep or discard the change. For readers arriving from crypto, his Bitcoin tutorial offers the same approach to understanding a system by constructing its parts, including keys, signatures and a transaction.
For the models behind coding agents, his Zero to Hero course connects small neural networks to a GPT implementation. Follow the earlier language-modeling and PyTorch lessons before the Transformer lecture.
Start with the original
Selected work
Post ·
Vibe coding: the original description
What does building through conversation look like? Karpathy describes a workflow in which spoken requests drive changes and the developer stops focusing on the code itself.
Code
autoresearch: agents running training experiments
How can an agent evaluate its own changes? Read the README and program.md to see a small experiment loop with one editable training file, a fixed time budget and a measured result.
Tutorial ·
A from-scratch tour of Bitcoin in Python
What is inside a Bitcoin transaction? This educational walkthrough builds keys, addresses, signatures and a testnet transaction in Python, connecting the cryptography to the data that moves through the network.
Course
Neural Networks: Zero to Hero
How do neural networks become language models? Karpathy’s course builds from backpropagation to GPT and tokenization. It asks for solid Python and introductory calculus; use the earlier lessons before the GPT implementation.
Projects & roles
autoresearch
A small training setup for experimenting with agents that modify code and compare results.
micrograd
A compact automatic differentiation engine and neural network library for learning the mechanics of training.