[nl]

Or from today, OpenAI's response to BERT: https://blog.openai.com/better-language-models/

Breaks 70% accuracy on the Winograd schema for the first time! (a lazy 7% improvement in performance....)