Scott Alexander examines AI-generated poetry produced by Gwern's GPT-2 model trained on classical poetry, highlighting its strengths and limitations.
Longer summary
Scott Alexander reviews Gwern's experiment in training GPT-2 on poetry. The AI-generated poetry shows impressive command of meter and occasionally rhyme, though it tends to degrade in quality after the first few lines. Scott provides numerous examples of the AI's output, ranging from competent imitations of classical styles to more experimental forms. He notes that while the AI sometimes produces nonsensical content, it can also generate surprisingly beautiful and coherent lines. The post concludes with a reflection on how our perceptions of poetry might be influenced by knowing whether it's human or AI-generated.
Shorter summary