All right, let's be real. Most AI explanation videos are either super technical and kind of useless because it's daunting, or very fluffy and useless because they don't actually give you new information. We understand metaphors, and I understand AI. So I'm going to give you some very simple metaphors that make all of this AI landscape approachable for you. And this is just a quick hit with Star Wars Halloween candy and game shows. Ready? Let's start with Star Wars. So there is not one kind of AI. And I don't mean the difference between, like, ChatGPT and Google's version of ChatGPT. Gemini. I'm not talking about that. There are AIs that are. Those kinds of things are the same kind of AI. They're just made by different companies. But then there's also very different kinds of AI, like the ones that do medical imaging in cancer centers.

00:49

And those are built by very different companies doing very different things. Let's think of this like Star Wars. We have some droids that are C3PO and they're designed to talk to people. And then we have some things that are like R2D2, and they interact with machines. So when you go to the bad guy spaceship, you plug in R2D2. When you're flying around in your X wings, you plug in an R2D2. And there's a bunch of R2D2s, and there's a bunch of C3POs, and that's how AI works. And all of them have different strengths and weaknesses. And understanding which one you want to use is a huge head start instead of just calling all of them an AI. So far, so good. Great. The main ones that you are thinking about are generative AI and built on a thing called a transformer.

01:33

We don't need to get into any of that. But just understand that all of these general generative AIs that you're familiar with basically work on the same underlying parts, and they have the same basic shape. And each company does things a little bit differently, just like cell phones, but it's the same basic idea. Cool. So now let's talk about these generative AI tools that you're hearing about and kind of intimidated by or frustrated by. And one of the main things people are frustrated by is it doesn't keep giving me the same answers. It's not right. It. Hallucinates is a big fun word that we're about to talk about. So why does that happen? Okay, there's different kinds of code. And like, when People are writing computer code. There's one version that is called deterministic. And think of this like a calculator.

02:20

When you type in 6 times 7 in a calculator, you always get 42 because it's deterministic. It is the right answer. But that's not what all these cool AIs are doing. These large language models, these ChatGPTs, these what have yous, they don't give you 6 times 7 equals 42. Every time they do something a little different that is called non deterministic or probabilistic probabilities. And think of that like a old game show that we all have watched the Family Feud and so they say we interviewed 100 people, top seven answers on the board and any one of those seven answers answer is considered a correct answer. This is how ChatGPT works, only it says, we have interviewed the entire Internet. Top 7,000 answers are on the board. And when you have the top 7,000 answers on the board, some of them are really bad.

03:11

The 3,000th best answer isn't a very good one. But this is also how talking to a person is. When you go ask a person a question, they have a top three or top 10 potential answers in their head and they try to give you the best one. But they don't have all the data, they don't have all the information. So they're not 100% right. This feature that it just talks kind of like a person in ballpark correctness is one of the things that is making these large language models simultaneously so disarming and approachable, but also frustrating when you need to get stuff done. So let's talk about why it's doing that and what we can do with that. So let's stay in the Family Feud answer.

03:51

A hallucination is when it just goes all the way down and grabs answer from way low on the list. So it's still on the. But it's just very low likelihood. There's some weird loose association between these words and it just grabbed it. Why did it do that? One easy way to understand this is a thing called temperature. Now that's the technical term, but you can just think of this like creativity. And sometimes software uses the term creativity, but it's just temperature. And that just means how comfortable the AI is going further and further down the list. And this is a setting that you can change. So you might say, I only want the top three answers. Or you might say, hey, for what I'm doing, let's be really creative and give me the top 20.

04:33

And you don't realize that's a setting that you can control, but it's a setting you can control. And knowing that means you can help influence your system to give the type of results you want. Cool, right? Let's do another one. So that's what a hallucination is. Just. It goes further down the list. Okay, but let's get away from those metaphors and let's add a third one, which is Halloween candy. Here's two more settings in a large language model like ChatGPT that you can sometimes play with. One is called Top P, and the other one is called Top K. You don't have to get into it in great detail. Here's all you need to know. Think of it like Halloween candy. With one of them, you say, I want 10 pieces of candy, and you just grab 10 and you don't care what they are.

05:16

And the other one is you say, I want the most popular candy. So you look at what you have the most of. Let's call it chocolate. And so you get all chocolate pieces. Pieces. So in one of them, you're only grabbing, like, the popular type, like chocolate. And the other, you're controlling for the count. And those are settings that you can play with in its data set. So you can say, just give me the smartest answers or just give me a few answers. And so now you're controlling how much information it has so that it can be more efficient. One of the reasons you want to be efficient is because there is one more thing, and this will be the final one that we talk about, which is called context window. And context window is just its ability to not get overwhelmed.

05:57

And these context windows have gotten pretty big. But the idea is, when you fill the AI's context window, its brain divides by zero, it gets overwhelmed, and it's just done. You know when you're just mentally done at the end of a long, hard, stressful day and you're just like, I'm not smart right now. I need a minute. That's your context window. So think of that like a single sheet of paper to take all your notes about a client. Every single thing you want to know about your client can only fit on this one piece of paper. The size of the paper that you have to take notes on is the size of your context window.

06:30

And so if we just look at the fact that there's different AIs, some of them are really good at medical imaging, some of them are good at talking to people, we can understand that. We can understand that we're dealing with the talking to people kind. And those have strengths and weaknesses. One of the strengths and weaknesses of the talking to people kind is that it's not really like a math machine. It's not really like a calculator, like we're used to computers being. It's much more like Family Feud. It's going to take the top few best answers, and it's going to just take a high probability. Now, here's the interesting thing about a hallucination. As soon as it gives the Family Feud answer, instead of giving the number one answer, it gives the number five answer or the number 25 answer.

07:09

But as soon as it gives that number five answer, it's going to, in all of its responses, work as if this was the best answer. Now, you and I know it was the fifth best answer right now, but in the world of that AI, it will say this was the best answer, and it will take the conversation down a path. That is assuming that's the best answer. Now, that also means if you ask it four things in a row and it gives you the fifth best answer to each of those questions, by the time you get to the fifth best answer, you're pretty far away from where you wanted to be. And that can get really frustrating because you're getting further and further off topic.

07:49

And that is just based on hallucinations, and that is just based on it grabbing things further down the list of options. So when that happens, honestly, the easiest thing to do is to start over with it. Not in that conversation. Start a new conversation. Because an AI is good at getting a song stuck in its head. You know when you get a song stuck in your head and you can't think of a new song because that song just keeps popping in and you kind of find yourself humming it. And when you're trying to do other things, AIs have that same problem. When that happens, it's best just to restart and just say, okay, we're going to start over. We're going to maybe even copy paste the parts that we liked above into a new conversation.

08:26

Say, hey, let's pick up from here, we got off track. And then you can start new and grab the highest answers again, reducing that temperature. The lower the temperature, the more it'll just be the number one to number three answers. You know, top three answers. The higher the temperature, the more likely it will be to grab answers further down that list. So if you want to control that temperature for some use cases or really open it up for others, and understanding that is something in your control is a useful thing to know. One final term that you might have heard getting thrown around is called a prompt or a chatgpt prompt. And think of this just like a hypnosis script. You just copy and paste it in.

09:09

So you just take their words and you say them, and you just copy and paste it into ChatGPT, and you are scripting it, you are prompting it to do what you want, and you're just copying somebody else's hypnosis script in. Only instead of saying it to a client, you're saying it to an AI, to a ChatGPT. And if all you want to do is copy ChatGPT prompts to get what you want, that's okay. But it's actually a lot harder in the same way that having to find the right script for the right client all the time is much harder than just knowing how this stuff works and being able to do what you want. And that is the point of this class. We're going to give you the tools and resources so that this stuff isn't overwhelming and technical.

09:48

It's stuff that you understand well enough that you can move with confidence in this new landscape that we're all exploring together. So let's make this thing no longer intimidating, but instead approachable and really understandable with simple metaphors that you can take with you. Yes, we will give you prompts and scripts. Yes, we will give you tools and resources. But more importantly, we will give you the confidence and the understanding so that you can grow with this technology. So if this is interesting to you, the link is below. We'll see the last Thursday in August for a really cool class that will be full of interaction and question and answer. See you there.