He's right about reporting bias. The vast majority of hard problems are not amenable to AI, at least not by naively prompting. You need to have a fairly good grasp of the math to make useful prompts. There are exceptions (the one-shot solutions), but these are hardly representative of research-level math as a whole.