Nothing we call AI today is intelligent. No machine exhibits the flexible modeling capabilities I described in the earlier chapters of this book. Yet there are no technical reasons preventing us from creating intelligent machines. The obstacles have been a lack of understanding of what intelligence is, and not knowing the mechanisms needed to create it. By studying how the brain works, we have made significant progress addressing these issues. It seems inevitable to me that we will overcome any remaining obstacles and enter the age of machine intelligence in this century, probably in the next two to three decades.
Machine intelligence will transform our lives and our society. I believe it will have a larger impact on the twenty-first century than computing did on the twentieth. But, as with most new technologies, it is impossible to know exactly how this transformation will play out. History suggests that we can’t anticipate the technological advances that will push machine intelligence forward. Think of the innovations that drove the acceleration of computing, such as the integrated circuit, solid state memory, cellular wireless communications, public-key cryptography, and the internet. Nobody in 1950 anticipated these and many other advances. Similarly, nobody anticipated how computers would transform media, communications, and commerce. I believe we are similarly ignorant today about what intelligent machines will look like and how we will use them seventy years from now.
Although we can’t know the details of the future, the Thousand Brains Theory can help us define the boundaries. Understanding how the brain creates intelligence tells us what things are possible, what things are not, and to some extent what advances are likely. That is the goal of this chapter.
The most important thing to keep in mind when thinking about machine intelligence is the major division of the brain that I discussed in Chapter 2: the old brain versus the new brain. Recall that the older parts of the human brain control the basic functions of life. They create our emotions, our desires to survive and procreate, and our innate behaviors. When creating intelligent machines, there is no reason we should replicate all the functions of the human brain. The new brain, the neocortex, is the organ of intelligence, so intelligent machines need something equivalent to it. When it comes to the rest of the brain, we can choose which parts we want and which parts we don’t.
Intelligence is the ability of a system to learn a model of the world. However, the resulting model by itself is valueless, emotionless, and has no goals. Goals and values are provided by whatever system is using the model. It’s similar to how the explorers of the sixteenth through the twentieth centuries worked to create an accurate map of Earth. A ruthless military general might use the map to plan the best way to surround and murder an opposing army. A trader could use the exact same map to peacefully exchange goods. The map itself does not dictate these uses, nor does it impart any value to how it is used. It is just a map, neither murderous nor peaceful. Of course, maps vary in detail and in what they cover. Therefore, some maps might be better for war and others better for trade. But the desire to wage war or trade comes from the person using the map.
Similarly, the neocortex learns a model of the world, which by itself has no goals or values. The emotions that direct our behaviors are determined by the old brain. If one human’s old brain is aggressive, then it will use the model in the neocortex to better execute aggressive behavior. If another person’s old brain is benevolent, then it will use the model in the neocortex to better achieve its benevolent goals. As with maps, one person’s model of the world might be better suited for a particular set of aims, but the neocortex does not create the goals.
Intelligent machines need to have a model of the world and the flexibility of behavior that comes from that model, but they don’t need to have human-like instincts for survival and procreation. In fact, designing a machine to have human-like emotions is far more difficult than designing one to be intelligent, because the old brain comprises numerous organs, such as the amygdala and hypothalamus, each of which has its own design and function. To make a machine with human-like emotions, we would have to recreate the varied parts of the old brain. The neocortex, although much larger than the old brain, comprises many copies of a relatively small element, the cortical column. Once we know how to build one cortical column, it should be relatively easy to put lots of them into a machine to make it more intelligent.
The recipe for designing an intelligent machine can be broken into three parts: embodiment, parts of the old brain, and the neocortex. There is a lot of latitude in each of these components, and therefore there will be many types of intelligent machines.
As I described earlier, we learn by moving. In order to learn a model of a building, we must walk through it, going from room to room. To learn a new tool, we must hold it in our hand, turning it this way and that, looking and attending to different parts with our fingers and eyes. At a basic level, to learn a model of the world requires moving one or more sensors relative to the things in the world.
Intelligent machines also need sensors and the ability to move them. This is referred to as embodiment. The embodiment could be a robot that looks like a human, a dog, or a snake. The embodiment could take on nonbiological forms, such as a car or a ten-armed factory robot. The embodiment can even be virtual, such as a bot exploring the internet. The idea of a virtual body may sound strange. The requirement is that an intelligent system can perform actions that change the locations of its sensors, but actions and locations don’t have to be physical. When you browse on the Web you move from one location to another, and what you sense changes with each new website. We do this by physically moving a mouse or touching a screen, but an intelligent machine could do the same just using software, with no physical movements. Most of today’s deep learning networks don’t have an embodiment. They don’t have moveable sensors and they don’t have reference frames to know where the sensors are. Without embodiment, what can be learned is limited.
The types of sensors that can be used in an intelligent machine are almost limitless. A human’s primary senses are vision, touch, and hearing. Bats have sonar. Some fish have senses that emit electric fields. Within vision, there are eyes with lenses (like ours), compound eyes, and eyes that see infrared or ultraviolet light. It is easy to imagine new types of sensors designed for specific problems. For example, a robot capable of rescuing people in collapsed buildings might have radar sensors so it can see in the dark.
Human vision, touch, and hearing are achieved through arrays of sensors. For example, an eye is not a single sensor. It contains thousands of sensors arrayed on the back of the eye. Similarly, the body contains thousands of sensors arrayed on the skin. Intelligent machines will also have sensory arrays. Imagine if you only had one finger for touching, or you could only look at the world through a narrow straw. You would still be able to learn about the world, but it would take much longer, and the actions you could perform would be limited. I can imagine simple intelligent machines with limited capabilities having just a few sensors, but a machine that approaches or exceeds human intelligence will have large sensory arrays—just as we do.
Smell and taste are qualitatively different than vision and touch. Unless we put our nose directly on a surface, as dogs do, it is difficult to say where a smell is located with any precision. Similarly, taste is limited to sensing things in the mouth. Smell and taste help us decide what foods are safe to eat, and smell might help us identify a general area, but we don’t rely on them much for learning the detailed structure of the world. This is because we can’t easily associate smells and tastes with specific locations. This is not an inherent limitation of these senses. For example, an intelligent machine could have arrays of taste-like chemical sensors on the surface of its body, allowing the machine to “feel” chemicals in the same way you and I feel textures.
Sound is in between. By using two ears and taking advantage of how sound bounces off our outer ear, our brains can locate sounds much better than they can locate smell or taste, but not as well as with vision and touch.
The important point is that for an intelligent machine to learn a model of the world, it needs sensory inputs that can be moved. Each individual sensor needs to be associated with a reference frame that tracks the location of the sensor relative to things in the world. There are many different types of sensors that an intelligent machine could possess. The best sensors for any particular application depend on what kind of world the machine exists in and what we hope the machine will learn.
In the future, we might build machines with unusual embodiments. For example, imagine an intelligent machine that exists inside individual cells and understands proteins. Proteins are long molecules that naturally fold into complex shapes. The shape of a protein molecule determines what it does. There would be tremendous benefits to medicine if we could better understand the shape of proteins and manipulate them as needed, but our brains are not very good at understanding proteins. We cannot sense them or interact with them directly. Even the speed at which they act is much faster than our brains can process. However, it might be possible to create an intelligent machine that understands and manipulates proteins in the same way that you and I understand and manipulate coffee cups and smartphones. The brain of the intelligent protein machine (IPM) might reside in a typical computer, but its movements and sensors would work on a very small scale, inside a cell. Its sensors might detect amino acids, different types of protein folds, or particular chemical bonds. Its actions might involve moving its sensors relative to a protein, as you might move your finger over a coffee cup. And it might have actions that prod a protein to get it to change its shape, similar to how you touch a smartphone screen to change its display. The IPM could learn a model of the world inside of cells and then use this model to achieve desired goals, such as eliminating bad proteins and fixing damaged ones.
Another example of an unusual embodiment is a distributed brain. The human neocortex has about 150,000 cortical columns, each modeling the part of the world that it can sense. There is no reason that the “columns” of an intelligent machine must be physically located next to each other, as in a biological brain. Imagine an intelligent machine with millions of columns and thousands of sensor arrays. The sensors and the associated models could be physically distributed across the Earth, within the oceans, or throughout our solar system. For example, an intelligent machine with sensors distributed over the surface of Earth might understand the behavior of global weather in the same way you and I understand the behavior of a smartphone.
I don’t know whether it will ever be feasible to build an intelligent protein machine or how valuable distributed intelligent machines will be. I mention these examples to stimulate your imagination and because they are in the realm of possibility. The key idea is that intelligent machines will likely take many different forms. When we think about the future of machine intelligence and the implications it will have, we need to think broadly and not limit our ideas to the human and other animal forms that intelligence resides in today.
To create an intelligent machine, a few things are needed that exist in the older parts of our brain. Earlier, I said we don’t need to replicate the old-brain areas. That is true in general, but there are some things the old brain does that are requirements for intelligent machines.
One requirement is basic movements. Recall that the neocortex does not directly control any muscles. When the neocortex wants to do something, it sends signals to older parts of the brain that more directly control movements. For example, balancing on two feet, walking, and running are behaviors implemented by older parts of the brain. You don’t rely on your neocortex to balance, walk, and run. This makes sense, since animals needed to walk and run long before we evolved a neocortex. And why would we want the neocortex thinking about every step when it could be thinking about which path to take to escape a predator?
But does it have to be this way? Couldn’t we build an intelligent machine where the neocortex equivalent directly controlled movements? I don’t think so. The neocortex implements a near universal algorithm, but this flexibility comes with a price. The neocortex must be attached to something that already has sensors and already has behaviors. It does not create completely new behaviors; it learns how to string together existing ones in new and useful ways. The behavioral primitives can be as simple as the flexing of a finger or as complex as walking, but the neocortex requires that they exist. The behavioral primitives in the older parts of the brain are not all fixed—they can also be modified with learning. Therefore, the neocortex must continually adjust as well.
Behaviors that are intimately tied to the embodiment of a machine should be built in. For example, say we have a flying drone whose purpose is to deliver emergency supplies to people suffering from a natural disaster. We might make the drone intelligent, letting it assess on its own what areas are most in need and letting it coordinate with other drones when delivering its supplies. The “neocortex” of the drone cannot control all aspects of flight, nor would we want it to. The drone should have built-in behaviors for stable flight, landing, avoiding obstacles, etc. The intelligent part of the drone would not have to think about flight control in the same way that your neocortex does not have to think about balancing on two feet.
Safety is another type of behavior we should build into an intelligent machine. Isaac Asimov, the science-fiction writer, famously proposed three laws of robotics. These laws are like a safety protocol:
1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey orders given it by human beings except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
Asimov’s three laws of robotics were proposed in the context of science-fiction novels and don’t necessarily apply to all forms of machine intelligence. But in any product design, there are safeguards that are worth considering. They can be quite simple. For example, my car has a built-in safety system to avoid accidents. Normally, the car follows my orders, which I communicate via the accelerator and brake pedals. However, if the car detects an obstacle that I am going to hit, it will ignore my orders and apply the brakes. You could say the car is following Asimov’s first and second laws, or you could say that the engineers who designed my car built in some safety features. Intelligent machines will also have built-in behaviors for safety. I include this idea here for completeness, even though these requirements are not unique to intelligent machines.
Finally, an intelligent machine must have goals and motivations. Human goals and motivations are complex. Some are driven by our genes, such as the desire for sex, food, and shelter. Emotions—such as fear, anger, and jealousy—can also have a large influence on how we behave. Some of our goals and motivations are more societal. For example, what is viewed as a successful life varies from culture to culture.
Intelligent machines also need goals and motivations. We wouldn’t want to send a team of robotic construction workers to Mars, only to find them lying around in the sunlight all day charging their batteries. So how do we endow an intelligent machine with goals, and is there a risk in this?
First, it is important to remember that the neocortex, on its own, does not create goals, motivations, or emotions. Recall the analogy I made between the neocortex and a map of the world. A map can tell us how to get from where we are to where we want to be, what will happen if we act one way or another, and what things are located at various places. But a map has no motivations on its own. A map will not desire to go someplace, nor will it spontaneously develop goals or ambitions. The same is true for the neocortex.
The neocortex is actively involved in how motivations and goals influence behavior, but the neocortex does not lead. To get a sense of how this works, imagine older brain areas conversing with the neocortex. Old brain says, “I am hungry. I want food.” The neocortex responds, “I looked for food and found two places nearby that had food in the past. To reach one food location, we follow a river. To reach the other, we cross an open field where some tigers live.” The neocortex says these things calmly and without value. However, the older brain area associates tigers with danger. Upon hearing the word “tiger,” the old brain jumps into action. It releases chemicals into the blood that raise your heart rate and causes other physiological effects that we associate with fear. The old brain may also release chemicals, called neuromodulators, directly into broad areas of the neocortex—in essence, telling the neocortex, “Whatever you were just thinking, DON’T do that.”
To endow a machine with goals and motivations requires that we design specific mechanisms for goals and motivations and then embed them into the embodiment of the machine. The goals could be fixed, like our genetically determined desire to eat, or they could be learned, like our societally determined goals for how to live a good life. Of course, any goals must be built on top of safety measures such as Asimov’s first two laws. In summary, an intelligent machine will need some form of goals and motivations; however, goals and motivations are not a consequence of intelligence, and will not appear on their own.
The third ingredient for an intelligent machine is a general-purpose learning system that performs the same functions as the neocortex. Once again, there can be a wide range of design options. I will discuss two: speed and capacity.
Neurons take at least five milliseconds to do anything useful. Transistors made of silicon can operate almost a million times faster. Thus, a neocortex made of silicon could potentially think and learn a million times faster than a human. It is hard to imagine what such a dramatic improvement in speed of thought would lead to. But before we let our imaginations run wild, I need to point out that just because part of an intelligent machine can operate a million times faster than a biological brain doesn’t mean the entire intelligent machine can run a million times faster, or that knowledge can be acquired at that speed.
For example, recall our robotic construction workers, the ones we sent to Mars to build a habitat for humans. They might be able to think and analyze problems quickly, but the actual process of construction can only be sped up a little bit. Heavy materials can be moved only so fast before the forces involved cause them to bend and break. If a robot needs to drill a hole in a piece of metal, it will happen no faster than if a human were drilling the hole. Of course, the robot construction workers might work continuously, not get tired, and make fewer mistakes. So, the entire process of preparing Mars for humans might occur several times faster when using intelligent machines compared to humans, but not a million times faster.
Consider another example: What if we had intelligent machines that did the work of neuroscientists, only the machines could think a million times faster? Neuroscientists have taken decades to reach our current level of understanding of the brain. Would that progress have occurred a million times faster, in less than an hour, with AI neuroscientists? No. Some scientists, like me and my team, are theorists. We spend our days reading papers, debating possible theories, and writing software. Some of this work could, in principle, occur much faster if performed by an intelligent machine. But our software simulations would still take days to run. Plus, our theories are not developed in a vacuum; we are dependent upon experimental discoveries. The brain theory in this book was constrained and informed by the results from hundreds of experimental labs. Even if we were able to think a million times faster, we would still have to wait for the experimentalists to publish their results, and they cannot significantly speed up their experiments. For example, rats have to be trained and data collected. Rats can’t be sped up by any amount. Once again, using intelligent machines instead of humans to study neuroscience would speed up the rate of scientific discovery, but not by a million times.
Neuroscience is not unique in this regard. Almost all fields of scientific inquiry rely on experimental data. For example, today there are numerous theories about the nature of space and time. To know if any of these theories are correct requires new experimental data. If we had intelligent machine cosmologists that thought a million times faster than human cosmologists, they might be able to quickly generate new theories, but we would still have to build space telescopes and underground particle detectors to collect the data needed to know if any of the theories are correct. We can’t dramatically speed up the creation of telescopes and particle detectors, nor can we reduce the time it takes for them to collect data.
There are some fields of inquiry that could be sped up significantly. Mathematicians mostly think, write, and share ideas. In principle, intelligent machines could work on some math problems a million times faster than human mathematicians. Another example is our virtual intelligent machine that crawls around the internet. The speed at which the intelligent Web crawler can learn is restricted by how quickly it can “move” by following links and opening files. This could be very fast.
Today’s computers are probably a good analogy for what we can expect to happen. Computers do tasks that humans used to do by hand, and they do them about a million times faster. Computers have changed our society and have led to a dramatic increase in our ability to make scientific and medical discoveries. But computers have not led to a millionfold increase in the rate at which we do these things. Intelligent machines will have a similar impact on our society and how fast we make discoveries.
Vernon Mountcastle realized that our neocortex got large, and we got smarter, by making copies of the same circuit, the cortical column. Machine intelligence can follow the same plan. Once we fully understand what a column does and how to make one out of silicon, then it should be relatively easy to create intelligent machines of varying capacity by using more or fewer column elements.
There aren’t any obvious limits to how big we can make artificial brains. A human neocortex contains about 150,000 columns. What would happen if we made an artificial neocortex with 150 million? What would be the benefit of a brain one thousand times bigger than a human brain? We don’t know yet, but there are a few observations that are worth sharing.
The size of neocortical regions varies considerably between people. For example, region V1, the primary visual region, can be twice as big in some people as in others. V1 is the same thickness for everyone, but the area, and hence the number of columns, can vary. A person with a relatively small V1 and a person with a relatively large V1 both have normal vision and neither person is aware of the difference. There is a difference, however; a person with a large V1 has higher acuity, meaning they can see smaller things. This might be useful if you were a watchmaker, for example. If we generalize from this, then increasing the size of some regions of the neocortex can make a modest difference, but it doesn’t give you some superpower.
Instead of making regions bigger, we could create more regions and connect them in more complex ways. To some extent, this is the difference between monkeys and humans. A monkey’s visual ability is similar to a human’s, but humans have a bigger neocortex overall, with more regions. Most people would agree that a human is more intelligent than a monkey, that our model of the world is deeper and more comprehensive. This suggests that intelligent machines could surpass humans in the depth of their understanding. This doesn’t necessarily mean that humans couldn’t understand what an intelligent machine learns. For example, even though I could not have discovered what Albert Einstein did, I can understand his discoveries.
There is one more way that we can think about capacity. Much of the volume of our brain is wiring, the axons and dendrites that connect neurons to each other. These are costly in terms of energy and space. To conserve energy, the brain is forced to limit the wiring and therefore limit what can be readily learned. When we are born, our neocortex has an overabundance of wiring. This is pared down significantly during the first few years of life. Presumably the brain is learning which connections are useful and which are not based on the early life experiences of the child. The removal of unused wiring has a downside, though; it makes it difficult to learn new types of knowledge later in life. For example, if a child is not exposed to multiple languages early in life, then the ability to become fluent in multiple languages is diminished. Similarly, a child whose eyes do not function early in life will permanently lose the ability to see, even if the eyes are later repaired. This is probably because some of the connections that are needed for being multilingual and for seeing were lost because they weren’t being used.
Intelligent machines do not have the same constraints related to wiring. For example, in the software models of the neocortex that my team creates, we can instantly establish connections between any two sets of neurons. Unlike the physical wiring in the brain, software allows all possible connections to be formed. This flexibility in connectivity could be one of the greatest advantages of machine intelligence over biological intelligence. It could allow intelligent machines to keep all their options open, as it removes one of the greatest barriers human adults face when trying to learn new things.
Another way that machine intelligence will differ from human intelligence is the ability to clone intelligent machines. Every human has to learn a model of the world from scratch. We start life knowing almost nothing and spend several decades learning. We go to school to learn, we read books to learn, and of course we learn via our personal experiences. Intelligent machines will also have to learn a model of the world. However, unlike humans, at any time we can make a copy of an intelligent machine, cloning it. Imagine we have a standardized hardware design for our intelligent Mars construction robots. We might have the equivalent of a school to teach a robot about construction methods, materials, and how to use tools. This training might take years to complete. But once we are satisfied with the robot’s abilities, we can make copies by transferring its learned connections into a dozen other identical robots. The next day we could reprogram the robots again, with an improved design or perhaps with entirely new skills.
When we create a new technology, we imagine that it will be used to replace or improve something we are familiar with. Over time, new uses arise that no one anticipated, and it is these unanticipated uses that typically become most important and transform society. For example, the internet was invented to share files between scientific and military computers, something that had previously been done manually but could now be done faster and more efficiently. The internet is still used to share files, but, more importantly, it radically transformed entertainment, commerce, manufacturing, and personal communication. It has even changed how we write and read. Few people imagined these societal shifts when the internet protocols were first created.
Machine intelligence will undergo a similar transition. Today, most AI scientists focus on getting machines to do things that humans can do—from recognizing spoken words, to labeling pictures, to driving cars. The notion that the goal of AI is to mimic humans is epitomized by the famous “Turing test.” Originally proposed by Alan Turing as the “imitation game,” the Turing test states that if a person can’t tell if they are conversing with a computer or a human, then the computer should be considered intelligent. Unfortunately, this focus on human-like ability as a metric for intelligence has done more harm than good. Our excitement about tasks such as getting a computer to play Go has distracted us from imagining the ultimate impact of intelligent machines.
Of course, we will use intelligent machines to do things we humans do today. This will include dangerous and unhealthy jobs that are perhaps too risky for humans, such as deep-sea repair or cleaning up toxic spills. We will also use intelligent machines for tasks where there aren’t enough humans, perhaps as caregivers for the elderly. Some people will want to use intelligent machines to replace good-paying jobs or to fight wars. We will have to work to find the right solutions to the dilemmas some of these applications will present.
But what can we say about the unanticipated applications of machine intelligence? Although no one can know the details of the future, we can try to identify large ideas and trends that might propel the adoption of AI in unanticipated directions. One that I find exciting is the acquisition of scientific knowledge. Humans want to learn. We are drawn to explore, to seek out knowledge, and to understand the unknown. We want to know the answers to the mysteries of the universe: How did it all begin? How will it end? Is life common in the universe? Are there other intelligent beings? The neocortex is the organ that allows humans to seek this knowledge. When intelligent machines can think faster and deeper than us, sense things we can’t sense, and travel to places we can’t travel to, who knows what we will learn. I find this possibility exciting.
Not everyone is as optimistic as I am about the benefits of machine intelligence. Some people see it as the greatest threat to humanity. I discuss the risks of machine intelligence in the next chapter.