Where generative AI will move to

- [Presenter] With all the technology that we spoke about in the previous presentation, there's been more and more communities being spun up and these communities have allowed development to be distributed for different tools, different codebases. They've allowed for interesting things like embeddings. So embeddings can have a few definitions but in terms of generative AI, embeddings are a way to be able to find different patterns in prompts and share those with other people. So it's a very unique type of tool where it's very small but extremely powerful. So they allow for images to be generated with a certain style. So you could have realistic images or certain types of art or drawings, they're a very small file, perhaps maybe 100 or 200 kilobytes. But people can discover these in the large networks and share them, and I think that's really fascinating. You don't need a super computer at all. So this is the original paper where the concept of embeddings came from, and they show a few examples of how they've been able to capture certain styles of images and then use them on a number of different media. As you can see, it's quite fascinating. There's also a lot of extended media being produced so it hasn't happened yet but it'll be great to see the idea of movies on demand. So there's a large amount of research being done by movie studios such as Disney and others using generative AI solutions to change exactly how a movie looks. So in this example, they're swapping people's faces in and out and there's been other examples where they've changed things like someone's age, if you remember the movie, the Irishman, where they were able to dynamically do that. This paper explores the approaches that they took to actually swap those faces out, it's really fascinating. I'm sure that there'll be more and more examples like this in the future. And this one was only possible with StyleGAN2, so using generative adversarial networks. So there's more and more people developing niche applications and sharing that online. There's more accessible tools with no large compute costs and there's more and more detailed media being created.