SIGNAL No.5 — Mechanistic interpretability is turning AI safety into an audit
SIGNAL by SORAH by SORAH INC.
Episode notes
Mechanistic interpretability — MIT Technology Review's pick for a 2026 breakthrough — is patiently turning AI safety from a guess about behaviour into something a regulator can actually inspect.
Tech (interpretability and heritage instruments) continues with the American Ceramic Society's review of computer-based methods for archaeological pottery analysis — structured-light scanning, cluster and factor analysis, sherds as data fabrics — and an arXiv paper on a two-arm robotic rig that automates 3D scanning of cultural heritage. Climate & Nature moves through North American spring bird migration, where radar networks show peaks arriving earlier and farther north, and onward to ESA's Sentinel-1 marking ten years of continuous radar velocity records over the Greenland and Antarctic ice sheets. Ar ...