Generalization beyond training data defines AGI, not task breadth alone
The key difference is the learning and generalization capability, not breadth of tasks. A system that does 95% of human jobs by copying millions of training examples is not AGI, even if economically valuable. True AGI must take a leap beyond what it was prepared for and generalize from what it knows to do new things. If you describe a novel alien civilization with bizarre customs never in training data, and the system reasons about it from multiple perspectives, that shows real inductive and abductive reasoning. AGI should figure out how to achieve unfamiliar goals in unfamiliar environments through creative imagination and trial and error, not just pattern matching against a massive catalog of prior examples.
— Rethinking the AGI Race, with Benjamin Goertzel · The Keynote on AI