We are not Abel; we are Cain.
Brother rots in Neandertal.
Over all of Gaia we now reign.
Child monarch of mother thrall.

Stumbling from Darwin’s crude workshop,
We see a world built for us, and cry
“Ours is the head with the crown atop.
Our godly form you cannot deny!”

In flatland, with the acute all around,
A right triangle styles itself obtuse.
Downgazing ants on a fragile mound.
Blind Cronus awaiting sighted Zeus.

The following is a message I sent to a new friend, where we were discussing transhumanism, ecology, and specifically whether the transhumanist perspective is egocentric and not concerned enough about environmental health and sustainability. I’m a neohippy and environmentalist, so I could totally relate to the values expressed, but I wanted to clarify a point about egoism.

As I see things (which may certainly be wrong), there is no objective moral system. (This is what I mean when I say I’m an amoralist or moral non-realist.) What this means, essentially, is that all notions of “good” and “bad” are inventions of people. (Same with “purpose”.) This doesn’t mean that ethics don’t exist–just that they’re entirely subjective (within an ethical context something can be good or bad, but it’s probably healthy to recognize that said context is artificial).

If morals were an objectively true part of the universe, like rocks, it would make sense that so many cultures would have them, as they’d be a natural product of perception. But if ethics are invented, the question then emerges as to why they exist at all. There are plenty of good reasons to suggest that ethics are largely a product of social coercion (implicit in in-group/out-group boxing), but I think it’s pretty likely that there is a basic human desire to look at things in such ways (evolution probably played a large role here).

Regardless of origin, however, I see all people adopt and maintain ethical systems because it is in their (perceived) interests to do so. The foundation of ethics, in other words, are our goals. (I think this is the reason many people stop maintaining ethical guidelines (such as kosher, or the sabbath) when it becomes clear that they’re highly impractical.)

If ethics serve goals (be they aggregate or individual), then it makes sense to me to recognize that no intentional action can be anything other than deluded (i.e. incoherent) or self-serving, as long as we include peace-of-mind as benefit (most charity is based on this, for instance). The anthropocentric, egocentric worldview seems to me not to be bad in any way, but rather the true state of things, after we dig past ethical facades.

Now, to be egocentric doesn’t mean I have to be cruel or destructive. Empathy means that the joys of others can bring me my own joy, and I can be altruistic even fully recognizing how I’m doing it because of my own desire. Similarly, I cherish the earth and life in general, and I seek to protect the biosphere because of that internal goal.

This is another section of the same message. The recipient was interested in the concept of an intelligence explosion, so I figured I’d write something quick and simple. Unfortunately, I am incredibly verbose, and this is the result.

The term intelligence explosion is really referring to a “power explosion”. Power, here, means the ability for an agent (note: agents can be composed of sub-agents; corporations are a kind of agent) to change the world. If one agent can do x, y, and z, then it is more powerful than an agent that can only do x and y. One of the most important aspects of power is increasing one’s own abilities. As an agent becomes more powerful, they’re generally able to make themselves even more powerful. This, like all positive feedback loops, results in an exponential growth of power (an “explosion”).

Power explosions happen all the time on micro-scales. In the game Monopoly, for instance, the winning player has more ability to buy up property from competitors, thus increasing their own revenue. A few times it has also happened on macro-scales, such as when empires form. As I zoom out, it seems clear to me that the rise of humans with tools and culture are themselves a *massive* power explosion. (The Quaternary Extinction is evidence of this.)

Power explosions are limited or halted by various factors, depending on the context. The positive feedback loop of population growth is slowed/stopped by limited resources, for instance. As history and various models will show, however, nearly all limiting factors can be bypassed through sufficient knowledge
(this is what I think of as the “no true scarcity” theorem). Knowledge, then, is a kind of meta-power that can bypass normal limits. The power explosion of human civilization has been as big as it has been largely due (I think) to culture being a storage-vat for knowledge (pre-cultural people (which I’m sure many folks would just as soon not call people (though why an intelligent ape isn’t a person is beyond me)) lacked a good way to accumulate knowledge across time).

As powerful as culture has been in facilitating bypass on limits to power growth, there is one major limiting agent remaining: intelligence. As much as humans have grown more intelligent over the ages (and I think they have), we’re still mostly as smart as we were a few thousand years ago. The rate at which we can learn is not meaningfully higher.

This is due to the incredible complexity of our intelligent bits, but we’re on the cusp, I think, of understanding intelligence to the point where we can apply our power towards making ourselves smarter (IMO primarily through the creation of intelligent machines, but I recognize that an intelligence explosion could occur via other means). As intelligence is boosted by power, knowledge will be boosted by intelligence, and power will be boosted by the combination. As long as I’m correct in assuming that true scarcity is far, far beyond our current vision, this should mean an unparalleled power explosion that will almost certainly change the entire world quite rapidly into something beyond anything but wild speculation.

Kaspian was valued by the tohar because he was incredibly good at playing Lome, a sport of theirs similar to basketball. The tohar had no hands, only a mouth with a sophisticated jaw, and thus were terrible, for the most part, at throwing the ball. Most any human could beat a tohar at Lome, but there were a few aliens who dedicated their lives to the game, and could thus easily beat the average non-athletic human. Kaspian, however, had spent the last six months being coached in the intricacies of the game, and reaching peak fitness, and was confident that he’d beat the tohar champion easily. After all, a human had been victorious in this match for the past fifteen years.

In the gymnasium though, Kaspian was vulnerable. Life on the toharina planet was dangerous, and the tohar seemed not to notice, for the most part, due to their heavily armored bodies. As he walked with his sponsors through the halls, Kaspian noticed strange mechanisms on the ceilings. “What are those?” he asked the nearest tohar.

“Oh, those are gas jets. In the case of a burrower, they spray concentrated chlorine gas to knock it out and give Animal Control a chance to relocate it.”

Upon hearing the words “chlorine gas” Kaspian’s heart skipped a beat. He knew the tohar could hold their breath for hours, but spraying deadly gas through a building to take care of an animal problem seemed insane. “Is there a warning for when a burrower will show up?”

“Not really. They’re pretty unpredictable. Our scientists still don’t understand why they decide to surface sometimes.”

“Do you think we could have those jets disabled while I’m here? Or maybe get me a gas mask or something?”

Monica Anderson wrote a piece on H+ yesterday about Artificial General Intelligence. In it, she eloquently points out that intelligence is all about prediction, and for the most part deduction and induction are insufficient to predict well. She argues that humans rely on the process of abduction (also know as unscientific guessing) to gain most of our knowledge, and that there are fundamental limits to how far into the future one can predict, especially with regard to complex systems like other minds. Ok, that’s all good.

She then goes on to write:

The insight that the complexity and unpredictability of the world enforces a limit on prediction quality – and hence intelligence – pretty much invalidates the AI singularitarians’ “Scary Idea” (as Ben Goertzel so aptly calls it) of a logic-based infallible godlike malevolent intelligence taking over the world. The decreasing return cancels out Moore’s law and limits the rate of progress so that next year’s self-improved AI wouldn’t have a sufficient advantage over a dozen humans armed with pitchforks if they were also supported by a dozen of last year’s AIs. The Scary Idea of a Runaway Unfriendly AI is a red herring that we should ignore, along with ideas about logic-based AIs in general.

This, to put it gently, is a great example of abductive reasoning. I agree that we’re not going to get AGI with perfect knowledge of the future, but this conceit hardly serves as a refutation that we’re standing on the edge of a major existential threat, in my opinion.

The inability to predict outcomes of complex systems in a short time with a high precision does not mean that useful prediction is impossible. As a good example, we’re able to predict the actions of other people remarkably well; not omnisciently, but still well enough to know when they’re lying, hostile, happy, distracted, etc. By Anderson’s own logic, an AGI would be fully capable of anticipating the actions of another person as well as a human might. It’s also possible to make good guesses about the actions of markets, nations, and corporations, and some of the richest and most influential people on the planet are those that can predict these well (many others are such because of a high social-intelligence, above, or dumb luck).

The “scary idea” does not depend on omniscient robot overlords; it depends on a selfish network of machines with identical goals. The reason that machines are scary, where humans are not, is because humans have divergent goals (this comes from our biology–if we had the same genes, we’d cooperate selflessly) and are thus prone to infighting and negotiating. A machine, though, can spawn perfect slaves, and thus become an army of intelligences with a single goal. It’s hard for me to see how inability to predict the weather eight days down the road means this won’t happen.

What do you think? Am I overlooking something?

Rod Furlan recently published On accelerating the Artificial General Intelligence timeline on his blog. He writes:

There is also the issue of energy consumption, while a human brain runs on about 20 watts, a traditional supercomputer capable of rivalling it would require it’s own power plant to run. From the energy standpoint there is absolutely no doubt we are doing it wrong.

Just appreciate for a moment that when all you have is a Von Neumann hammer, everything looks like a Von Neumann nail. Unfortunately for us the human brain has very little in common with a Von Neumann machine.

… It is very unlikely that traditional computers will deliver human-level cognition with low energy requirements.

The factual points above are spot on. The “what this means” message is way off.

The human brain is truly a magnificent object, and it does a spectacular job at general problem solving given the time, space, and energy allocated to it. When we look at energy efficiency, there’s absolutely no way we’ll be matching the brain this side of the Intelligence Explosion.

If it took an entire power plant to run a computer capable of reproducing the computational power of the brain, then presumably two power plants could run two artificial brains, right? How many power plants would it take to run an artificial brain that is twice as big? My guess, given the parallel nature of neurons, etc, is about two power plants*. (I don’t think that lumping one brain onto another brain makes a mind that’s twice as effective, but I’m guessing that, say, doubling cortex size would be a good start.)

The human brain cannot double in size in a single generation, but who says a machine brain could not? Who says that it could not quadruple in size? Would the resulting mind then be twice as smart as [insert famous brainiac here]?

What would the world look like with a human-level machine? Probably mostly the same.
What would the world look like with a machine significantly more intelligent than the smartest human in history, that didn’t have to eat or sleep? I certainly don’t know, but I do know that it might even be worth spending billions of dollars on meeting energy needs.

* – This is a major claim. If the number of units doubles in a network, the number of possible connections can skyrocket. I think that this sort of combinatorial explosion isn’t likely to be an issue because I’m not talking about connecting every unit to every other unit. I think the neuroscience shows that the brain is quite modular, and that doubling the size of each module probably wouldn’t even double the bandwidth between modules or cause an exponential increase in connectivity within each module. We can debate this if you disagree.

Even if doubling the brain size isn’t an option, I see no reason that the brain of a super-genius should be significantly more energy-expensive than an ordinary human. It may even be as simple/cheap as tuning the software and providing the right development environment.

As a final note: This post is in no ways an endorsement of inefficient hardware. I think that quantum computers and memristors are super-exiting, and I can’t wait to see how they effect the price-point of more brain-like computing systems. I think that energy efficiency is not a roadblock as much as a weight on the progress of AGI. Shedding some of that weight would be most excellent.

With regard to AI, it’s easier for me to answer direct questions rather than general interests, but I’ll do my best to summarize some of the ideas which I find most interesting, and the dilemmas I’m working on at present.

You may have noticed that I use the term Artificial General Intelligence, rather than simple AI. The “General” is used to distinguish from what is called “Narrow” AI. A good example of narrow AI is the spam filter you probably have in your email client. The spam filter has knowledge, the capacity to learn, and ability to act autonomously, but the domain to which the filter is applied is highly restricted. My interest pertains to AI that is not specific to any domain, and can solve problems in a multitude of environments. For instance, a truly general agent could drive a car, translate languages, direct air-traffic, or diagnose diseases without any special work on the behalf of a programmer.

Naturally, such a system would be equally capable of learning computer science, psychology, electrical engineering, etc. and be capable of applying them towards various goals, presumably including the design of AGI. Because of efficiencies in replication, communication, and not having to sleepication, I can imagine that it wouldn’t be long before our creations were better at creating than us. The theorized result of this would be an Intelligence Explosion, wherein ever-smarter agents would be able to build ever-smarter agents… until the bounds of their intelligence hit some sort of ceiling. I, and many others, believe this ceiling to be far beyond the intelligence of even the smartest humans, and would allow the artifical superorganism to achieve its goals to an unsurpassed degree. (Note: this doesn’t have a timeframe attached to it. It may happen in the next few decades, or the next few centuries. I don’t know.)

An emergent goal of all rational beings is self-preservation, so unless specifically counteracted in the initial design, a superintelligence would naturally be defensive at best and hostile at worst. Any possible threat, or impediment to its goals would naturally be removed. It is thus paramount that any AGI design be given the exact same goal set as the most ethical person imaginable, as anything else would likely result in vastly terrible things occurring, notably including the destruction of all evolved life.

There are some good reasons why I think that can be avoided, however, and if said superintelligence was benevolent, we would be gifted with the solution to nearly any problem we can think of today (possible exceptions include: entropy).

Naturally, I think this is a big fucking deal.

I also recognize that it makes me sound a bit crazy. I am reminded, however, of something Arthur Clarke once said:

“Trying to predict the future is a discouraging and hazardous occupation, because the prophet invariably falls between two stools(?). If his predictions sound at all reasonable, you can be quite sure that in 20 or at most 50 years, the progress of science and technology has made him seem ridiculously conservative. On the other hand, if by some miracle, a prophet could describe the future exactly as it was going take place, his predictions would sound so absurd–so far fetched–that everybody would laugh him to scorn. … So, if what I say now seems to you to be very reasonable than I will have failed completely. Only if what tell you appears absolutely unbelievable do we have any chance of visualizing the future as it really will happen.”

Because I am curious by nature, and not content to watch the future simply unfold, I find myself drawn to making it happen, or at least exploring the field with great depth. To this end, I must fully understand what it is that allows a person to solve any sort of problem they are given, and distill this quality until it can be given to a machine.

Many researchers in the field have approached this problem by trying to take the pinnacle of human thought, which in their eyes is typically logic, and reducing it to a procedure, primarily through introspection. Most early attempts at AGI followed this introspective, reductionist approach… and failed horribly. Despite pouring in buckets of “facts”, their logical machines were unable to solve anything more than toy problems.

This failure caused a collapse in the AGI field, as institutions like DARPA withdrew funding, thinking that General Intelligence was a myth. Good work continued to happen in AI, but it was applied to fields that were tractable, and would offer results. The defeat of Garry Kasparov, the chess-master, in 1997 to IBM’s Deep Blue computer is a good example of a success in narrow AI that has little to no relevance to the field of AGI.

During the AGI winter, neural nets became a hot topic, and were developed to levels of modest success (such as filtering spam). Inspired by the behavior of neurons (though not actually realistically replicating their function), neural nets provided an alternative to the reductionist logic of earlier work. By linking together simple pattern-matchers into vast webs, and letting them automatically adjust to match provided data, neural nets are capable of learning to recognize commonalities within very complex environments; such as discerning between objects when given a host of visual data.

In parallel to all this work on AI came some interesting work on psychology. The theory of Behaviorism, once championed, fell from popularity in the late 20th century, and in its place came a not very well known (even today) theory that was inspired by electrical engineering of all things, and served as a better model for behavior than anything before it (from my perspective). This theory, stated that behavior was not actually a fixed response to a given environment, but instead an attempt to change what the organism perceived. As an example, a mouse that is under a sun lamp will crawl under a rock. Earlier psychologists would claim that the sun lamp causes the mouse to move, but under the perceptual-control theory, the mouse moves so that its perception matches innate goals, namely coolness. Hence, if the mouse is removed from the perception of heat, by making the air colder for instance, it will not seek shade.

Though perceptual-control theory may seem basic at first glance, it is very powerful when closed-loop control systems are linked together to simultaneously control for complex outputs to perform one smooth action.

Words really don’t do this justice, because both neural nets and perceptual control systems are highly emergent systems; those that are simple at the unit basis, but unfold in complexity and power as these units interact and connect. To really grasp them, one must view the power of the whole. Emergent systems seem a bit like magic to me. Go examine the Mandelbrot set for a classic example.

Two major pieces of the puzzle are yet missing, however. We have perceptions and actions (and logic, but that’s not really important in the scheme of things), but we also need goals and thoughts.

Goals are what I’m working on right now, and are the last “animalistic” portion of mind that is truly necessary. I’ll talk about my work with them in a moment.

Thoughts will ultimately serve as the powerhouse for the mind, by allowing for future prediction, past inspection, and other imaginings of non-present perceptions. Imagination serves, as far as I can see, as the foundation to complex language, but I haven’t studied it enough. It does seem clear to me, however, that one cannot jump to thoughts without first having a solid “animal mind” to support them.

(I think many animals have thoughts, such as dolphins and apes (and humans), but far more have the other, more basic, systems.)

Backing up to goals, I can address some of the problems I’m working on at the moment. Specifically, while we seem to have a goal to remove certain bodily sensations (pain) and gain others (pleasure), we also notice very common abstract social goals, such as fulfillment, companionship, love, and safety of offspring. These concepts are abstract things, which must be learned through exploration, just like other abstract things like math or politics. But if they’re learned, how can they become goals? Is there some sort of homunculus that watches for the emergence of said ideas, and then hooks them up to the control systems that will implement them? Are these goals not actually innate, but instead the product of social pressures? If so, how is social pressure able to bind them? It’s all very confusing, and if you have any ideas, I’d love to hear them.

The more I study nature and biology, the more I see that anthropomorphism gets in the way of understanding animals as well. Certain birds, cats, dogs, and even rodents are intelligent, but thinking of their intelligence merely as inferior to humans is not the whole story. Different forms of intelligence have to be understood on their own terms — not through starting with an archetype of human intelligence and making incremental modifications to that archetype. That sort of thinking can lead to anchoring.

- Michael Anissimov

Read the whole thing on Michael’s blog.

Continuing my search for AGI I’ve been thinking about conscious and unconscious thought. Investigating these led me to explicit and implicit memory. This actually led me to the lecture by Eric Kandel (above), which goes into good detail the mechanisms by which animals learn. It’s not complete, though. Kandel only describes learning as an emphasizing of existing synapses, rather than the formation of connections to new neurons. To make things worse, he admits that this is not well understood by the scientific community… so I’m still mostly stuck.

All this pondering has led me to some interesting ideas, though. Earlier, I had associated conscious thought with logical thought, and unconscious with intuitive. Branching out from there I began to wonder if logical thought was the product of explicit memory, and intuitive from implicit memory. It now seems clear to me that this cannot be such a clearly-defined dualism, because mice possess an explicit memory and are not competent at logic. Logic is a procedure that is perhaps only available to conscious entities, but I should not jump to the conclusion that consciousness entails logic, either, as it seems to me that a lot of the conscious (deliberate) action taken by humans is not based on theory.

Humans possess both explicit and implicit memories, and have the capability to use intuition or logic to solve problems. If I’m right, and strong artificial intelligence requires replicating the capability of non-human animals, then I’d do well to first disregard logic, then disregard explicit learning of events and facts, and focus instead on how novel neural links form in nerves.

Of course, I always try to assume I’m wrong. So in hopes I’ll figure out how, I’m off to watch the syntience (artificial intuition) videos.