Mech Interp Log
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, 2026Started 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.
Deciding on a path
May 29, 2026Watched 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.
Transformer foundations, first pass
May 30, 2026Breadth-first pass through the transformer basics — mostly review given the DL background, so moved fast. Started implementing a transformer from scratch in PyTorch.
QK/OV circuits and virtual heads
Jun 2, 2026Came 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.
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.
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.
The Linear Representation Hypothesis clicked fast via word2vec-style intuitions — though it's a hypothesis with known non-linear counterexamples, not a settled fact.
Superposition Hypothesis
Jun 11, 2026Anthropic'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.
Logit Lens, Direct Logit Attribution, Gemma setup
Jun 13, 2026Setting up TransformerLens on gemma-2-2b (RMSNorm complications), then logit lens and Direct Logit Attribution — both fast since it's just residual stream linearity.
Jacobian Lens
Jun 16, 2026Jacobian lens generalizes logit lens via local linearization — more expensive than expected, and a nagging worry about picking whichever lens confirms the hypothesis.
Reading Attention Patterns (first pass)
Jun 17, 2026First cold read of Gemma's attention patterns, mostly noise without a hypothesis to guide it.
Activation Patching
Jun 18, 2026First activation patching run on Gemma produced a suspiciously clean result on the first try — which is exactly when to double check, not relax.
Stalled — back to Direct Logit Attribution
Jun 21, 2026Stalled over the weekend trying to interpret patching results without DLA — went back and combined the two properly.
Attribution Patching and Path Patching
Jun 22, 2026Attribution patching as a cheap linear approximation of real patching, and path patching to isolate causal paths through specific components.
Reading Attention Patterns, take two
Jun 25, 2026Re-reading the same attention patterns from over a week ago, this time with a hypothesis from patching in hand — completely different experience.
Induction heads, first pass
Jun 27, 2026First pass at induction heads following Olsson et al. — found textbook induction patterns on Gemma via a repeated-token diagnostic.
Induction heads, K-composition
Jun 29, 2026Confirmed 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 heads and in-context learning
Jul 1, 2026Induction heads and the training-loss 'bump' — a mechanistic explanation for a loss curve shape I'd only ever explained behaviorally.
Refined attribution methods and self-repair
Jul 3, 2026Refined 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.
Causal Abstraction
Jul 5, 2026Causal abstraction formalizes what patching has been informally assuming all along — and most practice, including my own, doesn't actually run the rigorous version.
Probing: classifiers and truthfulness
Jul 6, 2026Linear probes on Gemma for sentiment and truthfulness, and the non-identifiability problem underneath 'the truth direction.'
CAA, LAT, attention probes
Jul 8, 2026CAA, LAT, and attention probes — CAA previews the steering block ahead, the other two felt less mature.
Probing in production, and where this leaves things
Jul 10, 2026Closing out everything before steering: probing in production, and an honest note about selection effects in what counts as a clean circuit-finding story.