Explore AI and Its Different Types

0:00Hello and welcome to the first skill in the AI Essentials with Microsoft Copilot course. My name is Jonathan Barrios and I'm excited that you're here. In this course, we'll explore AI and its different types at a high level. You can think of this as a crash course in AI literacy. This is so you can understand what's really happening when you use Copilot. You'll get a clear picture of the different types of AI that power tools like Microsoft Copilot 365 and why that matters for your work.

0:31We'll also dig into prompt engineering. This is the skill of writing clear, effective prompts so that the AI can give you exactly what you need the very first time. A prompt is simply what you enter into a chatbot. And you can think of it like learning to ask better questions so you get better answers. So what's included in this course? We are going to talk about AI and prompt engineering basics, meaning plain language that you type into a chatbot.

1:02And this is directly applicable to Copilot and Microsoft 365 applications. Really, it's all about Copilot and how we can use it for Word, Excel, PowerPoint, and much more. And we'll focus on hands-on training for not only better prompts, but for writing, summarizing, and analyzing. So what is this course not about? We're not going to focus on IT setup, licensing, or any admin settings. There are no tutorials on how to build actual AI chatbots.

1:34And we're not going to use any custom Copilot Studio agents or connectors. But we are going to have fun with AI and Copilot. Another cool feature of this course is that it's part of an AI mini-series of courses. And we are here. What's important to note is that most of these courses start off similarly with AI literacy. What is AI? What are the different types? And a lot of them also have prompt engineering. And this is a course that I highly recommend taking if you enjoy writing prompts.

2:08We're only going to scratch the surface of prompt engineering in this course because it's really about Microsoft Copilot. And if you have any interest in coding, I would suggest AI vibe coding and security. This is for those of you that don't have a programming background, but would like to learn how to program using AI or having AI do a lot of the coding for you. That's what vibe coding is. However, if you do have a programming background, we also have the AI agentic coding with ChatGPT, Cursor, Cloud, and Copilot.

2:43Here, we're really taking the lessons that we learned in prompt engineering and applying them for programming. So if you're interested in coding, I would suggest starting here with prompt engineering. And then if you don't have a programming background, then go here and then finally take the AI agentic coding. But if you do have a programming background, well, you can just skip the vibe coding and go straight to AI agentic coding. This course is also similar to AI productivity for professionals.

3:14But here, this is tool agnostic. This is more about the techniques that you would use. Whereas in this course, it's really about Microsoft Copilot. But we also have Google Gemini that is embedded in all the Google products. And we have many chatbots and we have many AI tools that are outside of different ecosystems like Google or Microsoft. And that's what this course is about. And it really focuses on tasks for professionals. And finally, these two courses are about AI compliance, privacy, and copyright.

3:47This is for the non-technical teams and this is for developers. So the more technical side. And the reason this is important is because a lot of the regulations and laws are being drafted as we speak. So there's a lot of new developments. And I want to catch you up in these two courses. So a little bit about me. I'm Jonathan Barrios. I teach data science, machine learning, and deep learning. And the term that I use for this is AI data science. It's a blend of AI engineering and data science.

4:17Before teaching, I was a full stack developer. And I worked with AI and built SAS applications, AI applications, and worked on streaming platforms. I've been teaching for quite some time at top platforms like Treehouse, Thinkful and Chegg. And now I'm happy to be part of CPD Nuggets team. If you want to find me online, you can find me here at LinkedIn at jonathan-barrios-ai or at my website jonathanbarrios.com. And finally here at X at AI underscore data underscore science.

4:50So a quick note about X. A lot of people either love it or they hate it. But my angle is a little bit unique. All of the AI community is basically on X. I don't know why, but I don't really treat it like social media. And I don't really use social media. I really feel it's a waste of time. But going to X allows me to keep up with all of the new developments very, very quickly. So I only follow people that are talking about AI and that are leading in that field. And that way it's a very valuable platform for me.

5:22But otherwise I ignore everybody else. I only follow select people. So if you want to learn more about AI, I highly suggest following AI individuals like myself and others and ignoring everybody else. That's my suggestion. And if you do that, then X is pretty cool. Otherwise, I do feel it's a waste of time. All right. So now in this skill, we're really talking about AI and it's different types because everyone thinks that a chatbot is what AI is, but there's much more. So let's talk about what we're going to do in this course, but more importantly, what we're going to do across the six skills.

5:57And that's what makes up this course. This is a six skill mini course, and here's the plan. We first learn AI basics. This is really important because AI has a tendency to lie and it does so with supreme confidence. It's kind of like an expert liar because it's not really trying to lie. It's guessing and doing all kinds of stuff. And I'm going to show you exactly why it does that by exploring the different types of AI and showing you how it works. This is an invaluable skill moving forward, especially if you're going to start working with AI.

6:31And then we're going to talk about prompt engineering. So this is the first skill. This is the second skill. And here we're just really learning how to communicate to AI. The takeaway here is that most people talk to the AI like a human. So you write a prompt and then you just hope for the best. But a chatbot is not even close to a human. It's very, very different. And that's what you're going to learn in the first skill. The second skill, once you understand how it works, I'm going to show you very simple tools that you can use to communicate to the AI in the way that really works for AI.

7:08It's very tempting to talk to it like a human, but it just really gives you a lot of bad outputs. And more importantly, it'll just lie. Not because it's trying to lie. It's just it's hallucinating. That's the term that we use in the AI community. And you'll understand why in that first skill. And then we'll apply everything that we've learned up to this point using Copilot and now Word. And by this point, you understand AI and you have prompt templates, meaning very simple, repeatable templates that you can use to enter into the chatbot to get exactly what you need.

7:45And then we'll segue into using Copilot with PowerPoint. And we'll take the ideas that we learned from Word and directly apply it into a presentation. So here we come up with ideas and then we take those ideas and turn it into a presentation. In this case, in Word, we're starting a business plan for a health and nutrition startup. And then we'll segue into PowerPoint and create a presentation. Maybe we need to get financing and we can make that presentation. And then in the fifth skill, we'll apply everything to Excel.

8:18And what's really great here is that if you are very familiar with Excel, this is going to really speed things up. You can do analysis by typing plain language. How cool is that? Also, if you don't like Excel because you find it complicated, well, it just got a lot easier. So regardless of where you're coming from in the Excel application, this Copilot makes it way easier and more powerful. And then in this sixth skill, we'll talk about images and more because this is a mini course, so we can't cover everything.

8:53But here we'll talk about Copilot Chat, Copilot Create and how to create images and more. If you want a full course on Copilot 365, please let us know in the comments. This is the data that we need precisely to know which courses to make for you. And by the end of this skill, you'll have a repeatable approach to getting high quality results from AI, no matter which Microsoft 365 app that you're using. Again, just keep in mind this is a mini course, so we can't get into full detail, but we are going to focus on the important parts.

9:29And by the time that you get to the sixth skill, you'll have a strong foundation on AI, prompt engineering and Microsoft 365. All right. So what are we going to talk about in this skill? As a skill title says, we'll definitely explore AI and its different times. But first we're looking at AI as a whole, and it's not just one thing. That's the first takeaway. But we'll also talk about hallucination, which is a very common term in the AI community, meaning that the model can be super wrong and super confident at the same time.

10:03And then I'm going to simplify AI to help you build a mental model for productivity. And here's why. Once we talk about all the different types of machine learning and deep learning and generative AI, large language models, what is that going to do for you when you're talking to Copilot and Microsoft 365? It can be very helpful, but it's a whole course unto itself. So instead of doing that, I'm going to break it down in such a concise way that we build a mental model so that when you're in Copilot 65, you understand, okay, I know what I'm doing and I know what the AI is doing, generally speaking at a high level.

10:44And that's what we're doing in this part. Super important. And the way that I'm going to talk about it is different types of software, software 1.0, 2.0 and 3.0. We'll think of it as really simple, low level types of software. And then we'll briefly talk about prompts because prompts is really a type of software. So speaking in plain language is prompt engineering. And there's a famous quote on X from one of the top engineers and their tweet said the hottest new programming language is English.

11:18So you might, there's all these programming languages like Python, JavaScript, and so on, but English now is a programming language, but really it's plain language, meaning it's many different types of languages. So we're using plain language to program. And once you understand this, you'll really have a powerful tool when you're writing prompts in plain language. And I promise you're not going to learn programming. This is super easy. Once, you know, you'll get that aha moment. You'll be like, okay, I know how to talk to the chat bot now, and it's going to do exactly what I say.

11:53And it's not going to make stuff up. That's the takeaway. And on that note, I will see you in this first video. When we talk about AI and the different types, see you there. Bye.

0:00Welcome back. In this video, we're going to explore what is AI and how it's more than generative AI. But first, what is generative AI, right? That's the smallest sub domain of AI. So first AI here is a box. Think of it like that. AI is the box and there's a lot of AI inside, lots of goodies inside that box, but it's not the whole box. And that's the first misconception. When people are talking to ChatGPT and these different kinds of chatbots, and now with Copilot, there's Copilot Chat, but there's also Copilot built into all of the Microsoft 365 applications. That's generative AI. So everyone is using generative AI in that way. However, there are many types of AI that we're interacting with that are not chatbots. So if you think about it, you're probably going to come up with some different use cases. And we'll get into that in just a minute. But if you thought that the chatbots, because they came on the scene and they're talking like humans, that that's the only thing that AI is, well, that's the first thing that we need to clear up. It's more than just that. All right. So then we have machine learning.

1:15So machine learning is the first sub domain of AI. This is where the software learns from data. OK, so that's the first thing it learns by itself. And that's where the term artificial intelligence comes from. But it learns from data. And also here, the same thing, this one, deep learning learns from data. But what's the difference? Well, deep learning is software that's modeled after the brain. And machine learning uses statistical algorithms to learn data, to learn from data. So first things first, you don't really need to memorize all of this stuff at all. In fact, I'm going to make it very simple when we start talking about the different types of software. And we'll do that by using a hot dog and a hamburger and make predictions when we show it to an AI. That's the easiest way to understand how it works. And we'll be using deep learning for that. So you can think of machine learning as statistical learning and deep learning as software modeled after the human brain. It uses neural networks. And by neural networks, we're just saying that the brain has neurons and they're all connected. And that's a neural network. Everybody

2:30has one. So that's not a foreign thing. However, in deep learning, we use code to create artificial neurons and then hook them up together. And it works really well. And that's what powers things like chat GPT, and Microsoft Copilot. So they're built from deep learning. And that's why all of those tools are called generative AI. The easiest way to think of that is well, it's generating text, that's it, it's some kind of AI that generates stuff, it can create text, images, code, and much more. But the important part here is that chatbots are just one slice of the pie. They're not the whole thing. And a lot of people think that well, chatbots, that's AI. So now you know that we're thinking bigger than just chatbots, which is generative AI. Now let's take a moment to think about AI in everyday situations. So in certain cases, if you open up your phone, and you have the facial recognition instead of a password setup, that's facial recognition. If you watch movies, well, they have facial recognition to catch criminals. But it's also used in computer vision, where the car can identify not just faces, but other cars, people, trees, and even signs.

3:51These two are deep learning. So deep learning is not just chatbots, but it's also facial recognition, computer vision. And here we use deep learning as well, but also machine learning. It's a combination. So you can use more than one type, because the robot is using deep learning to see, but maybe it's using machine learning in different parts of how it moves the motors. And when we're thinking about intelligent document processing, we could be using generative AI, deep learning, and machine learning altogether. However, with fraud detection, which is often used in banks, they've been using machine learning for quite some time to find outliers, things that stand out. And that's one way that they find fraud amongst all of the transactions, millions of transactions. Using machine learning, it's great at finding these patterns. And you just give it all this data, and it learns from that data. And it says, look at these 10 transactions.

4:50I think that there is some kind of fraud here. So it can be very powerful and very helpful. But as you see, none of these are chatbots. So AI can be many different kinds of things. So AI is machine learning, deep learning, and generative AI. However, this is a simplified list. And the idea here is not to go further than this, just to get a general idea that AI has many different types. And here is where we try to make everything super simple. We're just thinking about three types of software, software 1.0, software 2.0, that applies to these two, and then finally, software 3.0. And we're going to go through each one of these. And I'm going to show you hands on how it works. It's pretty simple. For example, software 1.0 is coding, also known as programming.

5:40So it's traditional programming. So everything that you've ever known about programming, whether you see a movie where they're hacking into something, or you see a bank teller entering information, or you're using any application on Microsoft, that is all traditional programming. Only recently have they started to use machine learning and deep learning and generative AI. That's not 1.0. 1.0 is traditional programming. And I promise I will make it super, super simple. And we're not learning programming. So that's good news. So software 1.0 is basically just rules.

6:17If X, then Y. So you could also say, if the ticket is older than 24 hours, escalate. Or if your seatbelt's not plugged in, then the light goes off. So things like that. Very, very if X, then Y kind of thing. And that's just really rules. And sometimes we use analytics, but still, that's all software 1.0. Software 2.0 spans classic machine learning and deep learning. So here we have given past tickets, predict the priority one through five. So it learns from the data. And that's all you really need to know. And when we're thinking about software 2.0, we're doing the same exact thing. We're still learning from data, but it reads the email body and detects some kind of sentiment analysis, some kind of intent. Oh, this email is about a job promotion. So it figures out that intent. What is the goal of that? And that's just a different kind of learning. But as we know here, we're using algorithms. And here we're using neurons. So software modeled after the brain. That's not the important part. All you really need to know is that if the software learns from data, it's 2.0, even if it's machine learning or deep learning. So we know that software 1.0 is traditional programming. And software 2.0 is

7:37some kind of machine learning or deep learning. And then now we have software 3.0. And these are chatbots. And here I have something that says LLMs. And that stands for large language models. You don't need to know that as well. But now you know what it means. If you ever see that, all it means is a chatbot. The large means that it was trained on a large chunk of the internet. So large data. And it's using language primarily, well, text on the internet. And it's a model. Traditionally and normally deep learning model. And deep learning works great with large data.

8:12And it's great at language. And well, that's the type of model that it is. A large language model. But it's very, very specific. So it fits under the software 3.0 or generative AI. So here, summarize this thread for the next agent. Or draft a welcome email. These are things that you type into a chatbot. And it's great at that. So now we know that that's software 3.0. But the most important thing for generative AI is that you use plain language. So software 1.0, you have to learn Python or JavaScript or some kind of programming language. Software 2.0, you're also using like Python or some kind of a programming language to set up the model. But then you run that model or train it. And then it just goes and learns on its own. And then you can deploy it. So that's still pretty complicated, right? Software 1.0, 2.0, a lot of coding and different kinds of objective. One uses rules, the other one, it learns on its own. So that's why it's called AI.

9:14But the third one, you type in English, summarize this email, and then it just does it. So it's a kind of programming, but it's the best kind, because that means that we can all program without knowing any programming languages. And as promised in the next video, we're going to go through each type of software. And I'm going to explain exactly what it is using hands on examples. See you in that next video.

0:00Welcome back. In this video, it's all about that mental model. This is the first part of AI literacy that's literally going to be useful throughout the entire course and from now on into the future because the AI is here to stay. And this is really going to be part of your go-to as far as understanding these models. If you ever decide to go deeper, you have a great foundation. This is really solid thinking. And this mental model is super useful for this course.

0:32And in the next scale, we'll get into the other part of the mental model, which is prompt engineering. So for now, all you need to think about is software 1.0 is classic programming. We write the rules. If this, then do that. It's basically coding or programming. The behavior comes from the code. And if you want to change that behavior, then you change the code. That's programming. And when you're thinking about software 2.0, now it's machine learning and deep learning.

1:02The software moves into learning from data. We define a model, we give it some data, and then we train it. So it trains, meaning it figures it out on its own, and then it gives its own rules based on its own learning. If you want to change that behavior, that means changing the data or retraining it. Software 3.0 is where the prompt engineering lives. That's why we're learning that in the next skill. We program with natural language using prompts.

1:34And all I'm saying there is we're typing in English into a chatbot. That's it. We don't modify code. We don't train the model on large amounts of data. We simply prompt the model. We can use context and instructions. And that's what we're going to learn in the next skill. It's called context role expectation. And we can use small examples and maybe even snippets from an email or some kind of subject. Like look at this article and explain it to me like I'm five.

2:04Things like that. And really just think of prompt engineering as steering a chatbot to help you get the output that you want. So prompting conditions the model. Or even simpler, prompting is how you get what you want from the chatbot. Think of it like that. And again, traditional programming, there's no learning. You're telling it what to do, right? So you code and then you get a behavior. This is called programming by hand or coding by hand. And if you want to change that behavior, well, all you have to do is edit the code.

2:39And I'm going to show you how to do this in Google Colab. And if you're wondering what that is, is all you need is a Google account like Gmail. And then you can start coding in something called Colab. I'll share the link below in case you want to use it. But what we're going to do when we start coding, we can say something like if x, then y. And you're always going to get the same output. That's the goal. I could say one plus one equals, and then you're always going to get two.

3:09That means it's deterministic. And a really simple way to say deterministic is the same output every time. That's all that we mean by deterministic. The reason I'm saying deterministic is, well, generative AI is often described as non-deterministic. You type in the same prompt, you might get something else every time. Or maybe not every time, but sometimes. You can't really rely on it. It's unpredictable. That's really the easiest way to think of it. OK, so here's Google Colab. You can see that I'm logged into my account.

3:40So if I type in colab.research.google.com, you'll get a notebook just like this. Let me show you. You'll get something like this. I have a bunch of these. And so you would just click on New Notebook, and then boom, there you go. It's located inside of Google Drive under the Colab Notebook. So once you're done with this and you want to check it out again, well, that's where it lives. And you might be asking yourself, well, why aren't you using Microsoft 365? Well, there's no real programming application in that. And this is just the fastest, easiest way to program in a web browser.

4:15And it's very popular in data science. This is what we use to build our own AI models, even. But it's super shareable, and it's really easy to get started. There's no setup at all. And that's why I'm showing you. I promise we're going to get into Microsoft from now on. But this is just a browser type of programming environment. So I can say, like I said earlier, one plus one. And then I can hit this play button, or I can do Shift-Enter, and I'm going to get the output, which is always going to be the same, meaning it's predictable or deterministic.

4:47Let me show you. I'll hit the play button 10 times, and it's keeping track over here. That's 10 times that I did that. It's not changing. So traditional programming, let's give it some logic. So I can say my underscore number equals. If I run this, nothing happens. But if I enter my number, it's always going to tell me that that's two, right? Always. And that's all we're doing. We're going to save this output, one plus one equals two, into a variable, and it's always going to know that. So it's not really thinking on its own.

5:18We're doing the thinking. We're telling it what to do. So I could say if my number is greater than one, return true. We could print true. So let me show you how this works. And then it does it. So it's thinking if whatever number here is greater than one, and I can change this to two and run it, and it's the same thing. This is programming, so it's not really hard, and you definitely don't need to memorize this. This is just how it works. And the reason I chose Python is because Python is very similar to English.

5:51You can even create a simple game by saying if input and saying guess a number between one and five, and then I could say whatever number that we type in there, if it's greater than, or we could say equals, let's say my number is three, then we could say print, you win. And that's a game. Let me show you how it works. So I have to guess a number. I'll guess one. Nothing happens. So you can tell it. So else, so I'm going to say else, print, try again. So now we have a game, and it's not thinking at all.

6:22This is us saying if the number that we type in is three, then you win. Otherwise, try again. So let me do one again, and then I'll say try again. All right. Let's try it again. And let's add three. And it's still saying, okay, one through five, three, what's happening? And this is important to understand. Software 1.0, if we don't set it up correctly, it'll just crash. So in this case, you don't need to know why. I'll explain it, but we need to think about all of this stuff. We're the ones in charge.

6:52With AI, it's a little different. So here I'm going to say int. That just means think of it as a number. And this is why. When I type this in, it's thinking of it as text. You can't add text. That's it. So if I type this again, it won't work. But if I want to fix this, all I have to say is like, don't think of this as text. Think of this as a number. And the minute that you do that, it works. Because now it knows, okay, I'm thinking about this number that you type in. So three equals three, and that's it.

7:23So I'm not going to get into programming, but what we learned here is something very important. It returns the same output every time. If we did a good job programming, that's it. If we don't do a good job, it'll break. And that's how that works with generative AI. If we don't do a good job programming, it'll straight up lie to you. And lying is a strong word. That's why we say hallucinating. It's just making predictions. So I'll show you how that works. When we talk about the hot dog and hamburger classifier, that's where it gets really revealing.

7:55It'll sometimes think that I'm a hot dog. So it's kind of lying to me, right? It's hallucinating. I'm definitely not a hot dog, but it thinks it is. What's up with that? Here, it won't say, oh, you're, you know, you're a hot dog. It'll just break because it's not thinking, it's not learning. And so artificial intelligence is pretty special in that way. And on that note, I'll see you in the next video where we create a vision model. That means we're going to create a deep learning model using the hamburger and the hot dog. There's no coding.

8:25This is actually going to be a lot of fun. It's something that you can do. None of this stuff that we just did here. We're not doing this again at all in this course. This is just to set up how the AI thinks. Now you know that there's no thinking going on here. This is just coding. In the next video, we'll get into the thinking, the artificial intelligence.

0:00In the last video, we explored software 1.0 using Google Colab. And if you had any interest here, the good news is the vibe coding course can teach you programming and learning to code is a super powerful skill. If you've ever felt like, oh, I can't learn to code, vibe coding is really a game changer. Not only can you code using AI, it'll code for you, but it makes learning coding fun and super easy. It's never been more accessible. And the myth here is that, oh, you shouldn't learn to code.

0:33Well, that is not true. Learning the program is still super, super relevant. So if you have any interest, definitely check out the vibe coding course here with chat, GPT, cursor, and TDD test driven development. And that just means that we're going to add little tests. So it's test driven development. Again, you would start here and then go here. And then if you wanted to take it even further, you can go here. So we have a little mini path for you. And now we move on to software 2.0, and we're going to create a tiny classifier.

1:06We're going to teach it. In a sense, it's going to teach itself by training. And this is software modeled after the brain because it uses these neurons, these artificial neurons. And that's what we call an artificial neural network. And that's where artificial intelligence comes from. And so there's this concept of weights. These rules are created using weights. So each one of these neurons that you see in this picture are literally weights. How important is a piece of data compared to all the other pieces of data?

1:37So this one might be heavier than this one, and so on and so forth. So that's how it learns. That's not important. But instead of rules, it's using these weights. So what we do is we train the model, and then we get that behavior. And if you want to change that behavior, you have to train again or do something called fine tuning, which is training, but for a specialized task. For example, chat GPT and many of the models that we use for chat bots are general purpose.

2:07But if you wanted it to be an expert, a really good expert at one thing, you can fine tune it by giving it data on one thing. For example, let's say that we have a hot dog and hamburger shop and that we don't make real food items, but we make stuffed animal toys. And here this is a magnet that we make, and we need to find imperfections with our specific product that doesn't exist on the Internet. At that point, you would create a ton of these little images from every angle possible.

2:40And we're going to do that in this video, by the way. And once you do that, it gets really very specific. And then that's a form of fine tuning. Now you have an application that can find a flaw in your hot dog. Be like, oh, that smile is not just right or it's missing an eye or one of the feet is bigger than the other or the magnet is missing or there is no tomato in this thing or there's not enough sesame seeds, which is actually glued on rice, by the way. This is super interesting.

3:12So it can become a specialist at something. And that's what fine tuning is. And to build that classifier, we're going to use something called teachable machine. What we're going to do is create different samples, meaning photos of either hot dog or hamburger. And so each class means a hamburger is a class and a hot dog is a class. But you could also use your hand. So you could use your fingers to form a one or two. And that would be instead of a hot dog, you could do one instead of a hamburger.

3:42You could do two. So this way you can follow along and we're not going to write any code. The model is going to train and then learn what a hot dog and hamburger is based on the samples that we give it here. And that's called live classification, because once we're done, I could show it a picture and it'll tell me, oh, that's a hot dog or that's a hamburger. And that's what we're going to do. So in essence, software 2.0 learns from data. In this case, examples that we're calling samples. So that's how it does it.

4:13And it changes the weight during training. So there's tons of these neurons and it'll change the weight of a particular shape. Let's say that we're looking at the hot dog or the hamburger and it looks at this yellow piece right here. How important is that? How important is this? How important is that? So it looks at these shapes and features and patterns and gives everything a weight like, OK, the collection of all the weights means that this is a hamburger. Right. So that's sort of what it's doing.

4:44Again, it's not important to understand the very minutiae detail because it's going to be very clear when we do this together. So we're going to use something called teachable machine. You can see here that we're using images. So it's me or me plus a dog. But it could also be audio or it can be poses. And so audio is interesting because what you're doing is taking a picture of the wave and then there's still pictures here. You can have a sorter. Right. So wherever the arrow is pointing and here back to images.

5:14So let's do this together. Let's just click on get started. And I'll be sure to share this teachable machine dot with google.com link with you so that you can follow along. So I'm going to click on get started and we're going to choose an image project. But as I said, you see the audio, it gets turned into images. So it's really just images as well. So images, images, and these are images, too. The only difference here is that these are 24 frames per second or maybe 30 frames per second. So here we're just using one image at a time.

5:46This is like a ton of images. So that's the only difference. This is all considered computer vision. So let's click on image project. We're going to click on standard image. And now here we have our model. So I'm going to say that this is a hamburger. I'm going to say that this is a hot dog and now we need to give it images. So the easiest way to do that is to use our webcam. So I'm going to crop this and move myself out of the way. There we go. So now I can hold up the hamburger and we're going to record samples and I'm going to hold it in different positions.

6:20OK, so I'll start like this. Let me put it in the center, put it right here and I'll do 200. Why not? All right, 209. Not bad. So let's do the same number for hot dog. So I'm going to go over here and we have to do the same thing. We're going to do webcam and then I'm going to crop this and move it over here to the side and then hit crop. OK, so now we have the hot dog. Let's go ahead and record 209 samples. I'll talk about why that's important. OK, wow, 209. Perfect. OK, so let's close this.

6:51Now we have 209 and 209 of each class. In this case, a hamburger and a hot dog. If I had nine images of hot dog and 209 of hamburger, that would be bias. And so bias means that it's favoring one class over the other. So imagine if you have a lot of data on one thing on the Internet and not about another thing, then chat GPT or these general models that were trained on the Internet would have bias. So that's where that comes in.

7:22So not only is bias favoring one thing, but it could also guess and guess wrong or hallucinate or lie to you or give you information that is completely wrong. But it's doing it confidently. This is kind of how it does that. It's one of the reasons, at least. So now we're going to train the model again. I'm not teaching it anything. Once I hit train, it's going to start learning and it's going to do that by going over the images a bunch of times. So here's going to do it 50 times. All right. So now what I need to do is hit crop and move this up here and I'm going to click on crop.

7:57And now I'm going to hold something up. But you probably noticed that it's saying that it's a hamburger 100 percent and it's not right. OK, so how about if I get in there? I think I'm a hot dog. OK, hamburger. Perfect. Got that right. And hot dog 100 percent. It's really good at this. But if I take this away, then it's confident that that background is a hamburger. And that's not right. So let's go over here and add a class. And I'm going to just say none. And now I'm going to go to the webcam, do the same thing and crop this over here to the same place.

8:32Done cropping. And now I'm going to get 209 images of a background. And you might be wondering why we do that. Well, I'm removing bias because now it knows you take one off. OK, so I'm just going to grab this one and delete it. So now we have 209 samples because it only has two classes. Let's train that model again. And like we said before, it does this a bunch of times because it's learning. So when it's training, it's really learning the shapes and patterns of this. Oh, this background. It's just this way.

9:03OK, cool. So then it learns that. And if you click on advance, you can see this is kind of like what machine learning and deep learning is about. These are all the hyper parameters. If you want to learn more, if you're just curious, you can hover over this and it tells you more. Great way to learn. Now it says 100 percent background. OK, so we stopped it from hallucinating. So now you're learning quite a bit about AI. Boom. Look at that. It knows. How about if I do this? It still knows it's a hamburger. It's doing great. How about hot dog? Perfect.

9:33Look at that. Now our computer vision model is doing fantastic. If you don't have a hamburger or hot dog, you could just do one. Maybe move your hand back here so that it could see that. And two. When I do this, I think it really thinks I'm a hot dog. It's like, oh, you're definitely a hot dog now. And that's because we're showing it something that it hasn't been trained on. And that's another reason that we could have these models that don't perform well. And you can do other stuff like you can download all these samples or you can save them to drive or you can go over here and export this model.

10:05Here's a bunch of code that runs that model. And this is really what we're doing. But we did it with zero code just to give you an understanding of how this works. And these weights that I was mentioning before, it's just looking at the hot dog. Put a little mustard on there and it looks at this rounded shape. It sees that it's kind of long. It has a squiggly line for the mustard. Those all become features and the different features become more important than others. How about if there was a square one? Well, that would not be part of this.

10:37So if it saw that, it would not give it a lot of importance because it doesn't see that very often. Maybe there is a square somewhere or some pattern that is not part of this hot dog. It would just learn the important features of this image so many times and create really accurate weights. And that's what gives you that really high probability. So it was doing 99 percent and then even went to 100 percent. So it's never going to say false or positive, right?

11:07It's kind of like a scale and that's what it's called a probability. It's never 100 percent sure, even though it says 100. It's always kind of like shifting up and down. Let me show you like this. When you have partially the image, it thinks it's the background. Now it's like, oh, it's like 40 percent, 30 percent and then 70 percent, 80 percent. Those are probabilities. And in the real world, when we're dealing with these models, they're not always 100 percent. They're going to be OK. The next word in this sentence should be, for example, the cat sat on the airplane.

11:45Probably not. That would probably have a low probability. So that's why we get different words at different times. It's always looking for the probabilities, which are never 100 percent sure. Another reason why these models can make things up. All right. So enough of software 2.0. We've covered 1.0, which is traditional programming. And now we trained our own computer vision model. That's 2.0. And in the next video, let's talk about 3.0.

0:00All right. So now you know what software 2.0 is and specifically deep learning. It really just learns from images or data, really. And we label that data and then it learns patterns and assigns importance or weights to each one of those patterns. And ultimately it can tell us if it's a hot dog or a hamburger. That's a classifier. All right. So now what? We talk about software 3.0. This is the exciting part. This is what Copilot is. And the way that we interact with this type of software is by prompting or typing into the chat bot. And we use plain language. In our case, we're going to be using English and that's a form of programming. And that's the part that people usually overlook, but it's probably the most important part. If you talk to it like a human, that's where the trouble comes in. It's not going to understand all of the details, but when it talks back to you and it replies, you'll believe it. It's so believable. It's so confident, even when it's guessing. And I will show you how to program or how to improve your prompts, right? All we're doing is learning how to talk to it so that it doesn't make stuff up.

1:14So it doesn't just give us a bunch of weird sounding outputs, which is very common or outputs that sounds like everybody else who's using the same model, right? We want to be unique and there's a way to do that. So here, the rules are in prompts. And if you want to change the behavior, you need to prompt or steer that model using plain language. So this is pretty easy to learn. Prompts condition the behavior with context. We'll talk about context very soon.

1:45And that's how we steer it. We use a context and sometimes examples. These are patterns. So it learns from these patterns. And the goal is to have shorter, steadier, paste ready outputs instead of trying multiple times to get what you want. And also knowing if the model is guessing or not, we can program it in a way using plain language so that it says, I don't know. That's pretty cool. And actually that's a little tricky to do if you speak to it like a human, but when you're doing prompt engineering, it will tell you. So how do these chatbots or large language models, which is again, software 3.0 or generative AI, it's all the same thing. How do they produce these outputs?

2:26We talked about probability, meaning it's not zero or one, meaning it's not false or true. It's some kind of a percentage and somewhere in a range. And these LLMs or large language models pick the next token from a probability distribution. So that means it has choices. So the cat sat on the And then I'll have one, two, three, and it might say hat, Matt and rat. So each one of these will have a different probability score, but the highest one in this case is that 90%. And that's hat. So it'll pick that one. And so when I say tokens, it turns the words into tokens. Tokens is just the way that these models work. And a token is not always one word. It can be a word chunk. So just think of it like that. And it's not important to understand what a token is other than a representation of a word. And usually it's in the form of a number, right? So we're turning the words into numbers because that's what computers understand. They don't really read text. It's all numbers for them. So how do we turn the word into a number? You turn it into a token and that's it. Another concept is temperature. This is something that you won't see, but if you do, the higher that it is,

3:42the more variety you'll get. And the lower that it is, the more stability that you will get. These are things that are set behind the scenes and you can't really see it, but prompt engineering can help you adjust the temperature through these plain language prompts. And we're looking for consistency, which means you reduce the degrees of freedoms or reduce the choices, right? And that's where writing really good prompts come in. You want to specify the shape, meaning the labels, the count or the format, and then you can add one or a few examples to make sure that the output is exactly what you want. So we're getting there little by little, and here's where we're going.

4:23We're going to be programming with English, but I promise it's not programming like code, right? It's not something that you need to spend months learning or weeks learning or learning a new language. You already know the language that we're speaking. In this case, it's English. Whatever language that you speak, that's the programming language that you can use. It's pretty awesome. So programming with English, we're going to use something called CRE, and that's context role expectation. It's pretty straightforward. What is the context? The role, I want you to act like an expert writer or something like that, right? So that's what you want the model to be, and I want the output to be three bullets, and I want it to have a title.

5:06So when you're that specific, right, this is maybe a Teams update that you want to create, and you'd say, I want you to act like a project manager, and then the expectation here is a Teams post with a title and three bullets. If you do that, you're programming with English. Instead of saying, summarize this or make this look better or edit this, right, that's called naive prompting, and the shots are just examples. So there's something called zero shot, one shot, and few shot. We'll talk about that. And excerpts are just the facts and examples or the style and format. This will become very clear when we start to do prompt engineering in the next skill. And as we progress through the different skills, we'll apply different techniques to whatever co-pilot application that we're using. Maybe we're working with Word. Okay, let's use CRE and really get a tight output. Then maybe we move to PowerPoint. And at that point, you might want to add a length cap because you don't want to have paragraphs in a slide. You want short, concise sentences, maybe a title with only four or five words. You can add caps on those things. And while it does it automatically, it may not be what you're looking for. So if you want a higher degree of control, that's how you do it. Prompt engineering.