The Basis Of My Futurism

In a previous essay of mine (The Future Of Intelligence), I specifically disclaimed any attempt to provide evidence for my thoughts about the future, and about The Singularity in particular. This essay attempts to present some of my reasoning and evidence.

This mini-essay can be considered a somewhat less-well-written version of Xuenay's "14 objections against AI/Friendly AI/The Singularity answered". You should probably read that too/instead.

Beliefs

My conclusions are based on the following premises. I can offer only the barest evidence for most of them. Better still, several of them they are impossible to disprove at this time. Sorry about that.

The silly little tags are for quicker reference in the conclusions section.

SRNPossible

Self-replicating nanotechnological robots are possible. There's actually reasonably solid evidence for this; Google can help you with this, but you might want to start with Is Molecular Nanotechnology "Scientific"?.

SRNGraspable

Self-replicating nanobots (SRN) are within our grasp, technologically, in the near-term. Thirty years is, I think, a good upper bound. I would be unsurprised to see them in fifteen.

SRNEexistential

SRN are an existential risk to the human species. See Bostrom's Existential Risks Paper, amongst many others, for discussion on this point.

SRNWeapons

SRN-based weapons (or, for that matter, any nanotech weapons, self-replicating or not) are a fundamentally unstable military technologies, and have massive first-strike advantages, which makes them massively more dangerous than current WMDs. Ask google about {nanotechnology "first strike"}; this has been written about extensively. See in particular Nanotechnology and International Security.

StupidHumans

Humans are too stupid to be trusted with a military technology that is an existential risk and has a first-strike advantage. I argue that this is true prima facia, and that no sane, intelligent human would argue against it. I'm sure that once this is posted to the 'net, someone will try to prove me wrong.

AGIPossible

It is possible to have a computer-based general intelligence (at least as general as the average human), of at least human-level general intelligence. At this time, there is no direct evidence that this is possible. The only argument for it is anti-anthropocentrism, i.e. that it is arrogant to believe otherwnise because humans can't be that special. On the other hand, there's no evidence to the contrary, either. AGI stands for Artificial General Intelligence, by the way, and is used to distinguish an AI that can have a conversation with you from "AI" in the sense of "pathfinding algorithms in video games" and such.

IntIndependent

The stronger version of AGIPossible: There is no aspect of intelligence, in any form, that is dependent on its physical sub-strate. Any form of intelligence that a biological being can have, a silicon being can also have (even if that means emulating hormones or whatever) and vice-versa. This includes social intelligence, mathematical intelligence, sexual intelligence, aggressive intelligence (aka "cunning") and so on and so forth. Again, no evidence for this at all. Do, however, look at Xuenay's 14 Objections to see someone try to present such evidence anyways.

HumansUnspecial

The universe is capable of having beings fundamentally smarter than humans. What "fundamentally smarter" means is up for grabs; I have a friend that insists that humans are Turing complete with respect to intelligence, and that there is nothing that any intelligence could ever come up with that any fully-functional human could not understand if they had an indefinate amount of time to study it. I think he's probably right, but that doesn't change the fact that one Turing-complete machine can compute both faster and better than another, and I think that applies to minds as well. If you want me to talk about this one more, let me know. No evidence for this at all, again, except the anti-anthropocentrism argument. No evidence against, though, either.

BigIntPossible

There is little or no upper-bound on the maximum intelligence of a being, except for physical limits themselves. In other words, human-level intelligence isn't the ceiling (see point #8) and there's no hidden ceiling just above human-level. If there is a ceiling, it's at the point where the being's computational substrate is so huge (planet-sized or larger) that light-speed lag starts causing problems. Matrioshka Brains seem to be the absolutely insane outer limits of this idea. Again, no evidence. Again, no evidence against, though, either.

Conclusions

General

• From SRNPossible and BigIntPossible, we have that anything smart enough can invent SRN more-or-less immediately. By "invent" I mean "solve all the inherent engineering problems and develop a very short path from our current technological leval to SRN production".
• From SRNExistential we have that anything smart enough to invent SRN quickly is a clear and present danger to the continued existance of the human race.
• On the other hand, from SRNExistential, SRNWeapons, and StupidHumans we have that EITHER humans must entirely avoid creation of SRN, OR something smarter than humans must be the first being on earth to do so.
• From StupidHumans (more or less, plus general knowledge) we have that humans are too stupid to avoid creating a technology just because it might destroy them all.
• Therefore, from the last two conclusions, we have that a super-humanly intelligent being MUST be the first being on earth to develop SRN, because a small chance of such a being stopping us from destroying ourselves is better than the near-certainty that we'll manage to do so if left to our own devices.
• From SRNGraspable and the current state of the biotech industry, human augmentation is unlikely to arrive fast enough such that a super-humanly intelligent human is going to invent SRN. Keep in mind the gestation time of humans when thinking about this issue.
• Therefore, we must look elsewhere for our super-humanly intelligent saviour. The only other place I know to look is AI.
At this point, I've essentially explained the minimum set of reasons why I give money to the Singularity Institute, and further why I think that giving money to groups pursuing friendly general AI is the single most important philanthropic pursuit possible for a human today. For the more warm-fuzzy reasons, see my sysop scenario essay.

Hard Takeoff Conclusions

Hard Takeoff is the theory that once we have a near-human-level AGI, we'll have something far smarter than humans very quickly thereafter.

• From IntIndependent, HumansUnspecial, BigIntPossible, and my knowledge of computer software, I infer that a (human equivalent) AI created deliberately (that is, coded more-or-less by hand, without using genetic algorithms or neural nets or anything like that) will be able to make itself smarter by examining its own code, modifying it, and running test copies of itself, in exactly the way its human creators would. I'm not even going to postulate that the AI would be much BETTER at coding than its human creators would, but I do expect this to be true. General Intelligence And Seed AI's advantages section talks about this issue in some detail.
• From the above, the AI will be able to do this (improve themselves by teaking their code) more than once. From BigIntPossible, the AI will be able to do this almost indefinately. Each iteration will proceed faster than the previous iteration, because the AI is smarter now and because it knows more about modifying itself. Eventually the AI's intelligence will begin to be bound by the hardware ey is running on. At that point it will slow, probably dramatically, but by that time it is plausible that the AI will have been trotted out in front of enough adoring audiences (or whatever) that the project will have a substantial enough influx of cash to stave off this problem indefinately.
• From various previous conclusions, this process will continue until the AI gets smart enough to develop SRN (or something equally powerful) and smart enough to convince its creators to let it try its hand at its inventions. Note that the AI could be deceiving us about one or both points, and unlike when dealing with humans, we would have no secondary clues hinting at the deception. Once both the developing and the convincing is done, the AI is effectively the ruler of the world (unless other groups have already developed SRN-level tech, but from SRNWeapons, if humans have done so, we're probably dead anyways).
• It follows from the above that it is absolutely vital to be thinking about Friendliness (that is, the craft of ensuring that an AI still likes humans even after it is much, much smarter than humans) in every serious general AI project that is currently underway (and there are at least three that I know of). This leads right back to why I give money to the SIAI.