The Shape of the Future — Episode I: Artificial Intelligence
The world is developing technology at an increasing speed, and it can be hard to keep up with these developments. So when I decided to start a series of technology articles, I wanted to simplify the current trends to allow as many people as possible to understand them. Welcome to The Shape of the Future, where we will look at the technologies shaping our tomorrows through the lens of today’s world.
The term “artificial intelligence” is probably one of the most used and generally most misunderstood terms in our digitized world nowadays. So I picked it as a natural start for my blog. I’ll explain a little about what artificial intelligence, or AI, actually is and what it is not.
The first thing you need to understand is that AIs are nothing like what Hollywood depicts them. An artificial intelligence is a program, an assembly of code that is generally created to display the ability to solve problems and learn, which is how those who develop them define “intelligence”.
The lines here get blurry, as computers get more and more capable of doing things and there is a widespread discussion about which functions today actually are part of AI and which are only sophisticated routines by computers.
Let us not complicate things here, though, and begin with the basic understanding of what an AI actually can be. There are two types of AI as defined by programmers. One is a cognitive AI, also called AGI (Artificial General Intelligence), which is mostly associated with neural networks, and this one describes an intelligence that is not dissimilar to the ones of its creators. We talk about an intelligence that is able to learn freely and develop its own capabilities.
The second type of AI is what is generally referred to as Machine Learning, which means the program called the AI is trained to fulfill specific tasks and is enabled to understand data provided, analyze it and come to conclusions that allow it to make decisions.
There is a joke in the tech world saying Machine Learning is written in Python, but true AIs are written in PowerPoint.
That means basically that the main difference between an AGI and a Machine Learning AI is that the first does not exist yet and the second does and is used all around us.
So let us start by looking at Machine Learning and some examples of what it can do.
Given access to your data, they can analyze what you read online and sort content; through the appearance of certain keywords it can determine if there is a higher probability you are a potential buyer, terrorist or left-wing voter. For example, it can find out if you have friends who have already been identified as potential terrorists, if you frequently look up terms like “bomb”, “jihad” and “explosives”, watch videos which often contain the word “martyr” on forums known to host videos of known terrorist groups. So it comes to the conclusion you should be considered a potential terrorist. Not because it has thought about it. It cannot think in the sense we do. It has been told these criteria are a sign of your beliefs. Who told it? Oh, let me get to that one later.
Another example. A Machine Learning AI registers a certain sound from a machine and it also registers a 10 percent loss in productivity. So it knows from a database there are five reasons for the sound, but combined with a loss in productivity, it comes to the conclusion that the machine is damaged, which is the potential source of the sound and the only thing that could lead to a loss in productivity. So it initiates a process for the machine in question to be replaced by a new one and sent for repairs.
A third example: An AI is trained to communicate with clients. It has learned all the common kinds of questions and can identify them. When you ask when your package will arrive, it knows, because every way to ask this question has been fed to it. “When is my delivery to be expected?” “When is this fucking thing coming?” “Is this package coming before Tuesday?” It understands these questions as it analyzes them based on all questions provided to it that have ever been asked about the delivery time for the package; and once it knows the question, it enters a subroutine, tracks your package, gets the estimated arrival date and answers you. “Your package is expected to arrive Tuesday the 27th of February between 9 am and 11 am.”
The most famous example is probably AlphaGo and Deep Blue, machines trained to play Go and chess, respectively, to such perfection they beat masters of the game at it. Those seemingly superior machines also show the limits of this kind of AI. AlphaGo could not play a game of poker. Ever. It would not even recognize it. Why? It had not been trained for it.
Which is how these kinds of AIs actually come about. They are trained. Thus what is required for them to become fully functional is not only a group of computer scientists and information engineers, but also large pools of data which allows them to experience different kinds of scenarios, so they can be taught how to react to them. The larger the data pool, the more differentiated the response of the program can be to the challenges it faces. Take the customer service example above. Being fed millions of customer requests, an AI can basically learn to respond to any standard request, until it can operate at the same level as a human customer service employee. A human interacting with an AI could get any kind of support from it, it could gain from a human being interacting with another human being.
The places these AIs are trained are often referred to as “farms”, which is quite a fitting term.
So what is the big deal about these kinds of programs? Well, they are rather effective. Computers are by nature multi-tasking. They can run the same routine to analyze something as often as its computing power allows them. So our Customer Service AI, let us call him ED, can not only answer the question of one customer. It can answer simultaneously questions of hundreds, thousands, or maybe millions of customers at the same time, thus not replacing a single customer service agent, but the entire workforce. The computer will never get sloppy, get sick, have a bad day or get tired. The performance will always stay the same as long as its system is running properly.
With increasing computing capabilities, our dear ED can actually fulfill an increasing number of tasks. It can track your package, point out you haven’t paid, send you links, quote the relevant terms and conditions, reimburse you, advise you on similar products or special offers, offer additional services and initiate those once you accept them. It can also vary its answers and grammar to imitate more human-like interaction patterns. In other words, it can do anything like a human to the degree that you will have a hard time even saying if you interacted with a machine.
If you ask it out for a coffee, it can probably not help you, though. It can actually do nothing it was not trained to do.
The second kind of AI is what the term “Artificial Intelligence” originally envisioned. In its purest form, it envisions a computer becoming intelligent in a way that is very similar to human intelligence. A computer is simulating the brain, which is why the underlying technology is often referred to as “neural networks”.
These AIs exist, too. But none of them have reached the level of sophistication where it can actually perform any complicated tasks. Mostly, data scientists are still trying to figure out how to make these things learn. There has been progress over the last few years as computers have become rapidly more capable, but so far no AI has emerged that is “market ready”. The potential of these AIs is immense, though. A self-learning AI would obviously have greater capability to learn than humans, if its computing power is greater than human capacity to process information (which so-called computers probably have). There is a lot of criticism directed at this kind of intelligence nowadays and Hollywood portrays it regularly as a threat to humanity.
So how does it work? Well, a true AI consists of a set of abilities which all need to be there. First, it needs the ability to learn, therefore it needs a set of routines that allow it to store, search, acquire and probably prioritize information. To do so, it must be able to interact with information sources. There have been great advances in AIs understanding text lately. Yet, an AI that can analyze data without being told how is still not yet possible in a reliable way. The main challenges here are context and the perception of reality. An example: Children do not understand sarcasm. The same is true for self-learning computers. The ability to distinguish between fiction, comedy and fact is something a computer, who has no experience in being a human, still is struggling with. It is generally believed a more complete understanding of the human brain and thinking process will help unlock such abilities in artificial intelligences.
Nevertheless, recent developments have made it possible to give AIs a toolset unknown beforehand. Computers can now read, talk, understand speech and interact with humans on a level unknown a few years back.
So this is a quick run-down on Artificial Intelligence, a rough outline that is hopefully more easily accessible than most of the stuff out there.
