Artificial Intelligence (AI)

One thing that's become very prevalent nowadays is applications that use some sort of AI, or artificial intelligence.

What artificial intelligence means

The first thing to understand is that the term artificial intelligence might mean something a little different than what you think it means. There are several closely related terms and it can be a little confusing. For instance, there is the term artificial intelligence, machine learning and deep learning, which all mean a little different thing, except that deep learning is a type of machine learning and machine learning is a type of artificial intelligence. So let's take a little deeper dive into some of these terms.

The term artificial intelligence is a little more general than what some people think it is. It really just means that computers are performing tasks that are typically associated with human intelligence. For instance, if you're playing a video game and you're up against a non-player character, or an NPC, you're playing against artificial intelligence. That machine, that computer, that NPC is acting like an artificial intelligence, acting like a human would behave.

Another good example of this is chess. We've been programming computers to play chess for a long time now. That's a form of artificial intelligence, and to a certain degree a very basic form of artificial intelligence compared to what we have nowadays, but still a form of artificial intelligence.

State space search

There are a lot of different techniques to implement artificial intelligence. Some of these are more basic while others get quite a bit more complex. One example of artificial intelligence is state space search. The idea behind this is that there's a certain state, and what you have is several options, and each one of these options leads to a result. So option one leads to more options. Option A out of here leads to several more options. If we choose option two, then that is the end of it. If we choose option three, then we lead to a couple more options. Whatever our desired state is in the end, what is the best route to that state? The computer is going to choose whatever the best route is to that state.

A good example of this is using chess. Let's say the chess board has a certain state, which gives us options of what we can choose for our next move. Some of these options could lead to wins out here. And so we're going to determine which option it is that's going to lead to the greatest chance that we're going to get the amount of wins. We're going to choose that option that leads to the greatest amount of wins. That's what a computer is doing there. It's making that assessment, which can be difficult, because in chess the amount of moves that can happen on the board is 10 to the 120th power. That would be 120 zeros after that. So that is a huge number.

Logic

Another implementation is using logic. We have some sort of state, a decision that needs to be made, and then based off of that decision or that criteria that we create, we are going to make different choices. We can set up logic so that it behaves like a human.

The problem with something like logic is that it's a very basic form of artificial intelligence, and it could be very time-intensive to set up all of that logic to make a determination of what the computer is supposed to do. So that's where machine learning can come into play and might be a better solution. The machine can learn and grow, making decisions for itself.

Machine learning

Instead of using logic, we can set up computers to actually learn something new, and we call this machine learning. Let's use an example of a maze to compare logic versus machine learning.

Maybe we have a simple rule that just says follow a wall. Use the left-hand side of the wall and follow it till you complete the maze. Here's where we're going to come into the maze, and here's where we're going to go out. What a machine would do is use that logic and say, "Okay, I'm going to follow the left-hand wall. I'm going to take a left here," and then it's going to discover it's a dead end, come back out, and then it's going to follow this to make it to a left-hand turn again, and it's going to keep doing this to make it to the end. If you were to run through this exercise again, it would take the same exact path, because it hasn't learned anything new. It's just using that same logic to make it to the end. So that's an example of logic right there.

Let's contrast that with machine learning. Now we have a machine that's going to learn from its mistakes. We're going to send this machine into the maze and we're going to give it a little bit of logic and say, always choose to take a left-hand turn until you know differently. So it's going to go into the maze. It has a decision point right here. It's going to say, "Oh, well, I'm going to take a left right here." And it discovers that this is a dead end. So now it's got to reset here. It's going to go back into the maze. It's going to say, "Oh, well, I know last time that I took a left and that was wrong, so this time I'm going to go straight." It comes to this point right here. It is choosing between taking a left and going straight. We said it's always going to choose left, so it's going to take a left. It gets to this decision right here. It can go straight or take a right. It's going to prefer anything to the left, and so that's going to be straight in this case. And then it gets to this point right here. It makes another decision. I'm going to go left, and then we've got a dead end here. So now it's got to reset again, and it's going to go through this process, make it all the way to where it made the wrong mistake. And in this case right here, it will go this direction and it has a decision right here. Again, it's going to take a left, and then it's successful. Now it knows a quicker route to go through this maze than if it were to just follow a wall the same way every time to get to the end. Now it knows a quicker route to get out.

So with artificial intelligence, the machine might not grow over a period of time. It's just mimicking human behavior. Whereas with machine learning, it can actually learn from its mistakes. But there are some efficiencies in certain types of machine learning.

Deep learning

One type of machine learning is deep learning. Let's compare and contrast machine learning to what deep learning is. Deep learning is a type of machine learning. But in the example of machine learning that I gave, the machine keeps trying different routes until it finds a route to get out, and it would have to keep trying everything in order to understand whether this is the fastest route to get out or not. Let's say we present it with a different maze. Now, essentially, that machine learning example that I gave has to start from ground zero and learn a completely new maze before it knows how to get out.

Let's compare and contrast that to deep learning. I'm going to compare and contrast it to my brain. When I look at this maze, I can see the fastest route to get out. Even if I've never encountered this before, I don't have to go down every single route and say, "Nope, that's not it. Nope, that's not it." I just look at it and know. And how do I know that? I recognize certain patterns, like this is a dead end, and this is a dead end, and this is a dead end. So I recognize this, and there's some sort of recognition. What are the details of how my brain is functioning and working to detect that? There's a bit of obscurity to that. There's a bit of it that we can't completely explain. However, we do know that we can quickly discover what the route out is. That's an example of deep learning: that over time I programmed my brain to be able to look at a maze like this and discover the fast way out.

Neural networks

Deep learning really has a lot of roots in neural networks, and deep learning emulates how your brain functions. Your brain functions off of neurons, and that's why it's called a neural network. It's a network of neurons in your brain. What we've done is we've created artificial neural networks, and this is what artificial neural networks look like.

Artificial neural networks have some sort of input. This only has five inputs in here just because I am limited on the space here, but you could have as simple as a couple of inputs on here, or you could have many. And then we have some sort of outputs. Once again, we could have just one output here, or we could have many different outputs. And then we have everything in between here. These emulate, kind of represent, what neurons would be through this whole process. The more I have in each one of these rows, and the number of rows that I have, make a more complex neural network. So this is an example of kind of how your brain functions, but also how artificial neural networks work.

Our brains are very complex. In fact, it's said that we probably have around 86 billion neurons, where dogs have about 530 million. So this is a staggering difference. What happens is that it takes us a lot longer to program, or learn to program, our brains and to learn, versus dogs — it's a much shorter period of time. However, the capabilities of our brain are so much more than the capabilities of a dog's brain. We can use walking as an example of this. If we take an average human, we're walking around the 52-week mark, versus dogs, it's around the 2 to 4 week old mark. So they are walking a lot sooner, but they are much more limited in how much their brain can process, versus we take a lot longer to learn. However, we can do much more complex mathematical equations, and learning all sorts of new things and discovering things of the world. So our brain is so much more complex.

It's the same thing with artificial neural networks. If we have more and more neurons here in the middle, we can do more and more complex operations, although it takes a lot longer for computers to learn the more neural networks you place into it.

How learning happens

So how does learning happen with neural networks? Essentially, what we're going to do is give it a bunch of inputs on this side, it's going to get some sort of outputs on this side, and then it's going to be rewarded or punished for whatever the outcomes are. Maybe we're feeding it text on this side, or maybe it's going to be a picture. I'm going to use an example of a picture here in a second. And then it has some sort of output saying, okay, what is this? Maybe this is a dog, or maybe this is a cat, or maybe some sort of determination. And then we've got to train it to say how accurate it came out on the other side.

Now, how do we have inputs on one side and then some sort of output on the other side? And how do we determine, if we put in a picture on this side, how we get it to say this is a dog on the other side? The answer is that I'm not going to cover that here, because that one is going to require quite a bit more time and is not going to be the depth of this. But it just has to deal with mathematical equations and the correlation between all of these neurons, because what these are is a network. These are a network here, and depending on what it is that you have, it's going to put different emphasis on the network side of this. So the networks in between here, and the neuron side of this, which is all the in-betweens here. It really has to deal with just a lot of random testing with all this out, and then it discovers, when it creates this pattern within here, it can now recognize what's a dog, what's a cat, and so on and so forth.

Let's take a look and see if we were programming this neural network for cars. We would need a ton more neurons to do this, so this is just a really basic example of what it would look like. In fact, as far as the input perspective is concerned, as many pixels as we have right here, we would want that number of input neurons here. So we want a ton more input. On the output side, it could just be as simple as one neuron. It lights up if it is a car, or it doesn't light up. Or in this example, maybe we have a dog and a cat, maybe it's recognizing several things out of the picture here, but let's just use a car as an example.

We would pump it full of lots of pictures of cars, and every time it got the car correct, then we'd give it a thumbs up: yeah, you did that correctly. That would be the training aspect of this. Somewhere in between here, it's going to start recognizing basic shapes. Maybe it recognizes that there is a triangle here in the back windshield, and then there's a rectangle right here, and the whole car is a rectangle, and we've got circles for wheels. So it's recognizing these basic shapes here. And then somehow those basic shapes are put together to recognize certain parts of the car. It has a tire, and maybe it has glass. So there are certain parts within the neural network that kind of distinguish certain aspects, certain characteristics of this picture.

Once we have it trained, then what we can do is start analyzing pictures. We show it a picture, and then if the car lights up, then we know that it's a car. So now it's starting to train to recognize images, or things within that image.

Natural language processing and large language models

A couple more terms that you may see out there are natural language processing, or NLP, and large language models, LLM. These are a couple of terms that you'll see quite a bit out there. It just means that we've trained this neuron pathway. We train this model to words. So here, maybe we're doing some documents that we're pumping into here, letters and bulleted lists and communication. We're pumping it all into here. We're training this neural network to actually speak. And now we can talk with this, because it's been trained for this large language model. It just understands language and really can interpret language — any language we wanted to teach it: English, Spanish, Italian, German, even computer language, then it can learn it. So that's one of the reasons why one of the big applications to this is generating code.

Generative AI

One thing we've trained AI to do is generate new things. You might see this term generative, but what does generative mean? It just means that we can create something that's new and unique based off of really just programming this. So what happens on one side is we feed it a bunch of images of cars. This is what cars look like, and it starts looking at what are all the characteristics that it sees of what a car is, and based off of those characteristics, now it can understand and interpret what a car really looks like and how to define what a car is. And from there, then what it can do is create an example of what a car could look like, even though it's never seen this exact picture of this car before.

You can also associate generative AI with content. We're creating content, whether that's text or code or images, audio, music, video, maybe even some sort of actions like going through the maze, 3D modeling. So whatever it is that we're creating, we can create it in generative new and interesting ways, new and unique ways of doing this.

Another term that you might see is generative pre-trained transformer, or GPT. GPT is a common term that's out there. Generative is just what we've been talking about — it can generate something new and unique based off of how it was trained. And pre-trained means that somebody has gone through the process of figuring out how to train the computer to do this ahead of time. So now what happens is you don't have to train the computer with AI to create something new. There are different systems out there that are already pre-trained to give us new and unique content.

The hype cycle

Artificial intelligence and machine learning is crazy, and it is life-altering, in that this is going to have a disruption in technology in our lives. It's really going to change things, and there's going to be some drastic changes. But I'm going to put a little caveat out there. With any technology, there is a bit of a letdown between what the technology can actually perform and what it can actually do. And the Gartner hype cycle drives home this point.

There's some sort of technology trigger. In this case, it's AI. AI has been around for a long time, but now it's in the hands of the masses — most of us have access to some sort of artificial intelligence and can create things using AI. The other thing too is that not only is it in the hands of all of us, but it has more capabilities than it ever did before. This is amazing. But what it's generated is a ton of inflated expectations of how it's going to help us out and how life-altering this is.

At some point in time, we're going to be deflated in here. In fact, there's what's called the trough of disillusionment. That just means that we start realizing, oh no, this technology actually has a lot of downfalls and a lot of problems, and it's not going to solve all of our issues, and everything that we thought it was going to deliver, it's not really actually delivering. So we're going to go through this kind of downslope with this and with our sense of what AI can actually deliver.

Then there's the slope of enlightenment. This is where we start actually realizing, no, it actually is really helpful. It is life-changing, but it's not going to be this drastic everything-is-immediately-going-to-be-different and everything-is-changed. So we are going to have this slope of enlightenment, and then we go to the plateau of productivity. Now, the plateau of productivity is going to look a little different, I think, with AI. We're going to see it continue to rise and be a part of our lives and how we operate. So it really is a big drastic change from before. But I just want you to understand that a lot of people have these hyped up expectations, and historically speaking there's always been a letdown between what our expectations are versus what technology can actually deliver.

Downsides and concerns

There are a lot of downsides and warnings around what AI can do. It can have a lot of bias and prejudice involved with AI, and if we are living by what it's telling us, it gets things wrong. There's something called AI hallucinations, where it just gives you wrong information, and it gives you kind of strange behaviors and stuff. That will get better with time, but it'll probably never fully go away for quite a while.

There's lack of transparency: where is it getting this information from? There are implementation costs with this. There are concerns around job displacement, of what jobs are going to go away with this. There are concerns around privacy violation, that it's collecting data — when you do searches, it's collecting data on that. So there are lots of concerns that we have when it comes to artificial intelligence and how it's going to be implemented.