Week 12 is a fork, and either road is one session long

Entry 12 of the ARENA guide. The fork: part 1.3.3 §1, interpretability with SAEs, or part 2.1, intro to RL — exactly one of the two.

The mental model

The guide's only deliberate fork sits here, and the reasoning is worth studying as a

piece of planning. Depending on which kind of project you pursue afterward, one of

two skills serves best: if the work turns out interpretation-shaped, SAE literacy

(sparse autoencoders, the tool that decomposes a model's activations into readable

features) is the on-ramp; if it turns out environment-shaped, intro to RL (tabular

environments, policy evaluation, the vocabulary of episodes and returns) is. ARENA's

own structure makes the fork cheap: both candidates are light, single-session parts,

and no later part in this guide requires either.

What the two candidates actually contain. The SAE path: SAELens, latent dashboards,

and the exercise that matters most, finding interesting latents in a small GPT-2

model on CPU. The RL path: the definition of an MDP, tabular value iteration and

policy evaluation in NumPy, no GPU anywhere. Both are literacy plays, not

mastery plays; this guide says so out loud.

If you want to build this

Either candidate fits comfortably into a single sitting, which is the point of leaving the fork unpicked: take the road your project needs, when your project needs it. Quoted from ARENA’s pages:

• 1.3.3 §1 path: SAELens setup and intro, SAE dashboards, find interesting latents

(the part's graded exercise)

• 2.1 path: what is reinforcement learning, planning in known environments, tabular

RL and policy evaluation; first exercise budget 15–25 min

What it unlocks

Neither is a prerequisite for anything else in the guide; that is what makes it a

real choice instead of a chore. The SAE path opens 1.4.2 and the superposition

reading list afterward; the RL path opens 2.2–2.5, most of which assumes hardware

most readers will not have (the RLHF exercises expect an A100-class GPU).

Go deeper

• The ARENA 3.0 curriculum, chapter 1 part 1.3.3 §1 (SAEs) and chapter 2 part 2.1 (RL) — the two roads.

---

PREVIOUS Misalignment is measured, not assumed · NEXT Reasoning models say more than they do, so measure the gap · GUIDE HOME