"Theories that create theories" seems a completely arbitrary categorisation. We have software that derive predictions from records of past data. How is that not a machine creating a theory, albeit a simple one?
I believe the difference is same as fitting a curve to a data and then predicting the result versus figuring out the physics behind the data and using that to predict the result. I would consider the latter as the AI. The former is just deriving possible result from a set of rules and how to combine them, which might be smart but is still lacks AI in strictest sense.
If the prediction is good enough, and general enough, then I'd argue that is figuring out the physics behind the data.
What you're speaking of is a brute force mono-function program with no creativity or context.
The brute force approach is only feasible for small amounts of data. For larger data sets, often it's an iterative approach using general-purpose algorithms, combined with some degree of randomisation. I'd argue that describes how the human brain operates as well.