Generative AI

(electronic music) - Let's play a game that I think you're pretty familiar with. Take a look at the series of images. Could you predict which item should come next? Of course, it's this one. Now let's try again with a new set of images. Which image do you think should come next? Of course, it's the pumpkin. All right, one last time. Now you probably guessed that it's this one, but have you ever wondered why you're so good at this game? You probably also noticed that these images had a similar theme, Halloween. Your brain stores tons of information about familiar objects, but it also tries to create rules in order to predict what is coming up next. Lately, we've learned that we can not only teach computers to do the same thing, but we can ask them to use that information to create new patterns. We call that generative artificial intelligence. Let's take a look at how it works. Computers are great at detecting patterns in a useful way. So for example, phone cameras are great at recognizing faces and they can be trained to recognize if that face is yours. Now they can do this by looking for patterns in the image. They map out distances in different parts of faces, which is called biometrics. It's easy enough for them to make a prediction as to whether that pattern matches what it knows about you. The breakthrough in generative AI comes when you realize that you can train the computer to create a pattern its trained to recognize. Let's say you train a computer with some pictures of a series of noses. You can ask the computer to analyze the pixels in the image to learn what a nose looks like. With enough data, you can ask the computer to generate a series of new pixels that look like a nose. With enough additional inputs, you can imagine how this could be extrapolated to faces and then entire images. You can go to the site thispersondoesnotexist.com and have a generative AI create random human faces. Now the site claims that 90% of fakes are not recognized by an ordinary person and 50% are not even recognized by an experienced photographer. This type of analysis and generation can then be used on techniques like creating a deep fake where you asked the computer to replace an existing face by training it on a series of faces you provide. It's also not limited to images. Pattern recognition and generation is being used to create original music trained by learning from different music genres. Tools like Compose AI use a technology called GPT-3, which has been trained on billions of learning parameters and can write human-like text. Now that technology has been used by companies like GitHub in a tool called Copilot, which helps programmers write entire functions and cuts down the time it takes to generate code. Although a generative AI is pretty impressive, it does have some limitations. First, it requires massive amounts of data in order to get really good at generating new information. Second, it doesn't always generate desirable results. In the case of something like Copilot, you can't trust that the code it writes will actually work. Now third, it can't create anything new. This technology is only combining information from the patterns it already knows. It's a disruptive technology for sure, but right now it's best at helping humans process large datasets, generating tons of options, and shortening the time it takes to handle repetitive tasks.