Is Human Intelligence Simple? Part 0: Intro
Should we even be looking for a "key" or "secret sauce" to intelligence?
A question that’s been on my mind for a while, and is relevant to all sorts of issues about AI and brain emulation, is to what degree human intelligence is “simple”.
What do I mean by that?
Here’s some more specific questions clustering around that idea of simplicity:
Is there a single change or handful of changes, anatomical or genetic, that differentiates human cognition from that of our hominid ancestors? Is there a simple “special sauce” that makes humans cognitively far ahead of any other animal on Earth?
Are human cognitive and behavioral abilities more like “separate modules” that work independently, or is there something more like a single “intelligence module” that we apply to almost every task?
If “intelligence is not one thing”, if there’s little overlap in the mechanisms by which brains accomplish different tasks, then there’s effectively no such thing as “solving” intelligence, there’s no information-theoretically compact “key” to how we do it all.
That implies things like:
Progress in AI emulating human or animal capabilities is basically incremental and empirical. There isn’t a single discovery about “how intelligence works” that would suddenly unlock much more powerful cross-domain AI.
Humans don’t think in a qualitatively different way from our close primate relatives.
There’s no such thing as domain-general competence. The idea of an especially “competent person” who can do all kinds of things well is an illusion.
On the other hand, if there is a “key” to human/animal intelligence, then:
Discovering the “key” to human/animal intelligence would produce sudden improvements in AI capabilities
Discovering the “key” to human/animal intelligence would give us much more well-defined ways of measuring how “smart” people are, and likely make it easier to research ways to make people smarter
This is a hotly debated issue, and it sounds very theoretical, but we can start to look at some empirical questions to narrow things down. Stuff like:
What really does distinguish “behaviorally modern” humans from earlier hominids?
Are there any interesting sharp cutoffs in cognitive ability between other animal species and do they have physiological correlates?
How localized in the brain are the abilities to perform different tasks?
How similar or different, in terms of cell type, neural architecture, firing patterns, etc, are different parts of the brain? Does the brain, or large subsections of the brain like the cortex, look like it’s doing “basically the same kind of thing in parallel everywhere”?
I don’t think the question of “is human intelligence basically one thing?” is so purely philosophical that facts and experiments can’t help.
And I’m going to be super hubristic and start tackling it in future posts with the hope of getting discussion and feedback from people with a variety of backgrounds.
I've often thought there's a difference between Recognition Intelligence and Planning Intelligence.
If you threaten a mouse, it will remember the safest place it can reach, plan a route, and run there. If you threaten an insect, it will find the safest place within immediate reach, go there and stay. This makes mice *way* better at escaping. Even though they're only slightly better at telling which spaces are safe.
I think fish, lizards, and gpt-3 variants are all in the insect category, though I've never had to chase those out of my kitchen.
I think the point about complex intelligence implying gradual progress is not necesarily true.
Humans are able to do science, build a civilization and do lots of things our primate relatives can't.
Theres clearly some discontinuity in impact there whatever the source.
Even if it turned out that's a result of a combination of thousands of tricks that only work when you have all of them that doesn't seem like it would necesarily mean progress will be incremental.
Or at least it could be incremental but then at some point you get enough tricks to add up to being able to get culture and do science and basically you get a big discontinuous jump that in humans looked like going from not being able to take over the world to being able to.
Also it's posible there is some simple trick that makes it possible to generate all components of intelligence.
The world were human intelligence isn't simple is not necesarily the world were some simple algoritm like gradient descent can't assemble it's pieces, like evolution did.