arena-in-12-weeks
glossary about

Site feedback

What should we improve?

Your email app will open with this message and the current page URL.

a 12-week guided curriculum

ARENA in 12 weeks

This twelve week course follows ARENA chapters 0 through 4. It begins with machine learning fundamentals and ends with investigations of model behavior.

The course covers transformer interpretability, reinforcement learning, language model evaluations, and alignment science. Start at any time and work at your own pace.

how it works

Each week has two parts.

  • Read the explainer to learn the week's main ideas.
  • Complete the core ARENA exercises before starting the optional stretch work.

the 12 weeks

  1. 01

    Orientation and a first digit classifier

    Train and validate an MNIST classifier.

    Start Week 1 prework · 60 to 75 minutes

    arena 0.2, sections 1 and 2 · fundamentals
  2. 02

    Transformers from scratch

    Build and train a transformer.

    arena 1.1 · transformer interpretability
  3. 03

    Introduction to mechanistic interpretability

    Inspect activations and find induction heads.

    arena 1.2 · transformer interpretability
  4. 04

    Probes, steering vectors, and SAEs

    Read, edit, and reconstruct model activations.

    arena 1.3.3 · transformer interpretability
  5. 05

    Finding circuits in language models

    Test the IOI circuit with activation patching.

    arena 1.4.1 and 1.4.2 · transformer interpretability
  6. 06

    Understanding models trained on small tasks

    Reverse engineer algorithms in small models.

    arena 1.5.1, 1.5.2, and 1.5.4 · transformer interpretability
  7. 07

    Reinforcement learning foundations

    Implement planning and tabular temporal-difference learning.

    arena 2.1 Sections 1–2 · reinforcement learning
  8. 08

    Policy optimization and model training with feedback

    in progress

    Build PPO, RLHF, MCTS, and AlphaZero.

    arena 2.3 to 2.5 · reinforcement learning

  9. 09

    Designing language model evaluations

    in progress

    Design an evaluation and its dataset.

    arena 3.1 and 3.2 · language model evaluations

  10. 10

    Running evaluations, building agents, and AI control

    in progress

    Run evaluations, build agents, and study AI control.

    arena 3.3 to 3.5 · language model evaluations

  11. 11

    Studying misaligned behavior

    in progress

    Test explanations for misaligned behavior.

    arena 4.1 and 4.2 · alignment science

  12. 12

    Reasoning models and model behavior investigations

    in progress

    Investigate reasoning and model behavior with agents.

    arena 4.3 to 4.5 · alignment science

prerequisites

You should know Python well enough for functions, classes, and comprehensions, plus high-school maths. We teach anything beyond that when you need it, including matrix multiplication, derivatives, and log/exp. You need a Google account for Colab. You install nothing locally.