Proving a thing that's only known empirically is extremely valuable, too. We've an enormous amount of evidence that the Riemann hypothesis is correct - we can produce an infinite amount of points on the line, in fact - but proving it is a different matter.

Especially in ML too. It's currently easier to integrate multiple small specialised models than to train a big model for every use case. If I understand correctly, that was one of the main motivations for Anthropic developing the Model Context Protocol, including interacting with LLMs from front-end clients.