Explore AI & Prompt Engineering Fundamentals

0:00Hello and welcome to the AI Prompt Engineering with ChatGPT, Gemini, and Claude course.

0:08I'm excited that you're here. Before we get started writing any prompts, we need to have

0:13a good idea of what AI is so that we can leverage how the prompts are generated to get the best

0:20possible outputs. But first, my name is Jonathan Barrios, and in this AI Prompt Engineering with

0:26ChatGPT, Gemini, and Claude, I want to lay out the scope of this course. So what's in it and

0:33what's not in it. So what's in scope? We'll understand how to craft prompt inputs to get

0:39the best possible outputs. And we'll make sure that we don't add any sensitive information into

0:45the models. We'll talk about why that's not a good idea. That sensitive information is also called

0:50PII. We'll break that down, and I'll be sure to unpack that throughout this skill. Super important.

0:56And we'll develop a reusable prompt toolkit for you. And if you are doing this for your job,

1:02then your team, this is really valuable. And here's why. It might be tempting to just go back

1:08to talking to these language models like ChatGPT, like you're talking to another human. But when you

1:14do that, you go back to the not great outputs. And as you progress through this course, you'll

1:19understand exactly what that means. That could also include completely wrong information because

1:25these models are really confident when they're wrong. So what's out of scope for this course?

1:30We're not going to talk about AI engineering. This is all no code and no tutorials on how to build

1:37your own AI chatbots. And we're not even going to use APIs. That means no use of computer to

1:43computer connections. And this course is part of an AI miniseries. And here's where we're at now.

1:50If you're thinking about launching your own AI product or you're an executive or a leader,

1:56this is a great course to take. And we start with an idea and take it all the way to deployment.

2:01In this course, we're really talking about prompt engineering, meaning how to program in English.

2:07We'll talk more about that. But this goes hand in hand with vibe coding, meaning vibe coding is for

2:14non programmers that learn how to program or leverage programming using AI. But if you're

2:19already a programmer, you might want to take this course, which is agentic coding. So you can think

2:25of it as a coding assistant. Here, you're heavily relying on the AI to do the coding for you. But

2:32here you're doing the coding, but having the AI help you. They're different roles. Eventually,

2:38vibe coding should become more like agentic coding. So if either one of these are interesting,

2:43I suggest taking this one first. And if you don't have a programming background, this one second.

2:49And if you do have a programming background, take this one next. And if you want to learn AI

2:54productivity, generally speaking, this is a great course. However, if you're working with Microsoft,

2:59Copilot is embedded into the whole ecosystem. And finally, if you are deploying and building a model

3:06for your organization, you might want to consider the AI compliance and privacy for the non technical

3:13audience, HR, legal and it. And when I say it, I'm talking about the non technical it fields.

3:20But if you are in the technical field, then this companion course, which is the same course,

3:26but for developers is a better path. So now that you know that this course is part of an AI mini

3:32series, let's move on to a little bit about me. My name is Jonathan barrios. And I teach data science,

3:38machine learning and deep learning. And these days, that term is now called an AI scientist.

3:44When you combine data science with AI engineering, this is a term that's going around. I've worked

3:50with AI and built many different kinds of applications, including SAS websites, AI

3:55applications and streaming platforms. I've been teaching for quite some time at top platforms,

4:01such as treehouse, thankful and check. And now I'm happy to be part of the CBT Nuggets team,

4:07you can find me on LinkedIn, my website, or x.com. And the real quick note about x.com,

4:13a lot of people don't like x and a lot of people love x. So really, for us, when we're talking

4:19about AI science, or learning AI, or prompt engineering, this is where all the best minds

4:25of AI are. And this is an easy way to stay apprised of all of the developments. I don't

4:31follow anybody else, I only focus on AI. And that's it. Because social media in general is

4:39just a time suck. But AI on x is really a great is a fantastic way to stay updated. But you have

4:48to curate your social media or it will waste a lot of time. And this is the first skill in the

4:54course where we're going to explore AI and prompt engineering fundamentals. We're really focusing

5:00on AI here. What is it? It's not just one thing. There's many different parts to it. So if you

5:06already have a good handle on AI, you might want to consider skipping this one and then jumping

5:11into the next skill. But to make sure let me show you everything that we're going to cover in this

5:16first skill. Number one, we're going to talk about AI and how it's not only one thing. It's many

5:24things. Then we'll talk about LLM. That's a type of AI. And these are chatbots. But you need to

5:31understand how they predict the next word, how generative AI works. So this is really about gen

5:36AI. And this is what everyone thinks AI is. But as we discussed here, it's more than one thing.

5:42And then we'll build a mental model, making it easy to understand what AI is and how to apply

5:49to the different types of AI. And here we're going to explore the three main types. And you'll even

5:56build your own AI model to distinguish between a hamburger and a hot dog. This could be numbers or

6:05anything, but this is called a computer vision classifier or a simpler way. We can just call it

6:11an AI model. All right. I think that's it for this video. And now that you know what we're going to

6:15cover, I'll see you in this first video on what is AI.

0:01In this video, we're going to answer the question, what is AI?

0:05Because AI is a big umbrella term, meaning it's not one thing.

0:11Another way to say that is that chatbots are just one slice of the pie.

0:16So day-to-day companies still use a mix of different kinds of AI.

0:20But to the general public, a lot of people still think that chatbots is the only type of AI.

0:26And in the context of prompt engineering, understanding the other types of AI, not just chatbots, is really the secret sauce to better outputs.

0:37Here's why.

0:38Like I said, generative AI, this is the chatbot part, is only one part.

0:44But AI, in general, is that umbrella term that we were talking about, which includes rule-based AI, search, machine learning, deep learning, and generative AI.

0:56But we could just simplify it to these three right here.

0:59Because rules are an older version of AI, and search is something that's also done in these three terms.

1:06So really, you just want to remember machine learning, deep learning, and generative AI.

1:11In short, these models learn from data.

1:14They're not like traditional programming.

1:16For example, if X, then Y.

1:19When we have that kind of programming, it's always going to return the same answer.

1:24That's called deterministic.

1:26These models are non-deterministic.

1:28You'll get slight variations in the output.

1:31So that's the first hint.

1:33These models are non-deterministic.

1:35And that's why sometimes they make up information.

1:38More on that soon.

1:40Another interesting thing to consider is that they all learn in different ways.

1:44Here are different learning styles.

1:46Supervised, unsupervised, and reinforcement learning.

1:50And when I said we're going to build our own computer vision classifier, that's supervised learning.

1:55Meaning we give it different images of hamburger or hot dog.

2:00And we show it an image, and we label it as hamburger or hot dog.

2:05That's supervised learning.

2:07Unsupervised learning is when you give it a bunch of images like this or this.

2:12I'm not defining them.

2:13The computer has to figure it out.

2:15So that's the difference between supervised and unsupervised learning.

2:19And we'll build our own classifier using no code at all.

2:22So it's a simple, fun way to understand the complexity of machine learning and deep learning.

2:28Specifically deep learning.

2:30We'll get into that. It's really fascinating.

2:32And finally, reinforcement learning.

2:34Think of it like Google Maps and an AI agent trying to figure out the destination.

2:41If it hits a wall, it gets a penalty.

2:43If it gets closer to the goal, it gets a reward.

2:46That's reinforcement learning.

2:48You could think of it as a robot trying to navigate a maze.

2:52And then we have modalities.

2:53Meaning these one, two, three styles of AI can do a bunch of different things.

2:59They're adaptable.

3:01But the takeaway for right now is that chatbots, which are generative AI, are simply one slice of the pie.

3:09But instead of thinking about them as machine learning, deep learning, or gen AI,

3:16we can think about them in terms of software.

3:19And you might be asking yourself, well, why am I going to think of software?

3:22The easy answer is this is going to give you a deep understanding of how to write a good prompt.

3:28Because you may not know it, but when you're writing in English, that's programming.

3:33But you don't know it.

3:34So that's software 3.0.

3:37But let's start from the beginning.

3:38What is software 1.0?

3:40We'll explore these in individual videos.

3:42And I'll show you step by step what it is.

3:45Software 1.0 is traditional programming.

3:49These are rules by hand.

3:51Meaning if X, then Y.

3:53And I'll show you programmatically what that means.

3:56But you can think of a game.

3:58Once you get to number 10, then you enter level 2.

4:01That's it.

4:02That's traditional programming.

4:03That's the software that we've been using before AI showed up and took the world by storm with chat GPT.

4:10At that point, everyone started to understand, oh, there's this new kind of thing.

4:14A chatbot.

4:15It thinks like a human.

4:17It's artificial intelligence.

4:19That's all software 2.0 and 3.0.

4:22Let's break it down.

4:23Generally speaking, software 2.0 is software modeled after the human brain.

4:29And that's why we use terms like neural networks.

4:33Because the human brain has neurons.

4:35And they're connected.

4:37And that's a network.

4:38And that's how we think.

4:39We connect our neurons.

4:40And the neurons will just pass information from one to the other to the other.

4:44And that's how we remember things.

4:46That's how we classify things.

4:48Like if I show you this right now, what is the first thing you think?

4:51Hot dog.

4:52You just use a neural network in your head to classify a hot dog.

4:57I'm going to show you the software 2.0 of doing the same thing.

5:01But as you can see, they learned it from us.

5:04So when you're talking to chat GPT, that is based on software modeled after the brain.

5:10And the technical stuff is that we train it and you fine tune it.

5:13And then you get this behavior.

5:15That's not so important.

5:16Just remember that they learn by themselves.

5:20Here, software 1.0, we tell them by using rules.

5:24And we code it in programming languages like Python or JavaScript.

5:29If you've heard of those, those are programming languages.

5:32And that's software 1.0.

5:34Software 2.0, while it may use a programming language to set it up,

5:39the learning happens because we created AI.

5:43Artificial intelligence where it learns on its own.

5:47And that includes machine learning, deep learning, and generative AI.

5:51Machine learning is the older style.

5:53And this is mostly algorithms.

5:55Deep learning is really where it gets into neural networks.

5:58And generative AI is built on top of deep learning.

6:02And, again, this is not the important part for prompt engineering.

6:06But now you know the history.

6:07Deep learning is really the one modeled after the brain.

6:10And to get to generative AI chat bots, well, that comes from 2.0 deep learning.

6:15And to keep it simple, just think of it like this.

6:18Traditional programming is 1.0.

6:20That's the coding that everybody was doing.

6:22Software 2.0 is when we build these artificial intelligence models that think by themselves.

6:29So you might be asking yourself, well, what is 3.0?

6:323.0 is rules that are prompts.

6:36So here we code it by hand.

6:38Here we use something called weights.

6:40But this is how important a piece of information is.

6:44And then the model learns on its own.

6:47Let's say by itself.

6:48Here we're using plain language or prompts to steer the model into behavior.

6:55Here it learns by itself.

6:56But here we can tell it what the behavior or the output that we want by talking to it.

7:02So now it should start to connect.

7:03You're literally programming in English.

7:05And if you're using a different language, well, you're still programming in plain language.

7:10But you're programming.

7:12And that's the takeaway.

7:14That's why I spelled it out like this.

7:16When you talk to a chatbot like a human, you're making a huge mistake.

7:20And that's why you're not getting awesome prompts.

7:22It's not a human.

7:23You're programming it.

7:25So you need to think like a programmer.

7:27And I'm going to make it super easy.

7:29You're not going to learn coding.

7:31You're just going to learn the fundamentals of programming so that you can use that in plain language to get awesome prompts.

7:38And we're going to use things like CRE, shots, constraints, retrieve snippets.

7:43These are all technical terms for right now.

7:45But this is going to become very clear by the end of this course.

7:48And your prompts are going to be awesome.

7:50So let's review everything we've done.

7:52So what we're doing here is building a mental model.

7:56Just an easy way to think about AI.

7:59So AI mental model that I want you to carry through the course.

8:04Software 1.0 is the classic programming.

8:07If this, then do that.

8:09Traditional programming, this is coding.

8:11Software 2.0 moves from rules into that neural network or software modeled after the brain.

8:19That's all you need to know.

8:20It learns from the data by itself.

8:23That's why it's called AI.

8:25And while it does train and you can fine tune it to get that output or behavior, these are technical things that are not important for the mental model.

8:33Just remember, regular coding, software modeled after the brain for 2.0.

8:383.0 is where prompt engineering lives.

8:41We program with natural language and that's it.

8:44It's built from the software modeled after the brain.

8:48And that's where these chat bots come from.

8:50So you're programming in English or whatever plain language you're using.

8:55And in the next video, we're going to make software 1.0 concrete.

9:00We're going to unpack it and I'm going to show you some demos.

9:03Then we'll move on to 2.0.

9:05See you in the next video.

0:01In the last video, we clearly outlined the differences between 1.0, 2.0, and 3.0.

0:08The takeaway here is not necessarily that we're using machine learning, deep learning, and generative AI, such as chatbots,

0:15but it's to really understand that historically we had traditional programming, and now there is software modeled after the brain, which is 2.0,

0:24and we have to build the models and train them and do all this stuff.

0:27However, software 3.0, which is built on 2.0, is programming in plain language.

0:34That's really the takeaway.

0:35And what we want to do right now is to make software 1.0 as clear as possible.

0:41In this world, behavior lives in the code.

0:45That means that you create rules by hand, and you're programming, you're coding, and then it creates the behavior.

0:52So the output of your code is whatever you created by hand, by coding, and then you get the output, which is the behavior.

0:59If you wanted a different behavior, you would simply just edit the code, and that's it.

1:05So we edit with rules.

1:06That could be if X do this or else do this.

1:10There's loops, there's comparisons, all kinds of stuff.

1:13But the takeaway is that there is no learning.

1:16They just follow our instructions, just explicit instructions that the computer follows exactly.

1:23And like I said, you change the behavior by editing the code.

1:27And I'll show you in something called Google Colab.

1:30All you need is a Google account, and you get this for free, and you can run code.

1:35But you don't need to understand the code or you don't need to run the code.

1:39I'm going to do that just to make sure that we understand exactly what software 1.0 is.

1:44And we'll use a demo.

1:45Maybe we'll use the greeting based on the hour.

1:48But then again, I could code whatever I want because whatever I write is going to be software 1.0, meaning I'm writing the rules.

1:56So I could write whatever rules I want, and then I'll get the behavior that I'm looking for.

2:01If my code is good, you have to be a good programmer.

2:04You have to understand programming.

2:05That's the only caveat for 1.0.

2:07And here is Google Colab.

2:09And I'm signed in because you can see my photograph here in the top right, and that's all you really need.

2:15I could share this with you, and I am going to share it with you in case you feel like programming or running it just to play around with coding to get a good understanding of what coding is.

2:25However, that is completely optional.

2:27As long as you have a Google account, it'll work.

2:29But if you don't, there's no need to open this up and try it yourself.

2:34So here, and this is called a code cell, and the reason we're using Colab is that this is the easiest way to show you software 1.0.

2:42Normally, there's a lot of setup.

2:44There's just a lot of technical stuff that goes along with it.

2:47But with Google Colab, I can clearly demonstrate software 1.0, and that's why this is such a good tool for this part of the course.

2:55So the first thing that people do when they learn programming is this.

2:59Hello, world.

3:00And if I run this, I can just push the play button or hit Shift-Enter, and it executes this code.

3:06But first, it has to connect to the cloud, and then it executes.

3:10No matter how many times I run this, I ran it three times now, and you can see it's counting the number of executions.

3:18That means that this is always going to be the same.

3:21So I could add a comment, which is what programmers do, and I could say, software 1.0 is deterministic.

3:29And if you looked up that word, it just means that we're going to get the same output every time.

3:34So if you look at that definition in this context, that's what it means.

3:38ChatGBT, on the other hand, may output different text every time you run it because it is non-deterministic by nature.

3:46Sometimes it might do it, sometimes it may not.

3:49We'll explore that soon in software 3.0.

3:52But what if I wanted to do some programming and I wanted to create a greeting based on the time?

3:58I would literally have to input my name.

4:01So I could say name and this input function.

4:05You don't need to know what it does, but I'm going to show you right now so that it's no mystery at all.

4:10I'm going to run this cell and it's going to ask me my name.

4:13So this little line of code means ask the user the name, save it in this variable name, and then print it.

4:21And then it prints it down right here.

4:23So it didn't learn anything. It just followed my instructions.

4:26I created an input, meaning create an input form so that the user can type in a name.

4:31And I want you to save that under the variable name.

4:34That means that you can reuse that name.

4:36So I can print it. I could change it into uppercase.

4:40Let me show you. So I can say name.upper.

4:43I run that and it turns it into all uppercase.

4:46Again, whatever I tell it to do, it'll do.

4:49So it could also define an hour.

4:51The idea is that I'm going to write code and ask it to do something.

4:55So here I'm asking for the name and then I'll also ask for the hour.

4:59I want it as an integer.

5:01Again, this is not important, but this is how software 1.0 works.

5:04The reason I'm showing you this is because this is what you're sort of doing in English.

5:09Not exactly. That'll be clearer when we look at software 3.0.

5:13But here I'm doing 1.0 programming.

5:16When you do 3.0 programming, there's going to be some parallels.

5:20So that's why I want to show you at least one program.

5:24So you get a good idea of some of the little details that go along with it.

5:27You don't have to remember this or even fully understand what an integer is or how programming works.

5:33And here I'm just saying if the hour is less than 12.

5:37Now here you see this is explicitly telling it a rule.

5:41Then the greeting will be good morning because it's before noon.

5:44Else, if.

5:46And now here we're going to ask a different question.

5:49So if the hour is not before 12, then we could ask is it before 18?

5:53If so, say good afternoon.

5:56But if it's not before 12, if it's not before 18, then you can simply say good evening and you've covered all your bases.

6:03So how does this work?

6:04Now all we have to do is print the greeting and the name.

6:08So we know the name. It's whatever the user adds.

6:11So I'm going to add Jonathan.

6:12And then whatever the time is, is going to be based on whatever hour I enter.

6:16So let's try to do 11 to get good morning.

6:19So I ran this and I got an error.

6:21And that's really important to understand about software 1.0.

6:25When you make an error, it'll tell you.

6:27Chad GPT, when we're talking about software 3.0, if it doesn't know something, it'll lie to you.

6:32It'll guess.

6:33And that's called hallucination.

6:35That is the biggest difference that I can give you as far as the output between these two versions of software.

6:41So now I have to run this because I updated it.

6:44And then I'm going to say my name is Jonathan.

6:46And the hour, let's say it's 11.

6:48So that would be 11 a.m.

6:49And then I hit enter.

6:51And then now I can run this statement or this area to get the greeting and the name based on the hour.

6:57So my name is Jonathan and the hour is 11.

6:59Let's see if it works.

7:00Good morning, Jonathan.

7:02And it totally works.

7:03That's software 1.0.

7:05Not super exciting.

7:07Maybe a little boring if you're not a technical person.

7:10But if you are technical, I'll give you this code so that you can play with it, create your own.

7:15You can even go to chat GPT and have it write Python code for you.

7:19You can just paste it in here and run it.

7:22There's a lot that you can do that way.

7:24A lot.

7:25So if you like that concept, then take the agentic coding course or the vibe coding course, which is exactly what I just described, more or less.

7:34In the next video, we're going to talk about software 2.0.

7:37And we're going to use our friends, the hot dog and the hamburger.

7:41See you there.

0:00Welcome back.

0:01Now, let's contrast 1.0 rules with 2.0 learning.

0:07In software 2.0, you don't handwrite if this, then that.

0:12You collect samples.

0:14And we're going to use Teachable Machine to explore collecting samples.

0:19And the samples that I'm talking about are hamburger and hot dog.

0:24We're going to make this fun and super easy because this can be complex, but there's no

0:28need for it to be complex because we're just exploring software 2.0 and we're creating

0:35or teaching a tiny classifier.

0:38So when I say, what is this?

0:40You'll say hamburger.

0:42That's classification.

0:43If I show you the hot dog, you say hot dog.

0:46You're classifying what you see that sample as what you know it to be hamburger or hot

0:52dog.

0:53And what's fun about this is that we can train it and then we can test it.

0:56So we'll do live classification with a hamburger and hot dog.

1:00And the idea here, the real takeaway is that 2.0 learns from these examples.

1:06And we're talking about weights and what that means.

1:09We have these neural networks, we can call these neurons right now, and they're all interconnected.

1:15So each one of these has a weight, which just means how important is that information?

1:20And then these weights will just output a prediction, in this case, hamburger.

1:25So you show it a picture of a hamburger and then you do that a bunch of times, which we're

1:29going to do here with 20 to 50 samples per class, meaning one for hamburger, one for

1:35hot dog, and then you can create a prediction.

1:38And if that sounds a little bit like magic, well, it kind of is, but so is the human brain.

1:43The way that it works is really magical in that context.

1:47But let's demystify this.

1:48Let's make it super simple.

1:50We want to contrast 1.0 rules like we type it in, ask the name, ask the hour, output

1:58the greeting to learning.

2:00So now let's jump into Teachable Machine and build our tiny classifier.

2:04I should say teach our tiny classifier.

2:07All right.

2:08So here is Teachable Machine.

2:10This is from Google and you can do audio, video, images, which is what we're going to

2:16do.

2:17Or you can do audio or even poses.

2:19So let's click.

2:20Get started.

2:21We're going to do an image project.

2:22You could do audio or poses.

2:24The image and the pose project are very similar.

2:28This uses the same idea.

2:29This is just looking at a bunch of images in sequence, right?

2:33That's what video is.

2:34One at a time.

2:35Those are images.

2:36So let's click on this left one here.

2:38Standard image model.

2:39And now it wants us to gather samples, make or add some samples here or watch this video

2:45how to do it.

2:46Definitely watch this video if you want to learn more.

2:48I'm just going to take you right through it.

2:50This is going to be hamburger and this is going to be hot dog.

2:53So samples means give it images.

2:55So I cropped the image.

2:57That means that I can hold this up.

2:59Let me see right here.

3:00And then you can see a clear view of this.

3:02I'm going to turn it all around a little magnet, but this is how we're going to do it.

3:06And all I have to do is hold this down and it's going to take a bunch of images quickly.

3:10So we just need maybe 20 to 50.

3:16One more.

3:17OK, so that's good enough.

3:18And I've done the same thing here.

3:19We're going to add the hot dog.

3:21OK, so I'm just going to move it all around and give different images because it's learning

3:26little patterns.

3:27That's how it learns.

3:28All right.

3:29So I'm going to hold this down.

3:30Try to get the same number.

3:31I got two more.

3:32I'm going to delete two of these.

3:34All right.

3:35Now we have samples and then we train the model.

3:39But let me put it this way.

3:40When we're thinking about software 2.0, there's something called bias and this will translate

3:46to 3.0.

3:47And that's why sometimes chat GPT and other models give you incorrect information because

3:52of bias.

3:53So if you have a ton of images, hamburger, but very few of hot dog, it's going to favor

4:01this.

4:02It's going to say it's a it's hamburger when it's looking at a hot dog.

4:06Right.

4:07And that's incorrect information.

4:08So depending on its training data, which is a big chunk of the Internet, it could have

4:14bias because not all the information is equally represented on the Internet.

4:19That's why fairness is really important for AI.

4:22So just wanted to give you that background why it makes mistakes.

4:27Bias is a big part of that.

4:28That's why it hallucinates.

4:29All right.

4:30So let's train our model.

4:31But if you didn't want to learn more about what's happening here, how we're teaching

4:35this classifier, click on this video.

4:38For now, I'm just going to click train model.

4:41And here you can see that there's a timer and zero to 50.

4:45Now it's going pretty quickly, but it's doing it a bunch of times here.

4:49You could export the model, but now that it's looked at these images and trained or learned

4:54a bunch of times, we're going to look at the correct camera.

4:57And now I'm going to crop this.

4:59OK, so now it doesn't know what it's looking for.

5:02In fact, I should probably add another class.

5:04Let me do that.

5:05OK, so I've created a background here.

5:07And what I'm going to do is get 50 images of this background and delete one.

5:13And we're good to go.

5:14Now I'm going to retrain the model because we need to add the background without that

5:19background.

5:20It's continuously trying to predict one of these two, but it has no training data of

5:24the background.

5:25That's why it doesn't know.

5:26But now you can see that it's a background 100 percent.

5:30But if I look in here, it's like, are you a hot dog or you hamburger?

5:34It thinks I'm a hot dog.

5:35Well, it's good to know that look more like a hot dog than a hamburger, I guess.

5:38But now you can see that it's correctly predicting the background.

5:41You can see here 100 percent.

5:43Let's try out a hamburger, 100 percent hamburger, even if it looks at the magnet.

5:47Well, here we don't have a lot of training data for this angle.

5:51So we're so it's not as great.

5:53So we're introducing a little bit of noise.

5:55But if you look at the angles that we gave a training data on, it's learned this on its

6:01own.

6:02How about hot dog?

6:03Perfect.

6:04Moving around even this way, it's perfect on the hot dog.

6:08It has a bias for the hot dog, maybe because it's a cute hot dog.

6:11I'm not a hot dog, but I'll take that as a compliment.

6:14All right.

6:15So let's review what is software 2.0 software 2.0 learns by itself.

6:21So we were able to teach a tiny classifier, but really it's doing the learning by itself.

6:27All we did is give it these images and say, learn these patterns.

6:31And then I want you to be able to make a prediction.

6:35So that's what's happening here.

6:36And guess what?

6:37When you're talking to chat, GPT or any of these other models, it's predicting the next

6:43word.

6:44They call them tokens, which are like word chunks.

6:46So it doesn't do it by full word all the time.

6:50Sometimes it breaks it down into little chunks.

6:52So it's just making predictions.

6:54So when we say the cat sat on the and you probably would say hat.

7:01That's probably what chat GPT would say too.

7:03That is the most probable next token, but that's what it's doing.

7:06So instead of predicting hot dog and hamburger is predicting the next word after I sat on,

7:14well maybe we don't want to say it that way, but you get the point.

7:16We're predicting words, which are also called tokens.

7:20So software 2.0 learns on its own from examples.

7:24There's training fine tuning, and then there's behavior that comes out.

7:28And this is the output, which we said was predictions.

7:32All right.

7:33So in the next video, we're going to talk about software 3.0, and that's what we all

7:38know as generative AI or chatbots like chat GPT.

7:43See you there.

0:00Welcome back. In this video, we're going to explore the concept of plain language programming.

0:07And one of the innovators in the AI space is Andrej Karpathy, and that is the person

0:12that I'm borrowing the term software 1.0, 2.0, and 3.0 from. And a cool quote from Andrej

0:20Karpathy is, the hottest new programming language is English. Here's why. You can think of a

0:27prompt is a little program that you run at runtime, meaning you type it as a prompt and

0:34press enter. It's not code, it's English. Runtime is the moment of execution. For example,

0:42this is code, but I need to push this play button or run it. And this would be considered

0:47runtime. So then it outputs this hello world. And much like programming, this is the runtime.

0:55When you're talking about chat GPT. So you can think of prompts as runtime instructions

1:02without any code, but it's important to note that you're not changing the weights, meaning

1:07you're not changing the program that created this chat GPT, which is based on software

1:132.0. What you're doing is changing the behavior by giving it a context. You're steering the

1:21runtime to get an output or a prediction or a behavior. So you can think of it as constraints

1:27plus an example equals a pattern, meaning you can keep the facts in the excerpts or

1:33an example to teach the model, the style and the format that you want. So the goal here

1:38is shorter, steadier and paste ready outputs. You want to get the output the first time.

1:46Some people just prompt in hope and that's not the way to go. And you don't want to talk

1:51to it like a human. That's also not the way to go. You really want to think of prompts

1:56as plain language programming. First, let's talk about LLMs. These stand for large language

2:04models like chat GPT and others such as grok, Gemini or clod. And what they're doing is

2:11that they're picking the next token from a probability distribution. What does that mean?

2:17So token means word or it could be a word chunk. For example, R E L A Y. These could

2:25be two. But what we're saying is relay. So it could have two tokens to represent one

2:31word and it predicts just the token. So the next token from a probability distribution.

2:37So we could have the word or the token I, and then we could have a series of choices

2:43like I love or I ran or I met. And each one of these are going to have some kind of a

2:50percentage associated with them. And it usually picks the highest one. So there's concept

2:56of temperature, which we'll explore later. But for now, to give you an idea, the higher

3:01the temperature, the more variety, the lower the temperature, the more stability. Also,

3:07on top of variety, you'll get some pretty wild answers. It kind of gets more creative

3:12wild. And here you get a more plain and stable output, less creative and less wild. But when

3:18I say that consistency equals reducing the degrees of freedom, that means that if your

3:24prompt is vague, the model gets to guess and you don't want it to guess because it's probably

3:31not going to be what you're thinking. So the less freedom equals a steadier output, or we

3:37could say more consistent. So we want to reduce the guessing or the degrees of freedom. You want

3:42to specify shape and you want to use techniques like one or a few shot examples. And this is

3:48going to steady that voice. These are all techniques that we're going to explore in this

3:52course, but I'm just giving you sort of a preview of what the goal is. We want really good outputs

3:59and there's different things that you can do. Essentially, we want to reduce the degree of

4:03freedom, meaning we don't want it to guess. We want it to know exactly what we want. That's how

4:10it can give us what we want. In practice, this means you control the behavior by tightening

4:17the program. So here's where we're going. We're programming with English and that means using

4:23techniques like CRE. We want to give it a clear context, a clear role and clear expectation.

4:31Plus shots. These mean examples, no example, one example, or a few examples. And we can also add

4:40facts, which we can call excerpts or examples, which are going to relay the style in the format

4:47that we're looking for. So we'll start simple using CRE or context role and expectation. And

4:55then we'll dive into a schema, length cap, tone macros, safety boundaries for paste ready outputs.

5:02We're just going to increase the complexity. And when you do this, this is advanced prompt

5:07engineering. The goal of this course is to take you from the very beginning from AI fundamentals

5:13to understanding what software 3.0 is, and then give you the tools that you need a prompt toolkit

5:21so that you can reuse them for yourself or for your team. This is going to dramatically improve

5:27the quality of your work and the quality of the outputs of your models. All right. So let's talk

5:32about what a naive prompt is. Just to clarify, it's not a context role expectation. It's really

5:40just talking to a human, right? Something very casual. Should I update to iOS 26? This is the

5:47part of the problem. These companies are suggesting that we talk to these models like humans,

5:53but that's why prompt engineering really frames this as programming and gives you an advantage

5:59while everybody else is talking to it like a human, you can talk to it like a programmer,

6:04or at least with the foundational knowledge of programming and your outputs are going to be

6:09on fire. I'll show you why. So we're done with the video, but I want to show you how to do a

6:14naive prompt. Let's say that you need to summarize something. If you just say, summarize this, and you

6:20have some kind of a message, this is a naive prompt because you're not thinking about the context.

6:26You're not thinking about what the role is or expectation. So how is the model going to know

6:32in this case, it's at GPT-5, I believe, how is it going to know what you want? So the challenge is

6:38about doing a regular summary, so a naive prompt, and then strategizing. How could you give it more

6:45information to get the output that you exactly want? So the challenge is going to give you

6:51how the prompt should look after it gets summarized. So I'll give you a context or a scenario or some

6:57text data that needs summarization and the expected output. And without using any special skills,

7:04see if you can invent a way to talk to it. Maybe imagine that you're a programmer and you want to

7:10be very specific to see if you can get the exact output that we're looking for. And in the following

7:15skills, we're going to iteratively build on all of the techniques all the way to the very end of

7:21this course, where you'll have a prompt toolkit that you can reuse. You can just copy and paste

7:27so that you don't have to think like a programmer and reinvent the wheel every time that you're

7:32going to talk to these models. All right, so up next is a challenge. And until next time,

7:37I hope this has been informative. I'd like to thank you for viewing.

0:00Welcome to the challenge section.

0:02Here, we're going to use ChatGPT.

0:04Let's assume that you've never used ChatGPT.

0:07You could use another model if you like,

0:09but it would probably be easier if you use ChatGPT,

0:13which you can find by navigating to chatgpt.com.

0:17Let me make this a little bit bigger.

0:18All right, your task is to recreate a target output

0:22as close as possible to the example

0:24that I've given inside of the challenge.

0:26So just look at that example.

0:28And what you're going to do is just say, summarize this.

0:31And so anything that you type in here is a prompt.

0:34So I'm going to say, summarize this.

0:36And then I'm going to add this example.

0:39And this is what we're going to summarize.

0:41So let me run this.

0:42All right, so we get something.

0:44And this is something that just popped up

0:46that we're going to talk about in the next skill.

0:49It's really important to understand

0:50the terms and the privacy,

0:52and definitely don't share private

0:54or sensitive information in these models at any time.

0:58We'll talk more about that in the next skill.

0:59So I'm going to close this.

1:00And I haven't logged in.

1:02So this is very public.

1:03So all of this information is going to be stored

1:06and reviewed by OpenAI.

1:08But this is OK because you're just typing this in.

1:11And now your challenge is to get it

1:12to look like the solution, which is just below.

1:16And it has emojis and bullets.

1:19And so just try to type something here,

1:22create your own prompt that gets it to be closer.

1:25Because this does not look like the end example.

1:28And the idea here is that you're flexing

1:30your prompting muscles, right?

1:33This is the very beginning of our journey.

1:35You're going to try to get an output

1:38to be what you want it to be.

1:39When you just say summarize this,

1:41that's called a naive prompt.

1:43Well, you can see that the output is not even

1:45close to the final output that we're aiming to create.

1:49So try to do it.

1:50And as we navigate this course, I'm

1:53going to give you more and more techniques

1:55to help you create awesome prompts.

1:58So try to have some fun and enjoy this process.

2:00Be careful not to add any sensitive information

2:03into the chat.

2:04And if you get a message saying that you've

2:06run out of opportunities to use this,

2:08you can sign up for free.

2:10Or you can sign up for a paid account

2:13if you want to continue to use this.

2:15But if you're taking a prompt engineering course,

2:17you probably have used ChatGPT before.

2:20All right, good luck.

2:21And I'll see you in the solution video.

0:00Welcome back. So here is the example text that needs the summarization. So we grab this, copy it,

0:07go to chat GPT, and we paste it there. So we'll try the naive prompt first. Summarize this. That's

0:14a naive prompt. And let's see what we get. Okay, so we get this basically two sentence summary.

0:20But let's say that we want it to be for Slack, and we want it to look something like this.

0:25So how would you go about supplying a prompt that would look like this? And here is my solution,

0:31where I describe it in much more detail. So here I'm just saying summarize this. But here I'm saying

0:37summarize the following text into a short Slack friendly update. Use five to six bullets max.

0:43Each bullet should follow this format. Then I go in to describe the emojis. Then we have a bold

0:49label, a short casual description of each moment. Then above the bullets, we add a title,

0:55something like team or project update. And then I've supplied the text. And then we get something

1:00that looks pretty much identical. The only difference is the title is slightly different.

1:06It says pixel and beam update. And here it says pixel and beam launch prep. But that's close

1:11enough. All right. So that's it for the challenge. I hope you enjoyed it. And I will see you in the

1:16next skill. And until then, I hope this has been informative. I'd like to thank you for viewing.