1.1 Foundations: What is AI

Hello and welcome to the first module of the AI Leadership Masterclass. I'm Noelle Russell, and I am so excited that you are here to learn more about how to be a better leader in the era of AI. We are going to first start off in our very first module and lesson talking about what is AI. AI stands for artificial intelligence, and it gives us an opportunity to do more with the data that we are collecting as organizations. Maybe you don't know this, but companies and actually the world has lots of data available. And when you think about the world's data, only about a sliver of it really is data that's indexable by the Google and Bing of the world. The rest of this data, 70% of the data in fact, is actually corporate data. And most organizations, their corporate data is locked in their company and they haven't really figured out a way to leverage this data to solve business problems. I often will tell companies that, you know, I spent a long time in big data organizations. I worked for a company called Pivotal that launched— that was part of Pivotal Labs and Pivotal Cloud Foundry, if you're familiar with these. Initiatives and organizations. And I remember during the time of big data, we were telling companies to save everything, to store everything, all their data, log files, emails, everything they can because one day, a magical unicorn called a data scientist will come and make sense of it all. However, those magical unicorns, they came but they were just like unicorns, extremely rare. So, organizations struggled to use all the data they were collecting in order to solve business problems because they didn't really have data scientists to unlock the power and insights that was in the data that they had collected. So we now are in a world where we need to think a little bit differently. Thank goodness, right? We call it the democratization of AI. Because we can't necessarily have teams of data scientists in every single company, we've now made it easier for for organizations to leverage artificial intelligence. I wanna share with you first 3 types of AI that you might be familiar with or you might have heard of before. But if not, I wanna set the landscape kind of with them. The first one, of course, artificial intelligence is important. So AI, we should define that, right? Artificial intelligence. There's a great way to kind of think about the definition of this term. There is a guy, his name's Alan Turing, and you could do some research if you wanna go look him up and learn about him, but he had a really great way of identifying whether something was AI or not. It was called the Turing Test. And the Turing Test, basically what it was is that you'd build an AI system and you would have a user come and they would talk to a human and then they'd talk to the AI system. And depending on their response, so if the human could tell they were talking to a human and tell they could talk, that like when they talk to the machine, they're like, "Oh, that's a machine talking," it would've failed the Turing test. If the human couldn't tell the difference between a human response and a machine response, that was considered an artificially intelligent system. Now that's of course not extremely scientific, However, it became known as a way of easily identifying AI, that it could replicate human behavior so much so that a human themselves could not tell the difference. Now, inside of AI, inside of artificial intelligence, there's a couple different types. The first type, we'll go pretty broad, is called narrow AI. And this represents the lion's share of AI in the world. This means that we've built a, an AI solution that's very purpose-driven. It's built to solve a specific problem. Think like a computer vision program that you see every time if you've been on an airplane recently and gone through TSA. TSA has an application that is purpose-built to examine the contents of luggage. So narrow AI is an AI system that's built for a specific purpose, and it's not meant to solve a bunch of problems, but more a set of very unique problems. We then have what we're calling general or generalized AI, and that means that this AI is now capable of doing lots of things. So this is something that a lot of people are thinking that ChatGPT is evolving to become, where it can generally answer any question. It can act as a human in almost every capacity. I say almost, but general AI, right, also has been known as it evolves to represent a term called super intelligence. And this is, I think, where many of the statisticians, mathematicians, many of the brightest scientists of our time are becoming very concerned because they're worried that as we build a intelligence, we are not being careful and intentional in the building of that superintelligence. And we're not applying something pretty simple. I don't know. I was raised on the golden age of science fiction. My dad taught me about Asimov as a child. And I remember at least, the very least, the three rules of robotics. And the very first rule is that robots shouldn't hurt us. It shouldn't hurt humanity, shouldn't hurt the people that they were built to create, to serve, to help. And so superintelligence, as we create it, though we're not going to cover superintelligence in this course, it's still an evolving topic. We are going to cover narrow intelligence deeply, narrow AI, and we're going to cover at a high level generalized intelligence because that's what we're starting to see in tools like ChatGPT, Claude. You can go to a website. I'll give you a resource., right now, uh, Poe.com, and it allows you to aggregate— it aggregates a bunch of these models together so you can test them out in a single interface, which is kind of nice. You don't have to go to a bunch of different websites to test it. But those are 3 different types of AI: narrow, general, and super intelligence. Now what I want to also though talk to you about is that inside the world of AI there's a concept, uh, a category known as machine learning. And this talks more about the specific ways that machines learn. So I'm going to cover a couple of these topics so that you have some baseline understanding. When I use certain terms, you'll understand kind of the definition and background and context for the use of the terms throughout the course. So machine learning has a couple different levels of ways to train a machine. The first way I'll call supervised. It's called supervised learning. And what this means is that we are going to supervise, tell the model, we're gonna supervise how it learns, which means we are gonna create both a training dataset and a testing dataset. We are going to, in the training dataset, use a term called annotate. We're going to tell the machine in that training dataset, here's what I want you to care about, here's what I don't want you to care about. We are then going to create a testing dataset that doesn't include the same training information. It's a way to test when I told the machine very specifically what's right and what's wrong using the test data, does it actually make the right choices? If it does, great. If it doesn't, we go back and we go through this cycle. It ends up being a bit of a circular movement, a virtuous cycle, if you will, right? So supervised learning means we're supervising how it's learning. We're giving it data and then testing against what we gave it and detecting whether it did a good job. And if it did a good job, then we are going and retraining it I should say, if it did a bad job, we're going and retraining it so it makes better choices. So that's one. The next one is unsupervised, and this is more of a type of machine learning where you give it a dataset and you tell the model to look for its own patterns. And this is easy, but easier than from a human's perspective, like we don't have to do as much work here. Except we do have to have lots of training data. So that is, that is one of the things that make this a little bit more difficult because here we can build a pretty small dataset, relatively small, but we need in an unsupervised machine learning model, we're going to have to give it lots of data so that those anomalies and those patterns can be detected by the model. Now, oftentimes in an unsupervised model, we need a way to tell the model what's right and what's wrong. So we leverage another term called reinforcement learning, which is an opportunity for us to reward the model for making good choices and punish the model for making bad choices. This is done Typically, you probably have seen this before when you're in ChatGPT or inside Microsoft Copilot, and you might see a thumbs up or a thumbs down icon, and it asks you, "Did I do a good job?" You're giving it reinforcement learning. Oftentimes, this is known as reinforcement learning with human feedback, which means that we're providing this feedback. Now, thumbs up, thumbs down is not great because it's very binary, right? It's black and white, yes and no. Many times when you're training a model, it won't be yes and no. It'll kind of be like, well, yes, mostly, but no in this one case. And so this human feedback, often rather than it being yes or no or good or bad, you want it to be more of a Likert scale, right? Like, on a scale of 1 to 5, how good is this answer? And that way you get a little bit more variability in the feedback that you're providing. Now, another type of learning, which I use often when I'm moving between models is called transfer learning. And this is when we transfer what we've learned in one model to another. Now, I tend to use this in the concept of when I'm training a large language model, I might grab a model like Azure's OpenAI service, and then I'll start prompting against it, and then I will learn what prompts work really well and which ones don't. I will then take those prompts and use it to fine-tune a smaller model. And I don't want to get too deep because we're going to talk about this later, but I can use that to fine-tune a smaller model, and I'm transferring the learning I learned on this big model to a smaller model, which actually will be cheaper and likely less latent, um, than a larger model would be. All right, so we've covered some basic terms as we go through. We've covered narrow AI, right? Purpose-built models. We're still seeing this used today, even in large language models. We have general AI, which is AI that is more general in nature, knows how to be a lawyer and a doctor and a nurse and a teacher and a mom and a dad. So generalized intelligence is the type of intelligence that means that it's generally smart about a lot of things. Superintelligence is the one that we're a bit worried about, where it actually has the ability to replace humanity in many aspects. And there's some parts of humanity we might want machines to assist us with, but there's a lot of careful consideration that's happening in many rooms right now. And then different types of machine learning— supervised, meaning we provide the training data and test data; unsupervised, where we give it a lot of data and ask it to identify its own patterns. We reinforce those patterns for positive and negative impact through reinforcement learning. We don't just provide yes or no, good or bad. We then also add human feedback to that. And then finally, we do have the opportunity to transfer learning from one machine to another. And so there's a concept of transfer learning. All of these are like classes in and of themselves. But just to give you a sense of some of the terminology we're going to use, these are the basics of AI. And of course, maybe I've left some out. So I'll, I'll add additional definitions if I use terms I think you need to understand a little bit more context to. I'll add those definitions as we go along in the course. But that's it for the first lesson. I hope you enjoyed it. I'll see you in the next one.