Tagged interpretability · show all articles
Before any claim about what a model is doing, someone ran a forward pass, trained a probe, decomposed an activation into sparse pieces, and changed something on purpose to see what moved. This is that process, in the order it actually happens.
Long before the term existed, researchers tried to read meaning into a trained network's units. This is mechanistic interpretability's dated history: a naming paper, a landmark result, and three labs converging on one tool within months of each other.
Two questions decide how AI alignment and safety practice matures by 2035: whether verification moves beyond behaviour, and whether the field becomes an externally certified discipline or stays self-governed by labs. Four scenarios, each with a falsifier.
Two questions, not one, decide how Anthropic and the Claude line look in 2035: does its safety-training method stay on its current path, and does its scaling policy become everyone's floor or stay its own. Four scenarios follow, each with a falsifier.
Taking a mechanism apart tells you what each part was holding — but only once the part is off, and only for the mechanism on the bench. What interpretability has established, and what it has not.
Six problems that more scale has not solved, separated carefully from the many that are merely hard engineering, and the observations that would tell us any of them had been cracked.