Minimizing log-perplexity loss is equivalent to maximizing survival length in a Turing test. Assuming compute-loss scaling law, a scaled-up GPT that produces human-like science papers would cost ~200 years of global GDP.| Yuxi on the Wired
