logoalt Hacker News

api • today at 2:03 PM • 1 reply • view on HN

That's kind of what I mean.

I don't personally believe that intelligence is one dimensional, to the point of doubting that "general intelligence" exists or is even a valid concept. It's like saying "general locomotion." There is no such thing. There are wings for flying through air, fins and rotors for water, legs and wheels for land, rockets for space. A wheel doesn't work well in the air or space and a rocket is a poor locomotion method in the water.

This implies, BTW, that I don't think humans have general intelligence, because I doubt the existence of such a thing categorically. We get all up in our egos thinking we do because we feel so smart at all the things we're smart about, and of course we can't see the other things because we can't see them. Might not even be able to imagine them.

A chess algorithm is superhuman at playing chess. A human beats AI at writing prose or telling a joke. AI can beat a human at bug-hunting in a giant code base or constructing a large mathematical proof. Humans can beat an AI at dreaming up entirely new directions for mathematics. AI can beat humans at taking those new directions and fleshing them out and relating them to current art.

And so on.

We in fact tend to build machines to be hyper-good at things we are not good at, because that makes them useful to us. So we make AI that is better than us at fast thinking, huge context windows, huge scale associative memory, etc., just like we made classical algorithmic computers that were hyper-good at doing math.


Replies

muellero • today at 4:02 PM

I'm always amazed by this perspective. Despite the fact that a single principle component falls out of all analysis of abstract human tasks across individuals, there is a large contingent of people that still "doubt the existence of such a thing categorically".

There has been an open question in psychology for almost 100 years now: can you design a solvable abstract task for human performance that isn't correlated with this principal component? No one seems to be able to do it, and the AI models have only reinforced this. Making AI models smarter makes them better at almost EVERYTHING. The scaling properties of LLM models, and the scaling properties of neurons in the brain, ensure a general improvement of pattern matching that seems to be task invariant.

Personally, I am convinced by the speed hypothesis here. Since IQ in humans is very correlated with thinking speed, the ability to generate and evaluate more hypotheses per unit of time yields general improvement.