As a whole, an AGI is so general that we'll probably never be able to quantify its intelligence level. But we can quantify the intelligence of various "cross se

As a whole, an AGI is so general that we'll probably never be able to quantify its intelligence level. But we can quantify the intelligence of various "cross sections" of an AGI. As an extremely simple example, whatever the heck an AGI is, presumably an AGI can play chess. Therefore, we could instruct an AGI to compete in chess games and we could compute its Elo rating. This wouldn't be a measure of the AGI as a whole, but it would be a measure of the AGI as a chess-player. More generally, we can consider the AGI to be a reinforcement learning agent: not that the AGI reduces to nothing but a reinforcement learning agent (that would be ridiculous), but rather, whatever the heck an AGI is, an AGI would certainly be capable of performing in reinforcement learning environments. Various methods have been proposed to measure RL agent intelligence, ranging from more practical to more philosophical. Again, an AGI can be considered as a knowing agent, and identified with the formal set of mathematical theorems it knows in some language. Again, this isn't to say that's ALL the AGI is---that would be absurd. But it is one facet of the AGI. Whatever the heck the AGI is, it presumably knows some set of mathematical theorems in any given language (possibly the empty set, possibly a non-consistent set, etc.) There are various ways of classifying and even quantifying sets of mathematical theorems in a given language. See the slides I linked above for some details on this particular cross section.