2 ms·
We use variations on questions that are well known for duping inattentive humans, to test a system that we expect a priori to be incapable of such inattention.
by zahlman 2y ago
We use variations on questions that are well known for duping inattentive humans, to test a system that we expect a priori to be incapable of such inattention.
Unless "getting easy things wrong sometimes" is an inherent property of intelligence, we should expect that a properly "intelligent" computerized system would never err on problems far below its level of comprehension - unless we had some reason to believe it "wanted to", and as of yet I see no reason to believe this is even possible in principle.
Humans err, broadly speaking, for two reasons: genuinely reaching the limits of their comprehension, or trusting "system 1" (in Kahneman's analysis) too much.