SIGNAL No.5 — Mechanistic interpretability is turning AI safety into an audit

Today's Signals

Hosted by Oli.

01 — Inside the model, inside the material

• MIT Technology Review names mechanistic interpretability one of its ten breakthrough technologies for 2026 — the patient programme of opening trained networks neuron by neuron, turning AI safety from a guess about behaviour into something a regulator could in principle audit
• American Ceramic Society reviews computer-based methods for archaeological pottery analysis — multidimensional statistics, cluster and factor analysis, structured-light scanning — where each sherd becomes a data fabric a database can compare across centuries
• arXiv describes a hands-free heritage scanning rig in which two coordinated robot arms plan their own paths and cover every surface of an object, removing the operator's wrist from the loop and producing baseline scans future curators can compare against

02 — Climate, written down

• Daily Kos summarises radar tracking of spring bird migration across North America, where peaks are arriving earlier and farther north — phenological mismatch finally measurable in close to real time
• ESA marks ten years of Sentinel-1, the radar pair that has been measuring how fast Greenland's and Antarctica's glaciers flow every few days since 2014 — turning a debate about trend into a question of magnitude

03 — Listening rooms

• Hirshhorn / Smithsonian held Sound Scene 2026: A Distant Mirror on May 2–3, a two-day public sound art event that treats listening as a civic activity rather than an art-world one
• NPR previews Dataland, the first museum in the world dedicated to AI art — Refik Anadol's 35,000 sq ft palace opening June 20 in Los Angeles, where data itself becomes the pigment
• ScienceDaily covers Neural Resonance Theory — the claim that the brain and body don't decode music as a puzzle, they oscillate with it, with sideways implications for rhythm-based therapies in Parkinson's and stroke recovery