Physical World Brittleness

AI systems trained on clean data break when deployed into messy physical reality.

The pattern: data collection occurs in controlled environments to ensure model performance. Deployment occurs in what practitioners call “utter chaos,” where cameras move in wind, sensors drift, dust appears on lenses, lighting changes, and human behavior shifts unpredictably.

Models trained on pristine images mistake specks of dust for cancer. Camera movement from weather causes complete traffic detection failure. Models can’t distinguish vegetation growth from landslides.

The brittleness isn’t a bug in the model, it’s a consequence of optimizing for clean training data while deploying into environments that were never captured in that data.

Related: 04-atom—data-cascades-definition