Creative writing by OpenAI’s GPT-3 model, demonstrating poetry, dialogue, puns, literary parodies, and storytelling. Plus advice on effective GPT-3 prompt programming & avoiding common errors.

I continue my AI poetry generation experiments with OpenAI’s GPT-3 (released mid-2020), which is 116× larger, and much more powerful, than the 2019 GPT-2. GPT-3, however, is not merely a quantitative tweak yielding “GPT-2 but better”—it is qualitatively different, exhibiting eerie runtime learning capabilities allowing even the raw model, with zero finetuning, to “meta-learn” many textual tasks purely by example or instruction. One does not train or program GPT-3 in a normal way, but one engages in dialogue and writes prompts to teach GPT-3 what one wants.

Experimenting through the OpenAI Beta API in June 2020, I find that GPT-3 does not just match my finetuned GPT-2-1.5b-poetry for poem-writing quality, but exceeds it, while being versatile in handling poetry, Tom Swifty puns, science fiction, dialogue like Turing’s Turing-test dialogue, literary style parodies… As the pièce de résistance, I recreate Stanislaw Lem’s Cyberiad’s “Trurl’s Electronic Bard” poetry using GPT-3. (Along the way, I document instances of how the BPE text encoding unnecessarily damages GPT-3’s performance on a variety of tasks, how to best elicit the highest-quality responses, common errors people make in using GPT-3, and test out GPT-3’s improvements in NN weak points like logic or commonsense knowledge.)

GPT-3’s samples are not just close to human level: they are creative, witty, deep, meta, and often beautiful. They demonstrate an ability to handle abstractions, like style parodies, I have not seen in GPT-2 at all. Chatting with GPT-3 feels uncannily like chatting with a human. I was impressed by the results reported in the GPT-3 paper, and after spending a week trying it out, I remain impressed.

This page records GPT-3 samples I generated in my explorations, and thoughts on how to use GPT-3 and its remaining weaknesses. I hope you enjoy them even a tenth as much as I enjoyed testing GPT-3 and watching the completions scroll across my screen.