What is generative AI?
Been hearing a lot of buzz about large language models?
Answer
Generative AI is a kind of computer program that produces new text or images by running statistical analyses of its data pool based on the question you ask it.
It does this by relying on Machine Learning, a kind of programming where the computer doesn't have to be told explicitly what to do. Instead, it identifies patterns from its training datasets and then applies those patterns to new data.
This sounds big but at heart, it starts with a task you learned in middle school math, y=mx+b. In this equation, you are given the y-intercept and the slope of the line, and those together define the points on the line. It tells you which points make up the line, and which don't. And you don't need the slope to define the line, you can use any two points, assuming there's an actual line.
So what happens when the points aren't exactly in a line?
You can try to guess what the line is, using a technique called Linear Regression. Linear regression tries to minimize the distance between the points you have and a line you can define. Generative AIs apply this logic to human language, calculating line by line and phrase by phrase, which word will most likely follow given the prompts and the words that have come before.
Importantly, generative AIs do not know what the words mean. They know which words tend to go together. And while they are very powerful tools for generating texts and images, there are common computer tasks they cannot reliably perform. They cannot accurately solve the Hanoi towers puzzle with n rings. They cannot count objects you give it. If you gave it a picture of a 5-legged zebra and asked it how many legs were in the picture, it will tell you four, because "four" is the most common word associated with "zebras" and "legs".
So when a generative text AI gives you dates or amounts or names, use a critical eye. Check to see whether those can be verified by a different source. If it lists references, make sure the authors exist, that the titled paper exists, and that the paper has been correctly summarized.