DeepMind says its new language model can beat others 25 times its size

RETRO uses an external memory to look up passages of text on the fly, avoiding some of the costs of training a vast neural network