ajinkya.ai An experiment in learning with AI.
← All entries
21 June 2026 1 min read 30% human written Authorship mix

Honest estimate of who wrote what. The green bar is my human-written share — drafting, editing, structure, voice. The red stripes are AI's share — synthesis, scaffolding, first drafts. 30% human, 70% AI on this one.

Mechanistic interpretability — reverse-engineering a transformer

Mechanistic Interpretability Transformers Interpretability Ai Safety Tutorial Interactive

Most explanations of a language model stop at the outside: tokens in, probabilities out. Mechanistic interpretability asks the harder question — what is the model actually computing inside, circuit by circuit?

There is one reframe that has to land first, or the rest stays abstract: stop reading a transformer as a stack of layers. Read it as a residual stream — a shared channel that every layer reads a subspace from, computes on, and writes back to. Once you see it that way, attention heads and MLPs stop being a blur and become circuits you can name, probe, and reverse-engineer — much the way you'd reverse-engineer a piece of malware.

The guide below is a map of that territory: the residual stream, what attention heads and MLP neurons actually do, features and superposition (why a single neuron can mean many things), and how researchers pull human-readable structure back out of the weights.

It runs full-screen, with live diagrams you can step through.

Open the interactive guide →