We currently know of no form of intelligence more general than human intelligence. It serves as our default reference point. It therefore seems reasonable to test for general intelligence by administering identical tests to both humans and machines. If the machine performs as well as or better than humans, it appears reasonable to conclude that, at least for that particular test of general intelligence, the machine has not demonstrated general intelligence inferior to that of humans. This is the fundamental principle behind the Turing test and remains central to the Arc-AGI test.
I fully agree with François Chollet’s view that we can reasonably claim to have achieved AGI when it becomes extremely difficult or impossible to design tests where humans significantly outperform machines. However, it makes no sense to test for AGI using tasks that no human—or very few—could successfully complete. Gary Marcus’s definition of general intelligence sets inherently superhuman and therefore arbitrary requirements. Moreover, unless we succumb to a double standard bias (which I believe is precisely what’s happening here), this definition denies the possibility of general intelligence in humans, even though humans constitute our natural reference point.
I therefore fully endorse the author’s analysis, which is very well-written, thoughtfully argued, balanced, and engaging to read. I also completely agree with the author’s point that general intelligence should not be viewed as a binary property but rather as a gradient. Indeed, there’s little difference between what we mean by “general intelligence” and simply “intelligence.” While a calculator or classical computer can do calculus intractable for even the best mathematicians, computational capacity alone does not define intelligence—though some degree of it is certainly necessary. On this point too, I find François Chollet’s reflections particularly insightful.
Recent LLMs are already partial AGIs—entities possessing an intermediate level of intelligence between calculators and humans, though arguably already much closer to human intelligence.
We currently know of no form of intelligence more general than human intelligence. It serves as our default reference point. It therefore seems reasonable to test for general intelligence by administering identical tests to both humans and machines. If the machine performs as well as or better than humans, it appears reasonable to conclude that, at least for that particular test of general intelligence, the machine has not demonstrated general intelligence inferior to that of humans. This is the fundamental principle behind the Turing test and remains central to the Arc-AGI test.
I fully agree with François Chollet’s view that we can reasonably claim to have achieved AGI when it becomes extremely difficult or impossible to design tests where humans significantly outperform machines. However, it makes no sense to test for AGI using tasks that no human—or very few—could successfully complete. Gary Marcus’s definition of general intelligence sets inherently superhuman and therefore arbitrary requirements. Moreover, unless we succumb to a double standard bias (which I believe is precisely what’s happening here), this definition denies the possibility of general intelligence in humans, even though humans constitute our natural reference point.
I therefore fully endorse the author’s analysis, which is very well-written, thoughtfully argued, balanced, and engaging to read. I also completely agree with the author’s point that general intelligence should not be viewed as a binary property but rather as a gradient. Indeed, there’s little difference between what we mean by “general intelligence” and simply “intelligence.” While a calculator or classical computer can do calculus intractable for even the best mathematicians, computational capacity alone does not define intelligence—though some degree of it is certainly necessary. On this point too, I find François Chollet’s reflections particularly insightful.
Recent LLMs are already partial AGIs—entities possessing an intermediate level of intelligence between calculators and humans, though arguably already much closer to human intelligence.