Careful... you're on the verge here of realizing that intelligence is not a single-dimensional thing and that 'g' is bullshit.
“AI can’t even x, how would you expect it to ever possibly do y?”
It’s like saying:
“How could a baby, which can not even tie its own shoes, spell a simple word, or fill a cup with water, ever possibly grow up to write a book, manage a corporation, or formulate a strategy for fighting a war? Impossible!”
It’s almost as if they think babies can’t learn. The AI is not a baby though, they say. Yeah, it’s not. But it’s a constantly learning body of work that is not regressing and not slowing down. Why would it not get to ‘general’ eventually?
There are 0 examples of a man made construct reaching "general" intelligence (whatever fuzzy meaning you may attach to it); there is no reason to believe that because we can reach some arbitrary human milestone today ("be a world class Go player"), any other human milestone ("start and run a company, and be a mentor and inspiration to others") is plausible.
This is a very different case from e.g. "we built a prototype plane that can lift 15 ft off the ground for 10 seconds, we should be able to build one that can lift 10,000 feet for 10 hours", because in the former case we were beginning to have an understanding of aerodynamics that made that path plausible. We do not have such an understanding backing current ML work. That's well illustrated by the fact that only 6 years elapsed between the first manned flight and crossing the Atlantic by plane - practice caught up to theory really fast, as it tends to do in those cases. We are about 6 years in the current ML revolution - what is its equivalent to crossing the Atlantic?
The logic you're advocating for here is more along the lines of "I taught my dog how to sit - shouldn't be too hard to get him to sit in front of a computer and start writing Python code now that I've got the first half figured out".
(or - I taught my dog how to fetch - shouldn't be too hard to train him to be a full time food delivery employee, etc).
No reason? If you were an outside observer looking at the very first individuals of what you decided to call human, maybe. But after so many steps were made? You can argue that, before there was a plane, there was zero reason to thing that humans can ever create one. And you could've claimed that for pretty much anything else. And you'd generally see that they did get created. Why would artificial intelligence be fundamentally different? I'm not saying "we can do it in the next five years", but to say that there's no reason to believe that it will ever happen?
> We do not have such an understanding backing current ML work.
And we did not have the knowledge of aerodynamics before we started looking into throwing things, then keeping things in the air for longer than they would if they were thrown, and eventually flying them (yes, a bit simplified).
I don't think anyone has ever said that and meant it. Except to the very naive, it's pretty obvious that a child growing up into a functional adult is a function of training volume and adaptation. On the other hand, training volume for "AI" tends to be huge and yet there are definite strata that these machines seem confined to.
I agree that that form of argumentation is meaningless though. AI is obviously improved over time, but as far as I know that's due to external modification and not just more training.
What is bullshit is the formal theory of multiple intelligences, which is largely PC-inspired pseudoscience and does not account for the correlation between separate areas.
A car's speed and its ability to drive off-road are highly correlated?
Rocket thrust and specific impulse are highly correlated? (Hint: these are inversely correlated.)
I didn't come to this through PC. I came to it through a basic understanding of learning and optimization.
Burmese pythons are way stronger than humans yet I doubt anyone would call them more intelligent. Mantis shrimp have way faster reflexes and their vision is way more sophisticated than ours yet they're not more intelligent as well.
There is a lot of debate over the definition of intelligence but one thing most definitions have in common is that it is general. A human being can play the piano, solve an integral, write a novel, compose a photograph, and many many more things. We do all of these things using the same "muscle", our brain. It turns out that some people's brains are faster at being able to learn all of these myriad tasks. That's all there is to it.
I think that's the main point where we disagree. I think there's a hell of a lot more to it.
I remember noticing when I was in school that while I could learn a variety of things, there were definitely subjects that were a lot easier for me and others in which I'd struggle no matter how hard I worked. I also observed other people with very different profiles of easy and hard subjects.
I barely passed organic chemistry. I studied hard, but it just would not "take." I had a friend who breezed through it. Anything that involved writing though? I'd breeze through it. I could get As in any class that involved understanding concepts and writing about them with almost no effort. This same friend struggled to write an essay.
I actually dropped out of vertebrate anatomy. Tons and tons of clades and names for discrete things. I found a way to adjust my major requirements to avoid taking it. Otherwise I would have had to switch majors. There is no way I could have passed that class with any amount of tutoring.
It's like I have a photographic memory for concepts, but can't recall discrete facts. I kept looking for the underlying patterns in organic chemistry but it turns out we don't understand it well enough to have those... we just know a ton of reactions that do things. I flounder badly in subjects whose facts can't be attached to an abstract conceptual framework, but I can read about a concept once and I understand it forever.
As a programmer I really love IDEs with "code insight" and other features that basically consult the documentation for you in real time. I get everything about the concepts and algorithms but can never remember the order of arguments on memcpy() even after using it for 20 years. I know memcpy() is there and what it does conceptually though, so it's fine since the IDE will show me which argument goes first.
My point is that different people who are "smart" can differ radically in their ability profile.
Take someone like John Von Neumann. One of the greatest mathematicians of the 20th century (and perhaps all time), he could also converse in a multitude of languages (native Hungarian, English, French, German, Italian, Ancient Greek). He was also an avid history buff, reading through Wilhelm Oncken's 46-volume Allgemeine Geschichte in Einzeldarstellungen at the age of 8 [1].
Later, he studied and became a chemical engineer while at the same time studying for his Ph.D in mathematics at the age of 20!
there is only one single gene that gives rise to all of the variance in intelligence among humans.
This does not follow. The claim is that some brains are better at learning than others. This means they can model and describe the complex relationship between ideas across a multitude of subject domains. It doesn't say anything about a single gene being responsible.
[1] https://en.wikipedia.org/wiki/John_von_Neumann#Child_prodigy
Yeah a zero cost sum model is totally wrong. As are models which assume that everyone is born equally smart and all differences just come from training or whatever. Some people have better abilities and it doesn't imply that they are worse in other areas. But they might only have extraordinary performance in specific tasks. And others, like I've mentioned above, have golden hands, von Neumann being an example.
I've done some searching and the model of intelligence seems to be an open area of debate, but it seems that research has moved on from the "not much more than g" era. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6526477/
> For several decades, the question of whether measures of specific cognitive ability contributed anything meaningful to the prediction of performance on the job or performance in training once measures of general mental ability were taken into account appeared to be settled, and a consensus developed that there was little value in using specific ability measures in contexts where more general measures were available. It now appears that this consensus was premature, and that measures of specific abilities can make important contributions even if general measures are taken into account.
See also https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6480721/ and https://doi.org/10.1111/j.1744-6570.2010.01182.x
'g' is a scalar that tries to measure the relatively large chunk of correlation that tends to show up on aptitude tests.
On average, individuals who do well on one type of intellectual task tend to do well on most other types of intellectual tasks. Maybe someone is good at all forms of pattern recognition but just can't manage 3d-rotation tasks. It happens.
So an IQ test isn't perfect. There isn't one true value that exists in the world that is 'g'. Many tests now break out 'verbal' and 'mathematical' IQ, but I suspect even those have a correlation > 0.5
Whatever that special thing is, once you see it in action you realise it's a genuine facet of something. You may not want to call it an intelligence, but I'm ok with doing so. It wasn't a physical thing, they weren't physically special, ergo it was a mental thing. "Intelligence" covers that quite well, I think.
There is a nice paper on this topic.
On the Measure of Intelligence - François Chollet https://arxiv.org/abs/1911.01547
I believe generality in human intelligence is overrated. We can't do large multiplication in our heads, we can't solve the Salesman problem in our heads, and many other things we can't do because our intelligence is not in fact general. It has a set of skills and can combine them in many ways.
There are some direct biological 'g' factors that you can think of. One would be sleep quality, another brain volume. Also, brain damage events like concussions, and probably anesthesia, affect both athletic and mathematical performance. Thus, you should also expect there to be kinds of genetic and congenital developmental deficiencies that do the same.
Looking at relationships of reaction time/IQ as well as target practice/ASVAB score, they have a stronger relationship at the low end of the spectrum. Among kids in the 90% decile, you shouldn't see much of a relationship! They've got low difference in 'g' factor, I guess. But if you go down to the 20-50% range, you get, to increasing degrees, all those malcoordinated kids on the playground who seem like they were born with a pre-concussed brain.
Obviously, parental mating patterns affect some of the overall correlation, but the way the relationship is stronger at the lower end indicates the presence of true 'g' factors.
[1] https://en.wikipedia.org/wiki/Theory_of_multiple_intelligenc...
[2] https://en.wikipedia.org/wiki/Theory_of_multiple_intelligenc...
It suggests that such a meta-ability might exist, but doesn't prove anything. It's also possible that these abilities are discrete but are correlated because they're subject to selection as a package... sort of like how physical coordination and endurance are probably selected for at the same time.
In any case the existence of such a meta-ability doesn't mean intelligence is one dimensional unless you re-define intelligence so as to exclude everything that can't be tied to that ability.
It is also possible, but there's no evidence for it. There's no evidence for any of Gardner's work. The biggest criticism of his work is that there's no empirical basis for the classification of his "intelligences." Gardner even admits to this, saying the choices are the work of an "artistic judgment" [1].
[1] https://en.wikipedia.org/wiki/Theory_of_multiple_intelligenc...
Real world fitness landscapes are complex and multidimensional.
My comment was a little snarky because while I like this site quite a bit for many reasons I do get a little fatigued by the prevalence of the IQ cult. The notion that there is a single metric on which brains can be sorted is sort of a sacred dogma around here, and one which to me is obviously quite wrong. All it takes is a casual glance at intelligence as it actually manifests in the only real world example of an intelligent system we have: biology. It's clearly a multidimensional quality that can be optimized for a large range of things just like anything else.
If intelligence is multidimensional in all other living things, by what rationale would we conclude that human intelligence is one-dimensional?
... or ...
If virtually every engineered or evolved capability is subject to multi-dimensional optimization, then why not also intelligence?
This doesn't mean that AI won't get smarter than us at everything we can do, but based on the fact we are not getting much smarter we have to ask: is that just a limit of biology? Or are there scaling limits to Intelligence? Organizations and societies are a great example of the scaling limits of "Intelligence", they bypass a lot of human limits to make gesalt intelligent agents more reliable and less prone to certain biases, but generally not meaningfully "smarter" than any individual, just more patient and consistent. There are clear limits on the scalability Intelligence, just throwing more computational power at the problem isn't enough for runaway-agi