“The scientific community does not yet understand the theoretical basis for why scaling laws in AI pre-training work; they are an empirical observation.”