Oatmeal Agitante

if a model is optimized for making predictions, it may not be useful for e.g. producing counterexamples to other models

by "explanation" I mean a thing that acts on existing models to change/improve them

really though, the only reason I could say that of the former is that it is "under-specified" compared to the latter

so the former *might* be useful for explanations, or maybe not. but the latter seems optimized for something else

(also, everything I'm expressing here is based on my present contending with Popperian epistemology)

what if I specify that the latter learns; that its models continue to make good predictions in new situations?

what you're saying is rather abstract to me, can you give an example (tweet size permitting)

I've been going off the idea that an agent can make all correct predictions but have a very poor ability to "explain"