How is Generative Artificial Intelligence Different from "Regular" AI?

How People Learn

People learn through reading, listening, observing, sensing, feeling, playing, interacting, experiencing, reasoning, trial and error. And all those learning methods just scratch the surface.

Imagine you want to learn how to play checkers. You could do this in several different ways. You could hire a tutor. They would introduce you to the game, teach you how to move the checkers, and show you some winning strategies. You could practice by playing against the tutor, who would supervise your moves and correct you when you make a mistake.

As you increase your mastery of the game, you would then be able to play without supervision.

Another option would be to watch people play checkers in the park. By closely observing how others play the game, you could figure out how to move the checkers. Without any supervision, you could see how to jump an opponent's checkers, and how to earn a crown. You would probably also start to identify different winning strategies.

A third option would be for someone to show you the basics of the game and then set up a system of rewards. Every time you took one of the players' pieces you would get a small reward. Think of it almost as a cash reward (like a nickel or a dime). Over time you would start to develop your own strategies to maximize the reward.

In a sense, this reinforcement is like a mixture of the other two ways of learning. You would get a sense of what it takes to win and then it would be up to you to come up with the winning strategy.

How Machines Learn

The three ways to learn to play checkers are very similar to how machines learn:

Supervised learning: With supervised learning, you feed the machine labeled data. The machine can figure out how to classify the data. For example, you feed the machine a picture of an apple and label it, "This is an apple." You feed it a picture of a banana and label it, "This is a banana." The next time you feed the machine a picture of a piece of fruit, the machine can tell you what it is — a banana, an apple, grapes, cherries — assuming it was included in the training data. If it delivers a wrong answer, it learns from its mistakes, becoming more accurate with more data.

Unsupervised learning: With unsupervised learning, you feed the machine a set of data, and it discovers patterns in that data. For example, you feed it pictures of fruit without labeling it, "This is a banana. This is an apple. These are grapes." The machine determines how to cluster the pictures. Maybe it groups them according to color, maybe by size, maybe by shape, or maybe by some other characteristic you never imagined.

Reinforcement learning is a technique that involves rewarding the machine for its performance. With reinforcement learning, a reward is associated with a certain action or occurrence, and an environment is created in which the machine attempts to maximize its cumulative reward.

Generative Artificial Intelligence

These three methods for machine learning have worked well over decades when it comes to finding patterns in existing data. Unsupervised learning worked well for clustering data. For example, when you go to Amazon's website, it shows you which items are frequently bought together. Supervised learning worked very well for classifying data. E-mail providers could quickly classify messages as spam before they get delivered to your inbox. Reinforcement learning could be used when you want to discover something new. So an online music service can make recommendations for songs that you like based on your music taste.

All three of these learning methods work well reading data, but they're not very well suited to creating something new. It's almost like these machine learning models viewed the world as “read only.”?These models could find patterns and cluster together data, but they couldn't be creative.

To create something new you needed a learning method that could develop a model that's flexible enough to generalize data into something new. It wasn't enough to classify images of apples and oranges. Generative systems should be able to theorize and create an entirely new fruit.

To do that you needed a form of self-supervised learning that could identify patterns in the data and then take that extra step to transform that data into something new.?

If you were using supervised learning, you'd have to create a new model every time there was a change in the data. So if you wanted to teach a system how to play chess you'd have to train the system on the rules of the game and the different pieces it takes to play. Then if you wanted the system to learn how to play checkers you'd have to start from scratch and teach it the new rules and the different pieces.?

But newer self-supervised learning systems try to generalize the learning so that you don't have to start from scratch. In a sense it's much easier for the system to learn how to play checkers after it's already mastered chess.?There's a foundation model that you can use for many different games.

That means instead of training the system on all the different board games, the system is pre-trained and it can transform the model to apply it to new games. ChatGPT uses a very similar approach to generate new text. In fact, the “GPT” stands for Generative Pre-trained Transformer.

That's why when you ask ChatGPT to write a Shakespearean sonnet on making an egg white omelet, it's able to come up with something creative. It's not because the system has been trained on writing millions of sonnets about omelets. It's because of a generalized foundation model that can generate text in unexpected ways without any specific training. Then it uses probabilities to write the sonnet one word at a time. Of course to describe it this way shouldn’t diminish from the enormous technical accomplishment.

The next few years most of the focus around generative AI will probably be around text, pictures and video. But these self-supervised foundation models can be used to generalize a lot more than just this type of data. A system that knows how to play checkers and chess might use its foundation to invent an entirely new game. As these systems start to view the world as “read write” instead of “read only” they can develop new creative skills.

It might not be science fiction for patients to start preferring AI doctors and lawyers over their human counterparts. These new generative systems can crunch many lifetimes of data. It could be just a few years away before these systems do a better job diagnosing patients and writing ironclad contracts.

For now, this newsletter is still 100% human written (until your preferences change ??).

This is my weekly newsletter that I call The Deep End because I want to go deeper than results you’ll see from searches or ChatGPT. Each week I’ll go deep to explain a topic that’s relevant to people who work with technology. I’ll be posting about artificial intelligence, data science, data ethics and agile software development.?

This newsletter is 100% human written ?? (aside from a quick run through grammar and spell check).