On Intelligence

 

Intelligence is knowing when to quit

I used to participate in debate competitions in high school. They were the traditional parliamentary style debates where the effort was to convince an audience (or panel of judges) of your proposition. Imagine my surprise when I saw a U.S.-style debating competition where participants were speaking so quickly you could barely understand what they were saying, much less be convinced of their position. Here's an example. We are told the style helps debaters "squeeze long arguments into tight time frames." Maybe. But arguably, the essence of debate has been lost in this academic exercise.

I think something similar has happened over a much longer time frame to our understanding of intelligence. The concept is relatively new, but draws from older and related terms:

inter= between, within + legere =to bring together,
gather, pick out, choose, catch up, catch with the eye,
read; intellegere = to see into, perceive, understand

That could mean a lot of things. But today intelligence is thought of as (as various wags have stated over the years) whatever intelligence tests measure. We're not going to be satisfied with that obviously circular definition, so what then?

I won't quote the whole list of different approaches, but among them we have:

• the capacity to judge, reason and comprehend well
• the capacity to form concepts and grasp their significance
• the ability to adapt well to new situations in life
• the capacity to form responses based in truth or fact
• the capacity to inhibit and overrule instinct
• the ability to see relations and correlates

and so on. 

What can we draw from this? 

First, intelligence is not a thing, but a property of things, and specifically, a 'capacity' or 'ability'. This framing is remarkably dispositional; what I mean here is that we are being told not what intelligence is but rather what an entity that is intelligent does, and not what it does all the time, but what it does in particular (often counterfactual) circumstances. It is, in other words, a remarkably behaviourist characterization of what is intended to be a core cognitivist concept.

Second, there is a mechanism suggested that describes how this is accomplished, whether it is 'reason', 'forming concepts', 'adapt', 'inhibit', 'see', etc. These mechanisms mostly (but not exclusively) reduce to some mechanism of cognition; in Plato's day, it would be the exercise of the natural capacity to 'see' ideal forms of reason; in the Middle Ages the concept of 'nous' described the capacity of the intellect to apprehend truth. A modern approach might characterize it as the ability to perform calculations and construct representations in the mind.

And third, there is a success criterion which allows a definition of an entity being more or less intelligent. This is presumably what intelligence tests measure, though through history various candidates for success are presented, including comprehension of truth, ability to adapt, form responses, construct theories, and more. Whatever constitutes 'success' in human reasoning, we might say, 'intelligence' is the ability to produce it.

We grow uncomfortable with any of these elements as the cognitive abilities of computational entities begins to approach that of humans. We might agree, along with the critics of modern American debating, that going faster doesn't equate to greater success, but the three elements of traditional definitions may be challenged as well. Computers can certainly have properties or dispositions. They certainly have the ability to reason, and more recently, construct representations. And they are often (though not always) successful, for varying definitions of success.

Another approach might be to ask what, precisely, intelligence tests are testing for when they test for intelligence. Tests vary, but a standard list can be approximated as follows:

• verbal comprehension factor measures a person's verbal knowledge
• visual-spatial factor is designed to measure nonverbal concept formation, visual perception and organization 
• fluid reasoning factor measures the ability to detect underlying conceptual relationships
• processing speed measures the speed at which children can process simple visual information
• working memory is an indication of ability to hold information in memory to manipulate or perform calculations with it. 

I've taken intelligence at various times over the years and this approximates my own experience. There are also similarities and overlaps with other tests not specifically designed to measure intelligence but to measure something like intelligence (here I'm thinking of things like GMAT, LSAT, PISA, etc).

Numerous criticisms can be and have been made of such tests. 

For example, first, they are culturally specific (I remember being confronted with the term 'Ookpik' in a test, something that was totally outside my experience at the time; this tests asks participants to select the 'prettiest' woman from six drawings.

Second, it is not clear that they measure all and only factors related to intelligence. It's not clear, as we have discussed, that speed equates to intelligence. It's not clear that the ability to perform calculations involves 'holding information in memory'. And it would seem that there are many other ways of being 'intelligence', which would lead Gardner and others to postulate the existence of 'multiple intelligences'.  

And third, we still don't have a description of what intelligence is, only of (some things constituting) what an entity with intelligence does. It's a bit like, if we were asked to define a 'car', we responded, "it's something that goes forward, goes backward, turns, stops, runs without stopping for a certain number of miles, etc." We would not be happy with such a definition.

There are of course many ways to define things, and that is a part of the problem. Typically, when we define a thing, we identify what type of thing it is, and what distinguishes it from other things of that type. 'Intelligence' isn't a thing in the way a bird or a brick is a thing; it is a property, specifically, (a set of) capacities or abilities, but what makes 'intelligence' different from other capacities? It is not simply that there are success criteria, because other capacities also have success criteria, and are not forms of 'intelligence', such as the weight-lifter's capacity to lift weights.

Let's return to the car. Part of being a car is to be able to go without stopping for a certain distance (in certain conditions, such as, when filled with fuel, functioning properly, etc). The capacity is here expressed as a disposition, and in some cases (as in the case of a broken down old car) as counterfactually. In the right circumstances, the car would go a certain distance. 

In what circumstances? Here we have the key difference between our understanding of a 'car' and our understanding of 'intelligence'. A car would go a certain distance because it has an internal combustion engine, which means that, according to the laws of physics, when these conditions are satisfied, the car will successfully fulfill the function. If it is not fulfilling the function, then we can provide a story of explanation ("it is out of gas", etc) for this. We don't just say "well it's less of a car, then" or "it is not a car, then".

The car, in other words, is defined functionally. There are different ways to do this: we could say, for example, that the car was designed in such a way (which it most assuredly was), or that it is its nature to be that way, or that it evolved that way (so that the fulfilling of a function is a consequence of, rather than a cause of, the instantiation of the function), or that it could be seen that way (if we were, say, taking an 'intentional stance'), etc. 

All of these approaches play some role in the history of the definition of 'intelligence', and I think that we were perfectly happy to accept some sort of behaviourist or cognitivist definition so long as it enabled us to measure what we perceived as potential to function intelligently, and it is only with the arrival of (potentially) intelligent machines that this became more of a pressing problem. And it's a problem for two reasons, working together: 

First, we want to be able to say either that only humans can be 'intelligent' (i.e., it is conceptually possible for a there to be artificial general intelligence), or that a machine can be intelligent, but in such a way that we're not merely measuring similarity with human behaviour (as, in a way, the Turing test does). 

But second, both the behaviourist and cognitivist approaches are essentially black boxes that tell us (to some degree) what an intelligent entity does, but not how it does it. So then there is no principled way of determining whether an artificial entity can be intelligent; it can emulate intelligence, but there may always be the case (as eg. in Searle's Chinese Room example) that it does not possess the essential mechanisms ("a mind, understanding, or consciousness", you name it) for intelligence.

But without a story of  a mind, understanding, or consciousness, we don't have an understanding of intelligence, and it just becomes an empty word, standing for some of another set of black boxes, of which we have no understanding. That's why people like Chalmers call consciousness 'the hard problem' and why others tie consciousness to a 'sense of self' involving identity, agency, or will that a computer cannot possess. Whatever these are.

I think (and I haven't done the literature search to show this) that these sorts of responses break down or discussion of intelligence to a story in some way related to one or more of the following aspects of the self:

• will, or as it is described these days, self-regulation
• agency, which includes external factors involving capacity to act
• memory and retention, including what some describe today as 'content-knowledge'
• identity, including elements such as confidence, motivation, belonging, etc

To gt the full variety of such theories, we would also want to identify variations of each, specifically, those that reference some essential or innate property, those impacted by some sort of internal process or action, and those that are environmentally conditioned, including conditioned by culture, family or nationality.

Now it would be a full research program to trace these concepts as they appear in discussions of intelligence, weigh the various theories related to them (and I just know without looking that there is an entire literature around each). Now I don't want to engage in such a program; I'm too close to retirement. So I will just add my own fifth entry to the list (while acknowledging there may be more entries). 

• practice

So now I have two tasks: first, why would I reject explanations in terms of any of the previous four criteria, and second, what is the explanation with respect to my own criterion. 

My argument, in a nutshell, is that the previous four types of explanation are all circular. To explain a person's capacity to be intelligent, they postulate a thing, that embodies the person's capacity to be intelligent. Essentially (and I exaggerate a little here to make the point) each is a form of homunculus theory: "a concept is explained in terms of the concept itself, recursively, without first defining or explaining the original concept."

In short: I want to see a mechanism that describes how 'will' (or any of these) can lead to more (or less) 'intelligence'.

Now at this juncture it is a common response to suggest that what I am looking for is some sort of mechanism or mechanical process. For example, Martin Dougiamas responded, "There is a temptation to see the world as purely mechanistic, but there always seems to be some observer behind all our experience and even if it’s an illusion we need to nurture it.  Alan Watts and Sadhguru are useful here." 

In response, first, I would argue that there isn't really a good way to  distinguish between mechanistic and non-mechanistic, because when pressed, neither of these two terms makes sense. Do we mean by them 'digital' and 'non-digital'? Do we mean 'metal' and 'non-metal'? Do we mean 'physical' and 'non-physical'? In my experience, this distinction is frequently used to import into the discussion some property, entity or characteristic that is not in accord with scientific reality as we understand it, and at that moment, the possibility for useful discussion stops.

Second, and in recognition that a lot of what I believe is informed not only by western analytical but also eastern Buddhist and Taoist perspectives, I would say that Watts is useful, to be sure, but he doesn't exactly say"embrace the self".  Instead, "Watts argues with equal parts conviction and compassion that 'the prevalent sensation of oneself as a separate ego enclosed in a bag of skin is a hallucination.'"  Indeed, I see 'selflessness' as a major element of the eastern perspective.

And this leads to my third response, which is to say, with Hume, that we make this stuff up. It's an argument Hume offers over and over in his work, which is to say, "This connexion, therefore, which we feel in the mind, this customary transition of the imagination from one object to its usual attendant, is the sentiment or impression, from which we form the idea of power or necessary connexion." In other words: we feel a sense of contiguous self, and assume there must be an independently existing self (apart, obviously, from the merely mechanical), that must be the seat of all this, including 'intelligence'.

But even more to the point, and the strongest argument for believing that such theories are merely circular in nature, is that they cannot resolve, to any degree of satisfaction, some practical and pressing questions. 

Here is the question: can the theory of intelligence in question show, beyond grounds for reasonable doubt, that no one 'race' (whatever that is) or culture is of superior (innate?) intelligence than another? Now to be clear, I am personally very convinced that there is no such superiority; the commonality of human cognition essentially guarantees that this is the case. But - crucially - this cannot be shown by any of the four theories explaining intelligence I listed above.

In fact - as I comment wryly on Mastodon - we find the opposite: the vagueness of the theory allows those who would benefit from such arguments to make them using those theories. Hence, for example, we have an evolution of argumentation from the core concept of 'will' as seem in Schopenhauer and Nietzsche to the objectionable and offensive Triumph of the Will. Or we have David Brooks opining (or maybe pining) that All Cultures are Not Equal (meaning, ironically, that 'not all cultures are equal').

Those who explain intelligence in terms of will, agency, memory or identity have no good response to any of these arguments; and indeed, it is the continued popularity of such explanations that enable such views to survive and sometimes thrive. All that can be said against them is that they are morally repugnant, as indeed they are, but as we have seen in the recent surge of such views in some nations, their moral repugnance is nothing more than a point of view.

So what, then, of my own explanation (and therefore, account) of intelligence?

So: Now, drawing on my current experience, I have a pithy answer: Intelligence is knowing when to stop. Now I don't mean stopping in the sense of no longer cycling a vertical kilometer a day (though there is a Boltzmann-like quality to both). My answer is something like this: cognition is a neutral activity, basically pattern recognition. When I was younger I described this as relevant similarity. The 'relevant' part is important. You could go on forever, relevance tells you where to stop.

This approach requires not thinking of human cognition as a form of 'reasoning' and instead thinking of it as a form of 'recognition'. That's not to say that reasoning doesn't exist (that would be absurd). It is to say that reason is an emergent and public function of human intelligence that has many properties that are, as Hume would say, useful fictions.

In this discussion I am obviously drawing from recent work in artificial intelligence, and neural networks in particular, but there are clear and obvious parallels emerging from work in neural science - I find results pointing in this direction from people like Varela, Tversky, LeDoux, and many others. We may infer (correctly or incorrectly) that human brains 'think in language' and 'store mental representations' and 'perform calculations', but we can see that human brains form patterns of connectivity based on sensations, experience and other physical factors (such as nutrition or sensory deprivation).

So, by 'practice', what I mean is 'repeated sequences of experience and activity'. Activity may include physical activity (which may result in more experiences) or it may include mental activity (such as the production of sensations as pain, grief, elation, etc) and other embodied functions, which again may result in more experiences. Our conscious life (as I have argued elsewhere) is the ongoing sequence of experiences (not the having of experiences; consciousness consists of the experiences themselves).

This isn't the place for an extended thesis on cognition as connection and knowledge as recognition; I have made the argument elsewhere and people who have an alternative theory are welcome to try to solve the problem I set above. 

So how does 'practice' address the issue raised above? In a nutshell: there is more commonality between human brain mechanisms that there are differences between (so-called) races, cultures, nationalities, etc. That is to say, humans of different origins function the same way when it comes to matters of cognition and intelligence. There isn't an irreducible 'black box' feature in neural-network-based cognition that can be hypothesized to confer some advantage on some group of humans or another. It's the same set of chemical and physical reactions for everybody, and therefore, no subset has a built-in advantage.  

It is worth nothing that you can create what appears to be an advantage with the construction of intelligence tests. This is because intelligence tests are grounded in, and measure aptitude in, specific cultural phenomena. Each of the five elements (and I am by no means the first person to make this point) measures alignment with a specific cultural paradigm:

verbal comprehension factor measures language use common to a specific society (and often, to a specific (and usually advantaged) subgroup in a society.

visual-spatial factor measures a society's common ontology (in one case I was asked to count boxes on a store shelf, including 'hidden' boxes, which works only if we store food in boxes) 

fluid reasoning factor measures consistency with the sorts of patterns most prominent in a society (see Lakoff for a good discussion of how societies frame differently)

processing speed measures previously acquired and culturally specific mnemonics (such as the 'memory palace')

working memory (so-called) measures experience in specific (and often culturally-specific) forms of abstract reasoning and relationships

Intelligence, therefore, cannot be associated with s specific set of success criteria because the success criteria change. What counts as success for you might look like resounding failure to me (a good example is the recent case of GPT-5, which despite having 'PhD level knowledge', failed basic tests in spelling and geography - a clear case of substituting one set of success metrics (for which it was not trained) for another). We can do this with humans as well - hence the oft-cited case of a billionaire not knowing the price of basic grocery item.

Any measurement of intelligence on the 'practice' theory will be subject to the same phenomenon - and will be able to explain why as well. A person who has never seen tigers or images of tigers or descriptions of tigers will not be able to recognize tigers nor respond appropriately to them (one thinks of the case of English monks on the first arrival of Vikings. A person who has never worked with wood or anything similar will be a poor carpenter. A person who was never taught math or logic will perform reasoning operations poorly. That does not mean any of them is less intelligent. It just means their previous experience and practice do not match the current requirements.

So what is intelligence, then, on this account? Something like this:

For any set of previous experiences, and any set of present requirements, an entity comes to the best possible pattern recognition.

For example: if you have no experience of tigers, but experience of both cats and dogs, it is better to recognize the tiger as a cat, not a dog. In other words, your process of pattern recognition will trigger the most similar pattern (what I mean by that is, the net of neurons activated on presentation of the phenomenon is measurable as the 'most similar' to that which would be activated had you previous experiences with the current phenomenon).  

That said, as I mentioned above, there's no fixed definition of similarity. Indeed, for any given phenomenon, there are multiple different ways of defining similarity. For a tiger, the 'most similar' might be: cat-like things, black and orange things, four-legged things, animal-like things, things in cages, things Bill Watterson drew, etc. There is a potentially endless set of possible, and equally relevant, similarities.

This becomes most apparent when considering patterns in language. What is the best word to follow 'and' in a sequence of words? There is no end to the number of possible contenders, and no way, a priori, to determine the best possible candidate. But what helps is to have a wider context - not just "and <word>" but "<full sentence> and <word>" and to pay attention to the right context.

In the case of the semantics of counterfactuals, the truth of falsity of a proposition is defined by 'similar possible worlds', such that 'If P then Q' is true iff "if P then Q" is true in the most similar world. There are different ways of defining possible worlds (and hence, different types of modal logic) but when measuring 'similarity' of possible worlds the key factor is 'salience' - what possible world is most similar in the way most relevant to the current 'world' in which we are assessing the statement.

Being intelligent is picking the right possible world (not that people actually pick possible worlds when they recognize things; that's just a semantic analysis of what they're doing when they respond 'correctly'. In the world of language and models and representation, we could talk of picking the right 'context'. (So much humour is based on deliberately picking an incorrect context to answer a question).

In effect - what a person needs to do when presented with some experience or phenomenon is to consider a range of possible responses and 'settle' on the right one (again though: they're not actually doing this: this is just a description of a purely physical process that is happening in their brain). They could pick from any number of recognition - but they need to pick one (or, a few, as appropriate).

In other words, they have to stop recognizing and more on to the next phase, whatever it is. That means settling on the most appropriate context (also a form of recognition) to bring an end to the range of possible ways of recognizing something.

So what do we conclude from this discussion?

'Intelligence' is not definable in terms of 'the capacity to acquire, process, and apply knowledge and skills', at least, not in the sense measured by traditional intelligence tests; we need to know what intelligence is, not just what an intelligent entity does.

'Intelligence' is defined as (essentially) successful pattern recognition, which is typically context sensitive (which means that testing for intelligence might be more effective in unusual or even artificial context, and not contexts where previous practice and experience confer a significant advantage).

'Intelligence' isn't defined by (so-called) race, culture, or innate factors, except where these can be traced to physical impacts on pattern recognition, such as denial of a range of experiences, malnutrition, trauma, disease and genetic impacts, etc. specific to the individual.

'Intelligence' isn't something humans uniquely possess; indeed, any mechanism that successfully recognizes patterns has the potential to be intelligent (and the 'failures' of artificial intelligence can generally be explained in terms of inadequate or incomplete pattern recognition, including context recognition. For more of my thoughts on context, please see here.

Egilsstadir, Iceland