I recently spoke to a flow cytometrist in Boston about where AI will be most useful for science. Her baseline stance, as for most in bio, is of skepticism - the lack of data and verifiable tasks, real world experimental loops - mean that there are only specific areas of work where AI has been useful to biology so far 1
Her pov was this: AI will be most useful when targetted at problems that are 1) at the frontier of two complex fields that would require multiple human lifetimes of ongoing mastery 2) require very large scale data analysis
Navier Strokes is a fitting example that supports her hypothesis. At its most simplistic abstraction, it is a complex set of derivatives that explain every edge case of fluid dynamic movements. The sheer number of edge cases one has to consider to solve this would require ongoing study and mastery of nonlinear partial differential equations, functional analysis, fluid mechanics (among other things). You can imagine it is not a particular priority of how to spend a human lifetime (10,000 concurrent agents on the other hand…).
I am reminded of a line in the 3 Body Problem where the Chinese emperor tells Von Neuman “you may be clever but we have manpower which you don’t have. We can reach the same discovery through brute force.” The flow cytometrist appears to be correct - for problems like Navier Strokes that meet her two prerequisites: brute force mathematics, with human guidance and compute at scale is what proves most valuable.
Are we really to believe that superintelligence remains this constrained in generating novel scientific ideas? I can't help but think her argument is limited, but I do not have a strong counterargument yet. In fact, a lot of what I have read agrees with her. One of my favorite writers in biotech, Alex Telford, recently wrote about how <5% of new drugs entering the clinic are going after targets that haven't been tested clinically before, ~60% are going after targets with an approved drug.
Coding in contrast benefits from a large corpus of data from the open internet, which due to costs, bureaucracy, time are all much faster with software.
My current hypothesis as a result is that the emergence of discovery maps to the ability and pace of experimental validation.