François CholletVerified account

Generality in AI (which a matter of degree, not an absolute) is an v. important problem. Yet, strangely, the people who say they work on "AGI" don't seem to be interested in understanding the problem of generality, & instead focus on achieving task-specific skill by scaling up DL

It seems to me that the greatest obstacle to making progress towards AI of increasing flexibility & breadth (up to some 'generality' threshold) is the fundamentally incorrect conceptualization of intelligence that is propagated by these sci-fi memes, which is now pervasive

This is making it nearly impossible to have a conversion on this topic. Many people now think of "intelligence" as an intrinsic property of an algorithm, synonymous with "power over the world", that can be tweaked to reach arbitrarily high values. Like the height of a building

And many people are, accordingly, waiting for some kind of philosopher's stone, or lamp genie, of infinite intelligence and therefore infinite power. They don't quite know how to create it, but presumably all you need to do is figure out the correct series of incantations