> JP: As a lab, our objective is to match human intelligence. We're still very, very far from that, but we think it’s a great objective.
Also JP: AGI is a “bogus concept”.
??? How can you possibly square these comments?
That's what the term means.
Anyway, this idea that human intelligence is ‘not general’ is incredibly short-sighted coming from someone in such a post.
I'm mashing a bunch of pieces of carefully-refined material to convey a bunch of squiggles to convince people I've never met a point about a person I've never met. This sort of unstructured problem-solving is exactly what generality is.
Human intelligence being general is not the same as human intelligence being built of unspecialized pieces, or the idea that we should be able to solve all problems equally well, regardless of difficulty of how abstractly we have to approach them.
Consider the comment about shuffling all our optical nerves.
1. This is a HARDER problem, not an equivalent one, despite the apparent isomorphism. The minimum solution for the unshuffled problem is shorter than for the shuffled one. One would thus expect that it would take more time to learn, and more computation to process in real-time, independent of whether the human brain is specialized.
2. Our brain does handle problems where the optical nerve has been shuffled without introducing entropy, as examples like the inverted vision experiment show.
3. It's widely understood that at a low level human brains are more weakly general than at higher levels. If you presented these two tasks to a human on a computer, such that they can consider them holistically, the shuffled task would only be harder to the degree that the problem is fundamentally harder—humans would not struggle much to unshuffle a camera, given a little extra time.
4. Humans have learned to echolocate. They've learned to use sensory-substitution devices that project images on to their tongues. LeCun carefully says “to the same level of quality” because, you know what?, humans are so general that he can't even rule out the brain partially solving this shuffled-neurons problem too.
You can answer the question of "Is this machine as smart as a human" by giving it the same data as a human and asking it questions, how would you ever determine if its intelligence is truly general?