Safety-Critical AI

TL;DR for operators AviationLMM1 is best read as a blueprint for how aviation AI systems might stop treating radio, surveillance tracks, telemetry, video, operational text, and sensor feeds as separate evidence streams. The paper’s central claim is architectural: before a system can reason across those inputs, it must encode each modality appropriately, align them across time, space, meaning, and reliability, fuse them into a coherent operational state, and only then generate task-specific outputs. ...

A car follows a cyclist on a narrow road. The double solid yellow line says: do not cross. The empty oncoming lane says: perhaps you can. The cyclist may feel uncomfortable being followed. The passenger may be late. The vehicle behind may be getting impatient. The automated vehicle must choose. A normal benchmark would ask whether the final maneuver is safe, legal, smooth, or close to a human reference trajectory. Useful, yes. Complete, no. ...

Inspection is not glamorous. It is not the robot demo, not the dashboard, not the moment a prototype obediently follows a traffic cone across a test track. Inspection is the slow, expensive discipline of asking whether the thing that worked once will behave acceptably when the weather changes, the path bends, the sensor gets confused, or the requirement was written by a tired engineer using the phrase “successfully complete” as if English were a formal language. ...