Lot’s of people doing work here, but what are the biggest questions to ask? What parts of a scientist can already be automated? What does it mean to be better than a human at a literature review? Does this mean you can synthesize the information better and answer more questions?

There’s this thing called the The Black Spatula Project which is trying to use AI to find flaws in papers. This is something interesting. I think part of science is being able to find these flaws extremely quickly and formulate the holes in them.

Why haven’t we fully automated this already? It seems kind of plausible that you can do some minimum criterion thing here where AI is able to find where new papers contradict old ones, and it can bring about the Kuhnian revolution faster than just humans can. Maybe part of Planck’s principle of science progressing by the old guard dying is that it isn’t clear enough what is breaking, but could you visualize this? Even if AI could just visualize all the holes in a field and connect them, this could help people develop new theories at a breakneck pace.

Then obviously the next step is figuring out the holes, and not trying to draw hard boundaries around any specific topics. Perhaps scientists just think about why things are, find unsatisfactory answers, then think very hard about what other observations tie into that to solve this problem.