Art, Obscenity, and Intelligence
What is the intelligence in artificial intelligence?
Do We Know It When We See It?
Like art and obscenity, intelligence is a very general word and one that's hard to define. We feel like we know it when we see it, yet like art and obscenity, what it is and whether it's perceived to be present is very much dependent on the observer. Even in the scientific community there are still significant questions as to exactly what is meant by the term intelligence, and some scientists and philosophers question whether intelligence is even a meaningful concept at all.
But to be able to discuss AI and AGI in a useful way, it seems that nailing down what we mean by intelligence, artificial or not, is a necessary precursor. That’s what I’d like to do here before moving on to other posts.
Modern concepts of intelligence have their roots in the study of psychology, neuroscience, and biological systems in general. Typically what’s meant by intelligence in these contexts involves the cognitive functioning of the brain, i.e. the processes within the brain that involve taking in data and then acquiring knowledge and understanding of that data as well as the domain in which it was obtained.
So our starting point is this: the intelligence being referred to in artificial general intelligence is equivalent to the end product of the cognition taking place in our brain. In other words, cognition is the process occurring in the brain that we're interested in exploring and intelligence is the result of that process.
Defining the Intangible
Historically, psychologists have defined intelligence with variations of the following:
Intelligence is the ability of an organism to solve novel problems, reason, have knowledge of the world, learn from experience, think rationally and abstractly, deal effectively with one's environment, and be imaginative.
This definition isn't bad and likely makes sense to most people. However, there’s an ambiguity to it, a lack of specificity that keeps us from really nailing down exactly what we mean by intelligence. What does it mean exactly to think rationally? If we do something that most people would consider to be irrational, are we no longer intelligent? What sort of problems are we solving? What do we mean by reason? By imaginative?
This definition is also somewhat circular, in that it mentions thinking in the definition when thinking is pretty much what we're trying to define. We run into more problems when we try to apply this definition to the non-biological world. While it might be possible to evaluate the intelligence of a dog or a cricket with this definition, what about a calculator or chess program or an autonomous vehicle?
In the early days of computer science, intelligence was considered to be centered around logical thinking, symbol manipulation, and building models of the environment. Eventually, the idea of a rational agent began to grow more prominent in discussions of intelligence. The rational agent model involves an entity which attempts to optimize the attainment of goals. It has a utility function that measures how close it is to reaching these goals, and so optimizing attainment of goals involves maximizing the output of this utility function.
Modern definitions of intelligence from the field of AI/AGI tend to stick fairly closely to this concept. Here are a few of the many similar variations one can find with a quick Internet search:
Intelligence measures an agent’s ability to achieve goals in a wide range of environments.
Intelligence is the ability to accomplish complex goals.
Intelligence is achieving complex goals in complex environments.
A system is intelligent to the extent that its actions can be expected to achieve its objectives.
These definitions are attempts to offer functional definitions of the word rather than a technical or systemic description. In other words, they describe the effects of possessing intelligence rather than the processes which give rise to it. Since we don't really know what those processes are at this point, this seems like the best route to take for the time being.
However, the overall utility of a definition is closely tied to what you can infer about the thing it defines. Given that, these definitions all seem to run into issues when considering the intelligence of humans and other entities.
Deconstructing Definitions
One of the problems is the use of the word goal in the definition. A goal is defined by Merriam-Webster, Cambridge Dictionary, and Wiktionary, respectively, as:
The end toward which effort is directed.
An aim or purpose.
A result that one is attempting to achieve.
These definitions would seem to imply that a goal is driven by intelligence, and that makes it problematic to use goal in the definition of intelligence itself. In other words, if intelligence is the ability to achieve goals and goals are the aim of intelligence, we're not really getting a lot of information out of these definitions.
Definitions like these attempt to address the central focus of intelligence, yet they really only describe a vague subset of what intelligence can accomplish. Defining intelligence as the ability to achieve goals is somewhat like defining taste as the ability to identify food. Identifying food is certainly one aspect of taste, but our sense of taste involves a broad set of processes and responses. One result can be, but isn't always, the identification of a particular food item. There's much more to it than identifying food, and we still taste things even if we can't identify the food.
On top of that, goal is about as mushy a word as intelligence, and it leaves many questions unanswered when used in a definition; what exactly a goal is could vary quite a lot depending on one’s perspective. Also, if an entity doesn't attain its goals, does that mean it's not intelligent? Should the choice of goals and random chance affect our assessment of intelligence, since they so frequently play a large part in attaining goals? It would seem that there are many ways to exhibit intelligent behavior that don't involve goals unless one dilutes the definition of the word goal to the point of near meaninglessness.
Definitions such as these also don't specify who or what chooses the goals in question, and this has led to some questionable applications of the word intelligence. For example, a mapping app can achieve the complex goal of routing a car to a destination in a complex environment. Given that many environments can be and are mapped, this process works well over a variety of such environments. That certainly seems to fall under the definition of the ability to accomplish a complex goal in a wide range of complex environments, yet most wouldn't consider Google Maps or Apple Maps to be intelligent in the way people or raccoons are intelligent.
The above example also highlights the problem with yet another hazy word: environment. Environments are typically the physical or relational surroundings in which an entity operates. However, as mentioned above, a navigation app can function in a number of complex environments. So can a cockroach. Neither is particularly intelligent.
What we really want to imply in our definition is that an entity can operate in a number of domains, where domain implies an area of activity or knowledge. Swapping out environment for domain in the definition above immediately causes our counter examples to go back down to the bottom of the intelligence scale, as navigation apps are poor at filling out income tax forms and cockroaches fail miserably at composing symphonies.
The last issue I’d like to address with these computer science-oriented definitions of intelligence is that they’re all fairly action-oriented. While our actions frequently have intelligence motivating them, we can engage in intelligent thinking without engaging in any actions.
Sitting in the dark in quiet contemplation with no particular goal, no outward actions or discernible behavior, still involves intelligence. So is ruminating over past actions, imagining future possibilities, admiring art, and questioning the nature of one's existence. These need not be goal-oriented activities of the mind nor do they require any interaction with one's environment. On the other side of the coin, many actions of our bodies don't actually require direct oversight by our conscious cognition and yet are still quite complex in nature.
Scoring Points with Brevity
Perhaps part of the problem with defining intelligence is that there seems to be a competition in the artificial intelligence community to come up with the most succinct definition of intelligence possible. While these definitions excel in minimizing word usage, the price they pay in accuracy, precision, and utility suggests that brevity may not be the most important factor when defining intelligence. They all seem to leave unaddressed many important facets of intelligence observed in biological entities. They don't help much in categorizing or measuring intelligence.
Most importantly, they're not very useful in defining the qualities we’re most interested in replicating in a machine, qualities that would cause us to consider that machine intelligent in the way humans are intelligent.
Perhaps the most crucial aspect of intelligence that these definitions fail to address is comprehension, the ability to understand what one is doing and why one is doing it. As we've seen with the recent success of machine learning systems like OpenAI's GPT-4, it's quite possible to gain competence at a variety of complex tasks. This kind of competence could be considered intelligence using the computer science definitions.
Yet no matter how competent machine learning systems are at these tasks, no machine learning system currently has any comprehension of what it's doing or why it’s doing it.
Distilling Intelligence
So let's consider a potentially more useful functional definition of intelligence, one that will not only help get around some of these issues but also make much more concrete exactly what sort of intelligence we would expect from AGI. What we want in a definition for intelligence is something that describes those aspects of biological intelligence that we're interested in replicating. We can then consider the degree to which other entities exhibit all or some of these qualities and set a target for what we might strive to achieve in a machine.
There may be aspects of intelligence that fall outside this definition, maybe even aspects we're unaware of because they lie beyond the bounds and capabilities of human intelligence. But this isn't really pertinent; what we want is a definition of intelligence such that if a machine's capabilities matched that definition, we would say it's intelligent in the same way that people are intelligent.
So, for the purposes of discussions on this blog, a functional definition of intelligence is the following:
Intelligence is that quality which allows an entity to solve a wide range of deductive and inductive problems, extract and prioritize information from the environment, infer causal as well as correlative relationships from both small and large data sets over many known and novel domains, generalize knowledge from a known domain to another known or novel domain, extrapolate probable outcomes from both factual and counterfactual circumstances, recognize in its own cognition both the potential for fallacies and the fallacies themselves, synthesize existing knowledge to form original concepts, and acquire awareness of its own cognition and of itself as an independent and unique entity distinct from other entities and from its environment.
Briefly breaking down this definition:
Solve a wide range of deductive and inductive problems: This is the core problem solving capability that humans and all other animals have to some degree.
Extract and prioritize information from the environment: This is what allows animals to assess their environment and navigate it as well as figure out what information is important and what is not.
Infer causal as well as correlative relationships from both small and large data sets over many known and novel domains: This is the ability to assess how data points are related to each other across many areas and to make meaningful conclusions about these relationships with a very small to very large amount of data.
Generalize knowledge from a known domain to another known or novel domain: This is the ability to take knowledge from one area and use it to develop knowledge in another area, even an area in which there is little or no previous knowledge.
Extrapolate probable outcomes from both factual and counterfactual circumstances: This is the ability to evaluate current facts and make reasonable predictions as to what will happen in the future based on those facts. It's also the ability to make predictions about future events based on the absence of certain facts.
Recognize in its own cognition both the potential for fallacies and the fallacies themselves: This is the ability to realize that one's thinking is subject to error and to recognize when there is a discrepancy between actual data and one's expectation of that data.
Synthesize existing knowledge to form original concepts: This is the basis of imagination and creativity. It's the ability to conceptualize something that doesn't exist or has not been personally perceived.
Acquire awareness of its own cognition and of itself as an independent and unique entity distinct from other entities and from its environment: This is the basis of consciousness, of self-awareness and identity, and it’s closely tied to comprehension. It’s also likely to be one of the more controversial aspects of this definition. Why this may be a crucial aspect of anything that has what humans would consider true intelligence is worth discussing in a future post.
Each of the above traits is a critical aspect of human intelligence, and all, or at least most, are apparent to some degree in other animals as well. How well any particular entity is able to demonstrate these traits corresponds fairly closely to how intelligent we consider that entity to be. The definition provides a rough way to judge ability on a scale that applies to those entities we consider to have some degree of intelligence. On the flip side, those things that we generally don't consider intelligent fall to or near zero on that same scale.
Real and Fake Intelligence
Before finishing up this post, it’s worth talking about the artificial part of artificial intelligence.
When we use the term artificial intelligence, are we talking about actual intelligence or some derivative and possibly lesser form of it? While the word artificial doesn’t necessarily mean something of reduced quality, in common usage it certainly has that connotation. Yet if something is intelligent, one would assume it has actual intelligence regardless of how that intelligence is generated or in what form the entity possessing it exists.
This brings me to the name of this blog, which is short for Synthetic Cognition. What we label as intelligence is the result of cognition, the mental process of acquiring knowledge and understanding. Synthetic doesn’t imply that the cognition is inferior; it simply implies that it’s not naturally occurring. A synthetic object or system can be as good or better than a natural one. For example, synthetic diamonds are just as diamondy as any that occur in nature with the added benefit of generally having fewer impurities and flaws.
What this means is that the term synthetic cognition removes the quality of intelligence from being questioned. The process that can be either biological or synthetic is cognition, not intelligence. Any resulting intelligence is genuine intelligence regardless of the specific process that demonstrates it or the material in which that process operates.
The term synthetic cognition has the added benefit of making it very obvious that we're not talking about calculators or navigation systems or any machine learning system, such as Alpha Go or GPT-4. None of these systems are cognitive systems; none have the ability to understand.
Unfortunately, the terms AI and now AGI have become standard and aren't likely to go away anytime soon, so I'll continue using them to avoid confusion. But it's worth keeping in mind the term synthetic cognition when discussing AI and AGI, because it's a term that emphasizes the crucial difference between what we are currently able to achieve and what we are not.