“OpenAI's 2017 "unsupervised sentiment neuron" paper was the first time the lab observed semantic understanding emerging from a language model trained simply to predict the next character.”