I Tried "ARENA," a Free Curriculum for Learning AI Safety and Mechanistic Interpretability

Hello, this is Kaieda.

In my previous article, I introduced what mechanistic interpretability (commonly known as mech interp) is as a field. The idea was that you can extract a feature corresponding to the concept of "Golden Gate Bridge" from inside Claude, and directly manipulating it changes the model's behavior itself. Whereas traditional XAI was a tool for observing from the outside, mech interp is a tool that lets you touch the switches inside the model.

Series: "Trying Mechanistic Interpretability (mech interp)"

1. What is Mechanistic Interpretability (mech interp)? — The Story of Research That Turned Claude into a "Bridge"
2. I Tried "ARENA," a Free Curriculum for Learning AI Safety and Mechanistic Interpretability ← you are here
3. GPT-2 That Doesn't Break Even After Removing an Attention Head — Chasing the True Nature of Self-Repair Through Experiments (in preparation)
4. The Story of How Two Tools for Answering "Is This Head Important?" Gave Completely Opposite Conclusions (in preparation)
5. The SAE Hyperparameter Trap — When Reconstruction Loss of 0.0000 Doesn't Mean "Readable" (in preparation)

When it came to "so what should I actually start learning," the free curriculum "ARENA" is what I've been working through. This article introduces what kind of curriculum ARENA is, what each of its five chapters covers, and how best to proceed through it.

What you'll learn from this article

• What kind of curriculum ARENA is and what you can learn from it
• What each of the five chapters covers
• What order to proceed in

Please note that this article is not official content from ARENA or Anthropic, but a personal, unofficial learning log. I will not reproduce exercise text or solutions code, and I have summarized each chapter's content in my own words. Please refer to the official website and official repository for accurate exercise content.

What Is ARENA?

ARENA is a practical curriculum for learning AI safety and mechanistic interpretability, created and operated by Callum McDougall. Its full name is Alignment Research Engineer Accelerator.

The GitHub repository has a complete set of exercise notebooks and solutions published publicly, and as of August 12, 2026, it has 1,223 stars and 777 forks, making it widely referenced as educational material in this field.

The core of ARENA is a 4–5 week in-person bootcamp, but the curriculum is all freely available online and can also be used for self-study. In-person participants work on a Capstone Project at the end (a graduation project that dives deeper into a topic of interest).

The in-person program has been held 8 times as of August 2026 (ARENA 8.0), and the alumni page lists specific career paths of past participants. Reading about real examples such as working at Google DeepMind and UK AISI, participating in fellowships like MATS and LASR Labs, and Research Engineer positions at Equistamp collaborating with METR conveys that this program functions as an actual gateway into professional roles, beyond its positioning as a "free self-study resource."

The curriculum is divided into 5 chapters in total, all of which were published as of August 2026. A summary of each chapter's content is as follows:

Since Chapter 1 is the main scope of this series, I'll also list the breakdown of its 12 sets (please see the previous article for an explanation of mech interp itself).

• 1.1 Transformers from Scratch
• 1.2 Intro to Mech Interp
• 1.3.1 Linear Probes
• 1.3.2 Function Vectors & Model Steering
• 1.3.3 Interpretability with SAEs
• 1.3.4 Activation Oracles
• 1.4.1 Indirect Object Identification
• 1.4.2 SAE Circuits
• 1.5.1 Balanced Bracket Classifier
• 1.5.2 Grokking & Modular Arithmetic
• 1.5.3 OthelloGPT
• 1.5.4 Superposition & SAEs

Chapter 4 covers safety topics that don't fit into interpretability or evals, and all 5 sets were published as of August 2026. The breakdown is as follows:

• 4.1 Emergent Misalignment
• 4.2 Science of Misalignment
• 4.3 Interpreting Reasoning Models
• 4.4 LLM Psychology & Persona Vectors
• 4.5 Investigator Agents

Note that the GitHub README states "everything in Chapter 1 except the first 2 sets is optional," but the official website's Chapter 1 page does not explicitly indicate what is required or optional. Since the README may not have been updated, I will not assert this distinction in this article. If you're concerned, please check the official website.

How to Proceed

Faced with the volume of 6 sets in Chapter 0 alone and 12 sets in Chapter 1 alone, I think you'll first be unsure where to start. Here I'll present one approach based on the order I actually followed.

In terms of the order to tackle things, I recommend starting with the Optimization section of Chapter 0 (0.3 Optimization). Getting comfortable here with basic operations that come up repeatedly in later chapters—such as hyperparameter search using wandb—will make subsequent chapters considerably easier.

Next, move on to the from-scratch Transformer implementation in Chapter 1 (1.1), and then work through the mech interp introduction (1.2). In 1.2, you'll work through exploring induction heads, analyzing the OV/QK circuits of attention heads, and calculating composition scores between heads.

Once you've gotten this far, your understanding of the internal structure of Transformers should have deepened one level.

I myself am currently working through Superposition & SAEs (1.5.4) in Chapter 1. By following the guidance within the exercises and focusing on ① (the basics of superposition) and ⑤ (resolution via SAE), you can reproduce with your own hands the famous Anthropic phenomenon where feature vectors settle into a regular pentagon shape under sparse conditions.

My next plan is to move on to 1.3.3 Interpretability with SAEs, and I haven't started Chapters 2, 3, or 4 yet.

What I Learned by Actually Doing It

I'll write here about things that can't be grasped from just the curriculum table of contents, from the perspective of someone who has actually worked through it.

How Much Prior Knowledge Is Required?

By the time you enter Chapter 1, the following are required:

• Python: Essential. This is treated as a prerequisite.
• PyTorch tensor operations: You need to be able to trace changes in shape yourself. Whether you can articulate in your own words what a multi-dimensional tensor like (batch, seq, d_model) represents each time will make a big difference in how smoothly you progress.
• einops / einsum: These are used extensively in Chapter 1. If you've never seen them before, getting comfortable with them in Chapter 0 will help.
• Linear algebra: At the level where you can explain in words what a matrix product and dot product "are doing." You don't need to follow proofs.

You don't need to solidify deep learning theory first, but it's worth training yourself to trace tensor shapes beforehand. Honestly, that's also where I got stuck at first. Even if I could read the math itself, I'd spend time moving my hands without understanding what was happening because I couldn't mentally reproduce how dimensions were being rearranged inside einops.rearrange or einsum. Most of the times I got stuck in exercises came down to shape comprehension, not mathematics.

How to Run It (Environment Setup Is Not Required)

Let me say this upfront: you can start without cloning the repository. The official Setup page presents three methods, and the first one listed is Colab.

Furthermore, for cases where local or free Colab isn't enough, the official documentation even covers GPU rental procedures on VastAI, RunPod, and Lambda Labs.

If you just want to try it out, Colab is sufficient. In fact, while the GitHub README lists clone + install.sh as "Install Instructions," this procedure is primarily intended for in-person participants. You don't need to stop there just to start self-studying.

I chose to clone it because I wanted to keep the code locally, but I think trying one set on Colab first and then deciding whether to continue is the rational approach.

Notes for Local Installation

If you proceed by cloning, some packages may not get installed properly if you install manually from requirements.txt. In fact, sae-lens and sae-vis used in 1.3.3 were not installed in my environment, despite being listed in requirements.txt.

Checking your imports for the libraries used in a chapter right before you start it will prevent you from getting stopped mid-exercise. When in doubt, following the official install.sh is the safe bet.

Proceeding Together with Claude Code

Personally, I found that working through the exercises together with Claude Code suited me best. I can ask questions directly about things I don't understand, and when questions arise that go beyond ARENA's scope, I can actually set up experiments on the spot and verify things.

ARENA's exercises are in Jupyter format and hints are provided, but there are also moments where you can't quite get to the bottom of "why this implementation is correct" with that alone. Having someone to consult with while writing code made a big difference. Many of the experiments I'll introduce from the third installment onward in this series were also things I verified by stepping one foot outside of ARENA's exercises this way.

Proceeding on a Mac

Even on Apple Silicon Macs, torch.backends.mps is available, so the content of Chapter 1 ran without issues. For small models like those in 1.5.4, CPU alone provides sufficient speed.

However, since 1.3.3 has sections dealing with larger models like Gemma, if you're low on memory, you may need to make judgment calls such as skipping those sections, focusing on sections that use GPT-2, or using the GPU rental options mentioned earlier.

For Those Just Starting Out

First, even just the 2 sets of Transformer from-scratch implementation and mech interp introduction in Chapter 1 will considerably deepen your understanding of the internal structure of Transformers. I think it's best to aim for these 2 sets first, and then proceed to whichever sets interest you from there.

If you're considering participating in the in-person bootcamp, it's also a good idea to start getting a sense of your topic with an eye toward the final Capstone Project.

Also, ARENA has no explicit OSS license file (the license is also null on the GitHub API). It's safer to avoid directly publishing or reproducing exercise answers or code, and instead to summarize things in your own words as in this article, or publish only the experimental code you yourself wrote.

Finally, since Chapter 0 is required as prerequisite knowledge, if you're not confident in your deep learning basics, touching it rather than skipping it should make Chapter 1 and beyond easier to understand.

Summary

As a learning curriculum for AI safety centered on mechanistic interpretability, I feel that ARENA is a precious resource for how systematically it's put together at no cost. It's originally a program for an in-person bootcamp, but the fact that the entire curriculum is publicly available for free is something I'm truly grateful for as someone tackling it through self-study. I myself am thinking that once I finish working through this curriculum, I'd like to participate in the in-person program held in London.

Since there are still few Japanese-language introductory articles in this field, I hope this article can serve as a first step.

Reference Links

• Official website: https://learn.arena.education/
• Operating organization website: https://arena.education/
• GitHub repository: https://github.com/callummcdougall/ARENA_3.0
• Notion learning resources page: https://arena-resources.notion.site/

Starting from the next installment, I'll be writing logs of experiments I actually conducted myself. First up: the story of "a model doesn't break even when you remove a single attention head."

→ GPT-2 That Doesn't Break Even After Removing an Attention Head — Chasing the True Nature of Self-Repair Through Experiments (in preparation)