Seems to me that it matters because it has implications for how we can train smaller models?