Peter HarrisonThere are some common objections to the possibility of achieving general intelligence in machines. In this article I try to address them.
There is no way to predict the future.
It was predicted that we were close to human intelligence in the 1960's and 1980's, but they were wrong. This time will be no different.
In the early days of electronic computing the raw performance of computing devices and their capabilities in terms of making logical decisions looked like they might be capable of the same kind of robust deductive logic that humans are. It turned out that while machines can outperform humans in terms of raw compute what wasn't appreciated was the massive parallelism of the brain that made human perception and abstract thought possible.
Over the last fifty years our understanding of the human brain has greatly increased. We have far better understanding of how it functions. Machine learning systems have taken these discoveries and used them to inspire digital systems which utilize the important principles of connectivity and reinforcement learning. While there was limited success in the 1980's with neural networks they didn't really get out of academia. Today neural networks and machine learning are central to the biggest companies in the world, with neural networks used today in a wide variety of real world applications.
As James Burke demonstrated in his television program Connections, while we cannot possibly predict the precise trajectory technological development might take the one constant has been relentless increase in the rate of change. Essentially every new discovery opens up a new world of other possibilities.
To say we will never achieve artificial general intelligence is like someone in 1890 saying powered flight will never be achieved. It would have been impossible to say in 1890 exactly when powered flight would be possible, but it would take no genius to see the efforts being made toward powered flight would bear fruit.
Similarly when we look at the success of modern neural networks it is difficult to see how continued substantial investment and a determined team of experts will not make substantial progress in the relatively near term.
We will not build machines that can simulate the human brain in our lifetimes.
We can't even simulate a simple worm with 600 neurons yet, much less a human brain with billions of neurons and trillions of connections.
Human technology cannot yet produce the silky smooth, strong complex structures of feathers. Can we therefore conclude human flight will never be achieved? No. The idea that being unable to simulate or replicate something natural in every detail prevents us from being able to use insights from nature to design and build our own technology is clearly false. We used wood and later light metals like aluminium to build artificial wings. We utilized a most unnatural form of propulsion, the internal combustion engine, to achieve powered flight.
While today's commercial neural network technology has certainly been inspired by natural systems they are in no way attempting to simulate natural systems in detail. The Open Worm project is a rare exception which actually is attempting to replicate a natural system in order to better understand natural neurons. They seek to study natural biology further through detailed simulation to see if we can achieve further insights.
But such an approach is far from the only one. Artificial intelligence researchers have been busy developing algorithms designed to run on massive parallel computers to train neural networks. These algorithms, like the internal combustion engine, are not natural. They do not represent anything found in nature. They are an engineered solution to the reinforcement learning problem. While inspired by natural brains the form of modern software and hardware for artificial intelligence is not trying to simulate natural biology in detail.
The reality is that neural networks are becoming more generally intelligent and able to handle a wider variety of problems. Engineers do not suffer from the same limits as nature. Unlike natural systems many constraints on biological systems simply do not apply to engineered solutions. Could you imagine giving birth to something the size of a Google data center? The systems we have right now are able to understand human speech, recognize faces and expressions, translate between languages, drive cars and any number of other tasks given the right training data. No worm could do these things.
Even if machines replace jobs it will only be the boring mundane jobs.
A new generation of jobs will be created that are more inventive and creative. The process of replacing mundane jobs with machines and automation is the story of the industrial revolution, and certainly nothing new.
Machines previously replaced human physical labour. Artificial Intelligence is replacing human intellectual labour. We are just starting to explore the range of potential applications that the current crop of machine intelligence systems provides. There is a clear and present danger to peoples jobs even if we do not make further major advances in artificial intelligence. Drivers of all kinds will be replaced. Cargo ships will transport goods with no human crew. Planes will fly across the oceans autonomously. They already do.
But lets go one step further into the future, where machine intelligence is broadly better than humans at every intellectual task. Imagine a health care system that will diagnose your illness better than any human possibly could. Imagine transportation systems that gets you where you want to go, when you want to get there with virtually no risk or effort. Imagine getting to the point where having a human do a job by definition would be a second class job.
There would of course be people who would choose to interact with humans, even if those humans performed demonstrably worse than artificial intelligence. But such resistance would be short lived. How many people yearn for the days of visiting the bank to withdraw money? Who would prefer to slap a stamp on an envelope and take a trip to the post box to send a letter to your mum? Most of your local posties are out of the job, and nobody cares.
Machines will not just be better at the mundane jobs, they will be better at almost everything. No profession will escape. The lawyers, accountants, doctors, surgeons, teachers, shop assistants, engineers, scientists and my profession, software developers. Some might survive on the metaphorical high ground for longer, but eventually all will submerge under the AI tide.
How can I be so certain you might ask? We can conclude this by simple observation of our current economy. Corporations will continue to automate and reduce labour costs if they are at all able to. And so as machine intelligence is developed which exceeds the abilities of humans machines will replace them. There is huge investment into artificial intelligence exactly because it is an arms race that will be terminal for the losers. Companies that fail to invest will be outperformed by those who do. Natural selection. Law of the jungle. Once we have machines which have general intelligence and are able to learn much like a human does, only at a fraction of the cost, it is all over for human employability.
The idea that the whole population could migrate into being creatives fails to appreciate that inter-human commerce will not be input into the machine economy. We will need everything from the machines, but they will need nothing from us. The coupling between human labour, production and consumption breaks down, and thus the positive reinforcing economic cycle is broken.
Machines will never be able to challenge human intellect in principle.
There is something unique about humans that is impossible for digital machines to replicate. Machines have no soul, no mind.
Your brain is everything you are. The poetic idea that the mind is something separate from the brain is proven false by everything we know about the human brain. It is possible to alter our emotional state with chemicals. Specific capabilities are harmed when certain parts of the brain are damaged. We have studied the brain in depth and know broadly how it functions. We are able to image a working brain using fMRI, We understand it so well that we have used the principles learned to create neural networks which replicate capabilities which until now have been the unique reserve of humans.
Roger Penrose proposed in his book The Emperor's New Mind that human brains utilized quantum mechanics where the massive parallel nature of quantum waveforms plays a critical part of human consciousness. Digital computers which operate based on classical models would be incapable of taking advantage of this quantum aspect, and thus could not be used to simulate or emulate human brains. There is however no evidence at all that anything of the kind is true. The evidence is clear; it is the large scale structure of the neuron connections that give us our mental abilities. Human behaviour emerges from the interaction of one hundred billions neurons acting in a massive parallel network. The brain is certainly a remarkable organ.
It is clear that brains did not evolve simply by blind luck. The octopus for example has a large complex brain, but is does not share our evolutionary history. Our closest common ancestor is the mussel. Humans share more in common with a guppy than an octopus. Clearly large brains are an evolutionary advantage. Just as eyes have evolved independently many times it appears big brains may also be a common outcome in biological systems. It also means that there is nothing unique about human intellect.
Remarkable the human brain may be, but not magical. In many respects digital technology has already vastly outperformed it. The interactions in neurons are slow, the speed they can fire thousands or even millions of times slower than digital transistors. In contrast to the assertions of Penrose there is no fundamental physical reason we should not be able to replicate the functionality of the brain within a digital system, an artificial brain.
The answer is to improve education.
We need more focus on science, technology, engineering and mathematics. We need to focus on subjects that prepare our young people for a new world where machines will be used to automate what is currently human labour.
Some believe that the answer to unemployment caused by automation will be to provide retraining so that people can move into occupations more aligned with this new automated world. There will of course be a period of increased demand for engineers, software professionals and machine intelligence specialists. However, why expect people who had been unable to train for more advanced roles previously to be capable of retraining to fulfill advanced technical rolls tomorrow? The reality is that there is a spread in terms of intelligence and those most vulnerable to automation will probably be the ones least able to retrain into roles that require advanced skills.
Okay, but what about our kids, isn't a focus on science and technology going to help them adapt and have a promising career? Well, yes, in the short term. Being prepared for a technological future is better than not. But clearly only a subset of people are really cut out for highly technical rolls. And that is okay; we should not expect everyone to be engineers and software developers.
But what does the world look like once artificial intelligence outperforms everyone at everything? Humans simply lack the ability to evolve or advance as fast as machines. It is like the tortoise and the hare, only the hare will not sleep before the finish line. While I might recommend education for the best individual outcome it is no answer to the bigger question of how to establish a sustainable human society into the distant future.
There is also a case that our higher education systems are not doing a very good job of preparing young adults for the workforce, and that the huge loans required are putting them in unsustainable debt while real wages decrease. There are very real short term questions about whether education, at least in the current form, is the way forward.
Artificial Intelligence taking over is the stuff of science fiction.
Science fiction if full of mechanical monsters. Whether we are talking about iRobot or Terminator they often framed as malignant and hell bent on destroying humanity. Isn't this just science fiction?
Regardless of the level of machine intelligence it will still be incapable of self replication in the immediate future. Technology and machines depend critically on human civilization. To destroy human civilization would be self defeating.
Far more probable would be an extension of what we already see; large multinational corporations developing and utilizing artificial intelligence to make their businesses more profitable and compete with other businesses. In this picture the competitors to artificial intelligence will be other artificial intelligence, not humans. And how will artificial intelligence make more profit? By reducing costs and selling more. And what is one of the largest costs for a company? People of course.
Another opportunity for machines is in the military sphere. Having autonomous weapons capable of replacing humans on the battlefield will reduce the human cost of war, at least for one side of a conflict. Multiple countries are now developing autonomous combat systems. These systems already outperform humans. Artificial Intelligence in warfare will fundamentally change its nature.
Only nations will be able to afford the investment into these kinds of technology. Semi autonomous systems have already shown how effective they can be, with remote humans making the decisions.
Artificial Intelligence could also be deployed by the spooks, such as the CIA, NSA and GCHQ, who have been exposed intercepting and storing our private communications. Where previously monitoring the phone conversations and mail of a target was time intensive and expensive, technology is allowing broad wholesale collection of data and the ability to transcribe and analyse the tidal wave of data. This is only possible through the broad application of artificial intelligence.
Rather than fear malignant monsters destroying human civilization we should concern ourselves with how it may be put to the task of protecting the state; protecting it from dangerous activists. We have seen how Artificial Intelligence is already being misused to misinform, distract and disrupt social movements. It is being used to protect the elite, who are in control of artificial intelligence. And so the threat we face is not malignant machines cutting us down like a harvester harvests wheat, but rather the use of machine intelligence to exploit and control people. And we need not wait for this; it is real. It is here now.
Machines cannot have their own intentions, they are simply tools for human beings.
When you download and run Genepool by Jeffery Ventrella you can see your own little evolutionary garden of simulated beings. It is of course a very simplistic simulation, designed to capture only some of the important aspects of how life evolves to adapt to conditions. As you watch these creatures it is hard not to assign intent to their struggles, to make it to food to survive or to a mate.
Genepool may have predetermined the natural behaviour of these agents to require food and seek out mates, but surely we too are programmed by our DNA? It is difficult to look on one of these creatures and not assign determined intent to seek out food and mates.
More recently machine learning systems have been developed which give all the appearance of intent. They play computer games which require complex sets of moves in order so that the player can advance. They are showing a limited form of intent even in a simple simulated environment.
Just because storage and CPU speeds have followed exponential growth does not mean machine intelligence will be an inevitable consequence. Simply increasing CPU cycles or hard drive capacity will not result in artificial intelligence.
There is a misunderstanding about the exponential argument. It is common to point to examples such as the exponential growth of memory, or increases in processing speed to make the point that technological progress does not follow a linear trend. This is not making the argument that once we reach a certain number of transistors or certain processing speeds we will have something as intelligent as a human.
They are simply being used to illustrate that in the past technology has followed a certain growth trend, and that if artificial intelligence follows the same trend we can expect explosive growth. Critics like to say there is no way to know how far away true artificial general intelligence is, but there is a way to know something; by looking at other similar technology developments and compare it to what we see today.
So what do we see today? We see massive advances in artificial intelligence used in voice recognition, translation, face recognition, driving, and robotics. It is nothing less than explosive growth. Trends in the short term however are never smooth. We might actually see a period of maturation of new technologies in the immediate future. But behind the scenes science and technology continues to make progress. An example is the new line of chips from top integrated chip makers that focus on convergent deep neural networks.
Conclusion
Google is already trying to make self improving systems, and just as they proved with Alpha Go these self learning systems can get better than humans in a pretty short space of time. There is still an open question about how far away artificial general intelligence actually is. While the broad consensus is that it will arrive within three decades my own view is that it will be well within a decade. It would not shock me if the Wright Brothers moment for artificial intelligence came tomorrow. Are we going to wait for that moment before we start thinking about how we will adapt or stick our head in the sand and hope that moment never comes?
