The Singularity Is Never Coming
Futurist Ray Kurtzweil believes we’ll hit the singularity by 2045. Stephen Hawking warned that it could mean the end of our species. Elon Musk believes it’s such a threat that he’s invested a billion dollars into mitigating what he sees as the coming apocalypse. Like those who promote Peak Oil or the streamlining of English orthography, proponents of the singularity seem to think that it’s inevitable.
There are many different conceptions of the singularity, yet they all seem to agree that it will involve some sort of runaway computer growth which would rapidly surpass human intelligence, with catastrophic consequences for us.
Although imperfect, the benchmarks for public understanding of the singularity are HAL 9000 in 2001: A Space Odyssey and Skynet in Terminator.
For normal folk and STEM visionaries alike, the science fiction dream of superhuman machine consciousness seems to be as far away as when it was first popularized in the 1960s, which is to say, “any minute now.” Herbert Simon, who is often regarded as the father of artificial intelligence, claimed in 1965 that “machines will be capable, within 20 years, of doing any work a man can do.”
There is no evidence to suggest we should not dismiss similarly grandiose contemporary claims.
When we were told that A.I. had “decoded the Voynich manuscript,” that turned out to be incorrect, and when we were told that Facebook shut down two A.I. programs who were speaking to each other and invented their own language to do so, that turned out to be a well-understood consequence of programming dialogue scripts in a certain way; they certainly weren’t “evidence of the singularity’s arrival.” Every step closer to the singularity is exposed as an over-interpretation of the evidence, wishful thinking, outright lies, or a category error.
How Did The Idea Of The Singularly Gain Such Traction?
The idea of a coming singularity was borne out of a combination of the following:
a persistent anthropomorphism (the inability to relate to anything that’s not in some way human);
fear of the unknown (predictions, however unjustified, give us the illusion of control over the future);
an inability to contextualize available data (as there is only one global computer technology, we have nothing to compare it to);
Moore’s Law (the observation that computing capacity doubles every two years, which has held more or less since 1975, although Moore himself said in 2015 that he sees “Moore’s Law dying here in the next decade or so, but that’s not surprising”);
an extrapolation fallacy (the common mistake of determining the trend of available data points and continuing that trend to infinity);
confirmation bias (adopting much lower standards of evidence for information which confirms one’s theory than for information which contradicts it);
and the propensity of many top I.T. CEOs to find themselves inspired by science fiction (there are many entertaining science fiction movies where computers take over and try to destroy us all, but the general rule with science fiction is that it always tells us much more about the time it was written than about the future).
There is a well-understood trend that involves describing brains in terms of the technology of the time. Years ago, brains were described in terms of hydraulics, and later, clockwork. In the computer age, we describe our cognitive ability using computer metaphors. Unfortunately, this becomes back-to-front reasoning when modeling artificial intelligence. In other words, if you are forced to model Y using X, and the prevailing metaphor you have to conceptualize X is Y, then you’re really modelling Y in terms of Y.
The advent of cloud computing has given some commentators a counterargument to the simplest solution to any computer threat: unplug the machine. Yet the idea that computer code can upload itself to the “cloud” and operate independently of physical machines ignores the basic reality of computing, which is that everything is based on a server somewhere. The “cloud” is nothing more than a terrible metaphor for a series of back-up networks, all of which exist on real, grounded computers. This bears repeating: 100 percent of the cloud is on physical servers, and every single one of those servers is plugged into a wall.
Those who fear a “glitch” which might cause the whole thing to “go live” and take over should comfort themselves with everything they know about computers. To significantly understate it, glitches never increase performance or productivity. The good news is that you can fix most real-world glitches by turning the machine off.
Is this too facile a response, as some claim? Can it really be that our great response to the looming possibility of a humanity-vanquishing singularity event is to unplug everything before that happens?
STEM people—in particular, computer programmers — find this kind of talk extremely frustrating because it’s a low-tech solution, and they spend their lives working on high-tech ones. But I haven’t come across a convincing counterargument to my proposal that we simply unplug or disconnect before the imagined threat gets a chance to seriously affect us.
The fear of a machine takeover is underlined as a nonsensical when one considers how our lives have been immeasurably improved by allowing machines to handle the messy stuff. Among countless other examples, driverless cars are much safer than human-driven cars, fly-by-wire systems enable safer, more efficient flights, and machine monitoring of medical devices has saved countless lives and streamlined health services everywhere.
In 1970, Buckminster Fuller said:
We must do away with the absolutely specious notion that everybody has to earn a living. It is a fact today that one in ten thousand of us can make a technological breakthrough capable of supporting all the rest.
Perhaps, rather than a robotic displacement of the human, technology’s continuous advancement will bring about the conditions for an unprecedented rise in quality of life. For example, if technology ever gets to the point where it has to replace all the “normal” jobs, then perhaps governments will finally be forced to consider Universal Basic Income, and as a result standards of living will improve across the board. Why should we fear this future rather than look forward to it?
Yet before we rule out the singularity, let’s get clearer on the very notion of intelligence.
We might think of intelligence as immutable and measurable: an ability to perform certain benchmark tasks, or a demonstration of capability in various discrete subjects. In reality, human intelligence exists as a messy soup of apprehension, discernment, insight, close-enough solutions, preferences, and passions.
There is no way, even in theory, to code any of that. Computers do exactly what you tell them; human intelligence tends to veer off at crazy angles without warning for reasons even the humans involved cannot explain.
The most popular and enduring method of measuring intelligence in humans is the I.Q. test, such that having a high I.Q. score is generally taken to mean that you are intelligent. However, I.Q. tests are influenced by a number of factorsoutside the ability of the people conducting the tests to account for, such as race, social status, etc., but suggesting that black people or poor people might be less intelligent because of the color of their skin or their income level is self-evidently infra dignitatem.
Despite the insistence of right-wing conservatives, claimed correlations for I.Q. in areas of crime, education, and health make no allowances for common-cause variables and the fairly banal fact that correlation does not even imply causation. In any case, as with all aggregate statistics, even if a causal link could be demonstrated, it would tell us nothing about any particular individual, and it might mean nothing at all.
While the I.Q. test measures technical aptitude in a number of well-defined, narrow areas, it ignores creativity, practical ability, and more nebulous concepts like morality or integrity. It takes a certain kind of intelligence, for instance, for someone to make a table, which might not be reflected in his vocabulary.
The same person can get wildly varying results from an I.Q. test over relatively short periods of time, for no apparent reason. If your individual result depends on your mood, or the person conducting or programming the tests, then we might not need a test at all.
To sum up, the I.Q. test decides what “intelligence” is without consulting you, and then correlates high scores with positive hits on the thing it just made up. There is a very real chance that doing well in I.Q. tests demonstrates nothing more than your ability to do well in I.Q. tests.
The most popular and enduring method of measuring artificial intelligence is called the Turing Test. The idea is that we will know that a computer is “intelligent” when, through normal conversation, using natural language processing, its responses become indistinguishable from those of a human.
So far, the closest any computer program has come to passing the Turing Test was a fairly unsophisticated response algorithm. Here is a deeply unsophisticated response algorithm:
1. Ask “what is your name?” and store the input as X.
2. If someone types “hello [anything]”, then respond “Hello, X!”
While this targeted response might scare the hell out of someone from the 1920s, the rest of us understand that it’s not very impressive. Yet every single Turing Test candidate has been a more sophisticated version of this, from the brutal psychotherapist simulator of the 1960s, ELIZA, to Sophia, a robot who was recently granted citizenship of Saudi Arabia.
There is no good reason to assume that computers should be intelligent when they successfully imitate human responses to fairly anodyne prompts in controlled environments. Conceptualizing those responses entirely through the medium of discrete packets of human language seems positively arrogant.
There is no way to program contextually-accurate, sloppy imperfection into a computer program. Humans can be outed by something as simple as common spelling errors, but the true failure of the Turing Test is that it’s far easier to convince someone that a human is a computer by mimicking awkward, clunky or overly-accurate responses. After all, developing a human-style intelligence, which is optimized for meat and neurons, would be deeply inefficient for computers: like learning French through Chinese.
If Turing Test programmers spent more time studying linguistics, they would learn that they are using an incorrect model of human language. Language is not a framework of grammar into which you slot vocabulary, although it can seem like that if you’ve ever had to learn a foreign language.
When you were learning your native language, you spent no time learning grammar rules. Although you might find it difficult to describe the grammar rules of your own language, that doesn’t stop you from expressing yourself perfectly fluently on a daily basis.
For instance, ask yourself why “the three small, black dogs” sounds a lot better than “the black, small, three dogs.” There’s a solid grammatical reason for this, but you probably don’t know it and more importantly, if you’re a native speaker, it will never matter.
Language is more like an organized but flexible network of semiotics (the meaning of language), pragmatics (the context of language), phonology (the sound structure of language), morphology (the grammatical changes of language), heuristics (the short-cuts of language), intentionality (the focus of language) and under-determinacy (the unspoken aspect of language), which is constantly evolving to meet the changing needs of people who use it to communicate.
Most of these aspects of language are again impossible to program, even in theory. For example, under-determinacy is everything required to understand sentences but which is never referred to directly. We routinely say things to each other which will only make sense if the hearer already has some information about what we are saying. This includes irony, metaphor, and hyperbole, which are so under-determined and context-dependent that trying to model them without that context is incoherent.
Linguists have barely scratched the surface of language, and if we don’t understand how language works, a comprehensive simulation is impossible. Rather than determine how language actually works and work on simulating that, programmers synthesize the most superficial aspects of the symptoms of language use. Painting red dots on my face does not mean I have measles.
Those who feel that the singularity will represent the dawn of machine consciousness (by which we usually mean a computer’s self-awareness) will face many of the same problems presented by machine intelligence. The leading experts in human consciousness can’t agree on what it is or how it works, with some even taking the view that consciousness doesn’t exist at all.
In 1975, Julian Jaynes (in The Origin of Consciousness in the Breakdown of the Bicameral Mind) promoted the revolutionary idea that consciousness is not a personal interior awareness we can use to make decisions and judgements, but a modeling system for the behavior and actions of others: an essential part of the brain machinery of a social species. Although Jaynes admitted that we have no access to the information needed to prove this theory, recent research could support his thesis, as summarized by Chris Paley in 2014’s Unthink.
Paley argues that in order to function socially, we need a model of ourselves that roughly responds to how others are modeling us. Or, as Canadian poet Thomas Cooley put it more confusingly in 1992: “I am not who you think I am; I am not who I think I am; I am who I think you think I am.”
As we are nowhere near defining or understanding what “consciousness” is, the idea that we could replicate that almost total lack of understanding in machines (a fortiori that it could magically appear on its own) is naive.
While simulating intelligence or consciousness in machines is going to be more or less impossible if we can’t determine the nature or function of intelligence or consciousness with any reliability, this does not mean that A.I. is useless, or that it won’t improve.
For the vast majority of intents and purposes, machine intelligence will be limited to specific functions within very limited parameters, such as a thermostat. A thermostat “knows” to click off at 27 degrees and click back on again at 22 degrees, and in the future, any improvements in thermostat A.I. will be limited to working out quicker and more efficient ways to regulate temperature. This is the future of A.I. in general: responding to existing problems with more efficient solutions.
The real danger will not be that A.I. “goes out of control” and tries to kill us all, but that humans will intentionally try to exploit the normal (correct) functioning of A.I. for nefarious purposes.
For instance, last September, Mark Zuckerberg admitted that Russian operatives had manipulated Facebook’s newsfeed algorithms to influence the American election. The source of the threat in this case was clearly the Russian hackers, and not the Facebook A.I., which is optimized to maximise engagement, ultimately to sell advertising.
There will always be a dynamic relationship between people with bad intentions who are willing and able to abuse technology and those who find solutions to those threats. An awareness of this relationship cannot substitute for a cost/benefit analysis of any particular technology.
As a species, we are facing some genuine existential threats, such as climate change, finite resource exhaustion, non-sustainable agricultural practices, pharmaceutical-resistant diseases, and so on. We’ve made it through ice ages and sabre-toothed cats and wars and famines and plagues, and, as long as someone’s standing near the plug, we’ll make it through whatever happens with A.I.
Original version published on Arc Digital, 18 April 2018
