From the course: Introduction to Artificial Intelligence

Generative systems

From the course: Introduction to Artificial Intelligence

Generative systems

“

- So far, you've seen that predictive AI systems can be used with natural language processing, robotics, and internet-connected devices. You can also use predictive AI to find patterns in your data. This is exactly the type of application where predictive AI does very well. You could train the system to work on a specific task. Then the system will make predictions based on limited data. That way, you'll have a robot that can learn a few new things, or you could train a system on how to translate a few languages. But remember, with enough data and computing power, you might also use a generative AI system. The last few years, it seems like these systems have emerged out of nowhere. A lot of it has to do with the fact that there's now access to massive amounts of data. Plus, in the last few years, the cost of computing has gone down. This mixture of massive datasets and low cost computing helped create a bunch of new GenAI systems. In 2022, the company OpenAI released one of the very first public generative AI systems. It was called DALL-E 2, and it was designed to create generative AI graphics. The system searched through billions of online images. It was able to see patterns in these images and generates something completely new. Many people were shocked that the system could create photorealistic images of astronauts riding horses or cats in flying saucers. This was possible due to having access to billions of photos, plus the ability to perform extensive, inexpensive computation. A few months later, the same company released a generative AI chatbot called ChatGPT. People marveled at how the system answered questions. It was trained using machine learning on a deep learning artificial neural network. But this was different from a predictive system. It was flexible and it could answer very general questions. That's because it trained on trillions of connections between words. It gathered all the information that it could find online. It looked through the text of books, social media posts, news sites, and informational sites like Reddit. It looked through so much more data than what you'd get from a traditional predictive AI chatbot. In fact, it's still being disputed whether it could legally look through all this information. Once the system was trained, humans spent time practicing with the chatbot. Since it was gathering data from everyday people, there was a lot of misinformation. These humans helped correct the system to keep it from repeating things such as obvious errors or hate speech. The generative AI system was able to have a conversation by predicting each word in a sentence. It would respond one word at a time by analyzing billions of conversations. It could create new content by identifying patterns in the characteristics of everyday items. If you wanted to create a song about peanut butter and jelly, it would understand that jelly is fruity and sweet while peanut butter can be smooth or crunchy. Because these systems need a lot of data and computing power, they've mainly been run by large companies. These dozen or so companies have gotten a huge headstart in developing generative AI systems. So it's likely that most businesses will use these systems as a service. They'll tap into the power of their GenAI without the challenge of building a massive system.

Contents