How LLMs work
Next-token prediction, attention, training, and why models make things up
11 stopsFull map
Learn 10x faster with interactive learning maps
Turn the docs, papers and RFCs you need for work into short maps you play a few minutes at a time. Quick reviews bring each idea back before it fades.
Free and open source. Runs on your laptop. No account.
Made from RFC 6749 and RFC 7636
Why not share your password?
Four roles, one login
The redirect dance
Checkpoint
A code, not a token
PKCE
Asking for just enough
Staying signed in
You read it. A week later, it's gone.
You saved the article, skimmed the doc, nodded through the talk. Then someone asks you to explain it. Learncroc is for the things you need to actually know for work.
New team
Forty pages of onboarding docs
Paste them in and walk into your first design review knowing why the system works the way it does, not just where the files are.
The paper everyone cites
Attention, Raft, the new model card
Play the map before the meeting, so you can join the argument instead of nodding along.
Before the review
The RFC you're expected to have read
Get the trade-offs straight before someone asks you to defend them.
Built so it sticks.
Every part of a map comes from learning research, not from what looks busy.
Guess first
Every stop opens with a question you can't answer yet. A wrong guess you cared about makes the answer stick.
Use it, don't reread it
Drag, sort, predict. Checkpoints test you on situations the cards never showed, because that's what work does.
Relight it before it fades
What you've learned fades over days, then comes back in a five-minute review, timed by FSRS, the scheduler behind modern Anki.
Try a stop. It takes a minute.
This is stop 4 of the How LLMs work map. Guess first, play with it, then find out if you were right. The croc sits beside every stop: ask it to explain another way or go deeper. In a checkpoint it only gives hints.
How LLMs work, stop 4 of 11
- Question
- Try it
- Answer
You ask a chatbot the same question twice and get two different answers. Why?
Pick the one you'd bet on.
Make a guess first. No pressure, guesses are how this works.
Safe for the docs you can't paste anywhere.
Learncroc is open source and runs on your own machine, so your company's docs stay where they are.
- It runs on your laptop. Your docs and your progress stay in ~/.learncroc. Nothing is uploaded.
- Fully offline with Ollama. Point it at a local model and nothing leaves your machine at all.
- Or bring your own key. Anthropic, OpenAI, OpenRouter or any OpenAI-compatible server, with a monthly spend cap and a log of exactly what was sent.
- No account, no tracking. There's nothing to sign up for and no analytics. It's open source, so you can check.
Terminal
git clone https://github.com/learncroc/learncroc.git
cd learncroc
bun install
bun devNeeds Bun 1.2+. Then open localhost:5173.
Prefer Docker? Build the image from the Dockerfile in the repo.
Fits between meetings.
No course to block out a weekend for. Play a stop while the build runs, and review in the time it takes to make a coffee.
A few minutes a stop
Each stop is 2 to 4 minutes and under 500 words. A full map takes an evening.
No streaks, no guilt
No XP, no leaderboards, no nagging. Add a calendar reminder if you want one.
Review from your terminal
Relight without leaving the shell:
bunx learncroc relightTake your cards with you
Export your flashcards to CSV or Markdown whenever you like.
Start with a map we made.
Made by hand, and they work with no AI key. A full map is eleven stops; a taster is a complete five-stop arc you can finish in twenty minutes.
Bayesian thinking
Why a positive test usually isn't what it seems
5 stopsTasterColor theory
See like a painter, and like a screen
5 stopsTaster
Got a doc? Get a map.
Paste a link or the text itself and the croc finds the big ideas and turns them into a map of three, five or seven stops. It uses your own AI key, or a local model with Ollama.
A preview with sample docs. In the app, paste any link or text and the map is built from it.
Ramping up a team?
Maps are plain YAML files you can keep in git next to the code. Our Claude Code skills turn your docs into a map, and learncroc lint checks it before you share it.
What's the catch? There isn't one.
Learncroc is free and open source. You can read every line, run it at work and keep using it however things change.
Free, for good
The code is MIT. Our maps are CC BY-SA 4.0, so you can adapt them and share them on.
Our promise
The app and every map we make stay open source. A hosted version will sell convenience, never learning features. If it works on your laptop today, it will never move behind a paywall.
Want it hosted?
No install, sync across your devices, our AI key. We're letting people in from a waitlist.
Join the waitlist
Star on GitHub
Stars help other people find it. Issues and new maps are welcome.