It seems we've achieved the same thing using large language models. Essentially by learning the sequential structure of data, we have somehow created a machine that generates that intermediary representation, that we don't understand but that seems to be good enough.