The Problem with Intelligence
oreilly.com
oreilly.com
I suspect the idea that we need definitions first comes from education, where much of what we know is presented in this manner. This is rarely, however, the way our initial understanding was achieved - just consider how our concepts and definitions of 'energy' and 'matter' have changed over time.
In these situations does the creature try the eager solution despite it obviously not working near the end?
For example, does it take the close bridge that gets it 99% of the way across the river or does it not even bother and travel to the farther one that actually gets it across.
If it uses small sticks in nature will it employ artificially introduced tools to solve artificially introduced problems?
Etc.
The key here is to apply the test without breeding for it first. That is how you tell the individual is intelligent rather than the breeding mechanism. For example, I don't think the individuals are at all intelligent in OpenAI's[0] Hide and Seek. It's basically burned in instincts since the introduction of a button that would, say, swap positions with the farthest enemy agent wouldn't be utilized for hundreds of rounds of play. The learning is the brain and replication together. Reminds me of a story about aliens on LessWrong[1] about baby eating aliens.
[0] https://openai.com/blog/emergent-tool-use/
[1] https://www.lesswrong.com/posts/n5TqCuizyJDfAPjkr/the-baby-e...
You can very easily shoot a porn, even without being able to define what is pornographic and what is not (a proverbially notoriously difficult to define thing, often times debated legally in various countries, for example "movie where sex acts are performed" or "movie meant to arouse the viewer sexually" both don't cut it).
Not exactly, because both sex and filming could be done in a non-porn setting - including those two together (e.g. a sexual art film). And the question mentioned wasn't what is a "film that has sex", but what is porn - and even more so, what is it's definition (the same thing you asked for in the original comment, as necessary to "marks the goal").
The answer here (but how do we know we made a porn if we can't give a strict definition of what porn is) is that we can tell intuitively. Well, the same holds for intelligence.
That's perhaps the case for many things. Famously, we don't need to know what quadratic equations are to catch a baseball. Even though formally predicting its trajectory means solving one.
1. Reaction speed https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5608941/
2. Dendritic tree arborization, synaptic density, electrophysiological properties of pyramidal neurons https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6363383/
There is a problem with sophistry and competitive virtue signaling around these phenomena. Instead of ignoring biology we could as well embrace it.
I generally go for a slightly wider "The ability to create models of the world around them, and make predictions based on those models" - where "knowledge" would be "models" and "making predictions based on those models" is a sort of proto-skill.
Sure, the kinds of things that different people find easy to model vary. One person might find it easy to model mathematical theorems, others the internal working of car engines. But in both cases there's an underlying ability to make a mental model of the thing you're learning about, and use that to predict how it will function, and work out what you can do with it.
A human in a primitive society, even with the same IQ with a modern man, would be much less intelligent about solving a variety of tasks because they lack the mental "furniture". Most of our intelligence does not come from the brain, it comes from the culture, which is an evolutionary process.
And what is knowledge? When I say I know p, then p is a proposition. Thus, p is intelligible. Anything intelligible is conceptual. Indeed, if we analyze what a proposition is, we see that it entails predicates and predicates correspond to concepts. However, concepts are abstract and universal, that is, they are not concrete or particular, they are not mere images. Triangularity is not this or that triangle, but that which holds of all triangles, and you will not find triangularity out and about in the world on its own, but only instantiated in particulars, and any particular is not any of the other particulars for which the same predicate holds.
When we implement propositions in computers, we really only simulate the formal via mechanical manipulations. When I create a negation operation on a string of symbols, I am only moving symbols around in a way that corresponds to what negation would produce. But the computer is not strictly speaking negating anything. Furthermore, the symbols that stand for predicates are just placeholders at best. There is no concept in the machine. Deep learning does not somehow magically transcend this limitation.
> "The ability to create models of the world around them, and make predictions based on those models"
Modeling and prediction is not intelligence, but a consequence of it. It is also central to modern science because the purpose of modern science is, to a large degree, less about understanding nature than it is about mastering it for practical purposes (prediction is presupposed by control).
Also my oral communication tends to be more around vague ideas, not specifics, and I have trouble only communicating one idea at a time linearly. I’m sure I would score differently on the ability to model something depending on how the test was conducted.
In general I would say communication, focus and executive functioning will all get in the way of measuring raw intelligence.
You could also say that mathematics is a model of how Sets behave. It is a generalized model of reality. We don't create models of mathematical theorems, the theorems are a property of the very generalized and abstract model which is mathematics.
I'd challenge people who think definitions are necessary to grapple with coming up with a definition of a chair, in the form of a list of necessary and sufficient conditions. Make sure you don't exclude anything commonly thought of as a chair, or include anything not commonly thought of as a chair. For an extra challenge, come up with a definition most people would agree on. This endeavour will be a struggle.
And yet we manage to discuss chairs without difficulty. We have a generally understood concept of what we mean by a chair, with central examples like dining chairs or lounge chairs commonly held to be chairs. In the case of ambiguities of communication, we can clarify on a case by case basis ("Did you want me to take the ottomans to the other room too?")
Pattern recognition is close, as it's abstracting things into symbols and then comparing the symbols, but it's not sufficient. "Smart" I define as the ability to effect intentions, or to get what you want. In this sense, a lot of intelligent people are not very smart, and a lot of very smart people are unhindered by intelligence. Animals are perfectly smart for their environment without needing much intelligence. Humans are poorly adapted to our physical environment, and require a great deal more intelligence to have survived. Language is useful for many things, but the things it isn't good for are anti-smart (e.g. I think our ego as a refining filter for experience is an artifact of language)
Anyway, the Turing Test as a thought experiment isn't really a measure of intelligence so much as it is an economics model of an indifference curve, which is how much the observer cares (or not) about whether they are dealing with a machine.
I'm actually more bullish on the possibility of AGI for some admittedly very strange reasons, even though I am harsh about people who anthropomorphisze code and fall for animism. My view reduces to a kind of theistic argument where if we can create conscious life from rude materials, it is logical evidence that we ourselves may also have been the expression of some similar intention. If AGI is demonstrably impossible within our physics (like incompleteness theorem level proof), then we exist within a hard ontological boundary, and the best we can do is infer what that boundary is made of (probably time/gravity). The reason I think AGI is plausible is because I have theistic axioms that create a kind of circular reference where if we can Create life, then we could also have been Created with the intent to discover and appreciate the meaning of that, and if we cannot Create, we were not meant to experience that Creation. Maybe even if there is something on the other side of death, we may still just be programs or epiphenomena that aren't intended to reflect or apprehend our substrate, sort of like the one-way directional relationship between an instrument and a song played on it. An AGI would make the leap from song, to software, to operating on its environment using abstraction, generalization and feedback. It wouldn't be "us," but I think it could certainly become a them that could eventually exist independently of us.
Apparently whales can hear each other over distances of several thousand miles -- entirely across oceans.
So would a good thermostat be considered intelligent, assuming it takes great decisions? Intelligence needs to be defined with respect to a range of tasks and a range of priors. It's skill acquisition efficiency, not just taking good decisions.
I recommend "On the Measure of Intelligence" by François Chollet
>We argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to "buy" arbitrary levels of skills for a system, in a way that masks the system's own generalization power. We then articulate a new formal definition of intelligence based on Algorithmic Information Theory, describing intelligence as skill acquisition efficiency and highlighting the concepts of scope, generalization difficulty, priors, and experience.
Yep, you’re right. I forgot to mention that in my “loosely put” sentence, although the work by Hutter and Legg does mention that and discusses algorithmic information theory heavily.
I think approximately nobody is studying AI with the end goal of making a machine that is (say) as smart as a dog. It is a good intermediate goal but the hope has always been to surpass humans. If we want more doglike intelligences we can just breed more dogs.
Generally intelligent is only the start, and not a very interesting part. Animals can obviously learn things and they are "generally intelligent", but most are not all that bright. We are not all that clear on why either. Elephants have larger brains than us, so clearly it's not raw size but rather the complexity of the brain that makes us so much smarter. Which parts of brain complexity matter and which are accidents of evolution is almost entirely unknown.
Life is just a series of different kinds of tests, one after the other, every single day.
A bench press is simply a test of how much you can lift. But no one will say that it doesn't measure how strong your arms are.
A foot race is simply a test of how fast your legs can move. But no one will say that it doesn't measure how fast you are.
I think the thing that trips people up is that they often confuse knowledge tests with intelligence tests.
Reciting the periodic table is not a sign of intelligence, it is a sign of knowledge. As is knowing geography, languages, etc. If it requires a recitation of knowledge, it's most likely a knowledge test. Like most of those Facebook quizzes "Only the smartest can answer all these questions!!!". And you go to the test and it's just trivia. Trivia tests knowledge. Knowledge tests can be biased, they can be 'gamed' (except the best way to game them is to memorize the knowledge), etc.
Intelligence tests are something different. They test our ability to reason. Which is harder to measure. Because often there is a baseline of knowledge required. People like to dump on the old "brain teaser" and less old "leetcode" style interview questions. But they're probably better than quizzing people on esoteric language features. Language features are trivia. Questions with no real answer or no known answer are better proxies for intelligence. Not great and if you stick with the same questions, you'll eventually turn them into trivia as the answers become codified. But, you know, proxies are never the real thing.
I also happen to agree with the article. We do not have a clear definition of intelligence. Or consciousness. Or sentience. So unclear, we kind of use all three interchangeably. We have a fuzzy feeling. We don't think Lamda is intelligent/sentient because we don't feel it is. We feel that each other is intelligent/sentient because of course we are. We can't prove that to each other. And we can't come up with a definition that both includes everything our fuzzy feeling says is intelligent/sentient and excludes everything we think is not.
I, for one, definitely think of cats as unintelligent. A cat is like a potted plant compared to the companionship and relationship a dog offers.
Guess which one!
The answer is found here: https://www.nationalgeographic.com/animals/article/dog-cat-b...
Great apes are smarter than small apes. Big birds are generally smarter than small birds. Elephants are smarter than deer.
Well said. Here's another expression I often use: Stupider than Jupiter!
Jupiter is big but not very intelligent as far as we know.
You can't necessarily compare bird brains to whale brains, but within birds (or really any similar class of animal) neurons seems to directly correlate with perceived intelligence.
You can teach your cat to use the toilet, come when you call its name, roll and jump. They can even open doors which requires understanding how a door knob/rail system behaves. We had a cat that would follow us everywhere even if we went for a walk outside the house for several miles.
Additionally, cats are not nurtured like puppies are. There are considerable amounts of information and know-how from breeders on how to socialize puppies to develop their ability to follow instructions and respect their master. Kittens on the other hand are mostly raised in small cages in pet shops and treated as glorified hamsters that can jump high and purr. Going further, canines are pack animals which are naturally predisposed to work with others and understand complex hierarchies. Felines are mostly solitary and only really get in groups during mating season.
Another counter-example (unrelated to cats) are birds. Birds are incredibly intelligent, they can learn to solve puzzles, talk, and answer to their name. Yet if you ever get a budgie (or similar) you'll notice that your animal seems dumb as a rock. That's because birds require a lot of work in order to learn those tricks because they are naturally distrustful of humans. If you put ~1hr a day on training your budgie as you would with a young pup, you would see incredible improvements.
As a final note, I would concede that a cat's intellect is no match to a dog's, but using "follow instructions" as a metric to define intelligence and concluding "cat are dumb" is an oversimplification.
There are animal handlers and zookeepers who have trained cats to perform complex tasks on obstacle courses repeatedly & reliably, just like dogs and other animals. I saw this first hand at a zoo myself (I think it was San Diego but I'm not positive).
On top of all this, if you have interacted with a lot of cats, it becomes obvious that there is a ton of diversity in their personalities & behavioral patterns, even between siblings. There is no point in discussing 'smarter or dumber' here, their intelligence is just different-- it's unique to their species, as with pretty much every other animal capable of being trained.
With that in mind, dolphins have 2x and elephants have 3x the number of neurons as humans.
The thing I've heard is that cats are more like humans, in that cats operate on consent. Cats have to want the thing you are offering and they will make their desires known. They are conditional.
If you want worship, get a dog. If you want a low maintenance roommate/pal, get a cat.
I think you both overestimate dogs and underestimate cats.
A dog (or cat!) can show impressive signs of intelligence without having much inclination for tricks. Our dog knows who we are talking to on the phone, our daily routines and many "intelligent" things, like spoken instructions on how to behave the next morning, but hardly know any "standard" dog commands at all except "sit" and "stay".
I never have, but I've also known cats most of my life, starting when my mom decided the best way for me to learn their manners was to let them teach me themselves. It worked well enough to make me one of those people who gets along perfectly well even with somebody's cat who "doesn't like anybody!" - I can't count how many times I've heard that or variations on it. So maybe I'm biased, or maybe I'm speaking from greater knowledge. Make of it what you like, I suppose.
But really, we could say the same about almost any other animal on Earth, because almost no animal on Earth has been as extensively modified to suit human preferences and purposes as has C. familiaris. Absent the most extreme provocation, and sometimes even in its presence, they like us no matter what we do. Why not? We've spent several thousand years seeing to it that should be so.
So I'm not really sure what you're expecting you're going to prove with this whole line of discussion beyond that you like dogs and have some very strange notions about cats, neither of which points I think by now requires any further elaboration.
40 years of academic cognitive science would beg to differ.... This sounds like it suffers the common tech dunning-kruger approach to pre-existing literature