[jal278]

The fundamental premise of this paper seems flawed -- take a measure specifically designed for the nuances of how human performance on a benchmark correlates with intelligence in the real world, and then pretend as if it makes sense to judge a machine's intelligence on that same basis, when machines do best on these kinds of benchmarks in a way that falls apart when it comes to the messiness of the real world.

This paper, for example, uses the 'dual N-back test' as part of its evaluation. In humans this relates to variation in our ability to use working memory, which in humans relates to 'g'; but it seems pretty meaningless when applied to transformers -- because the task itself has nothing intrinsically to do with intelligence, and of course 'dual N-back' should be easy for transformers -- they should have complete recall over their large context window.

Human intelligence tests are designed to measure variation in human intelligence -- it's silly to take those same isolated benchmarks and pretend they mean the same thing when applied to machines. Obviously a machine doing well on an IQ test doesn't mean that it will be able to do what a high IQ person could do in the messy real world; it's a benchmark, and it's only a meaningful benchmark because in humans IQ measures are designed to correlate with long-term outcomes and abilities.

That is, in humans, performance on these isolated benchmarks is correlated with our ability to exist in the messy real-world, but for AI, that correlation doesn't exist -- because the tests weren't designed to measure 'intelligence' per se, but human intelligence in the context of human lives.

[stephendause]

This is a good insight, but do you know of better ways to measure machines' abilities to solve problems in the "messy real world"?