The author reviews Christiano's "Prosaic AI alignment": says it was largely right, but misjudged supervised learning
"Prosaic AI alignment": 10 years later
AI summary
Paul Christiano's "Prosaic AI alignment," published ten years ago, reaches its tenth anniversary, and a retrospective article says its predictions have largely come true. Since then, powerful and general systems have indeed been built, yet no new understanding of the nature of intelligence has been gained; all that is known is that intelligence can arise from predicting the next token. The one point he got wrong is that superhuman capabilities also come from supervised learning, first AlphaFold and later LLMs.