I'm sympathetic to Terence Tao's view: AI solving outstanding mathematical problems without advancing a conceptual understanding may miss the forest for the trees. But I think this stance overanchors to current AI capabilities. Ultimately, I expect AI to generate useful insights too. A good analogy for Tao's view is neural architecture search. AI designed AI architectures. It achieved SOTA on benchmark after benchmark and seemed like the clear path forward. Why hire a bunch of AI researchers if the AI + a lot of compute can do it better? But it produced intricate spaghetti architectures that, as far as I know, yielded no obvious insights, unlike concepts like residual connections or attention mechanisms. It was hard for AI researchers to build on these. Further, a notable gap is that it found improved variations over the LSTM, but never architected the Transformer. But as we broaden ML objectives to not only solve problems but also generate useful explanations, I see no reason why superintelligent AI cannot also build a rich conceptual framework that mathematicians can build on.