How AI Works, for the Acutely Impatient

Everyone has heard of AI in the past few months, but not enough people understand the black magic which it runs on. This includes you, so stop writing essays with ChatGPT and read this.

AIs work in the same way humans do – when an AI is created, it does not know anything, just like when you were born, you did not know anything. The difference is that an AI can learn all there is to know in a matter of days or weeks, whereas you need at least 3 years to stop speaking garbage.

AIs used for different purposes are trained differently. For example, an AI made for recognising images will learn differently from an AI made for talking with humans, like ChatGPT.

When you want to train an AI to recognise what is in an image, you give it a huge list of photos, with the descriptions of what they show attached to each one. Imagine it like how when you were a little, you had textbooks with photos of houses, cars, animals etc. with their name next to it. After years of being shown photos of these things with a word next to it, you could see a picture with no words next to it, and tell what was present in the image. Similarly, after being shown millions and millions of images with text, an AI learns how to categorise images based on what it contains.

Another strategy which is used to train AIs is the one your parents use to teach you. During training, whenever an AI gets a correct answer, it is given a reward – like how your parents would promise you a piece of candy if you got an answer right. Imagine this as adding one point to its score. Whenever it gets an answer wrong, it gets a punishment – like removing a point from its score, or like getting a tight slap. Just like humans, an AI wants to rack up as many of these points as possible, even though they have no real meaning whatsoever.

Now, you may ask, where do you get millions and millions of images to train an AI with. Well, the same place where you are learning about AIs – the Internet. Remember when you had to select those traffic lights, buses and zebra crossings to prove you weren’t a robot? All the while, you were actually helping Google teach their robots to learn how to drive cars. Whenever you filled out those random squiggly characters, you were actually training Google Lens.

Now, you might be wondering how this explains how ChatGPT, or that character AI everyone is obsessed about, works. Well, to illustrate, fill in the blanks.

Merry _________

Happy New ________

Oh My _____

Corona ______

Dominos’ _______

Albert _______

Barack _________

I hope you got all of them. Now imagine, if instead of the two books you have read in your lifetime, you had read and reread and reread the entirety of the internet. That, my friend, is how an AI works.

This article in video form

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.