One of the endlessly appealing profoundly mistaken ideas found in science fiction is that idea that we human beings could take a pill or have a capsule or micro-chip inserted into our brains and then immediately have all kinds of faculties and capacities we previously did not have. This idea was prominent in The Matrix films, for example. But it certainly didn’t start there.
I’m going to argue here that this idea is based on a particular kind of mind-body dualism that is ultimately rooted in ideas put forward by Socrates in some of the Platonic dialogues (although the extent to which Plato embraced these ideas is very much questionable). And I’m going to conclude that is a profoundly problematic idea that encourages us to think of our lives in ways that leads us to (1) misunderstand and become despondent about our bodies and (2) fail to understand how important failure itself is to our intellectual and moral development.
On the traditional, and enormously influential, Platonic understanding of the soul, the highest part of our soul enables us to connect with abstract ideas. Those abstract ideas are separate from, and thus uncompromised by their necessarily imperfect instantiation in the physical world. Understanding them gives one a pure pleasure not preceded by pain. The lower part of our soul pursues the satisfaction of bodily desires which can only be accomplished by some change the physical world. The pleasure that comes with satisfaction of bodily desires is necessarily impure for two reasons. First it is preceded by the pain of bodily dissatisfaction. And, second, because our control of the physical world is always temporary and imperfect. If one can imagine the highest part of our soul as surviving the death of our body, then it would seem that the pure pleasure of understanding the abstract ideas could survive our death. And, of course, that is becomes transformed by some of the biblical religions that imagine heaven as our pure souls communing for eternity with God.
The notion that a higher, perfectible sphere of life that we approach with our intellect is radically separate from the lower, physical, sphere of life that is fundamentally imperfect is a major source of the alienation from our bodies those is so common in those of us in the West influenced by this tradition. This idea has many ramifications in our culture including, but not limited to the asceticism, especially with regard to sexuality, that is so deeply rooted in our culture.
There are powerful philosophical reasons, rooted in the work of Hegel, Heidegger, and Wittgenstein, to reject the understanding of intelligent human activity that is found in the Platonic tradition. I’ve discussed those ideas in some depth elsewhere. Here I want to take a short cut to them by discussing them in the context of a quick overview of by how contemporary artificial intelligence works.
For years attempts to create artificial intelligence followed a paradigm in which computer scientists tried to model our knowledge and reasoning processes directly. They looked for ways to register abstract concepts in computer memory and then manipulate them by means of some “rules of reasoning.” This approach to artificial intelligence was thought to mimic how human beings think. But given how much our understanding of human reasoning was shaped by the philosophic tradition, it was ultimately based on the Platonic ideas discussed above.
This approach to creating artificial intelligence failed, over and over again.
The deep philosophical explanation for this failure can, perhaps be understood simply in this way, which I borrow from Gilbert Ryle: There are things we know how to do even though we don’t know exactly how we accomplish them. And when we do the things we know how to do, we don’t explicitly follow rules at all, either consciously or unconsciously. Intelligent activity doesn’t work by applying rules.
Think about riding a bike or writing an English sentence. There are of course rules that determine whether a sentence is grammatic although there are no rules for determining whether it is clear, elegant, or beautiful. We give people some grammatical rules to follow when they learn how to write but most of the grammar we learn is not picked up by learning rules. And most of the rules we learned we then forget. Yet we can keep writing grammatically. When we teach someone to ride a bike we give them some guide posts, e.g, try to keep your balance by turning the wheel, etc. But while these rules and precepts may play some role in teaching someone how to write or ride a bike, they don’t cover the whole process of riding a bike. And we don’t explicitly consult the rules when we ride for the rest of our lives.
Moreover, if reasoning is a matter of following rules, then we would face an infinite regress of rule following. How, after all, do we follow rules? Think of some activities that we accomplish by a process that looks like rules or instructions, such as cooking with a recipe or putting Ikea furniture together. We have to learn how to follow those rules. Doing so is itself an intelligent activity that must be accomplished by some other kind of process. If you have no experience or skill in cooking, you can’t pick up a recipe and understand and follow it. And following Ikea instructions also requires require some experience and skill as well. You can’t hand a recipe or a set of Ikea instructions to a 9 year old and expect them to cook a good dinner or put a poang chair together.
So intelligent activity rarely involves manipulating abstract concepts with rules. And even activities that involve doing so ultimately rests on forms of intelligent activity that does not involve following rules at all.
What these reflections indicate is that most of what we do involves through some other kind of process that does not require the explicit manipulation of concepts with rules of reasoning. And, if you think a little more about how we learn to do most of what we do—from speaking to playing sports to learning an instrument to writing to understanding a particular subject—you will realize that they mostly require practice that take a long time and much repetition. Practice involves trying something, failing, and repeating the process over and over again as we get a little bit better each time.
We can imagine the end result of this long process. As kids, we can imagine writing a novel or being a great basketball star or playing our instrument in Carnegie Hall. And we sometimes get frustrated by how slow our progress is in reaching those goals or our failure to reach them. So we imagine somehow we could skip the long years of practices and have some talent and ability embedded in us with a chip or potion. We imagine, in other words, that we could grasp all at once all the rules and concepts and ideas involved in some intelligent activity.
But if you stop and think for a second you will realize that almost all of our abilities required fine-tuned physical abilities, you will realize that this idea is fanciful. Learning to play the trumpet or basketball or to ride a bike are not purely intellectual processes at all. They involve the intelligent use of our bodies which in turn requires training our muscles and nerves to develop connections and strength so that they can respond to our intentions. Implanting a module in our brain cannot make this happen overnight because even if it could modify what our brains due it would not have any impact on the muscles and nerves in our body that enable us to do these things.
But what of purely intellectual activities such as writing a work in philosophy or an op-ed on public policy ? Do they require the same kind of training?
The answer is yes because they, too, are not purely intellectual activities but are also embodied activities. Not only do they involve manipulating sound or putting ink on paper (or activating electrons on a screen) but those physical skills are intrinsically connected to our ability to think about time and history and abstract ideas. And those capacities are only possible only after we learn to talk because abstract thought or thought that ranges forward and back in time requires a medium of expression, language, that is instantiated in some physical form, whether sound or through some kind of writing. And the more our thoughts goes far beyond the immediate circumstances of our lives—the more they range over time and space and involve abstract or theoretical ideas—the more we need to put our thoughts down in a way that enables us save, review, criticize, amend, and expand them. And that ultimately requires some kind of speech and writing which, of course, are physical processes. Writing, in particular, is the physical tool through which we vastly expanse our ability to think in range and depth, and through which we give historical and theoretical accounts of our lives. Speech and writing are the tools that allows us to review our ideas, rearrange them, deepen them, refine them, and combine them with other ideas.
People could produce texts orally without writing by training their memories through speech. But the repetition of speech between people is a physical process too. And there is no question that the sophistication, depth, of the length of the texts we can produce is very much enhanced by the capacity to put words down in order through the physical process of writing. Writing is a physically embodied process. And, of course, so is reading the process by which we learn most of our ideas.
The ability to do talk, speak, read, and write all develop through a long process of training, which takes place slowly over may years as we do increasingly difficult intellectual work. These skills develop in fits and starts. And failure and correction, by ourselves or others, is central to that development.
We are only beginning to understand how this process of developing abilities and learning practices works in our brain / bodies. But with the rise of neural networks, a new paradigm replaced the rule-following approach of artificial intelligence, one that was modeled on our rough understanding of how our brain / bodies are wired. And that in turn helped us grasp more fully the nature of our physically embodied intelligent activity.
Roughly speaking, instead of modeling abstact thought processes by defining concepts and rules for manipulating them, neural networks try to model the way our nervous system works by creating interconnected nodes that are trained to respond to various phenomena to which it is exposed–pictures, sounds, etc. Now that notion of training is critical. A neural network that, for example, can recognize cars or traffic lights in a picture can’t be created just by writing computer code. Rather the code creates a neural network and then the network is trained by being exposed to the phenomena, responding to it, and then finding out whether it responded correctly or not. So to take one example, you can create a neural network that recognizes traffic lights in a digital picture by asking it to identify the quadrants of the picture in which there is a traffic light and then telling it whether it was succeeded or failed. The correct responses are encouraged in the future and incorrect ones discouraged by, again very roughly speaking, giving added weight or lesser weight to the nodes in the neural network that gave the correct or incorrect response. The training, in word taken from behavioral psychology, reinforces certain reaction and not others.
The neural network get better in two ways. First, the network itself is made larger and deeper as it adds more nodes that immediately respond to the stimulus and higher level nodes that act to reinforce the first level ones. And second, a great deal more training takes place. Training is not instantaneous. It takes a lot of time and experience. Of course, faster computers and computers specially designed to make neural network processing faster is also important as well. But the limit to improvements in neural network based AI is often having sufficient experience.
Now if the example of training neural networks sounds familiar, that’s because you are training a neural network every time you use one of the captcha applets to show that you are a human being. You are providing the experience and correction to a neural network’s output (or guesses) and in doing so you are helping it learn how to recognize items in pictures better.
Similarly, all of us who drive cars with some version of auto-pilot are training the neural networks (housed partly in our cars and partly in the larger networks on servers to which they are connected). Cars with auto-pilot have learned how to adjust steering and speed to stay in lanes better as the cars themselves recognized when they cross lines and because attentive drivers correct them. Every time we have to take over for auto-pilot because it does something stupid, we are training the system to do better by learning to make the correction in situations similar to that one. Improvements in Tesla’s self-driving is partly the result of improvements in the neural net software that is better adapted to deal with the challenges of driving. But even more it is the result of the millions of miles that Tesla cars have driven with auto-pilot on, which has trained the neural networks that do the work of auto-pilot. Tesla’s auto-pilot system is far ahead of other car companies in this field not mainly because its software designers are smarter but because it has sold more cars that have primitive auto-pilot systems than other companies. Because it has had more cars using auto-pilot on the road, Tesla’s systems have had far more training than those of other car manufacturers.
Another important implication of neural network based-AI is that it is hardware dependent. The neural networks are trained to respond to the stimuli that is presented to it—and that is based on hardware that interacts with the world. The neural network based software systems that enables auto-pilot to work in Teslas have been trained on the various radar and camera sensors in the cars and have been trained to direct cars with particular capacities to maneuver, accelerate, and brake. As a result, one can’t simply plug the neural networks that drive Teslas into another car and have it perform well–or perhaps at all.
Here are two signs of the hardware dependence of Tesla cars. First, Tesla has been criticized for not making dramatic changes in its models more frequently. Despite some superficial changes in bodywork, the Model S and Model 3 are basically the same cars they were when first released. Some capacities have been added, such as improvements in battery and engines that have added some capacities. But capacities have mostly been added not taken away. And the basic geometries of the cars have not radically changed. One reason for this limited change in car geometries and capacities is to limit the need for retraining Tesla’s auto-pilot system.
And second in the one case where Tesla has taken away capacities by making the, possibly foolhardy, decision to eliminated some of its radar sensors in new Model 3s, it has announced a regression in the capacity of the new cars to do things that the old ones could for a time while training of the new systems took place.
Let me sum up the implications of both the short philosophical reflection on the physical nature of intelligent activity and my sketch of how neural network based AI works.
We have seen first, intelligent activity is not a product of memorizing and applying rules about how to carry out that activity but is a skill that has to be learned over time through practice.
Second, these skills are not just carried out by our minds but by our bodies as well. Or, even more precisely, there is no clear distinction between our minds and our bodies. When we do things, whether we consider them mainly intellectual like writing a book or mainly physical like playing basketball, we are engaging in intelligent bodily activity made possible by the training and education we have undergone through practical activity. Even the activities we think that are most intellectual, activities we think best represent our minds at work—talking and writing—are made possible by the training of bodies that gave us the capacity to do intelligent bodily activity. Even the activities we think are most our most bodily, that best represent our bodies at work—playing sports or having sex—are made possible by the training of our minds and are also intelligent bodily activity.
Third, just as there is no action that is a product of our minds and bodies, there is no pleasure that is purely mental or physical. The goods things in life, activity that we find fulfilling and that brings sensual and emotional joy are the combined result of intelligent bodily activity. (There is more to be said here (and is said in parts of my book) about (1) how sensual and intellectual pleasure are both a product of intelligent bodily activity and (2) about how even the pleasure that comes from reflecting on our situation in life is registered somatically in the reduction in bodily tension and anxiety.)
Thus, fourth, the notion that there is a hierarchy of goods, from bodily to intellectual to spiritual is based on a misunderstanding of human life and activity. All human goods, from those that seem bodily in nature to those that seem intellectual in nature are based on developing and exercising skills and abilities that are both bodily and intellectual in nature. Making love as much as writing philosophy is an intelligent activity and we must develop the skills to bring pleasure to our partners and ourselves. There is no perfect intellectual activity uncompromised by connection to the physical world because there is no intellectual activity that is not instantiated in human activity in the physical world and that is not created through a long process of training.
Fifth that long process of training necessarily involves trial and error which means making mistakes and correcting them over and over. Repeated failure is thus an intrinsic part of human life and, especially an intrinsic part of our growth in skills, abilities, and talents, from those we think are mainly intellectual to those we think are mainly bodily to those we think are mainly emotional (like, for example, having empathy for others. Moreover, this process never ends. Any time the most advanced scholar or inventor or critic or artist in any genre takes a step forward and develops new ideas or styles he or she necessarily engages in this same process of trial and error, making mistakes and correcting them every stop of the way. Tolerance for one’s own failures and confidence in one’s capacity to overcome them are the thus the two traits are thus necessary for creative thought and practice. If you want to do anything in life well you must, a as Ms. Frizzle says, “take chances, make mistakes, and get messy.”
Sixth, those mistakes will be mistakes of the body and the mind—or more precisely the mind-body. Thus any notion that there is a perfection to intellectual activity or intellectual pleasure not found in bodily activity or bodily pleasure is nonsense. The hierarchy of goods and forms of life that people have taken from Plato is a way of thinking that encourages us to denigrate and despair about our “imperfect” bodies. It is long past time that we dropped it.
Seventh, the failure to understand how we become skilled leads to a kind of rush to perfectionism that leaves too many people despondent about their lack of skill and achievement. And that, in turn, leads to lack of confidence at best and giving up too soon at worst. Not only is likely that we make mistakes as we develop our skills and abilities, but mistakes are a necessary part of improvement. There is no improvement without mistakes because developing our skills requires us to try thing we cannot do and learn to refine our actions and reactions by seeing which ones work out and which ones do not.
And thus, eight, the notion that can insert of new skills by installing a mental program or that our minds can survive the demise of our bodies are fantasies based on a false understanding of the nature of human life. (Which is not to say that nothing survives of us after our deaths. The impact of our life’s activity does reverberate into the future but only through the intelligent bodily activity of other people who have been shaped by the impact of our own intelligent bodily activity either directly or through the physical products that have been shaped by our intelligent bodily activity—our writing, speeches, artistic product and so forth.
Rather than despairing about, denigrating, and fleeing our bodies and failure, we need to embrace them. For they are the source of all that is valuable in our lives.