A common test is something like ARC-AGI 2, where even the best models struggle to get into double-digit percentages. One thing I'd expect from AGI would be to

A common test is something like ARC-AGI 2, where even the best models struggle to get into double-digit percentages. One thing I'd expect from AGI would be to go beyond "Solve this ARC-AGI test" to "Come up with more tests like ARC-AGI." It can't have been lost on the humans who worked to create those tests that their own jobs should be automatable.