Problems of specificity

Eleizer Yudkowsky on the virtues of specificity:

When the unenlightened ones try to be profound, they draw endless verbal comparisons between this topic, and that topic, which is like this, which is like that; until their graph is fully connected and also totally useless. The remedy is specific knowledge and in-depth study. When you understand things in detail, you can see how they are not alike, and start enthusiastically subtracting edges off your graph.

This is a problem that is ever-present online. Very rarely are online debates actual arguments, they are bickering contests between people who have completely different ideas about the actual definition of words and the scope of the debate.

I was impressed by Yudkowsky's thoughts on the three schools of the singularity because it addresses directly the oft-ignored problem of what exactly someone means when they talk about the singularity.

This was the problem with PZ Myers' objections to the singularity - he argued against a few elements of Kurzweil's thesis and used these inconsistencies to dismiss the whole thing out of hand.

I agree there are problems with various aspects of Kurzweilian singularitarianism but they need to be adressed clearly and specifically.

Delicious specificity

In a similar vein I've been trying to work out what the best way of organising my Delicious tags is.

There are some tags, like "technology", "economics", "politics", "science", and "toread" which are so wide-ranging they lose all meaning.

However if my intention is to be able to refer back to a specific article when I need a reference too much specificity can hinder my search.

Tag bundles help solve the first problem of overarching vagueness by promoting "technology", "economics", "science", and "politics" to a well-earned retirement on the board of directors {?}.

Could it be possible to build a system into Delicious whereby it is possible to say that something vaguely reminds me of something else?

I don't mean a tag like "remindsmeofStephenFry". I mean a way of tagging a document that doesn't explicitly reference Stephen Fry in any way but still reminds me of him. Something like remindsmeof:"StephenFry".

In this context remindsmeof would be a command recognised by the Delicious API to note that the article isn't explicitly about Stephen Fry but nevertheless puts me in mind of him.

I suspect that this sort of thing is less useful in practice than I imagine, particularly as much of the utility of Delicious comes from its simplicity and intuitiveness.

And as to Yudkowsky: obviously this respect for the specific can be taken too far: a general knowledge encompassing many fields can also be very valuable.

It's best to learn a lot about a little and little about a lot.

What I think of the three kinds of technological singularity

Michael Anissimov has a post up on the three kinds of singularity. This is based on this post on the three schools of singularity thought by Eliezer Yudkowsky. This is relevant to many criticisms of the idea of a technological singularity as criticisms frequently focus on minor or ancillary effects of the singularity.

As I said before, I don't "believe" in the singularity. I think it's either irrelevant or a fairly trivial observation of technological growth trends. But within the context of the three schools I think I can express my thoughts more coherently.

Here are the three kinds of singularity Anissimov and Yudkowsky describe:

1. Accelerating change: advances in computer science, AI research, genetics, human augmentation, and biotechnology create a positive feedback of rapid technological growth. As the abilities of our tools improve and our own abilities improve through augmentation (both external in the form of personal AIs and internal in the form of intelligence-enhacements and nootropics) technological change accelerates exponentially.


2. Event horizon: advances in computer science, AI research, genetics, human augmentation, and biotechnology lead to the creation of a greater-than-human intelligence. It is impossible for a less intelligent mind to predict the actions of a more intelligent mind so it is truly impossible to make any definitive statements about what will happen after a superintelligence (whether pure AI or strongly augmented human).


3. Intelligence explosion: advances in computer science, AI research, genetics, human augmentation, biotechnology and neurobiology allow intelligent beings (either human or AI) to alter their own brains so as to improve their own intelligence. As intelligence is the source of all technological development this process will feed back on itself, as the slightly more intelligent beings develop slightly better ways of improving their intelligence, all the while creating amazing spinoff technologies.

Here's what I think of them:

1. The accelerating change school of the singularity is the one I find most compelling. This is because it is both logically plausible and reflects the experience humans have had of changing technologies in the past. Technologies like electronics combine with digital computer theories to develop fast computers that go on to have a major effect on other areas of development. I think the accelerating change argument is the most coherent and reasonable depiction of a technological singularity


2. The event horizon school is flakier. First, I have issues with the idea that greater-than-human intelligences are necessarily unpredictable, second, I don't believe that raw intellectual or cognitive ability is the primary driver for technological progress, and thirdly we have seen that it is already impossible to accurately predict all the outcomes of any technological development, let alone strong AI/posthuman superintelligence.
   


3. The intelligence explosion school is flakier still. It is based on the assumption that a sufficiently powerful general intelligence would necessarily be able to comprehend how it's own mind works and know how to improve it. I do believe that as knowledge of the workings of the brain increases it will lead to real gains in various intellectual capacities, through nootropics, brain augmentation, or through brain simulation on faster substrates. Gaining additional knowledge about the brain doesn't require us to be "smarter."

With reference to the last point: the knowledge of how the brain works will be gained through trial-and-error scientific experimentation and ongoing technological development of brain-scanning technology (itself developed by trial-and-error technological tinkering), surgery (again also developed through the inductive tinkering of the barber surgeons), and neural interface technology (which is being tinkered with as I write).

Anissimov believes that all technological progress must be judged on the basis of how much closer it brings us to the existence of a superintelligent AI, because then the superintelligent AI will take over the business of technological development and create an intelligence explosion.

Anissimov describes himself as a technological determinist, as such he presumably believes social change is caused primarily by technological development. I agree with technological determinism in general but I feel Anissimov's perspective is closer to cognitive derminism: he believes technological (and hence social) change in the future will happen purely as a result of the cognition of AI.

This is at odds with our experience because the component of scientific and technological development that relies entirely on pure cognition (e.g. Einstein's development of the theory of relativity or Newton's laws of motion) is quite small compared with those which required a substantial amount empirical study (Darwin's theory of evolution) or mechanical tinkering (Faraday's law of induction).

This is a similar criticism to Kevin Kelly's idea of thinkism, where Kelly highlights the fallacy of believing you can study the universe by simulating it, without recourse to experiment to attempt to falsify your belief.

To summarise: although the development of a smarter-than-human AI will be a huge aid to our understanding of the nature of intelligence, consciousness, and the human mind there is no reason to assume the effects (though unpredictable) will include an intelligence explosion - I agree that it may help lead to an acceleration in technological development - but it will only be one part of the general acceleration.

The quantum leap intelligence problem: are posthumans ineffable?

There is an idea prevalent in transhumanism that when posthumans or strong AI finally develop they will be to us as we are to the beasts that perish. They will be so much more intelligent that we will be incapable of understanding them.

They will possess super-science mojo and will live in mathematically optimal blocks of matte computronium, and they will have awesome tech that we puny baselines won't be able to distinguish from magic. They will be ineffable and godlike and we simply won't be able to understand them or their motives.

(Aside: for the sake of brevity for the remainder of this post I will refer to "them and us" to distinguish baseline humans from posthumans - not because I don't buy into the whole Kurzweil machine-human merger, but just because it's easier to write about)

I disagree with this idea of posthuman ineffability.

The idea suggests that there are other ways of being intelligent (i.e. possessing a highly accurate model of the outside universe and a highly accurate model of yourself and your fellows, thereby enabling self-reflection, communication, and culture ¹) that are an entire quantum-leap above human intelligence such that we won't be able to comprehend them or their actions.

Michael Anissimov has written an interesting article making the point that human beings are dumb. In fact we possess only the bare minimum of intelligence required to create the civilization we have now.

Michael Anissimov makes some good points in this article on the current state of human intelligence:

Hey, human philosophers — I’ve got some bad news. It turns out that Homo sapiens probably isn’t the qualitatively smartest possible being.

...

How do I know? Well, most other members of the genus Homo had plenty of time to build agricultural civilizations, but they were too unintelligent to get off the ground. Homo sapiens was just barely smart enough to do the trick. And like a self-replicating machine that moves from 99.9% closure to 100% closure, the payoff was big.

I agree with this as far as it goes. All it took to develop complex social technologies like language and complex physical technologies like bows and arrows was a small increment in intellectual capacity.

We made that quantum leap from animal to human around 100 000 years ago: in the intervening period we haven't evolved a great deal (in fact, some say we've stopped evolving at all).

Ergo we are possessed of the bare minimum intellect required to sustain and develop technological civilization.

Anissimov uses this as an argument in favour of the idea that there there are many more superior modes of intelligence that we haven't yet developed or encountered, or in his words:

The apparent magnitude of our accomplishments, including those of Einstein, is merely a side-effect of how low our standards are. To another species on another world whose intelligence was crafted in the furnace of selection pressures more intense than ours, quantum mechanics is obvious from the get-go. The only thing funnier about how dumb we are to take so long to figure it out is our self-importance at having finally figured it out.

This is where I disagree with Anissimov. I think that the miniscule quantum leap between pre-human animals and human beings is a one-time event. Improvement is certainly possible, but to claim that there is some qualitatively and quantitatively different perspective on the universe that is definitively superior in every measurable dimension to human thought and would result in beings that we are incapable of understanding is incorrect.


Here's why:

1. Human knowledge and understanding does not progress wholly through deductive reasoning or pure cognition. In fact a large amount of human knowledge and understanding comes from what Nassim Nicholas Taleb calls stochastic tinkering, and Eric Beinhocker calls deductive tinkering. Trial and error and accident has contributed enormously to the development of human knowledge. Presumably a posthuman would make mistakes: otherwise how would it learn? And if it doesn't learn how does it grow and develop?
   
2. I agree with Kevin Kelly that there is a fallacy in the idea of "thinkism." Thinkism is the idea that it is possible for a mind, completely ignorant of the workings of the physical universe, to consider a few small objects, like a rock, flower, a feather, and a model of a galaxy embedded in amber, and then use these items to deduce the workings of the physical universe without any recourse to experiment. It could well be that there are other universes with different physical laws that could generate those items and without recourse to experiment how would this mind know which universe it lived in?
   
3. We will share the same universe (they may go elsewhere, of course, but the chances are baselines will stay here): as such this posthuman entity will be subject to all the usual laws of entropy, conservation of energy, gravity and whatnot. As such this posthuman will need things and do things that are explicable to us. Only by creating a solipsistic alternate-reality computer bizarro world could a posthuman behave in a completely ineffable fashion: and even then a posthuman would still be subject to the axioms of a given logical or mathematical environment. N-incomplete problems would remain so.
   

Looking at point 1 "how does a posthuman learn" suggests an interesting counterargument: "posthumans develop in a way that doesn't involve learning, they use something different and ineffable."

The problem with this counterargument is that what I'm arguing is empirically refutable: the existence of a truly ineffable posthuman entity is something that is observable, so my point can be refuted by the observation of one posthuman entity whose motives and actions we do not understand. However transhumanist thinkers can continue asserting that true posthumans are by definition ineffable until the end of time. I predict that as posthumans emerge and their actions are studied they will eventually always be found to be explicable by baseline humans (if weird and peculiar - see below).

I agree that there are almost certainly better modes of intelligence, but I disagree with the idea that these modes of intelligence will ever be wholly incomprehensible to baseline humans.

They may be faster, cleverer, wittier, more attractive, stronger, longer-lived, instatiated within superior hardware, and better at poker - but it doesn't mean baseline humans would be incapable of understanding them.

The distinction between what I'm arguing and what Anissimov implies in his article is fairly subtle, and I could be accused of nit-picking, but I think it's important that we realise that there is no reason to assume posthumans will be completely and utterly ineffable to us, at least not if they want to survive IRL (which they may not).

That humans have accomplished what we have says more about the power of the evolutionary methods of stochastic tinkering combined with occasional deductive reasoning than it does about the brilliance of human intellect: and this is exactly what Anissimov is saying and why I agree with the premise of his article.

But once you possess culture (what Ian Stewart and Patrick Cohen call extelligence, or what Richard Dawkins might term "a memetic environment"), and a reasonable means of manipulating the universe it doesn't matter how "smart" you are. Trial and error and learning and robust heuristics take care of the rest.

I believe that some posthumans will be pretty weird, some may be charismatic, some may be frightening. But we can get to where they are, they are post-humans and they took a path we will be able to follow. Because of this and for the reasons given above I don't believe posthumans or posthuman civilization will ever be truly ineffable.



1: In fact it could be that this superior intelligence works on a completely different basis to creating a highly accurate model of the universe and the self, and works on some other basis that we can't comprehend. This non-intelligent "intelligence" would be truly ineffable and would completely disprove my point if it actually was superior in every possible way to human intelligence.

Further reading: if you do understand precisely what I'm trying to say I should mention that it isn't original, Greg Egan argues something similar in the opening chapter of Schild's Ladder.