Generative AI For The Rest Of US

Generative AI: The Next Big Thing

We all use computers and other digital devices in our daily lives. Our interaction with them has always been based on the input-output model. When a programmer creates a software application, they define the inputs and outputs. The program processes the inputs to produce the desired outputs based on user instructions. The software is designed to perform specific tasks, and its capabilities are limited to the programmer's instructions. Now, imagine a computer that can learn, think, and communicate like a human being. This is the promise of Generative AI.

Intelligence, as we understand it, is the ability to learn, understand, and apply knowledge. Generative AI is a form of intelligence that transcends the traditional structured input-output model. It can generate new content, such as text, images, and even code, based on the input it receives. Using Generative AI today is like having access to a vast, cloud-based intellect that can assist with infinite patience and a vast array of knowledge. It's like having access to Einstein, Shakespeare, and Da Vinci all at once. Generative AI can write blog posts, poetry, lyrics, create images, compose music, and generate computer code. But it can also help your child cheat on their homework, it can help malicious actors create deepfakes, and it can help a scammer create a fake identity. The possibilities are endless, and the risks are real.

Behind the scenes, a Generative AI model like ChatGPT predicts the next word in a sequence through a sophisticated understanding of language acquired during its training phase. During this initial stage, the model is exposed to vast amounts of text, enabling it to learn the statistical relationships between sequences of words. This deep learning process equips the models with the ability to gauge the likelihood of a word's appearance following a given string of words, essentially understanding the flow of human language.

Let's think about the process of predicting the next word in a sequence as a probability distribution over a vocabulary, based on the sequence of words that comes before it. Imagine we have a very small vocabulary with only four words: "apple", "banana", "fruit", "is". And we want to predict the next word in the sequence "The apple". The model's job is to find the probability of each word in the vocabulary being the next word.

Let's denote:

• P("apple" | "The apple") as the probability that the word "apple" comes after "The apple",
• P("banana" | "The apple") as the probability that "banana" is the next word,
• P("fruit" | "The apple") as the probability that "fruit" is the next word,
• P("is" | "The apple") as the probability that "is" is the next word.

When trained, the model saw that the word "apple" is often followed by the word "is". Let's suppose the model may have learned that:

• P("apple" | "The apple")=0.2