It is a bit like asking how exactly does the XGB matrix build in stable diffusion create images. We know the end result that is produced, but we can't pick a s

It is a bit like asking how exactly does the XGB matrix build in stable diffusion create images. We know the end result that is produced, but we can't pick a single number and talk about what it contributes to the overall algorithm. I would even guess that you could replace any single number with another random value within the appropriate range and you wouldn't be able to tell which AI was before and after. The links each item serves is too complex to work out. In the same way, any single 'act of science' can be hard to measure. They enable some experiments, which allow us to test some hypothesis. In some cases it is simple enough to point out when it is testing hypothesis that end up benefitting humanity, but that's only part of the effect. Many times a hypothesis is either rejected or fails to be rejected and slightly influences a larger theory which in turn is used to create new hypothesis. This creates a circular reference until some time later we end up with a theory that is useful for improving humanity. But which hypothesis were actually pivotal to forming the theory? Which acts of science allowed testing a hypothesis and receiving a clear enough failure that future effort was reallocated to other, eventually more fruitful, areas? Other than trivial cases, the cause and effect becomes such a tangled web we can't be sure. You even have the publicity side. Where sometimes the effect isn't directly contributing to the theory, but engaging more attention to a field which leads to more children choosing the route of becoming a scientist. Math has a similar issue. Much of pure mathematics has no known or foreseeable application. Yet, we can look back at what math does have very useful applications and see how many times the math predated the application being found, sometimes by long enough gaps that those who discovered the tools never saw them to be applied.