Wu Dao 2.0: 10 times the size of GPT-3

Stop making these glorified Markov-chain text generators bigger, and add some actual persistent memory.