About this video
AI can now create images, art and video that no one has seen before. The video explains two methods: generative adversarial networks, where a generator and an adversarial AI train against each other, and diffusion, where an AI learns to reverse the process of turning an image into noise. It then asks whether this is true creativity and raises copyright questions.
You'll be able to
- explain how a generative adversarial network uses two AIs to create images.
- describe how diffusion trains an AI to turn noise into images.
- discuss arguments for and against AI having creativity.
- explain why AI art raises copyright questions.
Key words
- Generative adversarial network
- A method where a generator AI makes images and an adversarial AI tries to spot fakes, so both improve.
- Generator
- The AI that tries to create images similar to the ones in its training set.
- Diffusion
- A method inspired by physics where an AI learns to reverse the gradual adding of noise to an image.
- Noise
- Random changes to pixels that remove detail from an image.
- Copyright
- The legal right of creators to control how their work is used.
Transcript
Hi, I'm Cristobal Valenzuela CEO and co-founder of Runway A company creating AI tools for video generation and storytelling. One of our most recent AI tools is an algorithm called Gen two allows you to turn text just language into video. I'm Mira Murati I'm the chief technology officer at Openai.
The company that created ChatGPT. Computers were first invented to do math and solve hard problems. And despite all the advances in computer science, most people assume that computers can never have creativity and imagination. But is that still true? AI systems can not only write stories.
They can create images that no one has ever seen before. A.I. is able to create award winning art like this. It can create social media influencers that don't exist in real life and even create complete video sequences with just a few words of instruction. But how does this AI work?
And does it have true creativity and imagination? Let's dig in. Early projects that taught AI to create art used something called a generative adversarial neural network, which is a way of using 2 AI models to generate images. The way this works is that what one AI called a generator is trying to create images that are similar to the ones in the training set.
A different adversarial AI trains to detect if those images are close to the originals or not. Both AIs train against each other until the adversarial can no longer detect what image is real and and what image was created artificially. At this point, the generative AI has learned to generate real looking imagery.
Now, this is great, but computer scientists have come up with something even better called diffusion. Inspired by physics, Diffusion involves converting an image to an unrecognizable, blur or noise and training an AI to reverse this process. Take, for example, this image with all of its detail.
We can introduce a little bit of randomness or noise to these pixels, which takes away some of the detail. If you repeat this gradually, 1000 times the image loses all details and becomes pure noise. Now, every step of this gradual transition is an opportunity to train AI to learn to reverse the process.
An AI neural network can train by looking at the noisy image and then try to generate the slightly less noisy version because the noisy version of the image doesn't have all the details. The AI is learning to guess what the higher resolution details should be, and that's how it creates new images.
Now to train an AI, we’ll repeat this with millions and millions of images, each of them labeled with their descriptions. If we do this enough times, the AI learns to take a text description and generate completely new images nobody has ever seen before or to alter existing images.
The same approach can also be used for videos. Now this raises multiple questions. Is the AI really learning creativity and imagination? On the one hand, if you look at art and video created by AI it can be beautiful, original and amazing. On the other hand, the AI only learns this by doing math at the pixel level while studying creations made by people.
Is that really creativity? Another question is the issue of copyright. AI learns by studying the creations of others, and the original creators may want to say in this. Of course, when humans learn to create, they also study creations made by others. So the legal questions here are not simple.
We're still in the very early days of teaching AI how to create new types of media. Today, AI can generate photos and videos. Soon, it will also learn to create music and 3D worlds. This will have an incredible impact on all aspects of society, especially in entertainment, not just movies and music, but also games.
We hope you learn more about how AI works and how to use it for your own creations. What will you create?
Video © Khan Academy, used under CC BY-NC-SA. Watch on Khan Academy. Puzzicle is not affiliated with Khan Academy.