Mech Interp Log

less than 1 minute read

A running log of self-study in AI alignment and mechanistic interpretability, working through learnmechinterp.com’s curriculum. Entries are dated to when the work actually happened, not polished after the fact — read in order, oldest first.

Falling down a LessWrong rabbit hole

May 27, 2026

Started reading about independent alignment research — embedded agency, agent foundations, the Hamming questions. Respect for the depth, but not sure it's where I should start.

alignmentorientation

Deciding on a path

May 29, 2026

Watched Welch Labs' Dark Matter of AI and an ARENA lecture, found Neel Nanda's quickstart guide, and decided to follow a structured mech interp curriculum instead of freeform reading.

alignmentorientationmechinterp

Transformer foundations, first pass

May 30, 2026

Breadth-first pass through the transformer basics — mostly review given the DL background, so moved fast. Started implementing a transformer from scratch in PyTorch.

transformersfundamentals

QK/OV circuits and virtual heads

Jun 2, 2026

Came back after a couple days off. LayerNorm was fine, but QK/OV circuits and composition/virtual attention heads took a lot more time than expected to actually get intuition for.

transformersfundamentals

BlueDot: RLHF, Constitutional AI, and dangerous capability evals

Jun 4, 2026

Two days into BlueDot's Technical AI Safety intensive — daily 2hr calls plus readings. Training safer models and detecting danger sections: RLHF's actual objective, Constitutional AI, and what dangerous capability evals concretely test for.

ai-safetybluedot

BlueDot: interpretability in practice, and wrapping up

Jun 8, 2026

Finished the BlueDot intensive. The interpretability-in-practice section was the one place the two tracks actually touched — SAE features as a monitoring layer, not just an explanation tool.

ai-safetybluedotmechinterp

What is Interpretability? / Linear Representation Hypothesis

Jun 9, 2026

The Linear Representation Hypothesis clicked fast via word2vec-style intuitions — though it's a hypothesis with known non-linear counterexamples, not a settled fact.

interpretability-fundamentals

Superposition Hypothesis

Jun 11, 2026

Anthropic's Toy Models of Superposition reframes capacity as combinatorial, not additive — though the jump from toy model to real LLM is an extrapolation nobody's directly shown.

interpretability-fundamentalssuperposition

Logit Lens, Direct Logit Attribution, Gemma setup

Jun 13, 2026

Setting up TransformerLens on gemma-2-2b (RMSNorm complications), then logit lens and Direct Logit Attribution — both fast since it's just residual stream linearity.

logit-lensdlagemma

Jacobian Lens

Jun 16, 2026

Jacobian lens generalizes logit lens via local linearization — more expensive than expected, and a nagging worry about picking whichever lens confirms the hypothesis.

jacobian-lensgemma

Reading Attention Patterns (first pass)

Jun 17, 2026

First cold read of Gemma's attention patterns, mostly noise without a hypothesis to guide it.

attention-patternsgemma

Activation Patching

Jun 18, 2026

First activation patching run on Gemma produced a suspiciously clean result on the first try — which is exactly when to double check, not relax.

activation-patchinggemmaioi

Stalled — back to Direct Logit Attribution

Jun 21, 2026

Stalled over the weekend trying to interpret patching results without DLA — went back and combined the two properly.

dlaactivation-patching

Attribution Patching and Path Patching

Jun 22, 2026

Attribution patching as a cheap linear approximation of real patching, and path patching to isolate causal paths through specific components.

attribution-patchingpath-patchinggemma

Reading Attention Patterns, take two

Jun 25, 2026

Re-reading the same attention patterns from over a week ago, this time with a hypothesis from patching in hand — completely different experience.

attention-patternspath-patchinggemma

Induction heads, first pass

Jun 27, 2026

First pass at induction heads following Olsson et al. — found textbook induction patterns on Gemma via a repeated-token diagnostic.

induction-headscircuit-findinggemma

Induction heads, K-composition

Jun 29, 2026

Confirmed the K-composition claim with path patching rather than just eyeballing attention patterns — the first real circuit-level causal claim in this whole curriculum.

induction-headspath-patchinggemma

Induction heads and in-context learning

Jul 1, 2026

Induction heads and the training-loss 'bump' — a mechanistic explanation for a loss curve shape I'd only ever explained behaviorally.

induction-headsin-context-learning

Refined attribution methods and self-repair

Jul 3, 2026

Refined attribution methods, and self-repair — a genuinely surprising result that maps onto failure-compensation patterns in multi-agent systems, with an open question about whether it's real or an ablation artifact.

attribution-patchingself-repairgemma

Causal Abstraction

Jul 5, 2026

Causal abstraction formalizes what patching has been informally assuming all along — and most practice, including my own, doesn't actually run the rigorous version.

causal-abstraction

Probing: classifiers and truthfulness

Jul 6, 2026

Linear probes on Gemma for sentiment and truthfulness, and the non-identifiability problem underneath 'the truth direction.'

probingtruthfulnessgemma

CAA, LAT, attention probes

Jul 8, 2026

CAA, LAT, and attention probes — CAA previews the steering block ahead, the other two felt less mature.

probingcaa

Probing in production, and where this leaves things

Jul 10, 2026

Closing out everything before steering: probing in production, and an honest note about selection effects in what counts as a clean circuit-finding story.

probingreflection