Beyond Zero and One (Andrew Smart)

The last few years have produced an increasing amount of popular discourse about the singularity, which describes events in technological development that would lead to human-level machine intelligence, and would lead to computers achieving superintelligence. Following this occurrence, there would come an explosion of intelligent machines with cognition so advanced that it poses an immediate existential risk to humanity. Machine intelligence would learn how to improve independently of human programmers; it would spawn other machines capable of improving their intelligences, resulting, some argue, in the end of the human race.

I am beginning to suspect that the singularity refers simply to an explosion of books about artificial intelligence.

hedge

cursory

There are a few who argue that consciousness is a result of the fundamental organizational principle of the brain and therefore that we cannot call any machine truly intelligent without it. In other words, human consciousness is not a useless phenomenon but confers some real adaptive advantage. Whether that advantage is enough to survive a mega-superfast-and-supersmart army of zombie artificial intelligences is another question. I argue that consciousness is the fundamental issue for superintelligence, and thus, in contrast to many writers on AI, I focus on machine consciousness.

Contrast this view with that of Duke University neuroscientist Miguel Nicolelis,5 who is actually building brain-machine interfaces, or BMIs. “The arguments favoring the emergence of a relativistic brain strongly suggest,” writes Nicolelis, “that the primate central nervous system, and the human mind in particular, may not be compressed into any type of classical computational algorithm.” Bostrom and Rothblatt are preparing for a future in which it is assumed that technology will achieve these dazzling heights of superintelligence and cyberconsciousness. Both are much more concerned with economic and social issues surrounding superintelligence than with the actual science, philosophy, and engineering involved. There seems to be a curious correlation between how closely an author works with actual brains and her views on AI—the more you work in neuroscience the less you believe in superintelligence.

But why does the human brain react to LSD the way it does? And, if a machine were to achieve human-like consciousness, would it also react to “digital” LSD and enter an altered state of machine consciousness? Recently, Google’s artificial neural networks for image recognition have been creating fantastic psychedelic distortions of images all across the web. The networks’ parameters are tuned to search for certain features – and this tuning is fed back into the network as it analyzes an image. In other words, according to Google, neural networks that are trained to discriminate between different kinds of images have a lot of the information needed to generate images too. Google calls this “inceptionism.”

many of the writers cited above would point to dozens of cool AI projects, such as Jürgen Schmidhuber’s Gödel Machine, which can self-improve by writing its own code after it proves the new code is more optimal. It is a stunning model, but has yet to be really programmed because of the serious technical challenges involved. (Of course these challenges will be overcome eventually.) Or they might point to the success of the artificial intelligence startup based in London called DeepMind, recently acquired by Google for half a billion dollars. Larry Page has said that DeepMind is the most exciting thing he has seen in a long time. DeepMind created an artificial intelligence system that can teach itself how to play video games—oftentimes better than humans. The algorithm is simply given the instruction to increase the game score and it learns the rest by itself. Google wants to apply these “deep learning” algorithms to many types of data, from healthcare to finance.

DeepMind’s algorithm is directly inspired by theories about how the human brain stores, retrieves, and learns information—called reinforcement learning. The founder of the company, Demis Hassabis, is himself a neuroscientist who began his career as a game developer. Using insights from neuroscience, he proposes to solve the problem of intelligence. Even though DeepMind’s AI can only play relatively simple video games at this point, the implications for autonomous computer systems are enormous. These theoretical breakthroughs and practical successes all contribute to the growing chorus of voices announcing the advent of the age of superintelligence.

Within contemporary cognitive science there is a blurry and awkward line between intelligence and consciousness. However, it seems clear to me that much of our intelligence derives from the fact that we are conscious. This is despite the fact that it is well known that most of our brain’s activity remains unconscious. To use the computer AI metaphor, we can to a certain degree inspect and change our code to improve our own behavior. This is in fact borne out by current neuroscience; the organization of our cortex is in constant flux. Rats are conscious too, but we would not claim that they are very intelligent relative to humans. The assumption in artificial intelligence is that a computer must first achieve intelligence in order to become conscious.

Neither brains nor computers carry out computations. Consciousness is a hallucination, albeit a very useful one. And information is an epistemic tool that is a cultural invention rather than a natural entity. Some philosophers, like David Chalmers, argue that the brain doesn’t have to be computational or informational to be simulated. Artificial intelligence does not have to be conscious to achieve either evil or benign demigod-level

powers of intelligence. One thing discussions on superintelligence are missing that is fundamental to human brains is hallucination. Human brains hallucinate frequently in many different circumstances, because of certain diseases and on certain types of drugs such as LSD. What’s more, our consciousness itself shares many properties commonly associated with hallucinations. In other words, our brains are probably more or less hallucinating all the time. Hallucinations are real perceptions, just not very practical ones.

During the course of writing this book, many of these ideas became more crisp, while others that started as crisp became more vague and fuzzy. The vague-to-crisp concept (stolen from computer scientist Leonid Perlovsky) plays a central role in this story. After researching and writing about these topics for more

than a year, and thinking about some for at least fifteen, I am more confused now about digital technology and consciousness than when I started writing. Likewise, after you read this book I hope you are more confused about computers, numbers, mathematics, the Internet, consciousness, hallucination, technology, artificial intelligence, and the ultimate nature of reality than before you started.

What I hope is novel is the synthesis of these areas in a particular way that might generate some further thinking about thinking9—whether that thinking is done by humans, machines, or some kind of freakish hybrid.

As technology and Big Data advance, they enable the near-total bureaucratization of life—despite the fact that the promise of technology was exactly the opposite, to liberate life from bureaucracy. This dovetails with the trend in job automation: the more bureaucratic a job becomes the easier it is to automate.

accruing

There is a strange relationship between digital ideology and the metaphysical belief that our brains also run as computations. It is as if one belief justifies the other. It would indeed be very convenient for Google if it turns out that creating artificial intelligence is just matter of bigger data and faster computers. But what is the ontology of the digital? Are computation and information natural entities upon which our own consciousness and selves can be built? Ontology is just a fancy word that means “the study of what is.” With the ubiquity of computers, information, and data, few would question whether information or computations are real, or even that our own brains are somehow based on computations. However, upon closer examination, these concepts turn out to be as brittle as the latest version of Windows.

pipedreams.

One fundamental problem is that we do not understand human consciousness. Many in science and artificial intelligence assume that human consciousness is based on information or computations. Several writers have tried to tackle this assumption, most notably the British physicist Roger Penrose, whose controversial theory suggests that consciousness is based upon noncomputable quantum states in some of the tiniest structures in the brain, called microtubules. Other, perhaps less esoteric thinkers, like Duke’s Miguel Nicolelis and Harvard’s Leonid Perlovsky, are beginning to challenge the idea that the brain is computable. These scientists lead their fields in man-machine interfacing and computer science. The assumption of a computable brain allows artificial intelligence researchers to believe they will

create artificial minds. However, despite assuming that the brain is a computational system—what philosopher Riccardo Manzotti calls “the computational stance”—neuroscience is still discovering that…

this is where LSD enters the picture. It turns out that human consciousness is likely itself a form of hallucination. As I have said, it is a very useful hallucination, but a hallucination nonetheless. LSD and psychedelics may help reveal our normal everyday experience for the hallucination that it is. This insight has been argued about for centuries in philosophy in various forms. Immanuel Kant may have been first to articulate it in modern form when he called our perception of the world “synthetic.” The fundamental idea is that we do not have direct knowledge of the external world. This idea will be repeated often in this book, and you will have to get used to it. We only have knowledge of our brain’s creation of that world for us. In other words, what we see, hear, and subsequently think are like movies that our brain plays for us after the fact. These movies are based on perceptions that come into our senses from the external world, but they are still fictions of our brain’s creation. In fact, you might put the disclaimer “based on a true story” in front of each experience…

idea that there is no direct relationship between the external world and…

As far as we can tell, certain patterns of activity in this distributed network give rise to conscious experience. But fundamentally, this network’s activity is self-contained and the feeling of a unified flow of consciousness you have is not just from the processing of sensory information. The experience you have right now is a unique creation of your brain that has transformed data from your body into something closer to a hallucination. To break down this seemingly obvious point that we will deal with very often in this book and that I myself struggle to understand: the existence of our experience is real, but the contents of this experience exist only in your brain. Some philosophers call this “irreducible subjectivity,” which means that no totally objective theory of human experience may be possible. The contents of your experience are not representations of the world, but your experience is part of the world.…

I began to wonder if I would ever return to reality, but the voice in the back of my head reminded me again, “LSD is a drug… it will wear off.” The incredible thing about these experiences is that even now, fifteen years later, I can still recall them vividly as if they had really happened. To me, they did actually happen—in my brain. The pseudo-hallucinations and visions I had were real physical events occurring in my brain in a combined response to the world, my perceptions, my internal thoughts, my feelings, and the drug. And in the moment there was a certain reality to these visions and sounds even though I retained at all times the awareness that what I was seeing did not exist in the world outside my brain. I became increasingly confused, which made me anxious. My body seemed to be cycling through every possible emotion at a rapid pace. I thought of my mother and missed her intensely, and then began to fear that I might forget what she looks like. The walls and ceiling appeared soft and started to close in on me.

Does LSD actually enhance the accuracy of our perceptions? Do we see, think, or hear better on LSD? The question has not been scientifically investigated. In an email exchange, the philosopher Riccardo Manzotti wrote me, “I do not believe that because of drugs one is more conscious or conscious of a different dimension. There is just a world and that is the world we experience every day. However, we may miss many things of the reality in which we live. Such drugs may ‘open the doors of our perception’ and if such doors were open reality would be … infinite. Like Blake said.”

Someone turned on the TV. The colors were brilliant, images seemingly jumped off the screen, and there was a fuzzy 3-D quality to the newscaster. I instantly felt I could grasp all her psychological problems, those which led her to being on TV playing the role of newscaster. I could see her need for approval, her struggle with professionalism, and her naïve faith in the objectivity of the news she was presenting. My “perception” could have been a projection of my own struggles. Nonetheless, I felt that I was seeing her for who she really was. I felt an overwhelming pity for her, and wondered why on earth she played this stupid game. We changed the channel, and each show appeared more absurd than the next. I wondered if I was suddenly seeing TV for what it really was—an absurd circus of unbelievable stupidity designed to trap us in the dominant ideology. Was this not obvious to everyone?

Robert and Kristian were using LSD so often during this time that they said they had a hard time trusting the non-LSD reality they experienced even when they were not on the drug. After only one trip, over the following days I too would sometimes wonder if the thoughts and perceptions I had were not still the result of taking LSD.

Eventually, I fell asleep. When I woke up, I thought about the fact that I was now a person who had taken LSD. I was no longer a pharmacological virgin. Taking hallucinogens is a public ritual among some South American indigenous groups, and these societies provide much social and cultural support for integrating the experience into life, which is obviously very different from how Western societies view the taking of hallucinogenic drugs. Hallucinogens have been used for millennia in cultures all across the world. The informal bonds of friendship that form from sharing an LSD trip are crucial. Yet we lack any

cultural references for these experiences, except for the counter-cultural movements of the 1960s, but this cultural framework is limited to our distorted memory and history of that time. All these thoughts were floating around in my head when I looked over at a calendar near the bed that featured large photos of fruit and flowers. The images looked so clear and the colors were so intense that I thought I could pull the fruit from the printed paper. I felt at peace, a contentment I had not felt since I was ten years old and waking up and smelling my favorite pillow. I was drained; exhausted.

Ketanserin is a powerful serotonin antagonist, which means it blocks the activity of your serotonin receptors—effectively ending your LSD trip. Ketanserin is used in clinical studies of LSD as an emergency countermeasure in case someone

really needs to come off the drug. However, users can usually be talked out of whatever anxiety they are feeling in these studies.

What is the connection of LSD with machine consciousness? The kind of seemingly random association coming from humans on acid may be one of the key missing ingredients that AI has yet to master. Our brains make weird associations constantly. Some people, like the late great Robin Williams, seemed to make more than others. But we can all relate to hearing or seeing something that vaguely reminds us of something else, and perhaps saying, “Oh, that reminds me of an episode of Seinfeld.” Often these associations are related to whatever it is you are talking about at the moment.

The massive amount of cultural or semantic information stored in a human brain is readily and appropriately accessible in the right contexts. What are these contexts, and how does the brain identify them? The key to this mechanism is likely the vague and fuzzy logic our brains use to learn. The previous examples illustrate that even though Watson can be lightning fast and accurate, when “he” is wrong he is absurdly wrong. Mental illness would be suspected if a human were to come up with these associations. The enormous leap from a system like Watson—which can mimic or exceed human performance in very constrained tasks like looking up answers to trivia questions—to a system that can mimic the comedic miracle that was Robin Williams is perhaps too much to ask of artificial intelligence. Achieving anything close to this will require a radical departure from how current AI research conceives the human brain and intelligence. While it could be that a system like Watson can operate successfully (most of the time) without consciousness, a system like an artificial Robin Williams could not.

Building an intelligent machine that can match the cognitive capabilities of humans has been the dream of artificial intelligence researchers since the middle of the last century, when Claude Shannon, Alan Turing, and John von Neumann built the theoretical foundation upon which modern computer science stands. Back in the 1950s, as the excitement around the birth of the first computers spread, an intelligent robot that could walk, talk, see, hear, feel, perceive, understand, think, and reason seemed just around the corner. In the 1960s Marvin Minsky famously declared that “within a generation … the problem of creating ‘artificial intelligence’ will substantially be solved.” What is artificial intelligence? If we go with the most ambitious definition, or strong AI as it is called, it means creating an artificial mind. This is opposed to weak AI, which is more or less what Watson can already do,

Rather than just a model or smart system, a strong AI system would be an entirely autonomous computer system in no way controlled or influenced by human operators. It could successfully adapt to its environment or even be part of its environment, making intelligent decisions, and for all intents and purposes interacting with humans naturally. It would have vastly superior memory and computational abilities but would also be able to reason and act accordingly. What all of this boils down to is that a strong AI would have to be conscious.

Yale University AI researcher and philosopher Drew McDermott points out that current AI systems tend to work in narrow domains and outperform humans only in areas such as chess and numerical computation, where there are clear ground rules. He says, “A robot that can walk into a room, spy a chessboard, and ask if anyone wants to play is still far in the future.” Simple things that humans take for granted, like walking down stairs, actually require enormously complex processes that the human brain solves in both parallel and real time. We only began to appreciate how complicated these things are when we tried to build robots that could do them. The simple act of reaching out to grab a coffee cup is in reality much harder to compute than how to fork the Queen and protect the Bishop.

Many researchers in AI and neuroscience choose not to grapple with what philosopher David Chalmers calls the “hard problem” of consciousness. The hard problem refers to subjective or phenomenal experience. Chalmers argues that solving perception or cognition is easy because we know how to study these problems, and can explain them with physical mechanisms that perform these functions. But, he says, even when we have explained all the functions of perception15 with physical mechanisms, we are still left with the question of why any of those functions is associated with a subjective experience. When we see a cup of coffee, “there is something it is like” to see a cup of coffee.16 We feel something, and we experience it. How and why? Nobody knows.

Initial excitement about creating machine consciousness in the field’s early days faded when it was realized that creating an artificial mind would be a far greater challenge than was originally believed. AI research has since focused on the development of specialized algorithms and solving technical problems. Only within the last decade has the field again turned to questions about consciousness.

This indifference to consciousness in practical AI research among those who are actually writing code and coming up with algorithms (and for whom, to be absolutely clear, I have a great deal of respect) is inherited from a long tradition in cognitive science and the philosophy of mind of trying to explain away consciousness. Consciousness and subjectivity are too vague for most people who deal with computer code and equations on a daily basis. There aren’t many people actually trying to program a conscious robot right now.

One of the goals of current AI research is to create “machines with minds in the full and literal sense,” according to Google’s Peter Norvig. Authors like Kurzweil, Rothblatt, and Bostrom think that machines with minds are inevitable, even if the problem of consciousness might still be too ill-defined, uninteresting, or even too daunting to tackle for academic AI researchers. Labs like Google’s attract the brightest scientists to work on Ray Kurzweil’s vision.21 The Italian philosopher Riccardo Manzotti,22 known for his radical “spread mind” theory that consciousness is not only in the brain, argues that full-fledged artificial minds invariably raise the issue of consciousness. And I would add that the issue of consciousness invariably raises the issue of LSD and hallucinations. There is some foundational relationship between the mind, consciousness, and hallucination.

pwned,

To take an illustrative fictional example, in an episode of Star Trek: The Next Generation, Data finally gets a bad joke a full seven years after hearing it: Data: Ha, ha, ha! I get it. Ha, ha, ha! I get it. Geordi La Forge: You get what? Data: When you said to Commander Riker, “The clown can stay … ” Ha, ha, ha. “But the Ferengi in the gorilla suit has to go.” Ha, ha, ha. La Forge: What are you talking about? Data: During the Farpoint mission. We were on the bridge and you told the joke. That was the punch line. Ha, ha, ha! La Forge: Farpoint? Data, that was seven years ago. Data: I know. I just got it. Ha, ha, ha, ha, ha! Very funny! Ha, ha, ha …

This science fiction example captures one of the salient differences between human and machine intelligence. Humans get jokes within milliseconds but might not remember them for seven years. Data remembers the joke for seven years, but it also takes him that long to get it. Data could preserve and replay the joke perfectly in his memory, but could not understand or experience it—at least not right away.

Jokes are bits of knowledge that can only be acquired through living a life as a member of the species Homo sapiens. The Architect in The Matrix possessed truly God-like omniscience. But even the Architect did not seem to have much of a sense of humor.

To reiterate, there are a few fundamental and implicit assumptions in current AI research that have to be made explicit. This is the only way that these assumptions can be acknowledged and improved upon before we can hope to create the world that Kurzweil and others imagine. I would call this a general principle of intellectual integrity.

Identify all of the assumptions upon which you base your theoretical commitments, then state these assumptions as clearly as possible, and finally examine these assumptions as critically as possible. For what it’s worth, I am quite ambivalent about this future. I love technology and I love brains, so what could be better than technology + brains merged into one thing!?

From a practical scientific perspective, the main benefit of the quest for artificial intelligence is what it will invariably teach us about the brain. But a note of caution is warranted here. As renowned cognitive scientist Douglas Hofstadter points out, the supercomputer programs that can play chess tell us nothing about how humans do it. Okay, we understand how to make a computer play chess; now how do humans really do it? Even when AI successfully created programs that could beat the best human chess players, it did not increase our understanding of how humans play chess. It might turn out that creating a truly conscious machine will produce a similar result.

Since Turing’s time, cognitive science has also assumed the mind to be a symbol-manipulating machine, an analogy to the abstract mathematical Turing machine. But as the example of computers that can play chess illustrates, how a computer does math and how a human does math may be two different things entirely.

Again in an email correspondence with Riccardo Manzotti he says: There is a confusion between experience and knowledge, which is the same as saying that there is a confusion between phenomenology and epistemology. Since we do both with our minds, a lot of people confuse the fact that we experience the world—namely consciousness—with the fact that we use concepts to describe the world. The former is not an outcome of the latter. The latter could happen even in the absence of consciousness and vice versa.

One may be aware of one’s concepts and mental operations but that is the result of one’s capability of being conscious of the world (and thus also of certain internal tools to describe the world). In other words, we should not mix the concepts that we use to describe the world (like mathematics) with our experience of the world.

However, when studying human thought and artificial intelligence we do not have something like aerodynamics. We do not know the right physical principles and models that allow humans to think or be conscious. Right now we are convinced it is through computations, but I believe this is mistaken. The problem of flight turned out to require generating lift. To extend the flight analogy further, we can’t try to figure out how to “generate lift” for an artificial mind because we don’t even know if generating lift is the right way to solve the problem. We do not know what physical principles to study in order to create an intelligent machine because right now we do not understand how or why the soggy, lumpy, scrambled-eggy substance in your skull creates your world. Another way to think of it is like this: several attempts at making wing-flapping machines failed, rather comically in hindsight, before people like the Wright brothers realized that flight is not about flapping wings. Artificial intelligence is still at the stage of making the equivalent of wing-flapping machines in terms of our quest for an artificial mind.

However, as neuroscientist Christof Koch has remarked:

“Historically, significant scientific progress has commonly been achieved in the absence of formal definitions. For instance, the phenomenological laws of electrical current flow were formulated by Ohm, Ampere, and Volta well before the discovery of the electron in 1892 by Thompson. For the time being, therefore, I adopt a working definition of consciousness and will see how far I can get with it.” We use computations and information now because these are the best tools we have ever had to study the brain, and it is the only way we know of to program computers. Computations and information turn out to be very good descriptive tools for studying nature, but they are not part of nature except insofar as our brains can produce these concepts. In other words, if there had never been conscious agents (such as humans), computations would not exist. Computations are not like gravity, which exists whether or not we do. What we need to discover are the deeper physical principles upon which the brain builds not only the concepts of computation and information, but also consciousness. Interestingly, the fact that human consciousness is noncomputational may not prevent us from using computations to create artificial consciousness. For example, the concept of lift is entirely a human invention. It does not exist in the ontologically objective physical world, yet we have successfully used the idea of lift to create man-made flight.

Our concept of the number three can be represented as “3” in base ten, but as “0011” in binary. It is the same idea as using the word “three,” three fingers, or “III” to represent the concept of three. We likely use base ten to count because we have ten fingers and it naturally corresponds to how we count using our hands. But other cultures have used different bases, like base five. For most of us, counting in base ten seems as if it is the only way to count because it is how almost everyone in the world counts, and we begin learning to count in base ten as soon as we can talk. But in principle there are infinitely many ways to represent numbers—just as there are infinitely many numbers.

You can begin to see why we refer to 64- and 32-bit machines to describe or classify computers—this corresponds naturally to the base two numbers. A 32-bit machine uses 32 bits to identify each address in its memory. This can also refer to the set of wires that connects the components within the computer like the CPU and the memory, called the bus. A 32-bit bus can transmit 32 bits in parallel. 32 bits might not sound like much until you note that 32 bits means that 232 possible values can be represented, which are 4,294,967,295 possible values. This would mean that we need 32 columns, each set to one or zero, to represent all these values in binary.

A binary digit can be in one of two states. And this is the basis of the “bit”—which is a shortening of “binary digit”: 0 or 1. The information contained in a bit is a measure of how likely it is that the digit is one or zero. With all the talk about infinity so far, it should not surprise you that there is an infinite number of infinite strings made up of zeros and ones. We will return to the importance of infinite strings of ones and zeros and what it might say about human consciousness in chapter 7. The incomparable engineer and scientist Claude Shannon of Bell Labs, who is known as the father of information theory, used the bit as a way to solve the problem of reducing noise in telephone communications and transmitting messages electronically. Shannon’s 1948 paper that first featured the bit—and essentially started the digital revolution and invented the field of information theory—bears the surprisingly pedestrian title “A Mathematical Theory of Communication.” Shannon introduces the concept of the bit33 (which ranks right up there with DNA and relativity as one of the most important ideas of the twentieth century) so nonchalantly that if you are not careful you might pass right over

the sentence that is responsible for the conceptual framework that gave rise to computing. Shannon writes, “If the base 2 is used the resulting units may be called binary digits, or more briefly bits.”

Returning to infinity, it turns out that there is at least a conceptual relationship between the infinity of values between one and zero, and the infinity of values that voltage can have on a computer switch. The binary logic that underlies the computer is an approximation of all this messy stuff between “yes” and “no.” One can think of binary logic as a model and a tool that human culture has invented to control electrical machines (otherwise known as computers) to an incredible degree of precision. However, assigning the logical values “true” to “one” and “false” to “zero” is simply a convention, much like driving on the right side of the road. The number one is not intrinsically true, nor is zero intrinsically false.

From an electrical engineering perspective, the messy stuff in between one and zero does travel down wires, through chips, and in packets. Computer switches are actually not either on or off. Both physically and electrically, computer switches, transistors, or logic circuits just have different levels of voltage to which engineers have assigned the values one or zero. Further up the chain of abstraction to software, one then starts to mean “true” and zero means “false.” But without human beings writing and interpreting software code, the computer itself is merely moving electrically charged ions around. The computer programs that are encoded as ones and zeros and “interpreted” by the circuits and logic gates on the chips do not actually mean anything to the computer.

As I mentioned, when voltage is measured and digitized using signal processing technology, the probability that the real voltage on a computer chip has a particular value is negligible. This is a subtle but very important point. It is the fundamental fact about moving from the analog to the digital. When we measure the voltage coming across a computer circuit, the computer assigns the voltage a certain value, like 1.3597 volts, which can either push the logic gate into Logical 1 or Logical 0 depending on where the thresholds are set or what kind of chip it is. The precision of this value goes out to however many places to the right of the decimal point the memory in the system can support. No matter what the precision of the number is, I want to emphasize that this is not the real value of the voltage in the physical world. In fact the unit of measurement “voltage” itself is a human invention.…

Analog signals are known as “continuous” and digital signals are called “discrete.” This distinction simply means that analog signals vary continuously in time, while discrete signals are chopped up into separate units at each time point—which is otherwise known as the sample rate. The basic idea when moving from continuous signals to discrete ones is that the continuous signal must be sampled at specific points in time and assigned at each time point a numerical value within a finite number system, so that it can be represented in binary numbers and with a finite number of digits, and therefore be used by a computer. This process is otherwise known as quantization. For use in a computer we must quantify things that do not have intrinsic numerical values. This means that instead of differences in air pressure, the digital sound is now made of thousands of binary numbers represented by…

On a deeper level, which we might call the ontological level, the distinction between analog and digital disappears, as both concepts like “analog” and “digital” are ultimately epistemic tools humans have invented to describe natural and man-made signals. On a computer chip the only real thing that exists in the physical world as a physical phenomenon is (probably) electric charge—which might really turn out to be something else as well. The binary numbers one and zero are not real like mass or electric charge. A computer is not really performing computations at all. A computer is nothing more than some special transistors and electric gates. Even the iPhone 6 is not fundamentally different from the clunky early computers that filled up entire gymnasiums and only had a few bytes of memory. If we were to look…

Circuit designers try to make the voltage jump as fast as possible from Logical Zero to Logical One by making the

function represented by the sloping line in the grey area very steep. Nevertheless, in practice the line between one and zero is actually rather fuzzy. This situation recurs on every circuit, which means that at any point in the computer’s processors, a Logical ½ signal can be read out. This is not such a big deal in everyday computers because we can simply turn the machine off and on again, like IT departments ask us to when our computers crash. This temporarily removes all the voltage in the computer’s processors. And there are many error-correcting codes that can fix corrupted bits in memory that might have had some Logical ½s hidden in them. But we can’t just turn the complex computer systems on the Space Shuttle or in the stock market off and on again when they crash. Computers, just like brains, are filled with thermal noise because they dissipate so much energy. Have you ever wondered why your computer crashes? Or why automated stock market trading algorithms suddenly go berserk? Have you ever wondered what a computer glitch really is?

do not share Kurzweil’s faith in the reliability of even these safety-critical systems. He is correct in implying that humans are primarily responsible for most airplane accidents. The software on modern commercial airplanes is probably the most reliable software ever written. However, it is not foolproof and in fact may be decreasing in reliability because of the hardware on which it runs—which I will get to shortly. However, it is simplistic to say that airplane accidents are caused by either humans or the automation. Most airplane crashes37 are caused by a series of unlikely events that the normal redundant safety procedures, both automated and human, are unable to catch.

The Microsoft scientist Leslie Lamport captured the fundamental problem of achieving reliable consensus in distributed control systems with a beautiful and intuitive metaphor in the early 1980s. In a landmark paper called “The Byzantine Generals Problem,” Lamport described what he called a “Byzantine failure” as follows: Imagine a group of generals from the Byzantine Empire whose

divisions surround an enemy camp. After observing the enemy, the generals must communicate among themselves and come to consensus on a plan of action. They must decide whether to attack or retreat. If they all agree to attack, they win. If they all agree not to attack, they live to fight another day. If, however, some attack and some do not, then the generals and their divisions will be wiped out. The Byzantine generals communicate either via oral statements or signed written messages. The problem is that some of the generals are traitors. These generals try to disrupt the loyal generals’ attempts to achieve consensus about a plan of action by sending inconsistent messages. For oral messages between generals, Lamport assumes that: A1. Every message that is sent is delivered correctly. No messages are lost. A2. The receiver of a message knows who sent it. A3. The absence of a message can be detected. Lamport and others have shown that consensus cannot be achieved with only three generals when one of the generals turns out to be a traitor. The minimum number of generals needed to reach consensus in this situation is four. If these assumptions were true in the equivalent computer system, for example, we can say simplistically that this would translate to having four redundant processors in an airplane to achieve what is called Byzantine fault tolerance.

The solution would be to have each of the four generals exchange information with the other generals and take a majority vote to select which data that has been exchanged is correct.

If the Byzantine Generals problem sounds familiar to you, it might be because Bitcoin relies on a solution to exactly this problem to verify the authenticity of transactions and bitcoins. A mysterious programmer (or group of programmers) called Satoshi Nakamoto released the open-source software that underlies Bitcoin digital currency in 2009. Since then, Bitcoin has caused an enormous amount

of excitement and controversy. According to Wired magazine, people have been trying to unmask the true identity of Bitcoin’s author(s) ever since. Silicon Valley titan Marc Andreessen recently wrote an op-ed in the New York Times explaining that the key innovation in Bitcoin is its solution to the Byzantine Generals problem.

Since the initial formulation of the Byzantine Generals problem in 1982, it has been the subject of intense academic study and several Byzantine-tolerant algorithms and architectures have been developed. The most famous real-world solution to the problem is Nakamoto’s Bitcoin software. Whether Bitcoin is truly Byzantine fault-tolerant remains to be seen. While bitcoins offer a welcome alternative to government-backed currencies, the lack of transparency about its origins does not inspire a lot of trust even if, technically speaking, the system is fault-tolerant.

Kevin’s point was that if you know what to look for, these kinds of faults are all over the place, but they are normally just treated as “anomalies.” A series of random anomalies might actually be a systematic pattern of faults with a single origin, like a crack. These kinds of faults occur far more frequently than one in one billion per operational hour. Furthermore, a common fallacy is that random events are evenly distributed in time. For example, if we have a one-in-five-hundred-year storm, we assume that it will be another five hundred years until the next one. However, the conditions that give rise to the rare event often persist even after the event has occurred, so that they tend to cluster in time.

Today, the stock market is almost completely automated. Complex algorithms execute trades in milliseconds, making decisions involving billions of dollars without any human intervention. Each innovation to these algorithms makes them faster, so that one investment bank can complete a trade three milliseconds faster than the competition. The computer architecture and algorithms that run the stock market are far too complex for any individual person to understand, much less control. These algorithms, despite their incomprehensible speed, exhibit all the dangerous properties of human investors. In fact, the algorithms may be “more than human” in an unexpected way. Dr. Didier Sornette, professor and the chair of entrepreneurial risks at the Swiss Federal Institute of Technology Zurich (otherwise known as Einstein’s old school, ETH Zürich), contends they have a higher propensity for herding behavior than human investors.

Furthermore, new algorithms that exploit the high volatility associated with distress and crashes are being aggressively developed. The presence of these “agents” in the market can cause self-perpetuating, reflexive hyper-volatility that can spiral out of control. Dr. Sornette compares this behavior to aftershocks from an earthquake. These internally triggered transactions (defined as those transactions that are not the result of some new external information) increased from thirty percent in the 1990s to at least eighty percent today. According to Sornette, “This proves, using hard numbers, that markets increasingly live a life of their own, disconnected from the real economy, activated by machines and the algorithms that compete to trade in milliseconds.”

is important to keep in mind the difference between accuracy and precision. Accuracy refers to how close a measurement is to its true value; precision refers to the repeatability of the measurement. If you have a bias in your measuring equipment, you might be very precise (i.e., get the same result every time) but inaccurate (i.e., you are not measuring the true value). Think of a faulty blood-glucose

monitor that overestimates blood sugar the same way every time.

But Leibniz’s real reasons for wanting a binary numbering system were much deeper than the philistine goal of counting money. Leibniz, like almost all scientists of his day, was also deeply religious. He dreamed of inventing a universal language of logic that could once and for all eliminate erroneous thinking and enlighten humanity about God’s perfection and creation. In fact, both calculus and binary numbers were components in Leibniz’s ambitious ultimate philosophical project, the characteristica universalis. What Leibniz really wanted to do with binary numbers was what he called propagitio fidei per scientiam (the propagation of faith through science).

Leibniz came up with binary numbers in about 1679, when he wrote a letter to a colleague about a “new science of numbers” on which he was working. Leibniz was soon invited by the Royal Academy of Sciences in Paris to contribute to their new annual publication, the Histoire de l’Académie Royale des Sciences, which was more or less the equivalent of being published in the high-impact journals Science or Nature today. Leibniz was excited by the invitation and felt that, of all that he was working on, the concept of binary numbers was by far the most important. He wrote “Essay d’une Nouvelle Science des Nombres”—the first scientific paper ever on binary numbers—in 1701. The Royal Academy expressed interest in the idea, but they could not comprehend what binary numbers could be used for and now awaited “its applications with impatience.” In other words, they asked Leibniz to revise his essay because binary numbers seemed like a neat idea with no practical use. Leibniz realized that he would not have time to elaborate on his calculating machine before the publication deadline and, amazingly, wrote the Royal Academy asking them to withdraw his essay, which likely would have been published regardless. He wrote: “I thank you, sir, for withdrawing at my request my binary essay from the Histoire de l’Académie Royale. One must accommodate public taste, which wants to see palpable benefits.”

Numerical cognition and its relationship to linguistic evolution is a large area of active research. In contrast to the Mangarevan example, where a complex cultural ecology involving a distributed trade network may have spurred the development of advanced math, the Mundurucu people in Brazil’s Amazon River basin have been extensively studied because their language has only a few words for numbers, even though they too trade with external groups in complicated economic ways. The Mundurucu have approximate number words for plus or minus one, two, three, and four. For larger quantities, they use vague words like “a few,” “a handful,” or “many.” Some researchers suggest that what prevents the Mundurucu from developing exact numbers is the nature of their language, which does not make a grammatical distinction between singular and plural, among other things.

development cognitive psychologist Elizabeth Spelke calls this the “compositionality power” of language. It allows us to take the approximate sense of numerosity we are all born with and make sets out of the individual objects we estimate.43 Noam Chomsky has long claimed that it is in fact language that gives numbers their power, and that exact numbers can only exist because of the recursive structure of language.

For all intents and purposes, Penrose is a Platonist when it comes to mathematical truth. This might seem at odds with his view that Gödel’s theorem shows that the mind in noncomputable. Penrose accepts the implications of Gödel’s theorem, which implies that human mathematical intuition is outside of the axioms; however, he believes that mathematics still forms the perfect laws from which the Universe derives its existence. Plato believed in an abstract world of ideal forms that was distinct from the physical world. For example, any actual sphere in the physical world is only an approximation of the perfect ideal sphere in Plato’s world. In one of Penrose’s other incredible books, The Road to Reality,45 he lays out his metaphysics of math in typical clarity and enthusiasm: Yet the Platonic viewpoint is indeed an immensely valuable one. It tells us to be careful to distinguish

the precise mathematical entities from the approximations that we see around us in the world of physical things. In mathematics, we find a far greater robustness than can be located in any particular mind. Mathematics itself indeed seems to have a robustness that goes far beyond what any individual mathematician is capable of perceiving. Penrose admits that the actual mode of existence for mathematical truths is mysterious; however, what he means by their existence is their objective truth.

Compared to Penrose, Dehaene’s approach to the issues of consciousness and math is much more empirical. Dehaene uses the tools of cognitive psychology and the latest brain imaging techniques to isolate precise brain regions, and to analyze exactly what these regions are doing when we do math, read, or become aware of something. Here’s Dehaene in his own words in a 1997 interview from The

Edge magazine: Many mathematicians are Platonists. They think that the Universe is made of mathematical stuff, and that the job of mathematicians is merely to discover it. I strongly deny this point of view. This does not mean, however, that I am a ‘social constructivist’ … [I believe] that mathematical constructions transcend specific human cultures. In my view, however, this is because all human cultures have the same brain architecture that ‘resonates’ to the same mathematical tunes. The value of Pi, thank God, does not change with culture! Furthermore, I am in no way denying that the external world provides a lot of structure, which gets incorporated into our mathematics. I only object to calling the structure of the Universe ‘mathematical.’ We develop mathematical models of the world, but these are only models, and they are never fully adequate. Planets do not move in ellipses—elliptic trajectories are a good, but far from perfect, approximation. Matter is not made of atoms, electrons, or quarks—all these are good models (indeed, very good ones), but ones that are bound to require revision some day. A lot of conceptual difficulties could be clarified if mathematicians and theoretical physicists paid more attention to the basic distinction between model and reality, a concept familiar to biologists.

Penrose believes that math is precise while the physical world is approximate; Dehaene believes that math is an approximation of the physical world. Penrose believes that the universe is mathematically comprehensible and fundamentally mathematical, and further that mathematics has a mysterious ontological mode of existence independent of both physics and human minds. In contrast, Dehaene believes that the universe is not based on math, because math is essentially a human cultural invention. I agree that mathematics is a human cultural…

Dehaene further likens our perception of math to how we perceive color. Color is not an objective property of the world that exists independently of human brains. Color does not exist in the world. Our brains create color when certain wavelengths of light hit our eyes, and they do this in a way that is very useful for navigation through our environment. It is also very useful to have a number sense because it corresponds to salient features of the world that we need to understand in order to survive and reproduce. This is much like our faculty for language, which is also part of our biological endowment. Advanced mathematics, then, is the result of rapid cultural evolution that began in places like Polynesia in fits and starts but then really took off in Mesopotamia with the invention of writing. For reasons that science has yet to work out, the human brain has an amazing capacity to recycle neuronal real estate and use areas that originally served other functions for novel functions that may arise due to culture. Reading and writing are primary examples of this. Advanced mathematics and digital computers are further examples of the incredibly rapid pace of cultural invention. To tie this into the broad theme of this book, I propose that both language and math are representations created by our brain based partly on external signals it receives through our senses. Language and math are based partly on intrinsic principles of our brain’s…

John Searle’s examination of social ontology. A professor of philosophy at University of California, Berkeley, Searle specializes in consciousness, philosophy of language, and what he calls “social facts.” For example, what kinds of things are corporations and money? They exist independently of any particular human, yet depend on human society for their existence. He is an outspoken critic of the functionalist view of consciousness, which says that nonbiological things can be conscious because what matters to consciousness is causal structure and causal role, not the physical substrate. In other words, consciousness is essentially an algorithm that is being run by the brain. All we have to do is determine what that algorithm is; then we can program a computer to be conscious.…

Searle makes a distinction between observer-independent features and observer-relative features. Observer-independent features are things like gravity, mass, chemical bonds,48 the solar system, and mountains. These things exist even in the absence of human minds. Observer-relative features are things like the United States of America, that football is a game with eleven players per side, and that a green piece of paper with cultish imagery and “twenty dollars” written on it is money. These things require a society of conscious agents believing in them to exist. A rough-and-ready test for whether or not a fact is observer independent is this: Could it have existed if there had been no conscious agents at all? Whatever we might call it in our physics, the gravitational attraction between the Earth and moon exists whether or not we as conscious agents exist. But Searle points out an important consideration. The existence of consciousness itself is an observer-independent fact. It is an objective fact about the world that consciousness exists, and that attitudes exist. As I have mentioned, there can be no doubt that I am conscious and experiencing things even if, as Descartes argues, the only thing that we cannot doubt is the act of doubting. Hence his famous proposition “I think, therefore I am.” Therefore, in a strange way, the existence of observers is observer-independent.

The most important fact, which Searle says that almost all of social science and philosophy going back to Aristotle take for granted, is language. It is only by assuming language that you can develop social contracts or analyze how controlling linguistic categories grants power in a society. What about the existence of language itself? How does it arise and why does it arise? How does it function at an even more basic level than social contracts and power relationships? These are complex questions that are beyond the scope of this book.

Searle states, “Observer relativity implies ontological subjectivity but ontological subjectivity does not preclude epistemic objectivity.” To translate this into English: the existence of things that exist only because we believe them to exist does not mean that these things cannot be objective. Money is a perfect example. It only exists because we believe it does, but nevertheless it has an objective truth. Mathematics, like language, is observer-relative: its mode of existence is ontologically subjective in that it depends on conscious agents for its existence. But mathematics has epistemically objective truths. To break this down, when we talk about observer-relative phenomena like money, this implies that money only exists as a subjective belief in individuals. But Searle argues that this does not prevent there from being epistemically objective facts about…

“We can have epistemically objective knowledge about money and elections even though the kinds of facts about which one has epistemically objective knowledge are themselves all ontologically subjective.” He points out that if epistemic objectivity did require ontological objectivity, the social sciences would be impossible. An example of a social science that would be impossible with the requirement of ontological objectivity is economics. Economies require societies of conscious agents believing in them to exist, yet we can determine epistemically objective results from economics. However, without conscious agents living in social groups there would be no economies.

Whether or not the field of artificial intelligence deals with this foundational issue could be decisive for its project of building superintelligent machines. It turns on whether computations are ontologically objective versus epistemically objective. If computations are only epistemically objective, the project of artificial intelligence can be thought of as an attempt to build a conscious robot using the theories of economics rather than the laws of physics.

In an email discussion I had with Riccardo Manzotti, he writes: We use logic and computation to build artifacts—from chairs to spaceships. We need logic and computation as an extension of our innate mental tools. They help us to achieve the precision and generality that we need to build complex artifacts. We need reading and writing to build cathedrals and armies. Likewise, we need logic and computation to build an airplane or a skyscraper, or to design and build a computer or a robot. But we do not have to believe that the computer or the robot is made of our logic and our computations. There is no logic or computation inside a computer, no more than there is inside another machine. I may say, metaphorically, that inside a machine there are a lot of computations because I used them to devise that machine. Likewise, I may say that I put a lot of time, sweat, and such, inside a wall. But we do not have to take such metaphors for realities.

“The nature and phenomenology of hallucinations can be explained more fruitfully within a framework that accepts that, similarly to hallucinations and dream imagery, normal conscious awareness of the world during wakefulness is a fundamentally subjective and dreamlike experience. Philosophical idealism predicts that normal perception, hallucinations, and dream imagery are principally manifestations of the same physiological process. Hallucinations and dream imagery differ from normal wakeful conscious experience only with regard to the extent to which they are constrained by sensory information from the external physical world. Hallucinations are similar to dreaming in that a lack of sensory constraints on the physiological mechanisms of conscious experience makes these forms of conscious experience maladaptive for interaction with the physical world.”49 —Ralf-Peter Behrendt

“Many people would accept that we do not really have knowledge of the world; we have knowledge only of our representations of the world. Yet we seem condemned by our constitution to treat these representations as if they were the world, for our everyday experience feels as if it were of a given and immediate world.” —Francisco Varela

Modern neuroscience, psychology, and artificial intelligence often implicitly assume that our senses determine the contents of our conscious experience. The assumption is that the brain is primarily reflexive and simply exists to respond to the environment. The task of AI is to create systems that respond to the environment like the human brain responds. It follows that AI tries to mimic how the brain represents the world. But what if the brain does not represent the external world at all?

Few associate Nietzsche with philosophy of mind per se, but his perspectivism is very similar to Kant’s idealism in that both hold that we cannot have knowledge of the objective world. Perspectivism even rejects the possibility of objective metaphysics outright. Like the relativistic brain argument from Miguel Nicolelis, perspectivism argues that there can be no God’s-eye view that perceives the Kantian concept of the “Ding an sich,” or the “thing in itself.” Knowledge is always situated in a specific human mind, from a particular vantage point, and is inherently subjective. There is no purely objective “view from nowhere.” In The Will to Power, Nietzsche writes: That things possess a constitution in themselves

quite apart from interpretation and subjectivity is a quite idle hypothesis: it presupposes that interpretation and subjectivity are not essential, that a thing freed from all relationships would still be a thing. Conversely, the apparent objective character of things: could it not be merely a difference of degree within the subjective?—that perhaps that which changes slowly presents itself to us as “objectively” enduring, being, “in-itself”—that the objective is only a false concept of a genus and an antithesis within the subjective?

In cognitive science Kant and Nietzsche are mostly ignored and the view of the mind as representational prevails.51 The closest we might find to Kant’s idea is sometimes referred to as top-down and bottom-up processing. Top-down processing includes our expectations, our attention, our prior knowledge, and our goals, which are all somehow represented by our brains, typically in the more frontal regions. This includes the model we have of ourselves. These then constrain the bottom-up signals we get from our visual cortex, auditory cortex, and olfactory cortex. The interplay of top-down and bottom-up processes determines whether we become conscious of something. There is the well-known example of the video with the gorilla and the basketball team. Viewers are given the task of counting the number of times a group of players pass the ball. Meanwhile, a person in a gorilla suit runs between the players in plain sight. Viewers are good at counting the passes, but most do not notice the gorilla.

we can think of it like this: our top-down attention to our task somehow focuses our eye movements on the ball only. Even though the person in the gorilla suit passes through our field of view, our brain is not looking for a gorilla. Our visual cortex and visual system do not allow enough information about the gorilla to rise through our perception into conscious awareness—even though the gorilla is perfectly visible. Our top-down directed attention simply filters out the person in the gorilla suit. This is a two-way street: our attention is influenced by the bottom-up signals. If that signal is strong enough, it will break into our top-down attention. Have you ever not noticed an ambulance blazing down your street because you were busy trimming the hedge? But what we end up seeing in our conscious awareness—the object in Kant’s terminology—is created out of some combination of the top-down and bottom-up processes, or “with the qualities which sensible intuition bestows upon it.”

The assumption in popular culture and even in much of psychology and neuroscience is that hallucinations are erroneous perceptions in the absence of identifiable external stimuli. When people hallucinate they hear or see things that aren’t “really” there. But hallucinations are real perceptions.

cameras that take pictures of this preexisting world. What we see is more or less what is out there in the world. What neuroscience has discovered, however, is that this is not really what is happening. Vision and conscious awareness of the world are constructions of the brain. As I mentioned above, this process turns out to be much closer to Kant’s philosophical idealism, which would predict that our perception of the world is no different from a hallucination or a dream. The world that we perceive is a creation of our own mind, and this creation is only partly based on input from our sensory organs and brain areas. Cognitive and vision scientist Donald Hoffman points out that roughly half of the neurons in the cortex are devoted to vision. This amounts to around fifty billion neurons, and trillions of synapses, engaged in taking pictures of the external world. Why…

Hoffman says: “When we open our eyes, our visual system constructs in a fraction of a second all the shapes, depths, colors, motions, textures, and objects that we see. The computational power required for such construction is massive, but the construction is done so quickly that we are fooled into thinking that there is no…

all the eye is doing is catching and counting photons. The millions of photoreceptors in our eyes only react to the number of photons they catch. Our brains then construct the colors, shapes, objects, depths, motion, or textures we see based partly on how many photons the photoreceptors are catching at any given instant. This is why computer vision is such a difficult challenge. If the input to a computer vision system is a video camera, all the computer gets is a huge array of meaningless numbers, just like our brains only get photon counts from our eyes. From this, the computer must construct the meaning of images it is receiving. The fact that our brains do this in hundreds of…

As mentioned earlier, color is a well-known example of something that does not exist outside of our brains. There are no colors in the observer-independent physical world. The photons that hit our eyes travel in wavelengths.…

However, an apple still looks red at the beach on a sunny day, and at your desk in your office on a rainy day. There is about a hundred times more light hitting your eye at the beach than in your office—yet the apple is still red. The redness of the apple is entirely constructed subjectively in your brain. It is, however, very useful to be able to distinguish things based…

Kant argued that all the things we think of as being objects in the world are only objects in our heads. And, as Searle would say in more modern philosophical language, these things have an ontologically subjective mode of existence. Another example could be your back pain. It only exists as experienced by you—therefore it is ontologically subjective. However, as I mentioned earlier, the existence of conscious experience is an objective fact about the world,

even if our own private experiences are purely creations of our brains. To paraphrase the late neurobiologist (and Buddhist) Francisco Varela, we are condemned to treat the representations we experience as though they were the world. Will artificial agents that we create also be condemned to treat their representations of the world as though they were the world? We think we have real, so-called veridical, perceptions of reality. But our perceptions are actually no different from hallucinations, dreams, or illusions. All of our normal waking experience is a construction of our brain. While it is not a true representation of the world, it is a useful representation of the world that is always being revised. Miguel Nicolelis calls this “The Plasticity Principle,” a proposition that states that “The representation of the world created by populations of cortical neurons is not fixed, but remains in flux, throughout our lives, continuously adapting itself to new learned experiences, new models of the self, new simulations of the outside world, and newly assimilated tools.”

Visual awareness is not simply a translation of what our eyes see. Our subjective reality is an internal model that our brain must check against the input coming from our eyes. Hallucinations result when our brain does not check the model thoroughly, or rather when the brain cannot check the model it has created because of disease, exhaustion, lack of oxygen, or drugs.

how do our brains achieve useful hallucinations? This is one of the outstanding questions for neuroscience in the coming decades. Changing the underlying assumption that the role of human intelligence is to translate information about the world into veridical representations of the world—changing this into a working assumption that human consciousness does not arise out of sensory perception but rather exists independently of it—would help to open new avenues of research in machine consciousness. It would require a shift in perspective from the current approach to creating artificial intelligence, an approach that currently rests on the idea that machine intelligence can be achieved by collecting ever-larger amounts of

The key to human intelligence is that our brain chooses the right data out of the environment, and then runs the right algorithms on that data. I say algorithms for lack of a better word, because I do not believe that our brains run algorithms at all.

The structures for hallucination are in all of us. For instance, even those who may have barely used their eyes may experience visual hallucinations on LSD. A 1963 study reported:

It is evident that a normal retina is not needed for the occurrence of LSD-induced visual experiences. These visual experiences do not seem to differ from the hallucinations reported by normal subjects after LSD. Such phenomena occurred only in blind subjects who reported prior visual activity. The drug increased the frequency of visual events such as spots, lights, dots, and flickers.

Human perception and consciousness often hide the true complexity of the world. Aldous Huxley, after his experiences with mescaline and LSD, saw the brain like a “reducing valve” that is actually protecting us from the vast amount of information that might otherwise flood and overload our everyday consciousness. LSD then frees us from this restriction and opens us to the larger “cosmic” experience. This view is remarkably correct from a neuroscientific perspective. LSD removes the sensory constraints on our conscious hallucinations.

The current thinking in AI is that consciousness is a self-model, which is a device (or mental model) that a machine or human uses to answer questions about how it interacts with the world. Dennett, for example, argues that the evolution of the self-model was driven by the evolution of language. So a system that has a self-model can ask itself questions about how it works and somehow represent knowledge. I would like to point out, though, that the idea of a self-model is also a model—the self-model model of consciousness.

Human consciousness then is something that the brain does, but not in response to the environment. Consciousness anticipates the environment. The environment only sets certain limits on the contents of our waking consciousness—or biases us to experience our representations in a useful way.

Donald Hoffman relates the following example about dragonflies to illustrate the idea that biological perception does not report the truth about the world: Many dragonflies, for instance, lay their eggs in water. For millions of years, their visual systems have guided them to bodies of water appropriate for oviposition. This is an impressive feat and might suggest that their visual systems have evolved to report the truth about water. Experiments reveal instead that they have evolved a quick and cheap perceptual trick. Water slightly polarizes the light that reflects from it, and dragonfly visual systems have evolved to detect this polarization. Unfortunately for the dragonfly, Homo sapiens have recently discovered uses for crude oil and asphalt, and these substances polarize light to an even greater degree than does water. Dragonflies find pools of oil even more attractive than bodies of water, and end up dying in large numbers. They also are attracted to asphalt roads. Pools of oil and asphalt roads are now ecological traps for these dragonflies. Apparently their visual system evolved a quick trick to find water: Find something that polarizes light, the

more polarization the better. In the environment in which they evolved, this trick was a useful guide to behavior and allowed them to avoid constructing a complex understanding of the truth.

believe that the objective mind-independent world does exist. First-person conscious experience is intrinsically subjective, relative and uncertain. How can the field of artificial intelligence use the objective, third-person, logical methods of computer science to create a relativistic, vague, and subjective artificial mind? Does an artificial intelligence system have to acquire subjectivity on its way toward superintelligence? Does the fact that we experience the world from a specific point of view enable human intelligence? Do machines need to have subjective and relativistic points of view in order to achieve human-level intelligence?

The fundamental problem arises because all our current physics, including the Theory of Everything, has no prediction for, or understanding of, how subjective states can be objective states—or even how subjective states could be based in an objective state. Here we can see the possibility of a strange loop52 as pointed out by Colin McGinn in his classic book, The Mysterious Flame: Conscious Minds in a Material World. Even when we have the best objective data and explanation about what subjective conscious states are, we are still subjectively experiencing this objective explanation. McGinn argues that we do not possess the right kind of intelligence to grasp the mysterious relationship between the brain and the mind. What hope then do we have of building machines that can overcome this mystery?

Even if our experience of the world is fundamentally no different from a hallucination, we should take our conscious experience of the world very seriously. Hoffman uses the example of seeing a bus speeding toward us as we walk across the street. We had better trust this experience, even though it does not require us to believe in the literal truth of the experience. To recapitulate Kant’s dictum: “If our subjective constitution be removed, the represented object, with the qualities which sensible intuition bestows upon it, is nowhere to be found, and cannot possibly be found. For it is this subjective constitution which determines its form and appearance.”

The project of machine consciousness is to create a computational simulation with such high fidelity to the human brain and human thought (if these are even separate things) that this simulation would also construct subjective experience. In order for AI to be successful in creating a conscious machine, some version of the functionalist and computationalist assumptions must turn out to be correct. However, the degree to which consciousness is necessary for superintelligent computers is an open question. Consciousness-as-hallucination seems to me the key to human-level intelligence.

third way proposed by David Chalmers, which is somewhat between the ideas of the functionalist and biological theorists. Chalmers proposes that consciousness is itself a fundamental property of the universe—just like

gravity. In other words, consciousness in this theory is a fundamental force in the universe, instead of a phenomenon caused by lower-level physics. We say that gravity, for example, is a fundamental force in the universe and that it is not really composed of even more fundamental forces (while that may still turn out to be the case). As we have seen, Chalmers argues that even when you have the complete physical explanation of consciousness, there will always remain phenomenal experience that has not been explained. In this sense even Chalmers admits that he is sympathetic to dualism. In his account, consciousness could still be physical but, importantly, it could also be made of information. He is sympathetic to artificial intelligence and has said that he sees no reason why his account of consciousness prevents machine consciousness. It would be somewhat akin to a gravity machine.

the global neuronal workspace, which was developed by cognitive scientist Bernard Baars.

The idea of the global workspace is motivated by the fact that the frontal areas of the brain and the sensory areas at the back of the brain are long-range connections.

Global neural network “ignition,” as it is sometimes called, means that neurons across your entire brain become active in response to a stimulus. This is in contrast to when a stimulus is flashed for only twenty milliseconds, for example.

Our current understanding is that the brain is a complex network of networks composed of billions of neurons. Some of these have long-range feed-forward and feedback connections that enable neurons in one brain region to communicate with other neurons in other regions of the brain.

Another prominent theory is the integrated information theory (IIT) of consciousness put forth by University of Wisconsin–Madison’s Giulio Tononi. This theory expands on the classical information theory of Claude Shannon and

applies it to consciousness. In a nutshell, the idea is that the level of consciousness you have is related to how many possible states your brain could be in and how tightly integrated those states are. Consciousness is, according to the IIT, the same thing as the brain integrating information and can be quantified in terms of the brain’s organizational structure.

Without going into the mathematical details here, Maguire shows that if we do in fact have completely integrated experiences, this form of information integration cannot be modeled computationally. Both the global workspace theory and the integrated information theory can in principle be implemented in a computer program. According to Ned Block there is nothing intrinsically biological about either of these theories. This is why they are called functionalist, because the underlying biology, or substrate, as it is called, is irrelevant. In contrast to the many varieties of functionalist and computationalist views of consciousness, the majority of which I have not mentioned here, there is a view of consciousness that is called “biological.” This school of thought holds that only machines having the right biology can have consciousness. John Searle and Ned Block are the most prominent theorists espousing the biological theory of consciousness. They answer the question of how unconscious pieces of meat produce consciousness by saying this takes place through specific, though poorly understood, neurobiological features of the brain. There is by now a vast literature on the neurobiological features of the brain that produce consciousness.

In the same spirit as “disorders of consciousness research,” scientists are also beginning to study the neurobiology of psychedelic drugs by using these modern brain-imaging technologies. Which theory about consciousness best supports the conception that consciousness is a form of hallucination? Each approach could potentially account for a “useful-hallucination” theory of consciousness. From the functionalist perspective, hallucinations could be seen as a particular type of self-model. However, the functionalists would have to expand their sense of how this model is generated. It is not generated from representations of the world, but rather from within the brain and it is much less causally related to sensory input than is typically thought.

What does the notion of consciousness-as-useful-hallucination imply for artificial intelligence? It requires a paradigm shift within artificial intelligence, from seeing intelligence and consciousness as being about perceiving objective truth with a “view from nowhere” in pursuit of rational goals, to seeing intelligence and consciousness as unique inventions of each human brain.

Hardly any cognitive scientists are dualists in the Cartesian sense. While the Cartesians believed that the mind represented the world, they also believed it was separate from the brain and made of nonmaterial substance. However, the idea that human thought is computational is a form of dualism.

“I am aware that in this information-centered era, many scientists and philosophers are bamboozled by the idea that information lies at the bottom of reality. It is nothing more than a resurgence of Platonism supported by the social and economic power of computer science and [the] Internet. If one gets back to the original formulation of information, one sees that it is just an epistemic tool.”55 —Riccardo Manzotti

In philosophy, cognitive science, and, more broadly speaking, in computer science, there is a divide between those who believe that computations underlie physical reality—and that human consciousness is made of information, reducible to information, or caused by it—and those who believe that computations are human inventions—useful models or epistemic tools that are convenient descriptive strategies. But crucially for the anti-computationalist, information and computations do not play any causal role in anything because information is not real. This division is often much more implicit than explicit.

This metaphysical assumption—that human thought is based on objective computational rules—determines much of the methodology used in research on human psychology and artificial thought. But what if this assumption is wrong? What if there is no such thing as a computation?

Within the more empirical fields of brain imaging and experimental cognitive neuroscience, the computational assumption is interwoven into most hypotheses about the brain. In fact, in almost every brain imaging study the concepts of information and computation are deployed without definition or reflection.

Oxford professor of philosophy Luciano Floridi describes three flavors of the word information in modern science: “information as reality (e.g., as patterns of physical signals, which are neither true nor false), also known as ecological information; information about reality (semantic information, alethically qualifiable56); and information for reality (instruction, like genetic information).”

The concept of information is useful; so is the concept of lift for flight. This does not mean that lift is real. Importantly, we do not include lift in fundamental physics because the phenomenon of lift is reducible to still more basic laws of physics like mass and gravity. Lift is a successful descriptive strategy that allows humans to build airplanes, but the idea of lift is a simplified model of a much more complex underlying reality. Lift is therefore not an observer-independent phenomenon. Without conscious agents like humans, there is no such force as lift in the universe. It is exactly the same situation for computations and information.

Again, Searle lays out in very clear terms the basic issue about foundational ontology as it relates to consciousness: We know independently that the world consists entirely of physical particles in fields of force (or whatever the ultimately correct physics tell us are the final building blocks of matter) and that these physical particles are organized into systems and that some of the carbon-based systems have evolved over a period of about five-billion years into a very large number of animal and plant species, among which, we humans are one of the species capable of consciousness and intentionality. This then seems to be the general starting point for most philosophers, artificial intelligence researchers, and neuroscientists: the broad metaphysical assumption that the universe is physical. While only philosophers discuss metaphysical assumptions openly and argue about them, the vast majority of those building AI systems and doing neuroscience experiments implicitly subscribe to an ontological view of the universe that can roughly be described as “materialism” or “physicalism.” This means that we assume that the universe is somehow physically comprehensible—i.e., knowable by humans using physics. And more fundamentally, that the universe and everything in it—including human consciousness—are physical phenomena.

Physicalism in its most general form is the belief that the only things in the universe that are observer-independent are physical. This is also what ontological objectivity means: once again, what things in the Universe exist without human observers? Physicalism further asserts that human observers or conscious agents, and consciousness itself, are ultimately physical phenomena. The problem as we have seen is to understand how the elementary forces and particles of physics organize into conscious agents and create subjectivity. A related issue is how this process leads to semantics. In other words, how do the neurons in our head create meaning? The epistemological problem is

how can subjective observers know anything about the observer-independent world? And as we saw in Chapter 6, we cannot assume that we can know anything about the objective world.

But the divide between subjective and objective could also turn out to be an illusion of our own creation. This division goes to the core of our scientific worldview. If some form of physicalism is true, then the division between the inner subjective worlds will have to dissolve, and our inner worlds too will be understood as physical phenomena. LSD seems to bring us closer to this dissolution, I believe. Or the physical world will be understood as conscious, a view called panpsychism.

The real bone of contention between computationalists and others is whether to include computations in the list of observer-independent phenomena. Are computation and information like mass or the elementary particles? Or more dramatically, are computations even more fundamental than mass? Much like Penrose’s Platonic view that the universe is an approximation of mathematical perfection, the computational view is that the universe is made of computations. The “computational move” is necessary for computationalists to build a theory of how elementary particles of physics organize into systems of carbon-based humans with consciousness and intentionality. In fact, it is computations that cause elementary particles to organize into conscious agents.

Manzotti suggests that there is actually no such thing as information. Therefore, neither machines nor biologically based nervous systems can process it more efficiently—because information may not refer to anything. He argues that the computational level may be removed from one’s description of reality and everything will keep behaving in the same way—the computational level has no causal efficacy. He writes: “Computations are therefore epistemic tools that summarize certain causal patterns and match them with certain logical structures human beings have learned to use. Furthermore, computations have no ontological status since they do not refer to something that is intrinsic to physical systems. They tell us much more about the cognitive processes of the observers

than about the mechanism itself.”

invoke the flight analogy, we understand enough about

physics to build flying machines even without a final “theory of everything”–style complete understanding of the physical laws that govern the universe. I am skeptical that the same is possible with building a conscious machine.

one of the more counterintuitive parts to this argument is that computers themselves are not really performing computations. Manzotti argues that computers are simply “transferring electricity from one place to another and dissipating a huge amount of it. Computations and information are not inside the computer.” For example, think of a comic-book character. In the book, he seems to be speaking—but he is not really saying anything. In the same way, we treat computers as computational and informational agents, but computation and information are not intrinsic to computers—or, more important, to humans.

The problem arises because we know that we are intentional systems—our goals are somehow real because we feel like we have goals. Other people might attribute different goals to us. For example, perhaps a Freudian might explain your neurotic behavior in terms of your unconscious desire to have sex with your mother. You are not aware of this goal, and yet your behavior can be explained in terms of pursuing this goal. The problem is that from an external viewpoint, there are infinitely many goals that can explain any person’s behavior equally well. You might explain your goals completely differently from the Freudian. Who is right? This is the fundamental problem with describing things, including one’s own self, as having intentions. Kant also argued this point when he said that we do not have privileged access to our own consciousness. All descriptions of intentions are derived from an observer-relative position. If that is the case, how did intentionality arise in the first place? Likewise, Dennett is extremely suspicious of our own powers of introspection about our own goals. In fact, he doubts that we are conscious at all.

Dennett does not share the view that consciousness is decomposable. In other words, he thinks that our feeling that we have a unitary conscious experience is an illusion. He calls this feeling the “Cartesian Theater”—the place in the brain where all of the information processing in the brain comes together to form our experience. He proposes instead that what seems like a unified conscious experience is a succession of “multiple drafts” that has no central

organization.

Roger Penrose argues that because humans can grasp Gödel’s theory, the mind is not a formal logical system. He argues that the source of our intuitions and consciousness must come from a new kind of quantum physics. Leonid Perlovsky says that within AI the consequences of Gödel’s result are mostly ignored. Fundamentally, AI attempts to build formal logical models of the mind, and this continues today. Even one of the most powerful AI models that we have seen, the Gödel Machine, does not escape the fundamental limitations of provability and computability that Gödel’s theorem revealed.

believe this logical bias might be the source of the “computational stance.” We are biased toward explaining the world with logic and computations because this is the only way that the brain can make sense of reality. Logic itself might be the greatest hallucination. Perlovsky points out that Aristotle, who invented formal logic, did not use logic in his theory of the mind. For Aristotle, forms are dynamic entities that evolve from vague states to crisp states in the process of “mind meeting matter.” Today, as we have discussed, we call this matching top-down and bottom-up signals.

These drugs reliably induce states of altered consciousness, much like dreams or religious ecstasy. Daniel X. Freedman, a pioneer of psychedelic research on LSD, wrote that LSD enables “portentousness” of the mind, so that it can perceive more than it can tell, experience more than it can explicate, and experience boundlessness and “boundaryless” events. Intense psychedelic experiences often induce what is called “ego dissolution”—where the user is no longer able to distinguish between the self and the outside world.

The decade’s radical movements that were influenced by LSD helped with the expansion of women’s and minority rights, with the most effective anti-war movement in American history, and with new genres of literature and music. The cultural, psychedelic, and geographical locus was San Francisco’s Haight-Ashbury neighborhood. The high point59 of this movement was January 14, 1967’s “Human Be-In” featuring the controversial psychedelic professor Timothy Leary and the poet Allen Ginsberg in Golden Gate Park. This was followed by the Summer of Love, when the LSD-inspired psychedelic counter-culture achieved critical mass.

LSD influenced users to embrace the new counter-culture and reject traditional mores, jobs, and relationships.

Albert Hofmann lamented the use of LSD as an inebriating drug rather than a therapeutic

treatment or spiritual tool, which made serious research on the drug impossible. He called the recreational use of LSD “careless and irresponsible” and felt it to be a tragic misunderstanding of the nature and meaning of these kinds of substances. He believed that we should treat LSD with the same reverence indigenous peoples have for these substances.

Hofmann strongly believed that there was a spiritual crisis in Western civilization, and argued that society must shift from “the materialistic, dualistic belief that people and their environment are separate, toward a new consciousness of an all-encompassing reality, which embraces the experiencing ego, a reality in which people feel their oneness with animate nature and all of creation.”

Hofmann thought LSD could do is allow us to become aware of the ontologically objective existence of consciousness as part of the universe.

This is in contrast to how we normally perceive the world; which is that the world exists outside of us and that we are separate from it.

Hofmann says: “Reality became for me a problem after my experience with LSD. Before, I had believed there was only one reality, the reality of everyday life. Just one true reality and the rest was imagination and was not real. But under the influence of LSD, I entered into realities which were as real and even more real than the one of everyday. And I thought about the nature of reality and I got some deeper insights.” This to me indicates that LSD has the ability to create awareness of different kinds of perceptions, in which we encounter the fact that our perception of everyday life is a creation of our minds. Hofmann continues: “I analyzed the mechanisms involved in the production of the normal world view that we call the ‘everyday reality.’ What are the factors that constitute it? What is inside about what is outside? What comes from the outside in and what is just inside?” If we adopt the hallucinogenic model of consciousness, the entire contents of your subjective awareness come from just inside. What comes in from outside is merely a constraint on what the inside reality can look and feel like.

There is no doubt that we each have private, uniquely subjective experiences that exist only in our heads. Furthermore, we can only perceive the world from our first-person perspectives, as separate from the world. This may follow from the fact that consciousness and its contents are creations of our brains. Yet, as we have seen, each of our consciousnesses is a part of the world. Hofmann believed that LSD could allow us to perceive this. Is it possible that LSD allows the brain to perceive aspects of this elusive “view from nowhere” that includes other viewpoints?

Soon after he ingested the drug, his entire world began to transform—the colors, the forms—and he felt his ego begin to dissolve. Hofmann grew afraid he had taken too much, and asked his assistant to help him home by bicycle. While trying to bike home, Hofmann pedaled and pedaled but he had the feeling that he could not move from the spot. Time seemed to stand still. This is how the bicycle trip came to be used as a way to refer to the LSD experience as a “trip.” Temporal integration of brain processes into a unified conscious experience is one of the most mysterious aspects of consciousness, and LSD directly perturbs this process.

The late 1960s made the drug controversial because of the hippie movement and its mass unsupervised drug experimentation. The public controversy, and Leary’s popularity in the media, contributed to a polarized view about the drug. People like Leary, who felt it was his historic mission to make LSD known worldwide, promoted LSD as a way to reject the perceived stagnation and emptiness of American life.

The controversial status of the drug led Sandoz to stop sponsoring clinical trials in 1966. Researchers found it more and more difficult to obtain grants to study LSD. Supplies of the drug for research also became scarce and prohibitively expensive.

The law did not ban the use of LSD in research, but now researchers would be required to receive approval from the Bureau of Narcotics and Dangerous Drugs, which would later become the DEA. Following suit, the UN classified several mind-altering drugs as Schedule I in 1971. This essentially stopped worldwide research using LSD. As a Schedule I drug, LSD is said to have no medical purpose, to have the potential to lead to addiction, or to have the potential to cause harm. This is despite the fact that on a drug-harm score developed in the United Kingdom, which sought to assess whether a drug causes harm to self or others, LSD is ten times less harmful to users than alcohol. What’s more, LSD is fifty times less harmful to others—people not even using the drug—than alcohol. In other words, there is no evidence that people get high on LSD and beat their spouses or cause car accidents; the same cannot be said for alcohol.

Despite revolutionary advancements in brain imaging over the last twenty years, there exist few studies on LSD and hallucinogens using modern neuroscience techniques. Researchers are now able to measure various aspects of the neural mechanisms of LSD action using objective methods. However, exactly how the interaction of LSD with neurotransmitters leads to profound alteration of our subjective conscious experience is a mystery. This is partly due to our poor understanding of what consciousness is and how it exists in the first place. Psychedelic drugs perturb consciousness in a unique and novel way, and they are therefore powerful tools with which to study the brain mechanisms underlying consciousness. The relative lack of modern neuroscientific research on LSD and hallucinogens means that the neural mechanisms mediating all of the incredible effects of psychedelics remain largely unknown. This is changing, however, and a team of brain scientists in the United Kingdom led by Robin Carhart-Harris and David Nutt are intensely investigating how the brain reacts to psychedelics. One thing has been established: all psychedelics are serotonin 5-HT2A receptor agonists. What does this mean?