A little light reading ;-)
A recent article of mine ridiculing popular-science oracle Stephen Hawking's preposterous suggestion that we muck about with human DNA to keep abreast of advances in artificial intelligence drew a vast number of fascinating e-mails from readers interested in the topic.
The pitfalls of dealing lightly with metaphysics in a 700-word news item can't be underestimated. Most readers, understandably, didn't accept my claim that it's impossible for an AI to achieve human intelligence.
Nevertheless, I believe I can make a rigorous argument to support that claim. This will involve several of my own postulates, which you will either accept, consider thoughtfully, or reject out of hand. But even there, I believe all the postulates are defensible and dwell within the realm of common sense and reasonable inference.
In a nutshell, I say that it's impossible to manufacture an AI which will compete equally with human intelligence. The elusive quality which human thought possesses, and which an AI can't possess, is something I call 'irrational insight'. Note the modified noun 'insight'. I'm not talking about irrationality per se. 'Insight' implies, and deliberately so, the qualities of pertinence and consistency.
1. The universe consists of both rational and irrational elements and qualities. Let's call the rational the r-set, and the irrational the i-set.
2. Elements and/or qualities of the r-set can either be expressed as a ratio, or can be described fully and pertinently in a human language. Elements and/or qualities of the i-set cannot.
3. The i-set is real and possesses elements such as structure, substance and force, and qualities such as beauty, pertinence and consistency; though its forces, substances, structures and qualities are not rationally expressible.
4. The i-set and the r-set are interdependent and mutually interactive, just as the inexpressible, irrational length of the side of a square with the area seventeen is a necessary element of the perfectly rational area seventeen, and vice versa.
5. The universe includes randomness, manifest in both the i-set and the r-set; but 'irrational' and 'random' are radically different qualities which may or may not converge.
6. DNA is the self-sustaining engine ('prime mover') of speciation; and natural selection (evolution) is a secondary process crucial to speciation which injects outside (but not necessarily random) inputs into DNA's development.
7. DNA's development can be influenced by substances, structures and/or forces belonging to the i-set, and may exhibit qualities belonging to the i-set.
8. Natural selection can be influenced by substances, structures and/or forces belonging to the i-set, and may exhibit qualities belonging to the i-set.
9. The most tangible evidence of valid irrational insight among humans is manifest in religion, art and literature.
10. The religious, artistic and/or literary products of valid interaction with the i-set resonate with large numbers of people over many centuries and across vast cultural boundaries.
All right; if you can accept or at least suspend rejection of most of those postulates, then crack a cold one, sit back comfortably, and let's get on with it.
Computer science professor (and valued personal acquaintance) Dave Touretzky wrote to remind me that "One can add random and even irrational elements to a computer program. In fact, genetic programming works by random mutation followed by selection."
Ah, but when I talk about irrational insight in humans, I mean something completely different from mathematically randomized behavior in artificial devices. Irrational insight refers to our access to a system (the i-set) which can't be expressed rationally -- which means that it's impossible to know the characteristics and limitations of random elements possible within that system.
And while we may apply 'random selection' (which is not the same as natural selection) to AI devices, we have to recognize that evolution in the real world isn't random. DNA develops as "an 'iterated function system' (which maps X -> X repeatedly, and from which, we hope, interesting long term behavior emerges)," in the words of reader and math researcher Richard Clegg.
Or in my own words: fractal-wise, the output of the last function (or generation in this case) is the input of the next. (Think of this as a handy metaphor, not as a rigorous description.)
Natural selection brings to each generation some input beyond the potential contained in the original 'seed', or produced by the natural action of DNA's concatenated mapping. Evolution sees to it that natural, outside (note 'outside', not 'random') inputs advantageous within the context of an organism's environment continue, and disadvantageous ones die out. So evolution (natural selection) keeps the pseudo-fractal (DNA) developing along lines beneficial to a species in view of its environment.
Evolution influences DNA's development -- that is, it gives DNA a context (so it can work in diverse environments) -- but I say that it's not the 'prime mover' of speciation as Darwin groupies seem to believe.
We've got to bear in mind the difference between 'natural' with its inherent limitations, and 'random', with its (probably) incompatible limitations, to understand why random inputs into an AI can't recapitulate natural selection, except by accident. An AI admits only of 'artificial selection'. Sure, you can make the randomization slick, but you can't make it natural unless you can first define and describe comprehensively what natural selection is and is not. Given Postulate Seven, I say 'good luck'.
Irrational, you say?
I took a lot of heat for asserting that the universe is irrational. It would have been better to say that it admits of irrationality to an unknown degree.
Now, 'irrational' doesn't mean crazy or imaginary, or even random. It simply means 'inaccessible by ratiocination', like the square root of seventeen -- used as an example in the Theaeteatus. There's nothing random about it; it's perfectly real; we just can't express it as a ratio. So we give it a label and move on.
Many things that are irrational are eminently real, and eminently delightful. Perhaps the i-set is immensely larger than the r-set (but of course there's no rational way of determining that). What makes human thought impossible to simulate is the fact that, as I claim, we have valid insight into the i-set -- insight which manifests itself recognizably and universally, in religion, art and literature.
I say the i-set is both real and not essentially random, though it admits of randomness as the r-set does. The i-set is only a normal portion of the universe which can't be grasped through ratiocination. Remember, the Greek 'logos' means 'ratio' or 'meaningful speech' (interesting that to the Greeks the two should be so closely related). 'Alogos', which uses the alpha privative, means not the opposite of rational, but simply 'lacking in rationality' or 'other than rational' (other than 'logos').
By 'insight' I mean a very real (if personal and unpredictable) window into, and experience of, a very real area of the universe, and not just some spooky gibberish.
One may devise a machine which spouts spooky gibberish like some fraud swami or televangelist, but its pronouncements will deceive only a few. The i-set has consistencies and rules -- ignore them, and you'll eventually be exposed, like Nostradamus. The products of valid interaction with the i-set resonate large.
I see no hope for building intelligent machines capable of higher human mental functions, because we as a species can never accurately describe what they'd need to behave that way. We would first have to rationalize irrational insight, which is impossible. I suppose we could stumble on it by random selection and trial and error; but would we even recognize it if we did succeed that way?
Sure, we can fabricate an artificial intelligence, but I say only natural selection (again, not 'random' selection) and biological engineering can produce a non-human intelligence equal to or superior to ours -- that is, capable of insight into the i-set, and maybe far better at computation to boot. But that's outside our discussion.
Surely, at a minimum, we will one day create machines so well routinized that a person might not realize they were communicating with one in a Q&A, like circus animals which can be trained to ape human characteristics.
Such a machine could even be programmed to pause appropriately when asked a profound question. But a clever person will always be able to expose it by choosing questions you have to be alive to grasp: e.g., "suggest a solution to the staging challenges in a live production of 'My Dinner with Andre'", or, "describe the last time you acted spontaneously on your spiritual convictions, in a manner contrary to your temporal interests, and tell me what insights into Abraham's action in Genesis 22 it gave you."
With the benefit of bio engineering and natural selection, AI may cross the line into real human intelligence at some point, but then it's moving into duplication rather than simulation, and going beyond the issues we're discussing here.
Yes, we might 'culture' an AI like a batch of alcohol-resistant yeast or 'farm' it like a field of GM soybeans, or 'breed' it like a prize pig. The result of a mechanical approach will be forever inferior, but moving into bio engineering could easily narrow the gap.
Of course if we take this to a sci-fi extreme, we can endow AI 'hosts' with biological properties, and ourselves with mechanical AI properties, as Hawking seems to advocate. So far, it looks to me like an engineering challenge which we can eventually overcome (though I may be mistaken here). If we should eventually succeed, over time the distinctions between 'us' and 'them' would attenuate, and one day our descendants will end up arguing natural rights with our creations as science fiction imagines.
To go down that path strikes me as one of the top five most unwise choices that we as a species could ever make, which is why Hawking got such a belly-full of bile in my previous article.
He deserved a severe rebuke for saying what he said. But if he actually believes it, then the little shit deserves to be hanged. ®
Sponsored: Global DDoS threat landscape report