Can AI solve Science?
Wow, I find myself agreeing with Stephen Wolfram much more than I expected!
It is also possible, as I am almost certain he would argue, that I fully misunderstood his points.
But I think he nails it on many fronts, such as when he says that a "central part of doing open-ended science is figuring out “what’s interesting”", and "what's interesting" is a very human concept (it is almost tautological, since we are talking about what's interesting to a human). "What's interesting" has to be computationally reducible to a human or a human-centric device, i.e., there has to be a way to compress the signal into a more concise human-readable description. I love that he is (or is becoming) so human-centric. AI can't "solve" science because we need a human in the loop (my interpretation).
I would posit, however, that Stephen's line of thought assumes that "relevant variables" have been identified. In theory one could argue that identifying the relevant variables is a computational issue, but if the number of possible variables is, for all intents and purposes, infinite, it is impractical.
Think of it like learning data science on Kaggle: the problem has been defined and a manageable number of potentially relevant variables isolated (plus the data is usually reasonably clean), so now you can apply all kinds of nice algorithms to perform the computational reduction. But it leaves the most important aspects of being a data scientists out of the equation: understanding the context, (re)framing the problem, (re)defining the objectives and identifying the potentially relevant variables, not to mention all the data wrangling and thinking about how the solution can be implemented in an organization.
Open-ended science requires this additional ability. I reach the same conclusion as Stephen, but with an even stronger case for why humans are required.
More on this soon. Once again, I will invoke Hod Lipson, who has some very interesting insights on the topic of variable discovery from data. How did Newton come up with the concept of acceleration, how did he figure out that it was a relevant variable?