Andrew Curran argues AI will accelerate scientific subfields with cheap verification loops, while other areas remain bottlenecked
Noah Smith warns some accelerated fields pose existential risks.
I talked this over with both Opus and 5.5 Pro. They got a little catty with each other, as usual, but eventually agreed on this ranking. Pro's actual response was extremely long, cited dozens of papers, and had to be greatly condensed for this table.
Andrew Curran@AndrewCurran_If we can suddenly field large numbers of high-level researcher minds, then the subfields that will be accelerated the most are those where cognitive labor can be applied at scale to already-available information. Where there is a fast and cheap verification loop, this acceleration compounds and things properly take off quickly. Where there isn't, the bottleneck simply gets moved. I was thinking about this after reading some of the responses to the GPT Erdos solution that talked about how attention-bottlenecked fields will see the biggest changes from attention-at-scale. GPT-5.5 Pro generated the image. Incredible improvement in text legibility, this would not have been possible a few months ago.