Intelligent … Enough?
There is no intelligence, there is only memory. Because of this, Artificial General Intelligence is not achievable, but we can make better progress if we re-frame the discussion…
Spooner: “You are just a machine — an imitation of life. Can a robot write a symphony? Can a robot turn a canvas into a beautiful masterpiece?”
Sonny: “Can You?”
Last week, the internet was set ablaze by the most unlikely of sources … a 94-year-old man, named Noam Chomsky. What did he do? He wrote a piece that landed in the New York Times, lambasting and decrying the current state of artificial intelligence research and engineering. This line was probably the most inflammatory, as well as the one that stated his position most cogently: “The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching…” At this, the optimistic technologists have cried foul. One particularly fiery, though representative, piece from Scott Aaronson goes so far as to put Chomsky in the place of the conservative church authorities who condemned Galileo’s work. Strong words have been answered with strong words.
This conversation needs to be reframed. We’re going to do so with a little help from Will Smith and Alan Tudyk — or rather, their characters in the 2004 adaptation of Asimov’s I, Robot. (We’ll come back to that in a bit.) The debate, thus far, seems to be centered around whether GPT “and its ilk” are capable of accomplishing what Open AI and so many other teams are seeking: Artificial General Intelligence. Can a “lumbering statistical engine for pattern matching” produce general intelligence? What does “general intelligence” even mean? A brief walk through the annals of experimental psychology reveals that we don’t have a good answer to this question … yet.
The Search for Intelligence
We might sum up the search for intelligence as follows: in the late 19th and early 20th centuries, the theory of intelligence was the subject of a relatively new science … Psychology. The “Jesuit astronomers” of the 19th-century disciplines weren’t always welcoming to it either, but its proponents persisted. One such proponent was a psychologist by the name of Charles Spearman. In 1904, he eventually published a work, “General Intelligence” — Objectively Determined and Measured. He spends the bulk of his first and second sections enumerating the far-reaching investigations of his peers. Speedy tapping, perceptions of weight, reaction time, and subjective assessments from instructors were variously employed as indicators of intelligence. He closes by stating that “the only thing so far demonstrated is that the old means of investigation are entirely inadequate,” before positing his own framework, a “Theorem of Intellective Unity.” This included the famous “g” factor and the assertion that human intelligence was a single, measurable quality.
The problem is that we never found it. Firstly, it’s not clear that Spearman was all that certain about the unity of the intellectual powers of man. In his 1904 paper, he makes allowances for certain “s” factors, representing domain-specific proficiencies. Furthermore, in a 1927 book, he began to tease at yet a third factor. Raymond Cattell, among others, responded to Spearman by introducing a two-factor model: fluid and crystallized intelligence, in the 1940s. A popular three-factor framework, known as “triachic intelligence” was introduced by Robert Sternberg in the early 1980s. This was followed quickly by Howard Gardner’s “multiple intelligences” theory, which consists of some eight different categories. (I don’t know about anyone else, but their descriptions read vaguely like Myers-Briggs types, to me).
I’m already passing over a huge body of work. For instance, Joy Guilford once proposed a model of intelligence that, by way of combination, would have produced over one hundred categories! To make the field even more confusing, you have statements (such as the one made by my PSYC 101 textbook) that claim “intelligence can also have different meanings and values in different cultures.” But if intelligence is really so subjective, what is it that we’re trying to measure? One hundred years of conflicting (and increasingly complex) models don’t seem to have solved the problem. If we can’t decide what we’re measuring and we can’t say anything for certain about it, can we claim to really know anything at all, about it? It’s for this reason that intelligence frameworks up to this point have failed. (Here I’ll humbly admit — I haven’t been acquainted with one that’s convincing, but that doesn’t mean there isn’t one out there.) In each progressive refinement, what we are seeing is that “intelligence” is being described by the different types of situations in which one might act intelligently. In each of the above theories, “intelligence” seems to be equated with such notions as “aptitude”, “ease”, and “accomplishment.” I submit that the very notion of “intelligence” is a vague and unnecessary synthesis of other qualities that are better described using a different mental framework.
An Alternative Framework
When one studies learning with an empirical approach, it is not possible to directly measure or observe whether something has been “learned.” All we can do is give someone a learning opportunity and, after the fact, present that same person with a battery of tests to which they can apply that knowledge. Whether or not someone successfully applies what they gathered in their learning opportunity is called a product. These products can be assessed in a variety of ways, which we see as different types of tests. This presents an entirely different model for talking about “aptitude”, “ease”, “accomplishment”, and “intelligence.” Instead of someone performing well at something because of some intrinsic predispositions, their performance can be examined as an outcome of learning experiences.
If we pop back over to Gardner’s theory of “multiple intelligence”, we see that each one can be expressed as a set of “learning events” and “learning outcomes.” For instance, he talks about linguistic intelligence. The learning events that best facilitate the acquisition of a first or second language are well documented, including — but not limited to — emersion and (when it’s available) prior experience. He also lists mathematical intelligence. This, again, is a learning outcome. Much of the mathematics considered in K-12 education and early college is devoted to simple symbol manipulation. If one is particularly attentive to their rote-learning studies, then these learning events will compound to leave you with someone who is thought to be “mathematically intelligent” or, as the kids say, “good at math.” The learning outcomes associated with higher maths are more nuanced and require deliberate exposure to and familiarity with a great many mathematical systems — something that is also built up over time.
I will take up one last example because it sticks out in my mind as potent in the case against intelligence. I am a music teacher. I have had the privilege to help many kids begin making music and not just consuming it. When these kids start, very rarely do they come to me with the sorts of musical abilities Gardner might have had in mind. In those instances where they already have musical ability, it is because of some kind of past exposure. Each musician’s distinguishing ear, their sense of rhythm, and their command of their body is something that is built up over time through various learning events. However, even after a musician has demonstrated their capability to do a certain thing, there can be issues. They may play a wrong note, their sense of rhythm might break down in a certain section, or their interpretation might come across as uninspired. This is because ability and execution are two very different things. A musician’s ability to perform on stage as they did in the practice room is affected by all too many factors, including caffeine intake, sleep deprivation, the direction of our focus, the temperature on stage, and even the absence of the practice room itself. Time would fail me if I listed all such factors.
Learning this really hit me because it explained so many of my experiences as a young musician. At one point, I came to a youth orchestra rehearsal that I (thought I) was totally prepared for. It involved a particularly tricky section. In the practice room, I nailed it every time. At the rehearsal, I bombed it. When I went to my conductor to express my surprise, she said “yeah, we’re all rock stars in the practice room.” Her intent was to imply that I hadn’t really been nailing it in the practice room, but I knew better. What I had really stumbled upon was the practice/performance gap.
It is this gap between learning events and learning outcomes that makes the use of the word “intelligence” inappropriate. We are really only measuring whether or not a person can perform according to expectations, under a specific set of circumstances and assumptions. I posit that memory encoding and retrieval are much better frameworks for discussing intelligence. These concepts directly map onto our discussion thus far. A musician must encode a piece (or the necessary skills for sight-reading) and then retrieve it in performance. A mathematician must encode the behavior of mathematical objects as well as be able to retrieve them to compare to new circumstances. Finally, a language learner must encode the vocabulary, grammar, and spirit of their target language, as well as be able to retrieve those things when they are needed. A failure at any of these points would be perceived as a lack of intelligence, were we to use the vocabulary of the “intelligence” frameworks. It might not be flattering, but it doesn’t make any sense to speak of this problem in such terms. There is no intelligence, there is only memory.
What about the “ilk”?
Intelligence is measured in situation/response pairs. If a human, machine, or other system gives one of a set of appropriate responses in a given situation, we proclaim it to be intelligent. If we ask a machine how it is feeling, today, and it dutifully responds “Hello, World!”, we proclaim it to be unintelligent on account of its situationally inappropriate response. The degree of a system’s intelligence is positively correlated with the breadth of situations addressed and the percentage of situations for which it provides an appropriate response.
This is the measurement of intelligence. This is also where our quote from the 2004 classic I, Robot comes in. In this scene, our protagonist detective derides the robot for all of the things it cannot do: “can a robot do X? … can a robot do Y?” (Sound familiar?) At the end of his rhetorical attack, the robot simply asks, “can you?” No, our detective cannot perform these typically human actions. He lacks appropriate responses in our situation/response pairs. However, the defender of human intelligence will point to Spooner and say: “well, he could do that if you give him time.”
Here, we’re talking about a capacity for intelligent behavior. This is the universe of situation/response pairs that an intelligent system could hypothetically address. Most of the naysayers out there — Chomsky and Marcus among them — are pointing to this capacity for intelligence when they criticize current artificial intelligence models. And it is true, they have demonstrated that they do not understand, nor can they apply causal, narrative, and temporal relationships. The pro-pattern-matching yaysayser will typically fire back: “most humans have trouble with that!” To which the naysayers will return fire yet again, “but they have the capacity to do it!” And so goes the battle, ad infinitum. But is it really impossible for them to do so? We can’t answer that because we’re not entirely certain how we do it.
In human intelligence, we acquire primitive memories. These primitive memories are combined in infinitely complex ways to form memories in progressively higher levels of abstraction. This is a familiar process to people who work with computers. However, before we can answer whether or not our current approaches to artificial intelligence can match human intelligence, we have to understand this abstraction hierarchy. What fundamental types of memories can humans store? What are their means of combination? At some level of abstraction, we have some hints at the types of fundamental memories we store — such as procedural and declarative memory. How are these memories constructed and stored? When we can answer these questions, we can answer whether or not current artificial intelligence solutions are sufficient to emulate human intelligence.
Objective Intelligence
In the meantime, taking the word “intelligence” out of the debate also liberates us to give credit where credit is due. Part of the reason Chomsky’s essay received such ire is that, to those working on the problem, his words are dismissive. Even those who are close to the problem, such as Marcus, are targets of this response. However, if we use this framework of memory and retrieval as a mechanism to satisfy situation/response pairs, we can satisfy both sides. We can say “yes, the pure-pattern-matching yaysayers are right” when they say that these systems are extremely intelligent. The universe of situations our machines can now appropriately address — generating aesthetically pleasing pictures and acting as pleasant conversation partners — is what we called science fiction just a few short years ago.
However, Chomsky and Marcus are also right. There are specific situation/response pairs that we have not seen from our machines, yet. This means that they are not intelligent. Chomsky points to one such generalized group when he speaks of making “infinite use of finite means” — induction. Marcus points to modeling involving causation. Beyond this, there is a whole world of more subjective situations that we might expect from “intelligent” machines — opinions, virtues, emotional intelligence, and so on. It seems to me that these shortcomings all exist within a particularly important realm of human intelligence called modeling. While our memories appear to provide us with certain primitives for relationships (causal, temporal, and others), these don’t seem to be present in our artificial peers, yet.
But we humans are also limited. Take the memory systems we have just discussed. We have developed systems that allow us to accumulate information at rates that far surpass biological evolution. Given language — any language, written, oral, or physical — and culture, we can discover and pass on what we know about our environment. We can expand our repertoire of “situation/response” pairs at a rate much, much faster than any other creature because we can do so independently of the slow processes of environmental testing and reproduction. The Information Revolution has extended our ability to store and transmit information and wisdom. However, it hasn’t necessarily improved our ability to generate it or use it, ourselves. For instance, as expansive as our modeling abilities are, most of us still aren’t good at reasoning on feedback loops and delays — two of the primitives of dynamic systems.
Therefore, what if human intelligence weren’t our goal? What if we designed systems that could do much more? What if the set of fundamentals that construct our memories, once identified, is proved to be a relatively narrow set? What would happen if our machines could form memories using a much larger set of primitives? What if we took the fundamental aspects of dynamic systems and systems sciences — arguably, the fundamental aspects of everything — and produced machines that stored, associated, generated, and abstracted these fundamentals? What if we designed an intelligence with a whole new way of seeing — where it could make “infinite use of finite means”, but those finite means were so much more expansive than our own? If we did so, we would be seeking truly general intelligence — perhaps, even, objective intelligence.
