An interactive book
Autoresearch
Novice to Ninja
Karpathy's overnight research loop, taken apart: one file the agent edits, one number nobody edits, a five-minute budget, keep or discard, and a program the human improves between nights. With the reported results, five real ledgers of our own, and a loop you can run on a laptop.
Start readingOpen the toy loop
8 chapters · about 1.2 hours · 7 interactives · no background needed
How to read this book
- In order. Eight chapters of seven to ten minutes; each uses only what the earlier ones taught.
- With your hands. Seven interactives. Guess first, then move the slider or press the button. Two of them are made-up models and say so; the rest use the numbers from the text.
- Every number has a source. Quotations are from the repository's README and program.md as read on 10 October 2026. Results are labelled reported, with who reported them, or ours, with the ledger they come from.
- Recall before you read on. Every chapter ends with questions and folded answers.
Part I · The idea
What autoresearch is, and the loop step by step.
01The night shiftWhat autoresearch is, the three files it is made of, and how many experiments fit in one night's sleep.7 min · interactive02The loop, step by stepThe setup, the seven steps of every experiment, the ledger's exact format, what to do with a crash, and the rule that the loop never stops.8 min · interactive
Part II · Why it works
The metric nobody edits, the ratchet, and the program the human improves.
03One number nobody editsWhat makes a metric fit for a loop, why the harness is read-only, what happens when it is not, and what to do when the scale itself turns out to be wrong.9 min · interactive04The ratchetKeep or discard, with a fixed budget and git as the memory: why a hundred small, mostly failed tries add up, and how measurement noise makes the ratchet slip.9 min · interactive05The program, not the codeThe anatomy of program.md, the simplicity rule, and how the human improves the standing orders between nights, with real additions from two of our own programs.9 min · interactive
Part III · In the wild
What people reported, and five loops of our own with their ledgers.
06What people gotThe results people reported in the first weeks, each with its source, what the reports disagree about, and how to read a claim from an overnight loop.8 min · interactive07Five loops of our ownA database, a memory layer, a voice cloner, a voice twin and this website, each tuned by the same loop with the file names changed, with real rows from their ledgers.10 min · interactive
Part IV · Your turn
A toy loop that runs on a laptop, and the checklist for a real one.