sammatuba

ESSAY · PROJECT 0003

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

Dataset the ARENA 3.0 curriculum (TARA vault submodule, pinned 81533b2) + the TARA Learning Plan v2 · Charts 0 · Written SEP 2026

Entry 12 of the ARENA guide. The W46 pick: 1.3.3 §1, interpretability with SAEs, or 2.1, intro to RL. Route position: program week 11, tagged [light], exactly one of the two. Reading thread: Chen et al., reasoning models don't always say what they think.

The mental model

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

piece of planning. After the W44 track choice, one of two skills would serve the

capstone: 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 the plan's dependency audit found no later core page that 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; the route says so out loud.

The exercises

Quoted from the pinned pages:

(the part's graded exercise)

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

What it unlocks

Neither is a prerequisite for anything in the core route; that is what makes it a

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

reading list after the program; the RL path opens 2.2–2.5, most of which the plan

defers because RLHF assumes an A100 this laptop does not have.

Budget and receipts

Either candidate fits the week's slot with room for the pilot smoke tests the route

schedules alongside it. The decision record belongs in the W44 ADR, not in this

guide: the entry gets filled in for whichever road was taken.

[TODO: receipts — which road was picked and why, hours vs budget, what the latents or the policies showed (fill from the TARA vault when the week closes)]

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