Search This Blog

The Nature of Intelligence

In my last post I wrote an epic review of two books that delved into the nature of intelligence and the limits of computation: Gödel, Escher, Bach and The Annotated Turing. Both books sparked all kinds of new ideas about artificial intelligence (AI), especially GEB. I tried to stick to the material in the books for the review, but now it's time to dig in and explore some of the ideas those books spawned in my mind about AI. These ideas boil down to three main questions that define the scope of issues surrounding AI.
  1. What is the nature of intelligence?
  2. Is artificial general intelligence possible?
  3. What could happen if we create AGI?
These are big, complex questions that plenty of smart people are trying to answer for various applications. The specific reference to artificial general intelligence is there to distinguish it from the numerous examples of artificial narrow intelligence that we already have, such as chess programs, simulations, equation solvers, and search algorithms that do things better than us humans, but only in a narrowly defined task. AGI is a type of intelligence that we have not yet achieved with computers as of yet.

The implications of these questions are fascinating. The answers to the first question will define how we would recognize intelligence and what we're aiming for with AGI. The answer to the second question is almost certainly yes, but much more is behind it than a simple yes/no answer. The emergence of an AGI that meets the answers to the first question would show the positive result of the second question. The answer to the third question is extremely hard to foresee, and the possibilities get extremely gnarly when coupled with the property of exponential growth. The actual result will most likely determine our future as a species. Heavy thoughts. We'll dig into the first question in this post and cover the other two in subsequent posts.

What isn't intelligence?


One of the problems with AI research—and one reason why AI has been reclassified into ANI, AGI, and ASI—was that researchers continually made claims that once we had created a program that could do X, we would have created AI. They came up with so many variations on X that were inevitably achieved once the challenge was proposed that Tesler's Theorem was coined to express the phenomenon: "AI is whatever hasn't been done yet."

Of course the fault was not that we were redefining intelligence or that the novelty of AI wore off each time the new definition was achieved, but that we were trying to define intelligence through specific tasks. Intelligence is not the ability to do any specific task, but the ability to perform a wide variety of tasks and deal with a wide variety of situations.

Our concept of intelligence is by definition, general. To lay out a task that could differentiate AGI from ANI, the task would need to involve the synthesis of enough knowledge from a large enough range of topics that the AGI could perform much more than the initial task it was programmed for. Maybe some of the current tasks we're trying to program computers to do—image recognition, language processing, and expert systems—will lead to AGI, but I'm still skeptical. I think we're going to have to get even more general than that, but these latest tasks are on the right track.

At least it's easy to define what isn't intelligence at this point. Intelligence doesn't do only one thing well. It does an array of things well, and it can adapt its skills to new situations. An AGI that plays chess better than any human, but can't do anything else—even things that we consider simple—is no AGI at all. In anything other than chess, we would consider it stupid.

This line of thought quickly leads to The Turing Test. This test was conceived by Alan Turing and basically states that if a machine could fool a reasonably intelligent human into thinking it was another human, then it would have achieved human-level intelligence. Such a machine would have to have at its disposal enough knowledge about the world and the ability to apply that knowledge in a believable way that a human could not distinguish it from a thoughtful person.

Are there wildly different types of intelligence?


The Turing Test may seem to limit intelligence to our concept of intelligence as human beings. What if other types of intelligence exist that we aren't able to identify as such because they don't conform to our preconceptions of what intelligence is? That seems plausible. Let's explore that line of thought. I can think of two major systems that could be construed as intelligent from a certain perspective, and they are both much grander and more diffuse than human intelligence. They could pass by unnoticed unless we work to understand how their behavior can be interpreted as intelligent.

The first example is biological systems, i.e. life. If we look at how the building blocks of life—amino acids, DNA, and RNA—behave over billions of years of evolution, we can see numerous signs of intelligence. We see that life is an excellent problem solver and is highly adaptable to a wide range of situations here on Earth. Life experiments through mutations with failed experiments being discarded through extinction and successful experiments carrying on when life flourishes. Life could even be thought of as learning what works and what doesn't work through the results of these experiments.

Life has a language in the genetic code, and it uses that language to record information and pass it on through self-replication. The genetic code is both a written record and a form of communication in a chemical medium. We can even think of life as having a goal, and that goal is to survive. Genetic code that is able to reproduce and spread will do so, and it seems to have an innate drive to fight for survival. Life will find a way, right? These characteristics are all things we associate with intelligence, but it is an intelligence that is happening at a different level than what we are normally consciously aware of. Our human intelligence is built on top of a biological system and depends on that system for its existence, but it certainly seems like biological systems are their own form of diverse intelligence.

We can take another step back to look at an even larger form of intelligence, cosmological systems. The universe itself might be thought of as a form of intelligence. On the largest scales matter clumps together over billions of years to form galaxy clusters. These clusters form galaxies held together with dark matter and gravity, and within them hydrogen clumps together to give birth to stars. These stars, especially the larger ones, burn through their hydrogen fuel, creating heavier elements. Then they explode in supernovas, creating even heavier elements.

Over generations of stars, enough heavy elements are created to form rocky planets that orbit around newly born stars, like our sun. Our own rocky planet Earth is large enough to maintain a molten core of iron that operates as a dynamo that creates a huge magnetic field, protecting the Earth from the sun's radiation. Without the magnetic field, the Earth's atmosphere would be stripped away, leaving it a desolate wasteland without oxygen, nitrogen, or liquid water.

All of these processes result in an enormously complex system with numerous hierarchical levels from galaxy clusters down to planets. The language of this system is the collection of four fundamental forces and twenty fundamental particles of the Standard Model, and elements of the system are constantly in communication using this language. Different aspects of the system can be thought of as solving different problems in the system: how to keep a galaxy together, how to create heavier elements, or how to put a planet with an atmosphere close to a star.

Am I actually proposing that the universe is a sentient intelligence that is actively constructing a place for us to live? No, probably not. What I am proposing is that I certainly don't know if the universe is intelligent because the behavior of the universe is beyond a level and a scale that I can possibly comprehend. If we're considering forms of intelligence that are completely different than ours, then we have to stretch our minds and be open to radical possibilities.

To bring things back to ground level again, does it really matter if DNA or the universe is intelligent or not? For the purposes of contemplating the nature of intelligence, I'm not sure that it does matter. These other forms of intelligence, if they actually are intelligent, are operating on a level so different than human intelligence that it may not have much of an impact on thinking about other intelligences like ours. Sure, we depend on the behavior of life, the universe, and everything for our own intelligence, but we're not interacting with it directly on the same level.

Characteristics of Intelligence


What is more pertinent for artificial intelligence is what characteristics define human intelligence so that we can recognize the difference between ANI and AGI. In an attempt to answer the first question, "What is the nature of intelligence," the following list will itemize all of the characteristics that potentially make up intelligence. Taken as a whole, these attributes are so interrelated that it is hard to tease them apart, and many will flow together in progressive levels from basic qualities to qualities of high-level thought. To determine if each characteristic is necessary, it is useful to consider if we would still perceive a being as having human-level intelligence without that characteristic. It ends up being a good test of including a characteristic on the list.

First, an intelligence must be observant. It requires some high-bandwidth way to observe the world that it interacts with, and the more observant it is, the more intelligent it will likely appear. This quality depends on pattern recognition so that the intelligence can build up a store of useful patterns to apply to novel situations. Along with that, it would need pattern appreciation to be able to make associations between relevant patterns, assign defining properties to patterns, and classify patterns into similar groups. The highest level of observation, at least that we have achieved, is likely a sense of beauty, and any sufficiently advanced intelligence might develop to this level. Without observation, an intelligence wouldn't react in any dynamic way to its environment, and so would not be considered intelligent at all.

An intelligence would also experiment. Observing its environment is not enough. Without attempting to manipulate its environment, an intelligence would be inconsequential. Once it's experimenting, it would need to understand cause and effect to begin to make use of what it observed through experimentation. Beyond causality, an intelligence needs to have planning abilities and a sense of forethought, otherwise it will never appear to understand the effects of its actions or have any sense of purpose. If an intelligence displays good planning as a result of understanding causality, it will appear to have developed judgment because it can discern what is a better course of action. Good judgment is generally considered a high-level thought process that most humans don't fully develop until their mid-twenties and continue working on and failing at their entire lives.

Combining the powers of observation and experimentation results in problem solving, an ability required of any intelligence. At a base level, an intelligence may work out solutions to problems using only its own facilities, whatever they may be. The next level involves discovering and using tools from its environment to enhance its problem solving abilities. The level after that is actually making tools of its own design to solve increasingly complex problems. Tool making is considered one of the defining characteristics of human intelligence. Even though some primates can use tools they find in their environments, The intricacy and complexity of human tools far surpasses anything any other animal has done. We have long since been using tools to make more intricate tools in an ever increasing hierarchy of complexity.

In order to solve problems, an intelligence needs to display a decision making ability. What problems does it want to solve, and what is the best way to solve those problems? This process also requires goal setting, for if there is no goal, then there are no problems to get in the way. An intelligence must also decide on which goals it wants to pursue. In order to decide on a goal and make progress, it will require motivation. Without some kind of motivation, there would be no action, and an intelligence that never does anything might as well not exist. Motivation is intriguing because it can be thought of both as a fundamental and a high-level characteristic. Motivation is the driving force of any living thing, and so it seems fundamental, but humans with stronger motivation appear to achieve more and seem more intelligent, making it a higher-level attribute.

Underlying any intelligence must be an ability to learn. Without learning, it would never advance or be able to cope with new situations. Once it has learned enough knowledge through observation and experimentation, an intelligence would be able to expand on ideas through logic, reasoning, and inference. These are all higher levels of thought that apply previous knowledge to new problems and develop novel solutions. Because of the vast amount of information and circumstances an intelligence will experience, it will need to tolerate ambiguity. An intelligence that is better able to deal with ambiguity can be more flexible at solving problems, and in general is better able to deal with the real world, which is full of ambiguity. Flexibility in the face of ambiguity starts to look a lot like creativity. To acquire even more knowledge about its environment in order to better solve problems, an intelligence may actively pursue knowledge with a sense of curiosity and wonder. These qualities of creativity, curiosity, and wonder are normally considered high-level characteristics of human intelligence displayed by the greatest geniuses in history.

Along with tool making, language is the other major differentiator between humans and the rest of the animal kingdom. The complexity of our language and the ability to write down our thoughts, both for future reference and for others to read, allows us to use our intelligence to even greater effect. Language and communication act like a feedback loop, amplifying our intelligence. Any high-level intelligence will likely also have the ability to communicate, and along with that ability comes other related ones. If it can communicate with and understand people, it may come to understand how other people think and appreciate their motivations. This is the essence of empathy, a higher-order outcome of communication. From this controversial component of intelligence comes two others, consciousness and free will.

I believe that any form of intelligence at least at the level of a human being will show these three characteristics of empathy, consciousness and free will. Without them, we could hardly consider an entity as being intelligent. Without empathy, we would think it dull and obtuse. Without consciousness, we would think it incoherent and mechanical. Without free will, we would think it devoid of original thought. These are all aspects of high-level intelligence that would make one seriously question the intelligence of an entity that was missing them.

This list of 28 characteristics of intelligence is quite extensive. The fundamental attributes of observation, experimentation, problem solving, learning, and language are fairly obvious to include on the list while some of the higher-order attributes of motivation, creativity, empathy, consciousness, and free will may be more debatable. However, it would be a challenge to classify something as intelligent that was lacking any of these characteristics. It is clear that intelligence can't be defined by any single attribute, nor by the ability to perform any single task well. The definition of intelligence is broad with a diverse set of qualities, and it is likely that the higher-level ones emerge from the basic ones. How this emergent behavior might happen is a topic that we'll explore next time when we take up the second question, "Is artificial general intelligence possible?"

No comments:

Post a Comment