Top critical review
AI Professor Critiques Hanson’s Predictions Regarding Emulated Minds (Ems)
August 10, 2017
I have had a difficult time deciding how many stars to give this book. From the point of view of the amount of work Hanson has put in (trying to give an encyclopedic coverage of what he considers the most-likely scenario), he deserves 5 stars. However, it took me over a year to finish his book because, frankly, the writing is extremely abstract and, in many cases, downright boring, and so I would have to force myself, time and again, to return to his book and slog ahead. From this point of view, the book deserves only 2 stars.
I try not to assign number of stars based on whether or not I disagree or agree with an author, but with how well the author has argued for his position. Unfortunately, in this case, I feel that the author missed a much more likely scenario. Granted, including this (omitted) scenario would have made his task (of trend prediction) much more difficult, but that goes with the territory. At the end of his book he argues against any scenarios that don’t fit into the one he has selected. I find these arguments weak (and not addressing my criticisms). So, from the point of view of the likeliness of his scenario (with millions of em copies, along with their clans), he only deserves 2 stars.
I am a CS professor emeritus at UCLA, whose research has been Artificial Intelligence — concentrating on symbolic and neural-network models of comprehension of the semantics of all forms of human language and thought, including language learning and also the evolution of simulated “animats”, in which populations of simulated creatures move about —sensing and affecting their simulated environments, including mating and recombining their artificial genes to produce new generations of simulated offspring.
I have been fascinated by what life might be like for emulated minds after the singularity. The term “singularity” refers to when humans will be able to download their connectomes (the connectivity structure of the neurons of their brains) into computing machines. (Note: Hanson’s index lacks both the terms “singularity” and “connectome”.)
To my knowledge, Hanson’s book is the only book which, in any sort of systematic way, attempts to predict what life might be like for emulated minds after the singularity. Hanson points out that, while there are thousands of history books, there are no “history books” written about the future, even though everyone agrees that the future is more important than the past. One might answer that such an attempt is not possible because the future has not yet occurred, but Hanson attempts to write a history-style book about the future, where the time period is immediately after the singularity and the creation of ems (emulated minds) is common. He accomplishes this feat by attempting to extend past trends (ages of human foraging vs farming vs industrialization) into his future scenarios. This attempt (at writing a “history” of a future age) is quite unusual and unique and, for this attempt alone, his book deserves 5 stars.
The trends that he concentrates on are mostly in the area of economics and social psychology. This is itself predictable because he is an economics professor. Unfortunately, the book is very light in other areas, such as the computational complexity of some of the tasks involved; the differences between being an em (emulated mind) that controls a robot vs being an em that lives with a VR (virtual reality) environment or the psychological effects of loss of biological bodies.
Don’t get me wrong. I enjoy reading economics (I read The Economist for many years.) but his concentration on economics has been at the cost of little or analysis in areas such as biology and physics (which is central, as I will explain below).
Hanson claims that his book is full of detail but, ironically, it lacks the kind of concrete details that any reader would be most interested in, which involves such questions as, What might a “normal” day be like for any given kind of em? What might being in a simulated body (that has no need for food or drink or sleep) be like? Due to various (from my point of view, critical) omissions on Hanson’s part, I have finally settled on giving his book 3 stars. OK, onward with my comments/critiques.
First, Hanson divides time into subjective vs objective time. He assume that ems will operate, minimally, at 1000 times faster than humans. Thus, when a human experiences 1 year of objective time, the ems will experience 1000 subjective years. An interesting result for his book is that the age that he considers will most likely only take around 10 years (or less) of objective time, so he is only predicting at most 10 (objective) years past the singularity. Thus, the many years of em societal development that he considers are em subjective years.
Let us examine this fundamental assumption. What is the likelihood that em brains will operate 1000 times faster that human brains? Well, the computational complexity of this assumption depends, not only on how quickly connectomes and bodies (that those connectomes control) can be modeled but also on the complexity of modeling the VR environments in which most ems will live. Right now, it might take a month of computing time to generate a minute of animation (as is done to created special effects in a movie). Even if computing power keeps growing exponentially (for at least a while) it must (as all exponential trends) come to an end; otherwise the universe would become full of grains of rice (e.g. the famous story of doubling a single grain of rice — after, say, just 600 doublings all of empty space in the entire universe would be filled with grains of rice). Even if quantum computing is achieved, a quantum computer would be able to completely accurately simulate only a small segment of reality and to simulate larger areas would require a slowdown, in terms of objective time. That it, to simulate an emulated mind in a really accurate VR environment, with a really accurate physics, would require that months or years of objective time would pass, for just minutes of subjective experience for a given em.
Why is this the case? For a highly accurate simulation of physics, we would want to simulate reality at the quantum level. The shortest moment of time is Planck time. There are roughly 10**44 Planck time units in a second. The smallest unit of space is Planck space. There are approximately 10**35 Planck units of space in a meter. To simulate the length of one meter of space for one second would take 10**79 calculations. Thus, the only way ems are going to operate at 1000 times the speed of humans is if (a) the physical laws and material objects within VR environments are very greatly simplified and (b) the brains and bodies of ems are also greatly simplified. Another alternative is to not embed ems within VRs but, instead, only have ems whose simulated brains control physical robots (because if your body is physical, it will be under the laws of physics “for free” — that is, there is no need to simulate those physical laws). I will not consider this alternative in any great detail because, for the most part, Hanson does not consider ems controlling physical robotic bodies.
The vast majority of Hanson’s book covers ems within virtual reality (VR) environments. He has really little to say about what life would be like if your emulated brain controlled some sort of robotic body. I guess he largely ignored this scenario because it doesn’t allow for the reader imagining any kind of “normal” life in such robotic bodies. In contrast, I think that ems will avoid becoming embedded in VRs and will prefer to control physical, robotic bodies. I explain why later.
The advantage of living within a VR is that, ostensibly, the em can experience a virtual reality with a modified, simplified physics (e.g. one that allows you to levitate or fly about) and be surrounded by em-imagined/created artificial objects (e.g. within some virtual reality an em could, say, spin a simulated coin and that coin might then turn into a sparkling, flowering-type object whose petals flap and it flies away).
It’s an interesting philosophical question (that Hanson never really poses to the reader): Which would you prefer? To live in a continual, Disney-style world (with abilities to shrink/grow, fly, teleport, be invisible to other ems, etc.) or in a world indistinguishable from our current natural world with real, accurate physics? Before you answer, consider just two of my own recent experiences. I live near L.A. Recently I drove up to Topanga Lookout (near Saddle Peak and Stunt roads). I walked along a dirt path and came upon an area containing “onion rocks”. When cut through, these large rocks look like onions. Apparently, a million years ago this area was a sea bed (but now above 1000ft) and by some geological process, these onion rocks (layers of concentric spheres) formed.
A day later I am waiting to see my mother-in-law and I happen to look at a plant outside her house. I examine a single leaf. It has exquisite vein structures, along with deformations (probably due to an early attack by some insect). If you look down at a concrete sidewalk you are walking along, it is full of incredible detail (colors, dust, insects, shadows of light/dark, etc.).
In a VR environment this will not be the case. Reality will be created “on demand” as you look in any given direction. Reality will cease to exist behind you and only come into being again if you turn around. Everything you experience will be em-made. If it hasn’t been simulated for you then it won’t exist. In this sense it is forever a Disney-style theme park.
Currently, all humans experience an “egalitarian" physical environment, in the sense that all humans are under the control of the same set of physical laws. This would not be the case if you are embedded in a virtual reality. Different VRs could have different forms of physics and different levels of physical precision/accuracy. Maybe you’ll be one of the lucky (powerful, wealthy) ems that gets a more desirable physics; or maybe not.
Now let us consider the simulation of an em’s body. Hanson has, for the most part, assumed that our bodies will be similar to our current bodies. He mentions in passing that we won’t have to eat or drink but never really follows up on this.
When humans compare themselves to other animals it is our brains/minds that distinguish us from them, so we tend to think that our minds are the most important things that make us human, but this is incorrect. In the case of ems, we are comparing ourselves with creatures with similar connectomes, so what really distinguishes us are our bodies. Consider how much of your humanity is actually due, not to your mind, but to the biology of your body.
Ems have no actual bodies and so will not give birth to babies. Ems also will not go through the process of biological aging. There goes birth and childhood (bye-bye). Think how much of our humanness is involved in the rearing of children (all those cartoons and animated movies and amusement parks for enjoyment by children; all those picture books and literature for the youth) — all gone. Ems do not replicate by means of sex. Thus, gone is all literature concerning romantic love. A major, driving force of life is an impending, inevitable death, preceded by the infirmities of old age. All of this will disappear. Consider how much of being human is due to biology: As children we are focussed on play, socialization and learning; as young adults we are focussed on sex and finding a mate; as middle-aged adults we are dominated by the needs of parenting and establishing careers. As older adults we hope to look forward to retirement.
Hanson, I think, confuses readers by talking about ems “retiring”. He believes that, instead of children, ems will generate copies of themselves and he spends many many pages discussing how copies of ems will form clans and the various economics of different types of clans. I will explain below why I doubt this particular scenario.
My main point here is: destroy human biology and you fundamentally alter what it means to be human. The moment you no longer need to eat or drink; to sleep; to age; to give birth — you will become something very different from what we consider human. A major point of Hanson’s book is that ems will remain very human in most ways. For the reasons I just gave I think he is quite wrong.
Let us examine what being an emulated mind really means. Currently, your connectome consists of real, biological neurons (with their thousands of dendrites and axons). To modify any neurocircuitry in your brain would require sophisticated brain surgery and current technology is at a very gross level. In the future, for example, if a patient suffers from sexual identity dimorphism (e.g., feels like a man trapped in a woman’s body) the neurocircuitry that makes that woman feel like a man could be altered so that this woman would feel like a woman. That is, surgeons could alter that specific neurocircuitry, rather than what has to be done today, which is to alter the woman’s body to become a man’s body, in order to conform to what the brain feels like, in terms of sexual identity. But the main point here is that a such brain modification would require very sophisticated surgery.
In contrast, the brains of ems are emulated. What this means is that an em’s connectome is a data structure (DS). For example, instead of real neuron N1 connected to real neuron N2, there will be a label (call it N30756) in a computer memory, with a pointer to the address (at another location in computer memory, labelled, say, N5429). This link represents the fact that simulated neuron N30756 is connected to simulated neuron N5429. The DS would be much more complicated, because it would include types of connection (e.g. axo-axonal vs axo-dendritic), length of dendrites/axons, state of the synapse at the connection site, etc.
A connectome simulator software would take an em’s connectome (call it CTM459) and simulate its activity. For example, if the retinal cells in CTM459 were stimulated (by placing, say, a vector of numbers representing firing activity across these retinal cells), then the simulator would simulate the propagation of simulated neural activity, by following and traversing the connectivity specified by the data structure CTM459.
When activity finally reaches motor neurons then the simulator would pass the information to a body simulator, which would move simulated muscles. A VR simulator would then calculate the effect that a part of an em’s body movement would have on the VR environment and update the state of that VR, which would include simulating the physics of that VR. For example, if a given em (em543) “sees” a simulated flower; em543’s brain activity is then simulated, which results in em543’s simulated “hand” reaching out. The VR simulator then simulates how the flower is moved, under the (simulated) forces applied by the simulated hand. If there is gravity in this RV and if the force of the hand broke off the flower from its branch, then the VR simulator would simulated the flower falling.
Again, this description is greatly simplified. For example, I spoke of numbers representing neural firing activity, but there are models at a more precise level. Rather than simulating overall firing rate, so-called “spiking” neural networks simulate each individual spike (a neuron sending out a single action potential). There are many different levels at which neural networks can be modeled. Hanson must be assuming very shallow levels, in order to get his assumed 1000-fold speed up. I am very familiar with a wide variety of neural models (having taught them in grad classes).
If you were an em, would you want your brain simulated in an extremely shallow way? Or in a much, deeper, more accurate (but computationally much more complex) way? Hanson briefly discusses how slower-speed ems might interact with fast-speed ems but he does not address the issue of the quality of an em experience, based on how shallow/deep is the level at which their simulated brain is being emulated.
The biggest problem with Hanson’s scenario is that he assumes that em connectomes remain largely the same over the em time period he considers (of a few hundred to a thousand subjective years). I completely disagree with his premise. Once a connectome is a data structure (DS) it becomes trivial to alter it (and also analyze its structure). To connect a simulated neuron N456 to another neuron N11141 requires only making changes to the DS.
At the very end of his book, Hanson briefly considers the possibility of ems altering their own connectomes and argues against this happening. His argument is that a brain is a complex device, with many modules and so it would be hard to alter any module (due to its effect on other modules). He likens em brains to air traffic control software (that has become so complicated that it is hard to modify it, because of potential, unforeseen, negative side-effects). I disagree. A major new field, recently forming in computer science (CS) is that of Big Data (and Big Data Analysis). CS and other engineering researchers are rapidly developing tools for manipulation and analysis of Big Data and the human connectome is a classic example of a Big Data structure. Once a human connectome is turned into a DS, Big Data techniques will extract structures/patterns from this DS and simulators will be used to map neurocircuits at very detailed micro-analysis levels. These advances will make it relatively easy for ems to modify their own connectomes to augment their own memories and intelligence. (The size of the human brain is restricted to fitting inside a physical human skull, which will not be the case for an emulated brain controlling an emulated body).
A major assumption in Hanson’s book is that billions of ems will be created, due to copying of current em connectomes (since data structures are trivial to copy). Hanson believes that ems will allow for copies of themselves to be created. This will occur, according to Hanson, due to economic forces. For example, as an em I want to maintain my economic status and I can perform more work, not by augmenting my own connectome, but by creating a copy of myself. This copying brings about many problems (that Hanson spends many pages on, throughout his book). Hanson discusses, for example, if em (John) has made hundreds of copies of himself and em (Mary) likewise, then what happens to their copies if John and Mary form a pair-bond? How does John interact with his own copies? etc.
Each copy will deviate over time. The connectomes might start out identical, but the neural simulator will modify the connectome being simulated, in order to simulate the ability of ems to learn. For example, the simulator might have a type of Hebbian rule — e.g. that the strength of connection between two neurons, that fire simultaneously, will increase over time. Hanson briefly discusses the possibility that the original John em might treat the copy ems as slaves, but does not really take this possible scenario into any direction. Hanson occasionally mentions, as though in passing, that em brains could be altered slightly (‘tweaked” is the term he uses throughout) but he fails to face this issue straight on because he has convinced himself that any major forms of “tweaking” is a low probability event (when it is actually an extremely high probability event).
Let us examine the potential PROs and CONs of being an emulated mind.
PROs: You can control different types of simulated bodies. You can avoid pain (by having pain circuits disconnected in the data structure that is your new connectome). You don’t need to eat or drink (except for pleasure). You will live forever; never aging. Hanson believes that ems “age” mentally and spends way too much time on what he calls “retired ems”. (Discussing ems as “retired” seems laughable to me and exhibits an overly anthropomorphic view of ems as still being something like humans.) As an em you can experience many different types of VR environments. You need not travel; instead, you can “teleport” to some VR event (since being “at” a virtual location involves the VR simulator stimulating your simulated sensors with signals from that simulated location, so no physical travel “to” the location is needed). However, if simulated eating gives you pleasure, why bother with the act of eating? Why not just stimulate directly that portion of your connectome that gives you the experience of the pleasure from eating (or from sex, or from … more about this problem later).
CONs: You no longer can escape work by claiming a need to sleep (however, there are some neural models in which something like sleep happens — the model is executed without external inputs, to allow connection strengths to be modified and re-organized before external inputs are turned back on). Given so many copies, the chances are that you are a copy (vs an original) and so might be enslaved by the original. With so many copies, the “worth” of any one copy will be near zero, so your own personal existence will probably not be highly protected. (Hanson talks about creating temporary “spurs” that do some job and then are eliminated. How would you like to be one of those spurs?) Of course, your connectome could be tweaked so that you would not mind doing that work and then being eliminated. You might have others use your connectome to control a body (say, the body of an insect) against your will. Remember, if you are embedded in a VR world of some sort, the ems that control the physics (simulators) of that world will be like gods to you. They can alter the physics you are embedded in (similar to a god “working miracles”); they can slow down or speed up the simulator; they could send signals to that part of your connectome that causes you intense pain. You could be trapped controlling the body of a worm and no one would hear you scream and you would live on forever as a worm (as long as there are computational resources to execute your simulation). Truly, the horrors of hell become real possibilities in a VR universe. (Note: There is no entry for “torture”, “pain” or “suffering” in Hanson’s index.)
For the above reasons I believe that ems will avoid becoming embedded in VR environments and will prefer to control robotic bodies that exist under the physics of our universe. Over time, technology will produce more and more improved robotic bodies but an em that sticks to controlling robotic bodies will not be able to enjoy the god-like powers of being embedded in a VR (such as defying gravity). The irony is that, if I am embedded in a VR, I might have god-like abilities but I will become the helpless subject of those who are like gods over me (i.e. those who control the simulated physics that I am embedded within). One might argue that, if I am embedded in a virtual reality I am first given a command that I can execute to exit that VR. However, once I am actually within that VR, the escape command could easily be made (by those running the simulation) to not work. Once you’re embedded you have to really trust those who have now become your “gods”.
Hanson recognizes that ems (being data structures) are easy to steal and he spends a lot of time discussing methods by which ems would try to avoid “mind theft”. I find this ironic, since at the same time he assumes that every em will be creating hundreds or thousand of copies of themselves and assumes also that these copies will be (at least mostly) autonomous.
As Big Data analysis of connectomes reveals all the functions of the micro-level neurocircuity of our brains, there will be a much greater understanding of how new (simulated) neurocircuitry might be added to em connectomes, to augment em memory and intelligence. Image you have a choice: (a) you can make an autonomous copy of yourself to do some task and then you have to deal with this copy (which will have experiences that you don’t) or (b) you can augment your own connectome in order to be able to do that task (or have that experience) directly.
Once connectomes are data structures and their organizational micro-structures and functions are well understood, connectome augmentation will always be preferred over the making of copies of oneself. (Potential negative side-effects of altering one’s own neurocircuitry could be ameliorated by running simulations first, concerning these potential augmentations and an older version of yourself can always be restored in case of serious negative side-effects.). The increased need for computational power (that Hanson postulated, due to millions of copies) will, instead, be used to run the greatly augmented connectomes of a much smaller number of ems. As ems become smarter and smarter, they will rapidly deviate from anything we might consider human.
Another danger (that Hanson does mention but does not elaborate) is that em connectomes could be altered so that ems would actually enjoy being slaves. They would get pleasure from working constantly to achieve the goals of “master” ems. In such a scenario, if you were an em you would mostly likely be a “happy slave”. (Is that what you would want?)
Near the end of Hanson’s book (in short section titled “aliens”) he briefly considers Fermi’s Paradox (this term is not in his index) which is the problem of “Why don’t we see any evidence of advanced civilizations, given that there are trillions of planets in our galaxy alone?” Scientists have proposed several answers, including: (a) advanced civilizations destroy themselves when they discover how to harness atomic power or (b) they destroy themselves when they master processes of life (e.g. creating deadly viruses) or (c) are destroyed by their robots, once they figure out how to create intelligent robots that then compete with them.
I will add “connectome as data structure leads to the VR-heaven pitfall” to this list.
As ems take over the planet (leaving surviving humans on small bio-reserves) ems will continue to modify and augment their own connectomes. Since connectomes are so easy to modify, they will rapidly discover that they can give themselves great pleasure (by stimulating their own pleasure neurocircuits) without all the negatives that go along with direct pleasuring of a biological body (e.g. injecting oneself with heroin causes many negative effects to one’s addicted body over time, which would not occur with an emulated body and mind). Once you have created your own VR heaven (of continual pleasure and bliss), why strive for travel to the stars? Perhaps the “connectome as data structure” pitfall is the fate that befalls all advanced civilizations (that have survived atomic, viral and robotic destruction). Once ems exist in a simulated “VR heaven” (and perhaps waited on hand-and-foot by “eternally happy” slaves) there will be nothing more to strive for.