My AI professor, in the early 90s, described AI like this: "In the 60s, we wanted to build computers that acted like people. Not just people, but smart people.

My AI professor, in the early 90s, described AI like this: "In the 60s, we wanted to build computers that acted like people. Not just people, but smart people. So what do smart people do? We play chess! So we spent a lot of time beating chess and learned basically nothing about AI." Beating the Math Olympiad strikes me as much the same. They're solving "hard" problems, but not solving easy ones. I want a robot that can clean a toilet. Hand it a brush, send it into the room, and get it clean. Then have it make the bed, without crushing any of the stuff strewn haphazardly about. Something humans do for minimum wage because anybody at all can do it. Physical manipulation of the real world isn't strictly required for AI, but as a test it precludes a lot of solving the hard problem without solving the 'easy' problem. The real world is very unforgiving of automata in uncontrolled circumstances, something that animals (not just humans) do with minimal effort.