Mechanistic interpretability
The study of what computations a neural network actually performs, circuit by circuit, rather than what it produces.
Learn first
A trained model is a large matrix of numbers. Mechanistic interpretability tries to reverse-engineer the specific circuits inside it — which neurons activate for which inputs, how information flows from layer to layer, what internal representation the model builds before it generates a token. The goal is to read the mechanism, not just observe the output. That matters for safety: a model whose reasoning is opaque cannot be audited for deception or systematic error. The hard part is that circuits are distributed, polysemantic, and do not map cleanly onto human concepts.
Where it came from
In megatrends
Finds citing this concept
Overall = breadth + depth + substance + ½·novelty + freshness (a small boost that halves each week)
The thought before it's spoken
Anthropic found a small self-organized channel inside Claude that holds its silent working thoughts — edit a concept the model never says aloud, and its spoken answer changes.
How this connects
Tap a node to open it