GPT-2: How to Build "The AI That's Too Dangerous to Release"

The sentence you just read wasn't written by me, the author of this article, nor was it written by the editor. No. What you just read was written entirely by OpenAI's GPT-2 language model, prompted only with the word "Today". Apart from another fancy acronym, GPT-2 brought along somewhat coherent (semantically, at least) language generation capabilities, some semblance of hope for zero-shot transfer learning, and a transformer network trained with approximately 1.5 billion parameters on a text corpus with over 40 gigabytes of internet wisdom. In this post, I'm not going to talk about better language models and their implications. As the great Stan Lee once said, "nuff said" about that.