Other Kinds of Minds Friday, Apr 13 2007
AI 3:06 am

A central aspect of the transhumanist project should be to imagine what we are trying to become or create. What kind of minds can there be? It may be that there is an abundance of possible minds we cannot even imagine from our current frame of reference, but on the flipside, some of the minds that we can imagine may in fact turn out to be forbidden by the laws of nature. We have only one true data point on the graph of mindspace – humankind. The psychological differences we observe between humans may be a useful guide in imagining other kinds of minds. Philosophy of mind, cognitive science, and computational complexity theory may help too. On a past post that went over some basics of the nascent field of Friendly AI – building positive AIs – a commenter, Raphael, said this:
I agree that most people anthropomorphize and that the Asimov’s Laws are risible (if they worked, then Asimov wouldn’t have had any drama to write about). It also seems highly plausible that the space of possible intelligences is much larger than the space of human intelligences (e.g. from low-IQ to Newton).
However, I’m not sure how much we can say beyond that. For instance, given that we don’t know what intelligence is (in any detailed way), it is hard to say exactly how diverse the space of intelligent minds is. I will substantiate this point by analogy. Without knowledge of computational complexity theory, it is easy to assume that the space of conceivable algorithms is the same as the space of physically realizable algorithms. In other words, that all algorithms we can devise are such that we could (with huge amounts of computing power) actually implement those algorithms. However, we now know that this is completely wrong. There are simple problems which can only be solved via algorithms that cannot be implemented in our universe.
The question is interesting. I have strong intuitions that a very large diversity of minds is possible, and can name specific qualities of different possible types of minds, but I lack the empirically verified theory that would be necessary to say that such minds are possible with certainty. As an effort to explore this space, I’m going to list some different types of minds I can imagine and invite people to criticize the plausibility of their existence. Are some mind variants more plausible than others? Is there a criteria for determining this other than their foreignness from a human-centered psychological perspective? That’s what we should try to find out.
Some relevant dimensions along which we can imagine minds that vary widely are:
1) Clock speed. Functionalism tells us that minds are defined as the interactions of the hardware on which they are instantiated. When the hardware is slow, the mind is slow. When the hardware is fast, the mind is fast. Humans have about 10^11 neurons firing at about 200 times per second (200 Hz). It should then follow that arbitrary changes to hardware performance in any given brain results in a corresponding change in speed of thinking, perceiving, experiencing, planning, creativity, and communication. In this model, minds with 10^11 neurons that run at 0.2 Hz could be expected to think a thousand times slower than human beings, whereas minds with the same number of neurons running at 200 KHz could be expected to think a thousand times faster. Invalidating this proposal would require that minds with slower hardware, do not, in fact, operate slower than human brains, or that minds with faster hardware do not, in fact, operate faster than human brains.
2) Distribution. Human minds are contained up in a 3-pound hunk of meat shielded by a calcite cranium. If functionalism – that is, the idea that minds are defined as the physical activity in the brain – holds, then a variety of distributed or condensed mind-forms should be possible, as long as there is low latency and sufficient bandwidth between cognitive nodes. Electronic signals travel at the speed of light, meaning that an artificial intelligence should be able to exist as a program distributed across computers on opposite sides of the planet and still function effectively. At the same time, highly miniaturized computing machinery, such as nanocomputers, should, in theory, be able to hold human-equivalent minds with a volume and weight much smaller than our current 1450cc brains. For size scalability among brains to be impossible would require that functionalism is false, which, according to anyone that studies the brain, looks quite unlikely.
3) Communication. Imagine the evolution of homonid communication from the standpoint of natural selection if it were a intelligent actor: “If I lower the larynx a bit here, give better control of chest muscles to the neocortex there, and expand the auditory cortex like so, then we might have something.” Language: complex information transfer accompanied by high-fidelity memory storage of the data, is not something that comes naturally to evolution-designed biological life. Electronic systems offer the appeal of the discrete state and the regularity of distinct file types. They can transfer images, text, sound, CAD files, even virtual landscapes at speeds limited mainly by the bandwidth of the data pipe. We can imagine minds that store complex skills, observations, and analyses as independent data files – kungfu.mi, for instance. For nonbiological minds, we should expect the possibilities of communication to be much greater than among biological humans. This dimension of variation falls out automatically when you go outside the default human I/O.
4) Reprogrammability. Clearly the human mind is trainable to a certain extent and even partially deliberately reprogrammable. However, we have very limited control over the overall anatomical structure or chemical makeup of our brains. Our limbic system places games with our “higher faculties”. We have little control over the structure of neurons, the way our brains process visual or auditory data, or our fundamental instincts. A mind with access to its own source code could modify any aspect of itself it thought fit. It could observe the fine-grained structure of other such minds and take inspiration from their cognitive structures in the way that a modern painter might take inspiration from Renaissance artists. It could even build new sensory modalities designed for specialized purposes, like simultaneously visualizing the dynamics of thousands of moving parts in a complex engine, being able to tell one from another in the same way we can tell two faces apart.
5) Intelligence. The vast intelligence difference between primitive microbes and today’s humans give us a kind of clue at the type of variation that is possible. Rather than assume that human beings are around the limits of intelligent thought, or can “visualize anything if we put our minds to it”, we should assume we are typical – somewhere in the middle between microbes and the ultimate limits of intelligence – if not on the far low end of the spectrum. The actual quantity of processing power we have at our disposal is certainly one limiting factor on our intelligence, but perhaps even more important is the particular arrangement of our brains. Intelligence is a relatively recent evolutionary innovation, and our brains as a whole, most of which is based off of preexisting morphological complexity, is not specifically designed to accommodate or nurture general intelligence. We should expect that a mind designed deliberately to implement intelligence could go far beyond our ability to reason, imagine, plan, create, and experience reality.
6) “Morality”. The baggage-free version: “goal orientation”. An AI could begin with whatever goal orientation we give it. There are practically no limits. We could make AIs that are our obedient slaves and love it. Humans dislike being enslaved because being so tends to result in fewer reproductive opportunities, thus evolution selected against it. This evolutionary design would not be built into AIs. Synthetic minds start off with nothing: a blank slate. You could make a humanoid robot whose goal is to eternally hop in a circle while rubbing its stomach and patting its head, and it could achieve that goal, as long as it had a reliable power supply and spare parts. The concept of “self” in the sense of “I exist” is not the same thing as the concept of self as humans know it: i.e., “I’m a self-interested moral agent with certain inalienable rights which I continuously have to reemphasize because my evolutionary and cultural history consists of others constantly challenging those rights, and I am programmed by evolution to be concerned about them”. In an AI, all that self-centered goal complexity is simply not there. A certain degree of goal complexity is necessary to produce intelligence at all, but it’s quite minimal, and the room for variation in the ‘footnotes’ is immense. If we want AIs to do things we consider “normal”, like caring about others, we’ll have to program it in, and instill a desire to build upon those goals in ways that keep the spirit of the original intentions. Otherwise, there’s no telling what we’ll end up with.
Can you imagine other dimensions of variation? Can you think of any regions of the above possibilities which might be ruled out by physical law? If the potential space of possible moralities in particular is so large, do we have an ethical obligation to craft synthetic minds with human-empathic goal orientations?




Another awesome post, Michael. Excellently researched and written. It’s a bit presumptuous for me to think that after considering it for a few minutes I could come up with a significant addition to your dimensions listing. However, coming from a computing background, I think “Efficiency” could also possibly be a meaningful variable that determines a given mind’s potential.
With “Efficiency”, I’m referring to the utilization of the physical attributes that you have already listed. Also, I am specifically referring to efficiency in terms of energy expenditure and processing time. For any mind to reach its goals, the methods that it uses to exploit its intrinsic capabilities have a considerable effect on the final result.
“Optimization” might actually be a better term. Going back to my computing analogy, poor software will deliver slow/incorrect results, regardless of the specialized hardware it’s running on. I don’t believe it’s safe to assume that high-end hardware will naturally include the most efficient software and/or algorithms. In regards to an intelligent mind, the use of neurons and efficient thought processes will go a long way in avoiding dead-end logical paths and wasted time.
Obviously “Efficiency” as a metric could be applied to each of your main points, but I think it’s unique enough to stand on its own. All other things being equal, an advanced intelligence could simply keep increasing its efficiency to continue advancing without changing any other parameters.
How about things like Hive Minds or setting up derivative copies of one mind, sending it off to do a task, and then the original merging with it when it has completed said task? There’s got to be some variable about whether the mind is unitary / pluralistic. Possibly something to do with multitasking.
I agree, Justin, that’s a good point. Similar to “Communication” that Michael’s post had, but distributed across multiple instances/minds. Or possibly multiple minds that operate in shared mode working towards a common goal: “Collaboration”.
A better term for what you’re trying to state is “parallelism”. Much like how all neocortical processing is accomplished in 6 steps, thus allowing for near-instantaneous processing of all sensory input despite the vast complexity of the continuous feed of information, whether the mind is parallel or linear in nature can affect greatly its “processing capacity.” A hive-mind is simply a vastly parallel metaconsciousness, which qualifies as a mind unto itself.
I’ll whip out my qualia-meter and figure out which, if any, corporations or nation-states qualify.
How flawed the mind is.
Physical / structural flaws
error prone / error rates – how much it ends up in non-productive states
How biased
How open to interpreting data
relates to bias and rationality.
Many people do not have rational thought processes and reject data that should change their choices.
Lack of scientific method.
Quality of sensory inputs.
Not just our senses but using devices to gather information. It goes to question of being able to determine the details and the ways to control, amplify and use actual physical effects.
this goes to technology development. Really understanding physical and informational effects and determining how to amplify and exploit and what to exploit it for.
========
I think the last time I looked Friendly AI was looking at a mathematical and objective basis for morality.
The friendly AI site seems to have been reconstructed. Where do things stand in regards to mathematically constructed effective morality ?
Since May 2004, the most up-to-date published item on Friendliness theory has been found here:
http://www.singinst.org/upload/CEV.html
The math isn’t there yet – the objective here is to get a theory likely to work, then implement it over the course of a development project.
Further material:
http://www.sl4.org/wiki/WikiInterviewWithEliezer/EthicsAndFriendliness
http://www.sl4.org/wiki/FriendlyAIPAQ
The most insightful view into Eli’s thinking on FAI that is available online is probably this:
http://www.sl4.org/wiki/DialogueOnFriendliness
The original CFAI seems to have been removed recently, leaving a large hole in the canon.
Michael-Vassar: If that statement was directed at me I should like to point out that I wasn’t implying that all organizations of individuals make up hive-minds; only that hive-minds are made up of individuals.
The closest thing I know of to a hive-mind is ant (possibly termite) colonies; as they communicate via chemical exchange this could be seen as a form of direct “thought” relay from one individual to another, just with an extremely low “hertz-count”.
Minor nitpick “shielded by a calcite cranium,” bone is not calcite. It’s a calcium phosphate, specifically calcium hydroxylapatite.
If skulls were made of out calcite, there would be a lot more fractured skulls.
Keith Henson
7. Coherence. You touch on this under ‘Reprogrammability,’ but humans have disparate motivational and cognitive systems that can produce doublethink and motivational conflicts. A human can have a declarative belief that he will continue after death and be rewarded with 72 virgins, yet fail to adjust his behavior accordingly or make logical deductions from that belief. An AI with an integrated mind, capable of bringing its beliefs and actions into consistency, would be radically alien and superhumanly effective.
8. Transparency. An AI might record the evidence and inferences supporting its beliefs, and the history of each. Human consciousness seems to be, in some important ways epiphenomenal, a process of ex post rationalization. An AI could possess true self-awareness, and immunity to the ubiquitous self-deception that characterizes humans.
I have though in this direction the last year; I am pessimistic in that we as humans are limited to a severely limited cognitive thoughtspace; in other words – I think we can think only a limited number of possible things, limited by what medium we evolved in and evolution.
Non-evolved (constructed) minds, or minds evolved in different mediums (context) would be able to think completely different things. At lower ends there would be overlaps between, for instance, dolphins and humans, but even in these two we see a vast chasm in understanding, because of the way the brain is wired in dealing with the medium.
What I particularly fret about is if we go and create AI, will two AI’s, developed by two different companies, in two different countries, with different cultures and possibly even linguistic systems, by completely different? And worse, what if both have a small overlap with the mental thought sphere but each respectively has two vastly different mind-scapes wherein it can deal proactively with the world around them? Some such cognitive systems could produce seemingly magical results – we’d have no way to understand how the AI does what it does.
For instance, imagine an AI that can deal with chaotic systems, model chaotic systems in a way that is as intransparant to us humans as mathematics is to chimpansees? It isn’t just a matter of transcendance (as human mamthematicians are to chimps) but also aptitude. Chimps for instance have a physical aptitude of climbing and do so far better than humans.
I think AI’s may, once they start developing their own code and substrate, have access to cognitive mindscapes completely alien to ours, with all benefits that entails.
For human minds which have been uploaded in the Kurzweil sense I’m sceptical that such minds will be able to run at arbitrary speeds. The rates at which feedback loops in the brain operate are critically tuned to closely synchronise with the rates at which perceptual events and physical events in the world occur. Make the loops faster or slower and you would get a qualitatively different kind of system, which may not function coherently at all.
However, I do think that AGIs not directly based upon human neural architecture will be able to think at arbitrary speeds.
There are numerous examples of other kinds of mind in the animal kingdom. For an indication of what early AGIs may be capable of we need look no further than our mammalian cousins. It will be at least as difficult for us to understand the inner workings of an AGI mind as it is for us to understand the minds of other animals. We may have a complete understanding of initial AGI designs, but once self-modification begins these evolved mindspaces will soon become unfathomable to us.
I’ve been thinking about inspiration lately, or that “Eureka!” moment when the solution strikes. These instances feel truly non-linear in their connection to prior progress against the problem. Of course there are other problems where you slog it out and come to a workable solution, but I’m not talking about those cases.
My query is this. Does the human mind have an obvious non-linear solution functionality? Is it at all quantum in nature? Or could it be emulated through the use of a quantum computer (software) in parallel to typical computing power? It just feels like that is what is going on when there are so many variables and unknowns to a problem and then almost magically the “right” answer appears in your head… Inspiration.
Sorry, if I’m not thinking of quantum computing functionality correctly, but I hope you can see my driving point.
Keith, thanks for the correction. I’m so happy you’re not in jail and that you continue to post and participate regardless of the trouble that damn cult has been giving you.
Carl, nice additions. I really hoped someone would chime in with some.
Khannea, you really get it. The space of possible AI designs is so overwhelming we can barely begin to comprehend it.
Bob, the problem you point out could be overcome by running the brain in a virtual environment with events running at the same speed as the mind. For an accelerated mind in a cyborg body, it would need to interact with other accelerated minds in accelerated bodies to feel “normal”. Otherwise everything would look very slow. I doubt that this would cause your mind to spontaneously and suddenly break down – for example, people still survive throughout sensory deprivation experiments – but it could be annoying.
I see your point Eric. I know that human intelligence feels really powerful and mysterious sometimes, but the similarity of “non-linear” thinking processes to quantum physics is quite superficial. That surprise and feeling of non-linearity comes from our ignorance over the processes of our own mind. Like a great sports play – say you’re playing basketball and you do a magnificent drive past three guards that leads to a perfect 3-pointer. You don’t know how you did it, but you did it, and it’s not due to quantum physics – it’s because you couldn’t introspect on all the complex motor calculations your brain was doing for you automatically. Leaps of insight in the process of thinking are no different.
I just fell asleep while I was posting this comment…-..but then again the only eyes watching me was my cat’s…no-w she’s laughing at me again.