I haven't read this article, but the one place machine learning is really really good, is narrowing down a really big solution space where false negatives and false positives are cheap. Frankly, I'm not sure how you'd go about training an AI to solve math problems, but if you could figure that out, it sounds roughly like it would fit the bill. You just need human verification as the final step, with the understanding that humans will rule out like 90% of the tries, but if you only need one success that's fine. As a real world example machine learning is routinely used in astronomy to narrow down candidate stars or galaxies from potentially millions of options to like 200 that can then undergo human review.