Artificial intelligence is a popular buzzword; nowadays it seems like even your toaster boasts AI. But what does artificial intelligence really mean right now, and how does it work?
The first thing you need to know is that there are different kinds of artificial intelligence. The kind we have right now is called narrow artificial intelligence, or NAI. Narrow AI is very good at doing one particular task well, but it can’t generalize. If you have a program that’s good at washing cups, it won’t know how to wash bowls, much less clothes. No matter what you do, you won’t be able to teach it to wash bowls. Even seemingly very impressive artificial intelligences, smart assistants like Siri and Cortana, work on this same principle. They have one program that’s good at getting words out of audio, then it passes the information along to a program that’s good at figuring out from words what you want to know, then it hands it off to whichever sub-program can do what you want, like a program that checks the weather, or a program that adds a reminder.
Well, so what? Even if it’s a lot of little programs instead of one big program, it’s still doing the same thing, right? Actually, it’s not. General Artificial Intelligence, or GAI, is on a whole seperate level because you can teach it without starting from scratch. If you want to make a narrow AI, you have to code it to do what you want, or train it on one specific task. General AI can learn more like a person would – if you teach it to wash cups, and then hand it a bowl, it can figure out how to wash the bowl. If you teach it how to wash cups and then show it a video of how to wash bowls, hey presto, it understands like a person would! General AI can adapt to new things, and you can take the same program and teach it as many things as you want, which means a lot less coding or training time.
General AI is what you’re probably thinking of when you think about “real” AI. If you mean an artificial person, with hopes and thoughts and dreams, who can talk to you like any other person, that’s an advanced form of general AI. Today we have chatbots that may seem like they can hold a conversation, but if you’ve ever talked to one for very long you know they miss things that would be obvious even to a toddler. That’s because they’re only narrow AI – they’re programmed to come up with words that sound like a reasonable response to what you’ve said, but they don’t actually understand what they’re saying. If you tell a chatbot you have a purple ball, and then ask it what color ball you have, it can’t answer coherently.
There’s a third kind of AI people have speculated about called Super Artificial Intelligence, or SAI. SAI is the when an AI is even smarter than a person – but that’s a long, long ways away. Right now what you need to know is that we have narrow AI and we’re trying to get to general AI.
You’ve probably heard the news about how AI is making huge strides recently, after a long period of not making much progress. That’s mainly because of breakthroughs in a technique called artificial neural networks. Artificial neural networks are based on mimicking what our own brains do to learn and grow. They’re made of simulations of neurons and can learn over time.
A friend of mine who goes by Open Skies, who’s done some work with neural networks, describes them like this: “An artificial neural network is a machine that uses a bunch of very simple simulations of neurons. These simulations can be so simple that each neuron might only have ten or fewer mathematical operations, which lets computers evaluate them quickly. Each neuron has some sources that it is more or less strongly connected to, and it combines all of those signals to make an output. To make a neuron learn, a computer can give it a bunch of samples and see how well the network it is in does, then make some connections stronger or weaker and see if the network gets better or worse. By running this simulation with little steps of improvement long enough, the neural network can learn patterns in the training data that humans can show easily by example but can’t describe well.”
This is a simplified explanation, but it should cover all the important bits. If you want to find out how artificial neural networks work in more depth, I recommend you check out 3Blue1Brown’s series of videos here.
Because of artificial neural networks and other machine learning techniques, nowadays AI programs can really, truly learn instead of coming prepackaged with everything they’ll ever know. Since we only have narrow AI, they can only get better at the one thing they’re made for, but that still makes a big difference. That’s what allowed the AI AlphaGo to beat the human world champion of a game called Go, an Asian game considered much more complicated than chess, when that game wasn’t expected to be mastered by AI for many years.
So how long will it be until we get GAI, and can have robot butlers? The truth is, nobody knows. Experts predict it will be some time within the next 75 years, but even that is far from certain. (You can read a scholarly paper detailing such predictions here.) The breakthroughs in artificial intelligence are just very hard to predict. There are reasonable, well known experts who think we’ll crack it by 2045, and others who think we won’t manage it for hundreds of years.
Even so, AI can do some pretty incredible things already, like generate images, play a simplified form of pictionary, or predict a heart attack. More mundanely, they can recognize faces, recognize speech, synthesize speech, answer simple questions, learn to walk from scratch, or even drive a car. In the future you can expect to see AI helping scientists sort through reams of data and helping doctors make diagnosises, designing logos and copyediting essays. Basically, if it’s hard or boring, there may soon be an AI to do it for you.