Short answers, pulled from the story.
Chris Olah coined the term mechanistic interpretability to describe reverse-engineering neural networks. Before this moment, researchers mostly looked at what models produced rather than how they worked inside.
The field emerged as a distinct subfield within explainable artificial intelligence after Chris Olah defined its core purpose of understanding internal mechanisms in computations. Early work combined feature visualization and dimensionality reduction with human-computer interaction methods to analyze models like Inception v1.
The goal remains identifying structures, circuits or algorithms encoded in the weights of machine learning models. This approach seeks to analyze neural networks in a manner similar to how binary computer programs can be reverse-engineered to understand their functions.
Empirical studies suggest high-level concepts exist as linear directions within activation spaces where relationships between entities follow straight paths. The relationship between a country and its capital is encoded in a linear direction in examples.
Researchers employ formal causality theory tools to trace how internal components influence model outputs and identify potential risks before they become problems in real-world deployment. Verification becomes possible when researchers can inspect the internal logic rather than just observing results.