Scientists Found a Way to Peer Inside AI Brains

Mechanistic interpretability aims to uncover hidden model states, component effects, and intervention responses. This paper introduces "mechanistic tomography"—a framework for measuring these internal mechanisms through designed interventions. It describes a practical procedure: begin with the least costly measurements, test on held-out interventions, calibrate simple mismatches, and expand the measurement family when structured residuals appear. Because an estimate that guides an intervention acts as an observer, control settings provide a strong validation test. In experiments, observer error tracks control error, and sparse aggregate measurements can recover finite-effect maps with fewer interventions than coordinate patching. On GPT-2-small, a specific cross-group interaction is the largest held-out predictive term; on Qwen-2.5-7B, a calibrated additive map suffices without pairwise lifting.