AI: How it works(Part 1)

AI is a new hot topic in the modern world, with dozens of different AI's pooping out left and right, such as Chat GPT, Grammarly, Google Gemini, and so on. The AI's I've just listed are all examples of a specific type of AI called "LLM" AIs. How these Large Language Model AIs usually work is when the AI is fed a ton of information, such as the entire English dictionary, or what is trending on TikTok currently.

Despite only receiving digital information, these AIs are still able to generate extremely sophisticated answers to any question someone might ask them. This begs the question,

How do these AIs do that?

Let's start from the basics. There are multiple different types of AI, which you can learn more about in this link. However, we will be focusing on a special type of AI called "Reinforcement Learning AI." Any normal human gets smarter by learning. This type of AI mimics that experience, starting from complete randomness to very precise and accurate through trial and error. Not only that, but this specific type of AI also mimics the human brain through a neural network that is composed of hundreds of thousands of neurons. However, describing an AI is far easier than creating one, as it requires several complicated processes.

How do we make an AI?

We first need to create the neural network of the AI that helps it learn and progress. Every AI needs and input of information in order for it to make decisions, just like a human. In order to do so, all we need to do is set up a system of inputs that the AI might need. For example, in a game about racing, the AI might need some inputs such as its speed, its angle, how close it is to the center of the road, and how long it has been since it has started the race.

These inputs are extremely important, as it is the foundation of the AI.

Next, we have to create thousands of neurons organized into "Layers", which is essentially a row of neurons. Although they sound insignificant, these layers are the key to learning for an AI. An explanation to why this helps the AI learn will be provided after this section.

Finally, after creating your inputs and layers, you must create your outputs. There can be several outputs, as they are essentially the controls of the AI. If we go back to the example of a driving game, we can say that some outputs of the AI would be to increase car speed, decrease car speed, turn left, turn right, or even restart the race. Now, these outputs don't literally come out as "turn left" or "turn right." Instead, they come out as probabilities, which are expressed as a value between 0 and 1.

Congratulations! You have successfully created your first AI!(not really but at least its something.)

Why does creating inputs, layers, and outputs result in artificial intelligence?

Think about how you react, how you get stimulated, or how you respond to certain situations. For example, if you see a pencil flying at you, your intuition would be to duck. How does this work? A certain number of your neurons flare up, which in turn causes more neurons to flare, ultimately resulting in the output of "duck." Simply put, your neurons are activated by the stimulus of observing an object flying towards you and causes your brain to send a signal to the rest of your body to perform the action of ducking. An AI is similar; it takes the inputs it is given(the input above would be what someone is observing) and "turns on" some neurons, which causes a chain of neurons to react, giving a final output.

However, that is just a extremely superficial and basic explanation. The neural network can be easily visualized.

In the image above, the inputs would be the blue neurons, the layers would be the green neurons, and the output would be the yellow neuron. Its a very simply diagram and yet it is how we can make digital life. If we take a look at the layers(or hidden layers), we can see that there are 2 layers on 4 neurons each. In other words, a layer is almost like a stack of neurons unrelated to each other. If two neurons are in the same layer, they do not have any impact on each other whatsoever. However, if we look at two different layers, we see how they are interconnected.

How does this give AI the ability to learn? And what is a neuron in the sense of AI?

Its a hard concept to grasp, but its also a very clever one as well.

A neuron hold two different values. One is the "Bias" of the neuron, and the other is the "value" of the neuron. The value of the neuron is basically what turns the neuron on or off, which ultimately decides whether or not it should impact neurons ahead of it. The bias of the neuron is a value that the value of the neuron must exceed in order to be "on." The bias is very important, as without it, the neuron would be far too sensitive. Think of it like this: a bias would be how hot does it have to be before you realize that it is noticeably hot. If you are warm, then you won't realize that it's hot. This is because you have a threshold between warm and hot, and that threshold acts exactly how a bias would act in a neuron.

To be continued