Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models