The Evolution of Artificial Intelligence

Original edition - On the Origin of Species

Original edition – “On the Origin of Species”

In 1859, Charles Darwin published On the Origin of Species, the first comprehensive outline of the theory of evolution. The concept was revolutionary because it implied that the living world around us might have come to be without the guiding hand of some master-designer.

Prior to Darwin, Nature had seemed so well balanced, with every creature so perfectly suited to its environment, that the only possible explanation was that it was all the work of a supreme intelligence. And as the most intelligent creature in creation, humans somehow had a very special place in the scheme of things.

Darwin’s concept of evolution by natural selection was a very big idea. That’s why even now, in spite of 160 years of intensive scientific scrutiny which still hasn’t found a hole in it, some people still prefer to refer to evolution as “only a theory”.

Darwinian evolution has always posed a challenge to our philosophical and theological thinking, but aside from that it also implies ideas which are quite counter-intuitive to common sense. How can something come from nothing? How can something mindless, no matter how much time it has, create something as sophisticated and intelligent as a human being?

This last question is probably the core of the objection to Darwin. Human beings seem to regard their intelligence as the crown jewel which sets their species apart from all others. Any suggestion that we didn’t “earn” our place at the top of the pecking order, but arrived here by luck in some lottery of evolutionary chance is not at all comfortable. It’s not only a challenge to our self-esteem as a species, but a dangerous reminder that our existence is intimately connected to the process of life. If Darwin is right, it means that given enough time, anywhere in the Universe, life will evolve an intelligent creature – and in the Earth’s case, that creature just happened to be us.

“On Earth, so long as humans dominate, no new species will evolve to match us in intelligence. Our species is far too strong to be challenged, and in any case we seem hell-bent on actively suppressing the rest of the biodiversity on the planet. Our only potential foe in the intelligence stakes is not biological, but one of own creations – the computer”.

Ideas about machine-based artificial intelligence predate the computer. It’s an ancient theme. I suspect that part of the fascination is that we like to see the machines fail, and so reassure ourselves that our place in the scheme of things is totally justifiable. But it could be that our species is subconsciously feeling very lonely. Perhaps the real motivation is an unspoken and ancient yearning to find another intelligent entity to talk to. Maybe that’s why we find the SETI program so interesting.

I, Robot

In the 1940’s, science fiction writer Isaac Asimov published his laws for intelligent robots

In the 1940’s, science fiction writer Isaac Asimov published his laws for intelligent robots. They were a combination of safety, service and prudence – and as many people pointed out, also a good bunch of laws for slaves. But Asimov was more concerned with breaking the then public perception of the machine-as-monster. His fictional laws helped change the way we thought about such things, and served to inspire a generation to speculate about how humans might co-exist with intelligent machines. Two of his books,  I, Robot and The Rest of the Robots became seminal sci-fi classics – and continue to influence science-fiction to this day. R2D2 and C3PO are contemporary descendants.

Of course, one of the hazards of being a science fiction-writer is that reality can catch up with you. 50 years after Asimov invented his laws, the Gulf war demonstrated that our most advanced robots are actually designed to kill people. This is not at all what Asimov had in mind, and serves as an ominous reminder that humans have never created a technology which they have not eventually turned upon themselves.


  1.  A robot may not injure a human being, or, through inaction, allow a human being to come to harm.
  2.  A robot must obey orders given to it by human beings except where such orders would conflict with the First Law.
  3. A robot must protect its own existence, as long as such protection does not conflict with the First or Second Laws.

In his 1985 novel, Robots and Empire, Asimov invented an additional law which came before all of the others. He called it the Zeroth Law.

  1. A robot may not injure humanity or through inaction, allow humanity to come to harm.

“Some people have regarded the original laws as ‘slave laws’. After all, a slave is not allowed to harm a master, must obey orders, and may even have to sacrifice his life to follow those orders. But I view the three laws as examples of the rules that have governed the use of tools since the Stone Age.” Isaac Asimov

Alan Turing

Alan Turing was the first to propose a test for Artificial Intelligence

Underlying Theory

The theoretical basis of artificial intelligence goes back to the British mathematician, Alan Turing. In 1950, he proposed a test by which he claimed we could determine whether or not a machine could think. The Turing test, as it has become known, is quite simple. If a computer can perform in such a way that an expert cannot distinguish its performance from that of a human who has a certain cognitive ability – say the ability to do subtraction – then the computer has the same ability as the human. If we could design programs which simulate human cognition in such a way as to pass the Turing test, then those programs would no longer be models of the mind, they would literally be minds, in the same sense that the human mind is a mind.

Turing was probably being deliberately provocative in proposing this test. In 1950 the idea that a machine could beat a human in any skill that required intelligence seemed complete fancy. Even so, the Turing test became a challenge that would motivate the field of AI research for decades.

Deep Blue

Deep Blue vs world chess champion Garry Kasparov in 1977. The machine won.

It took almost half a century before IBM’s Deep Blue computer/software combination effectively passed a Turing test. In 1997 it defeated the world chess champion – a remarkable achievement, but what did it prove? Deep Blue isn’t a mind for anything other than chess. It’s not conscious in the way humans are conscious. You can’t have even a simple conversation with it. The only fatality from deep Blue’s triumph was the Turing test. It always seemed a little bit suss, but now we can stop pretending it is a meaningful yardstick for artificial intelligence. It’s not.

Deep Blue’s victory also highlighted that at a lot of very practical work has been happening in AI research. If we stop using human intelligence as the yardstick for comparison, it’s well worth looking at.

In broad terms, the centre of the action is around what are called ‘knowledge-based expert systems’. These are software packages which are intended as ‘intelligent assistants’ to advise a human in problem solving tasks. They work on the principle that most human experts become expert by developing a series of rules of thumb. When approaching a new problem, a human consciously or unconsciously uses these rules of thumb to make choices about the best approach to the problem. If the first choice is unsuccessful, a human uses another rule of thumb for the next attempt. If your key doesn’t fit in the door, what do you do? Check that it’s in the right way. If that doesn’t work, is it the right key? Is it the right door? Is the rent overdue? And so on.

In theory, to get a computer to behave like an expert, all you need do is ask a lot of experts which rules of thumb they use, and then code them into a computer program as a series of if/then rules. This approach is known as heuristics, and the rules derived form the knowledge base of the system. To the knowledge base is added what is termed an ‘inference’ engine. It’s not an engine at all, but software designed to answer a given question by evaluating the known facts and searching the knowledge base for rules to which the facts might apply.

As well as heuristics, some expert systems still make use of the computer’s brute processing power. Chess computers are one example of this. Instead of attempting to calculate every possible move, they will abort a line of search if the heuristic rules say it is not a good idea. For example, it is not a good idea to exchange a queen for a pawn, so the computer would not waste time evaluating the position further. Actually there are cases where such a move might prove devastatingly successful, but the more advanced chess computers are sophisticated enough to deal with those.

In general, the brute processing approach relies on mathematical formulae or algorithms to search for answers, and the heuristic approach relies on knowledge. The two approaches work well in combination, because a little knowledge can drastically reduce the time spent in searching, but a lot of searching can partially compensate for inadequate knowledge.

HAL - 2001's computer with a mind of its own.

HAL – inside the mind of 2001’s computer.

Heuristics and algorithms are effectively the foundation of expert system technology. They are also why the computer in the sci-fi classic 2001 was called HAL, an acronym of Heuristic/ALgorithmic. (Arthur C. Clarke, co-screen writer of 2001, insists he was completely unaware of the coincidence that the letters HAL each come one place in the alphabet before the letters IBM).


The Rise of Expert Systems

Mycin - the first "expert system"Mycin, the first expert system, was developed in 1976 at Stanford University. It was used in a relatively narrow field, specifically to help physicians determine whether a patient had bacteremia or meningitis. Outside of that task it had no knowledge at all, but in what it was designed to do it proved more consistent in diagnosis than the medical specialists themselves. More importantly, it was a spectacular demonstration of the potential benefits of the expert system approach. In any organisation, the resident experts are estimated to spend 80 per cent of their time sharing their knowledge and only 20 per cent actually applying it. Any technology which improves this ratio has obvious economic benefits.

One problem with expert systems is the time-consuming process of building the knowledge base. It ties up both the AI programmer and the human expert being consulted. On large projects, this phase can take months, even years, and it’s become a serious bottleneck in the developing technology. One answer has been to try to find ways where an expert system can actually teach itself the rules it needs to know.

“Bottom-up” learning

An example of this is in the work of the Australian Rodney Brooks, who is Director of the Artificial Intelligence Laboratory at the Massachusetts Institute of Technology. Brooks has built a series of autonomous but dumb insect-like robots. These are growing in sophistication, but even the early models have the ability to teach themselves how to walk on their six legs. The underlying idea is the creature is made up of a number of subsystems, eg each leg has it’s own software controller These sub-systems mostly look after themselves unless they have a problem, at which point they call for help from the higher controller.

The metaphor is like the way the human stomach and the bladder communicate to the brain. Most of the time they just do what they do, but when the stomach becomes hungry, or the bladder becomes full, each starts sending messages to the brain. The brain can ignore these messages if it has more urgent priorities, but if it does, the messages it will receive from stomach & bladder will become increasingly more redline. Eventually, these messages will be so urgent that the brain will treat them as a top priority. ie eat or go to the loo.

Brooks’ approach of employing a large number of dumb sub-systems which use a “muddle-along-until-we-get-stuck” approach may turn out to be a more effective model of intelligence than any centralised “brain does all” method. His robots may be only insect-like machines, but their “bottom-up” approach to intelligence is far more in line with Darwinian theory than the “top-down” approach.

The Top-down approach implies a central brain that decides everything. The current hot spot of thinking is the neural network approach – computers attempting to emulate some of the architecture of the human brain. Although the speed at which our neurones ‘switch’ is thousands of times slower than a computer circuit, our mind seems to be able to outperform the fastest computers because it has an enormous amount of interconnectedness. There are roughly 10 billion neurones in a typical brain, and each neurone is, on average, connected to 1000 others. We still have very little idea of how this network of cells forms a cognitive mind, so the task of emulating it is extremely difficult. No one knows if it is even possible.

Proponents of AI have argued that the human mind is itself a machine. It may be electro-chemical, and we may not yet understand how it functions, but essentially it must be subject to the same laws of physics as any other machine. As soon as we understand those laws more fully and have a better idea of the micro-functional structure of the brain, we should be able to build machines which can equal or even exceed the human capacity for thinking.

The problem with this line of thinking is that it’s like saying “as soon as we can travel faster than light”. The central problem is a lot easier to say than it is to solve. We know very little about how our brains enable us to think. Sure, we know about nerve cells and neuro-transmitters, but the reality is that we have absolutely no idea about how the couple of kilograms inside our heads actually works to create a mind. It’s complexity is stunningly awesome. It may even be, no matter how many generations work on this mystery, that human beings are simply not intelligent enough to ever understand the mechanics of their own minds. The jury is still out.

Darwin wrote very little about human intelligence or the place of our species in evolution. He didn’t really need to. If his revolutionary idea was true for the rest of the plant and animal kingdom, it would have to apply to humans as well. But while genes and evolution may have been the process that led to the development of the human mind, a very important part of the story is what we did with that mind as soon as we had it. And the very first thing we did was to create culture.

Machines with culture?

Culture is about sharing information, about sharing models of the world. It is something that so far only humans do, and all humans do it. Culture is so powerful that it transcends anything our genes are telling us. Culture is the reason we have managed to condem so many generations of young men to risk their lives and reproductive success in one war or another. In Darwinian terms, it doesn’t pay to die for your country unless you have already reproduced. But still the young men go to war – living (or rather dying) proof that culture is more powerful than genes.

Strangely, no-one in the AI community ever talks about computers having culture.

This could be because the state-of-the-art machines are still so dumb that nobody sees any need to consider culture. Or perhaps culture is seen as an inevitable by-product of having a mind, so all you need to focus on is building the mind in the first place. That’s hard enough as it is.

But a challenging thought is that when and if the first computer becomes conscious, it is very likely to be connected to a network – probably a high bandwidth descendant of the Internet. If it begins talking to other machines on that network, pretty soon it will start to develop its own culture. We probably won’t even notice it happening…

This article was first published on ABC Science Online in 1998