I like the breakdown into concrete attackable tasks! But I'm biased to believe in emergence from large-scale training rather than going after specific subtasks one by one (which will definitely increase productivity in the short term, and is worth doing just for that)
https://www.lesswrong.com/posts/9JbGq4t4ihDkXan5e/daniel-paleka-s-shortform?commentId=9coeTiaSv5CGxHTYN compare list of tasks
I like the breakdown into concrete attackable tasks! But I'm biased to believe in emergence from large-scale training rather than going after specific subtasks one by one (which will definitely increase productivity in the short term, and is worth doing just for that)