What is Generative AI?
In a nutshell, the fundamental function of text generators is to predict the next word in a given phrase. They do not "know" things in the way that humans know things, but rather make predictions based on all of the information that they have been trained on. For example, ChatGPT may look as though it "knows" the beginning of Abraham Lincoln's Gettysburg Address, but in reality, it produces the phrase "Four score and seven years ago" because that is the most likely string of words that is associatied with Lincoln's famous speech. Text generation tools, like ChatGPT, Google Bard, Claude, and others receive data from all across the Internet, make probability-based predictions from that data, and then take in user feedback to continue improving.
Because many of our readers have never seen ChatGPT in action, here we provide a short video of some of the things you can do with it including writing press releases, translation, automatic rewriting for specific audiences and automatic analysis of survey data with the results going into a PowerPoint presentation.
Below are two videos introducing the background to Large Language Models. The first is more introductory, the second is more advanced.
In this (hour long!) video, Sebastien Bubeck from Microsoft research gives an overview of the research they conducted in 2022 to try to get ChatGPT to act more and more intelligent. Systems have evolved dramatically since then but there are many interesting lessons in the video.
Text generators are not the only Generative AI tools available. There are also image generators that can take in prompts and create brand-new, never before seen images. There are several similarities between the way that image generators and text generators work. Like their text counterparts, image generators, such as DALL-E and Midjourney, must take in training data and provide prediction-based outputs to a user. The difference is that instead of learning word association and providing sentences when given a text prompt, the statistical model relates words to themes inside images and uses those to create new images from scratch when given a text prompt.
Below is a video that explores whether AI has creativity and imagination.