format & ground rules
How the course works
ARENA in 12 Weeks guides Python programmers through the current ARENA curriculum. It starts with machine learning fundamentals, then moves through transformer interpretability, reinforcement learning, language model evaluations, and alignment science.
The course arc
- Week 1: machine learning fundamentals and a first trained model.
- Weeks 2–6: transformers and methods for studying their internal computations.
- Weeks 7–8: reinforcement learning, from value functions through RLHF and tree search.
- Weeks 9–10: language model evaluations, agents, and AI control.
- Weeks 11–12: investigations of reasoning and misaligned model behavior.
Each week has two parts
- Read the explainer and change its runnable examples. The explainer supplies the context that ARENA assumes.
- Open the linked ARENA notebook and complete the listed core sections. Start stretch work after the core work.
Use the solutions notebook when you are stuck, then explain the result in your own words. You can study alone or share a screen with a small group and rotate who drives.
Where the code runs
Small examples run on the course CPU from your browser. The longer ARENA exercises run in Google Colab. Choose a GPU only when the notebook calls for one. You can also clone the ARENA repository and work locally.
FAQ
Do I need an ML background?
No. You should know Python functions, classes, and comprehensions. The explainers introduce the maths and ML concepts when they first matter.
Do I need local setup?
No. The published lessons work with the course runner and Google Colab. Local setup remains an option if you prefer it.
Does this replace ARENA?
No. The explainers prepare you for selected ARENA exercises. The notebooks remain the main source of implementation practice.
What if ARENA notebooks change?
Links point to ARENA's main branch for freshness. Record the commit hash if you want to return to the same notebook version later.