I keep a running note file of things I forget and re-learn while doing this work. Cleaned up a little, here's where it stands.

1. Circuits are messier than the papers suggest.

Every diagram makes a circuit look like a clean signal path from input to output. In practice it's a tangle of features that partially overlap, partially compete, and only look clean once you've already decided which five components to draw and which forty to leave out of the picture.

2. Correlation of activation isn't causation of behavior.

A neuron that lights up whenever the model is right about something is not necessarily the neuron responsible for it being right. The only way to know is to intervene — turn it off, turn it up, and see whether the behavior actually moves.

3. Ablation is the closest thing we have to a controlled experiment.

Zero out a component, rerun the input, measure what changes. It's blunt — you're asking the model to function with a piece missing that it never trained without — but it's the cleanest signal available for whether something matters or just correlates.

4. The best visualizations lie by omission.

A heatmap showing "where the model looked" is real, but it answers a much smaller question than it looks like it's answering. It's tempting to stop there because it's the part that renders nicely. The interesting part is usually in what didn't make it into the picture.

5. Interpretability is a product problem too.

A true explanation that no one can act on is barely more useful than no explanation. Part of this work is technical — finding the mechanism — and part of it is making the mechanism legible to someone who isn't going to read the weights themselves. I underrated this half for a long time.