Identify the Main Types of AI

0:00Hello and welcome to the first skill in the AI Productivity for Professionals course.

0:07My name is Jonathan Barrios and I'm excited that you're here.

0:10This first skill marks the starting point for learning how to use AI tools like ChatGPT,

0:17Gemini, Copilot, and NotebookLM to work smarter, faster, and safer.

0:24In this course, you'll learn to draft and refine documents,

0:27messages, and data-driven insights with AI.

0:31Turn ideas into polished deliverables using Google and Microsoft's ecosystem.

0:37And most importantly, apply prompt engineering to steer the AI reliably,

0:42not just experiment with it.

0:44I want to show you that this course is part of an AI series of mini courses.

0:49And we are here.

0:51And while I did mention Microsoft Copilot, there's a whole mini course on that.

0:57Again, it's a mini course, so it's not a complete course on Microsoft Copilot,

1:02but it will get you started with understanding AI, prompt engineering,

1:08and how to use the Microsoft ecosystem leveraging AI essentials.

1:13In this course, we're going to explore ChatGPT, Google Gemini, and NotebookLM.

1:20But again, it's also a mini course on AI productivity.

1:24And most of these mini courses include a first skill on what is AI,

1:30because it's paramount to understand what AI is and how it works behind the scenes.

1:36Then we get into prompt engineering, but it's a mini crash course on prompt engineering.

1:41So I highly recommend taking this mini course on AI prompt engineering

1:47to get the most out of your AI models.

1:50Again, here in the AI Productivity for Professionals, we're just scratching the surface.

1:55And if you're an executive or leader or interested in launching your own AI product,

2:01Artificial Intelligence for Executives and Leaders

2:04takes you from the idea stage to the deployment stage on AI best practices.

2:10And for those of you who are interested in programming, I suggest this one first.

2:14And then if you have no programming background and want to explore

2:18vibe coding, which is a new trend that is taking the internet by storm,

2:22which simply means the AI codes for you.

2:26But you definitely need to know a little bit more than just using AI, for example, security.

2:32And we have you covered in this course.

2:34However, if you already have a programming background,

2:37then a good next progression would be AI agent decoding,

2:40where you use AI agents to help you program even faster.

2:45But if you already have a programming background, you don't need to take this one.

2:49You can just skip right to AI agent decoding.

2:52And finally, if you're trying to deploy your product or you're part of a team

2:57and want to understand AI compliance, privacy and copyright,

3:01we have one here for developers, which is a more technical approach.

3:05And this one for non-technical audience.

3:09So the goal here is to get you up and running with AI in all of these different contexts.

3:14Again, if you find that you want to go deeper in any one of these,

3:18just tell us in the comments.

3:19And that is precisely the data that we need

3:22so we can make the courses that you are most interested in.

3:25All right. So now what is in scope or included in this course?

3:30First up, we need to identify the differences between

3:33the main types of AI because AI isn't just one thing.

3:39And we also need to understand the basic prompt engineering for AI productivity.

3:44These two are foundational concepts to build a mental model.

3:50And then we'll leverage AI tools using what we learned in prompt engineering

3:54to give us the optimal productivity using AI.

3:59So now what is not included in this course or out of scope?

4:03We're not learning about how to code.

4:06Even though we're going to talk about it briefly,

4:09I guarantee you that it's going to be fun and non-technical.

4:13The idea here is just to have a good understanding

4:16of how AI works in the background and what is not AI

4:21and no tutorials on how to build AI chatbots.

4:25If you want to learn more about AI chatbots,

4:27definitely check out the AI agentic coding course.

4:31Furthermore, this is a vendor agnostic focus,

4:34meaning there's no specific vendor focus.

4:37While we do talk about chat GPT and Gemini,

4:40this can be applied to almost all AI models,

4:44even Microsoft Copilot.

4:46And really the takeaway is that this is about general AI productivity for professionals.

4:52And a little bit about me.

4:53My name is Jonathan Barrios,

4:55and I teach data science, machine learning, and deep learning.

4:59I like to call this AI data science,

5:02which is the combination of AI engineering and data science.

5:06I've worked with AI and built SaaS websites,

5:10AI applications, and streaming platforms.

5:14And I've been teaching for quite some time at top platforms

5:17such as Treehouse, Thinkful, and Chag.

5:20And now I'm excited to be part of the CBT Nuggets team.

5:24You can definitely find me on LinkedIn at jonathan-barrios-ai

5:29or my website at jonathanbarrios.com.

5:32And finally here at AI underscore data underscore science on x.com.

5:38And a real quick note on x.com,

5:41some people love it, some people hate it,

5:43but I don't really use social media.

5:46So you might be wondering, well, then why are you on x?

5:49And here's why.

5:51I think social media is kind of a waste of time,

5:53but what I use it for is to keep up with the speed of advancement.

5:58It's a rapidly expanding field,

6:01and this is better than a newsletter.

6:03The entire AI community is on x,

6:06and you have access to Andrej Karpathy

6:09and many others who are leading in the field,

6:12posting directly on x.

6:14So if you follow just these people,

6:17and including myself,

6:18you're going to hear about AI and AI data science,

6:22as I like to say.

6:23You're going to get a feed full of AI developments

6:26and not all of that noise.

6:27All right, enough about me.

6:29Let's talk about this course and the first skill.

6:33Here, we're going to identify the main types of AI.

6:36Before we dive into any prompts or any projects,

6:40we need to understand what AI actually is.

6:44Not the buzzword version, but the functional one.

6:47You'll see how software itself has evolved.

6:51And to make this super accessible,

6:54we're going to talk about three different kinds of software,

6:57and I will make them very easy to understand.

7:00And furthermore, it's going to be directly applicable

7:04to prompt engineering and AI productivity for professionals.

7:08We definitely don't want to talk about all of the nuances of AI,

7:12which I personally love to do,

7:14but it's not going to help you be more productive.

7:17We're going to distill it into a very simple mental bottle

7:21that you can take with you from now into the future,

7:24because AI is not going anywhere.

7:26The sooner you build this mental model,

7:29the sooner you'll be able to navigate AI much better,

7:32and then we'll get into prompt engineering.

7:35And best of all, by the end of this skill,

7:38you'll understand the landscape and the mindset

7:41and the foundation of every AI workflow we'll build together.

7:45All right, so what are we going to cover in this skill?

7:48Number one, what is AI?

7:51Really?

7:51Before we get into prompts,

7:53we're building that mental model or that mental map.

7:56You'll see how AI evolved from simple rules

8:00to programs that learn from data.

8:02And to study that evolution,

8:04we'll look at it through the lens of software 1.0, 2.0, and 3.0.

8:11These are three types of software that are very easy to understand

8:15and will give you the intuition needed for the next skill,

8:19which is on prompt engineering.

8:20And then why modern AI can be so confident and also so wrong.

8:29This is probably one of the more important concepts of this entire course,

8:34and it's the foundation for improving your productivity.

8:38And number four, we'll lay the foundation for prompt engineering,

8:43and we'll do that with a challenge at the very end.

8:46So this is going to be a really fun skill.

8:49We're going to build an advanced AI model

8:52that can tell the difference between a hot dog and a hamburger.

8:55And to do that, there's zero code involved.

8:58It's just a simple website.

9:00And I have my two friends right here,

9:02hot dog and hamburger that are going to help us do this.

9:05It's going to be a lot of fun.

9:07And if you don't have any hamburger and hot dog toys,

9:10don't worry, you can just use the number one or number two as your classes.

9:15And this is really about giving you that foundation for your mental model

9:19before we get into prompt engineering.

9:21All right.

9:22On that note, I will see you in the first video on what is a I see you there.

0:00Before we start learning about prompt engineering or using tools like

0:04Gemini, Copilot, and even ChatGPT,

0:08we need to understand what's really going on under the hood.

0:12If you know how AI thinks and learns, you'll instantly get better results.

0:17Not because you memorize different prompts or you have a deep understanding of

0:22AI,

0:23but because you understand how to talk to it on its own terms.

0:29Many people will talk to AI like it's a human. And when you do that,

0:33it confuses the AI and then it starts to guess.

0:38And that's where it makes things up. And it does so very confidently.

0:42That's why we start here with what is AI.

0:45It's the foundation that we need for writing proposals and docs to analyzing

0:50data in Google Sheets or Microsoft Excel.

0:54And more importantly, AI isn't one thing.

0:57If you thought that AI was ChatGPT, it is,

1:01but that's not what AI is entirely.

1:04It's not all chatbots. AI is really an umbrella term.

1:08So let's check it out. When we say AI,

1:11we're really talking about this umbrella term, artificial intelligence.

1:15And that umbrella of different techniques is how the computer learns patterns

1:20and makes decisions, and sometimes even creates new content.

1:26That would be generative AI,

1:28but Gen AI is just one slice. At the simplest level,

1:33AI is software that improves with experience instead of

1:37relying on human written rules. For example,

1:41when we look at machine learning,

1:43this is one type of AI that finds patterns in data to make

1:47predictions like detecting spam emails or recommending products

1:52or forecasting sales.

1:54The key word here is learns.

1:56It learns on its own and it learns from data.

2:00We're not telling it what to do, how to think that's traditional programming.

2:05And when we look at deep learning,

2:07this is a subset of machine learning that uses neural networks.

2:12Neural networks may sound complicated,

2:14but we all have neurons in our brains.

2:18And when they're connected in a network, those are neural networks.

2:22And that's how humans think. That's how we process information.

2:27So it's very similar to how humans operate. In fact,

2:31it's software modeled after the brain.

2:34So instead of using human neurons,

2:37we use something called an artificial neuron.

2:40And the way I like to think about it is like Gandalf,

2:43thou shall not pass unless the data is important.

2:47And that's what a neuron does.

2:49It only allows important information to go through.

2:52And if it's not important, thou shall not pass.

2:55And when you have a giant network of these,

2:58that's how deep learning learns from the data.

3:01It may not make total sense right now,

3:03but when we train our hot dog and our hamburger models, it will.

3:08And we're not going to use any code.

3:10And in fact, you're going to build your own classifier.

3:12And I'm going to explain precisely how it works.

3:15And then finally, the newest layer of AI generative AI,

3:20also known as gen AI for short powers tools like chat,

3:24GPT, Gemini and co-pilot here.

3:28We can create language or programming code or media like

3:32images or videos all on demand. It's pretty exciting.

3:37And when you use Gemini or co-pilot,

3:40you're working with deep learning models that have been fine-tuned to

3:44understand and generate language.

3:48Those are called large language models or LLMs for

3:52short large, because they use a large amount of data language because, well,

3:57that's what they're using. And what are they models?

4:01So a large language model is what powers things like

4:06chat, GPT.

4:08And what's important to note is that AI doesn't know things like a person does.

4:13It predicts what is likely next based on patterns and it's

4:18training data. And that's why,

4:20how we ask a question is so important.

4:24We need to ask the right question,

4:27set expectations and even give context.

4:31And that's what we call prompt engineering. And that's in the next skill. Okay.

4:36AI is an umbrella term for machine learning, deep learning,

4:40and generative AI.

4:42They generally learn from data in different ways,

4:45which we're going to explore,

4:47but you don't really need to memorize the different learning styles,

4:50such as supervised self-supervised or unsupervised,

4:54or even reinforcement learning.

4:56These are just the learning styles in case you want to learn more.

4:59And the AI umbrella really spans all of these rules,

5:03search machine learning, deep learning, and gen AI.

5:07But we're just focusing on the top three machine learning,

5:12deep learning, and generative AI. That's all we need for our mental model.

5:17And if you're thinking about modalities, well, vision,

5:20that's what cars use to drive themselves speech.

5:23That's how chat GPT talks to us and other models that use

5:28speech NLP. This just means natural language processing.

5:32That's what powers generative AI chatbots recommendations.

5:38You might see this on YouTube and robotics uses a combination of these

5:42three. Let's explore how, when you look at these AI use cases,

5:47you might be overwhelmed saying, well, how do I know what AI does?

5:51What, but you can think of it in really simple terms. For example,

5:56when I explain what I do for a living,

5:58I say that I teach the computer to read, to see, and to speak,

6:03but how does that actually happen? Well,

6:06I use different kinds of AI like machine learning, deep learning,

6:09and generative AI, but that's all we really need to think about.

6:13Is it seeing, is it using language to communicate like text,

6:18or is it doing something else? For instance,

6:21when you're thinking about facial recognition,

6:23this is often deep learning because it's great at learning from the data.

6:28Meaning if you give it a bunch of pictures with labels and gets really

6:33good at identifying them and making predictions.

6:36So you could say that facial recognition is a form of deep learning,

6:39but you could also say that it's computer vision,

6:43which autonomous vehicles use because they're identifying patterns.

6:48So here it sees a tree, a car, a person assigned here.

6:52It sees an older person, a younger person, a person with Brown hair.

6:56They're wearing orange shirts and so on.

6:58These are all instances of computer vision,

7:01which includes facial recognition. But how about fraud detection?

7:05This is often machine learning.

7:07Machine learning uses algorithms to make predictions and learns from the

7:12data by using algorithms. So in this case, fraud detection,

7:17it looks at a bunch of data and it says, well,

7:19the values are from 50 to 100, like 99% of the time.

7:24But if you see a value that's 10,000, well,

7:27that's going to be a red flag and that could be fraud detection or

7:32some kind of an outlier. And algorithms are great for that.

7:36But when you're looking at all of these patterns,

7:39deep learning is better for that because it's software modeled after the brain.

7:43Well, how about robotics?

7:45That could be a combination of deep learning and machine learning,

7:50but it could also be gen AI because it uses deep learning to

7:55see, and maybe it speaks using generative AI,

7:59and maybe it's calculating some kind of an outlier when it's moving its robotic

8:04motors. So AI can be combined. For example,

8:07here we could use deep learning and gen AI.

8:10If we're looking at an image of text,

8:12turning that into characters and then creating new variations on those

8:17characters. So that would be deep learning and gen AI. And again,

8:22it's not really important to know what use case uses,

8:25what it's really to understand that AI means different kinds of

8:30things. Generally it learns from data in different ways.

8:35In the next video,

8:35we're going to really dial this down and simplify it by looking at the different

8:40types of software, 1.0, 2.0, and 3.0.

8:44See you there.

0:00To really understand how AI changed the way we work,

0:04let's step back and take a look how software itself evolved.

0:08The model of thinking about software as 1.0, 2.0, and 3.0

0:13comes from Andrej Karpathy,

0:15who's one of the leading figures in the AI field right now.

0:19They helped start OpenAI,

0:21and also played a huge role in Tesla self-driving vehicles.

0:26And what I really like about Andrej Karpathy's framing

0:30is that it gives us a clear picture

0:32for everything we'll do in this course,

0:35from writing prompts to generating full documents

0:38and presentations.

0:40Here's a story in three simple steps.

0:43When we were talking about software 1.0,

0:46we told the computer exactly what to do, step by step.

0:50Every condition, every formula, every loop.

0:54If you wanted to change how it behaved,

0:56you had to change the code.

0:58So the code changes the behavior,

1:00and these are rules that we coded by hand.

1:03So this is traditional programming, also known as coding.

1:06The way that I like to explain software 1.0,

1:09or traditional programming for that matter,

1:12is if X, then Y.

1:16So you can tell the computer,

1:18if the number is above 10, then do this, Y.

1:23So if X, then Y.

1:24If some condition, then do some action.

1:28This is logic that is exact and reliable,

1:32but it's also rigid.

1:33It's not like software 2.0 and 3.0.

1:37So that being said, let's explore software 2.0.

1:41Instead of coding the behavior manually,

1:44we feed the system data

1:46and train it to find the patterns by itself.

1:49So here, the AI learns from data,

1:53and we have machine learning and deep learning,

1:56which are both part of software 2.0.

1:59And the rules are written in weights.

2:01When you think about neurons and neural networks,

2:04we have all of these individual neurons in a brain,

2:07but here, we can think of it as a computer.

2:10All of these individual neurons get assigned a weight.

2:13That means, how important is this piece of data?

2:17And that's how it learns.

2:19It just says, oh, that's not important, this is important.

2:22And when you combine all of those different weights,

2:25it doesn't need traditional programming or rules.

2:28It just predicts, it learns from the data.

2:32You could also say that it learns the rules from the data.

2:35Unlike software 1.0,

2:37when we're coding the behavior manually,

2:39here, we're just giving it data

2:41and training it so that it learns.

2:44And that's machine learning and deep learning.

2:46The difference between machine learning

2:48and deep learning, again,

2:49deep learning is software modeled after the brain,

2:52and machine learning uses statistical learning

2:55or some kind of an algorithm or statistics.

2:58But the end goal is the same.

3:00You see this in things like recommendation engines

3:03or fraud detection or email spam filters.

3:07And here's this training thing.

3:09What is that about?

3:10You're gonna find that out very soon

3:12when we get into the hamburger hotdog model

3:15that we're gonna build without any code,

3:17and it's actually quite fun.

3:18And then fine-tuning is how you create a specialized model

3:21like hamburger hotdog.

3:23We could fine-tune a model

3:25to distinguish between those two items.

3:27And if you wanna change the behavior,

3:29well, you have to retrain it with new data

3:32or fine-tune it with new data.

3:34It's not code, that's the takeaway.

3:36All right, so now how about software 3.0?

3:40This is Gen AI.

3:41So now we've covered all of the different types of AI,

3:45but what's so special about software 3.0 or Gen AI?

3:50Here, we guide the behavior with plain language.

3:54The rules are now prompts,

3:56and this is why we need prompt engineering,

3:59because talking to the AI by typing

4:02is much more powerful than it seems.

4:05It's really a kind of programming.

4:07Here, we use code to program.

4:10Here, the model learns by itself.

4:12It doesn't even need us.

4:14Sure, we do write some initial code and give it the data,

4:17but beyond that, it creates its own rules.

4:20It learns the rules.

4:22But here, the rules are really the prompts that we give it.

4:26So now everybody has the power

4:29to do some kind of programming.

4:31So if you wanna steer that model and change the behavior,

4:35you do that at runtime.

4:37That means whenever you type it into a chatbot

4:40and hit Enter,

4:41you're steering that AI model and changing its behavior.

4:46So it's really like a type of programming,

4:49and that's what I'm getting at.

4:50The sooner you can start thinking about generative AI

4:54as a form of programming, the better.

4:57Andrej Karpathy himself said,

4:59"'English is the hottest new programming language,"

5:03which is pretty bold.

5:05Why would English be a programming language?

5:07How's that even possible?

5:09We have programming languages such as Python,

5:12which are very human-readable and easy to learn,

5:15but it's still a programming language.

5:18There are data types,

5:19all kinds of symbols that you need to understand

5:22how they work in order to build

5:24some kind of a software application.

5:26But English?

5:27And that's the cool part.

5:29Yes, not just English, but plain language,

5:31because AI works in many different languages.

5:35So here, for software 3.0, we don't code rules

5:39or train the data directly.

5:42We simply steer the powerful pre-trained models

5:46using plain language.

5:48How do you do that?

5:49You describe what you want,

5:51who it's for, and how it should sound.

5:54That's more or less prompt engineering,

5:56which we'll get into in the next skill.

5:58So just think about chat GPT.

6:01Every time you type a prompt into chat GPT,

6:04or even Gemini or Copilot,

6:06you're not programming in Python,

6:08you're programming in English.

6:10In essence, plain language.

6:12And that's why we start with the mental models

6:14before we write any prompts.

6:16When you realize that the words are a type of code

6:21or a new kind of code,

6:22you start to see why clear, structured communication

6:27is the core skill of modern productivity.

6:30All right, so we learned that prompting

6:32can change the behavior of these models,

6:35and you can change the weights by fine-tuning or training.

6:40What does all of that mean?

6:41You're about to find out.

6:43So what are the actual levers or the tools

6:46that we use to program in plain language?

6:48That's gonna be CRE, shots, constraints,

6:52retrieved snippets, or more simply said, facts.

6:56These are all components of prompt engineering.

6:59CRE just means context, role, and expectation.

7:04Shots are examples.

7:06Constraints are less than 90 words, for example.

7:10And then we give it facts.

7:12This helps it stay on point and not guess.

7:16All right, enough of the theory.

7:18In the next video, we're going to check out Google Colab.

7:21You don't need to use this yourself, just follow along,

7:24and I'm gonna show you how traditional programming works.

7:27See you there.

0:00Welcome back. In this video,

0:02let's make this idea of software 1.0 real,

0:06and we'll use something that every professional understands.

0:10Excel formulas and email rules.

0:13You either know about Excel formulas or you know about email rules,

0:17like sending something to spam or even just deleting it.

0:21In software 1.0,

0:23everything the computer does comes from instructions that you

0:27explicitly define. And we're going to do that above.

0:31This is where you add the code.

0:32What we're going to do is tell the computer what to do line by line.

0:37And each one of these are called code cells.

0:40This is just to let you know what we're going to do and how it works.

0:44Let's say that we want to think about Excel. In that case,

0:48you might write a formula that says, if this is very common in Excel,

0:52and then you would use parentheses. And inside of the parentheses,

0:56you would write some sort of condition. So you could say something like sales.

1:00If that value is greater than 1000, then do X.

1:05If not do Y. So what are X and Y?

1:08We can say bonus and it would classify that as a bonus and it would change the

1:13text to bonus. So instead of showing you a value,

1:17it's going to change it to that bonus because it's above 1000. However,

1:21if it's not, then it's going to say no bonus. If X,

1:25then Y else Z. So you can have multiple cases,

1:29but the takeaway here is that it's a rule. In this case, if this,

1:34then that, then that, or in Gmail or Outlook,

1:38you might create a rule that says if the subject contains invoice,

1:43move to the finance folder. It's the same thing. It's a rule,

1:48but instead of Excel, we're in some kind of a mail application.

1:52Every outcome depends on manual logic and that's not how AI works.

1:59And we can do this when we're programming.

2:01Let's say we take sales and we give it a value of 5,000 or let's say 50,000 are

2:07really big number.

2:08And then I'm going to just say if sales is greater than 10,000,

2:14I want you to do something. But in Python,

2:17it's a little bit different than an Excel formula.

2:19So you simply do this and you type in bonus.

2:22So I'm going to run this cell and it connects to the cloud and then immediately

2:26gives us an answer bonus. But we could take this a step further.

2:30We can say else print no bonus.

2:33Now we have the same exact thing that we have in Excel.

2:37So here it's above 10,000. So it's always going to print bonus.

2:42If this, then that, and it's simple as that. However,

2:46if I change this to 100 and then we run it again,

2:50you get no bonus. So to change the behavior, you change the code.

2:55So let's say that you're looking for it to be equal. You run this,

2:59no bonus, or you can say not equal in this case. So we're changing the code.

3:03So it's not equal to the bonus,

3:06but we'd say 1,000 or 10,000 in this case. And it says bonus.

3:11Maybe it might make sense to say something like bonus amount.

3:14What I'm doing is changing the code, changing the behavior.

3:18And that's it. This is precise, but limited.

3:22The system can't generalize. It only knows what we're telling it here.

3:27If we needed to do something else,

3:28we need to actually change the code that's located here. Same thing with Excel,

3:33same thing with a Gmail rule. And this is where AI steps in.

3:37Instead of hard coding, every rule AI learns patterns.

3:42And that shifts from your logic to the system learns the logic.

3:47And this is what brings us from software 1.0 to software

3:512.0 and 3.0. Just before we get there,

3:56just remember this mental image in software 1.0,

4:00you have to think of every scenario in advance and software

4:042.0 and 3.0, the models learn from the data.

4:09And more importantly than 3.0,

4:11you describe the goal and the AI figures out how to get there.

4:16So that means that you don't need to learn how to code. Instead,

4:20you need to learn how to prompt. And in the next video,

4:23we're going to demystify how models learn from data.

4:27That way you understand how generative AI works,

4:31and that's going to give you the secret sauce for prompt engineering.

4:35And that way you can leverage this, not just for productivity,

4:40but for any AI application moving forward.

0:01In the last video, we explored traditional software, also known as software 1.0.

0:06And that's where we write the rules.

0:08Rules by hand means that the code generates the behavior.

0:12And if you want to change the behavior, then you need to edit the code.

0:16So the code just means explicit rules and there's no learning.

0:19That's the important part.

0:21And in the demo, we could have created a greeting based on hour, but we chose Excel and Gmail.

0:27And now that we've seen how software 1.0 works, where we use explicit rules and every scenario must be hard-coded,

0:34let's move into software 2.0.

0:37This is where the system learns the rules.

0:40And to do that, we're going to build a tiny classifier.

0:43And by build, I mean teach.

0:45And we'll use Teachable Machine from Google to do just that.

0:49This is a wonderful application that's in a browser and you don't need any kind of coding experience at all.

0:56All you need to do is just have a webcam and that's it.

0:59And furthermore, you don't even need to do this.

1:02You can just watch me do it and you'll understand how the AI learns.

1:07Not at a deep level, of course.

1:09You're not programming and building this out, writing code, which is what Teachable Machine abstracts.

1:15And that way, we can focus on how these models train and then make predictions.

1:20And that's exactly what ChatGPT does.

1:23Okay, so the rules are in these weights.

1:26But the secret sauce is when you train the model or fine-tune the model,

1:31these are two ways that the models learn the patterns and that's what creates the behavior.

1:37To do this, we're going to collect 20 to 50 samples.

1:40In this case, that could be 20 to 50 images for hamburger and 20 to 50 images for hot dog.

1:48But if you don't have those toys lying around, then you can use your hand.

1:52With a number 1 or a number 2.

1:54And then once you've created the data, then you train and then the model learns from those images.

2:02And then you can preview it and do live classification.

2:05What does that mean?

2:07After you train the model, you show it either a hamburger or a hot dog or in your case, 1 or 2.

2:13And it's going to tell you if it's a 1 or a 2 or a hot dog and a hamburger.

2:18And it's going to do so using a probability.

2:20It'll say 100% hot dog or 10% hot dog or 90% hamburger and so on.

2:28And that's an important detail because the AI doesn't always know 100% what something is.

2:35And it has to make a choice and that's why sometimes it's wrong.

2:38So what we're saying is that instead of writing a list of conditions like if this, then that,

2:45we give the computer examples.

2:48It studies the data, looks for patterns, and builds its own internal logic.

2:54And that's what we call a model.

2:56So now let's make that concrete and fun in a simple demo using Teachable Machine.

3:02Again, you don't have to do this.

3:04But if you are curious and you want to follow along or you want to play with this on your own,

3:09you can go to teachablemachine.withgoogle.com.

3:13You can use audio.

3:14You can use video.

3:16You can use poses.

3:18You can use images.

3:19Let's get started.

3:20The first choice we need to make is whether we're working with images, audio, or poses.

3:26But in reality, they're all images.

3:29Here you can see that the audio has been converted into some kind of a visual representation of sound.

3:35And here in the poses, these are just individual images stacked one after the other.

3:41The images, we're just using single images one at a time.

3:45But they're really all image projects, which means we're doing computer vision of some kind.

3:50All right.

3:51So now click on the standard image model.

3:54And here we need to gather our samples.

3:57So it says make or add some samples here or watch this video on how to do it.

4:02So definitely click on this to learn more.

4:04But I'm just going to show you how I do this.

4:06First, I'm going to name this hot dog.

4:08And I'm going to name this hamburger.

4:10Next, we need to click on the webcam.

4:12Okay.

4:13So now I'm going to click on crop and make this smaller.

4:16Move it over here.

4:17Try to get out of the way.

4:18And then click on crop.

4:20Now you can just see this area.

4:22And I'm going to get the hot dog and then hold it up there.

4:26And then I'll move this around.

4:27But I need to hold this down and record these images.

4:32So let's aim for 100.

4:33Perfect.

4:34100 images of the hot dog.

4:36Now it's time for hamburger.

4:38Same thing.

4:39Click on crop.

4:40Move this to the top right to get it out of the way.

4:42And say I'm done cropping.

4:44And then here we have the hamburger.

4:46I'm going to do the same exact thing.

4:48I'm going to hold this down and move it around to get different images, even the back.

4:53Perfect.

4:54We have 100 images of hamburger.

4:57Let's close this.

4:58But you might be asking, well, what happens if you only have 10 of hot dog and 100 of hamburger?

5:05And that's called bias.

5:07We favor hamburger 10 to 1 because it has 10 times more samples.

5:12And that's why fairness is so important.

5:15You need to have equally distributed representations of data.

5:19So generative AI can be really tricky.

5:22Now we train the model.

5:24Now that you have two classes, you can train your model here or add more classes, which we're going to do.

5:31But first, let's try to see why we need more classes.

5:34Click train model.

5:35And then it does this quite a few times really, really, really quickly.

5:39You can see that it does it 50 times.

5:42These are 50 training epochs.

5:45And then here it wants to preview it.

5:47But I need to get out of the way.

5:49Let me crop this and try to get out of the way because now it thinks that I'm a hamburger.

5:54Now it thinks I'm a hot dog, right?

5:56It's super confused.

5:57So let's click on crop.

5:59And there we go.

6:00I'll get out of the way.

6:01And now it's saying that the background is hot dog.

6:04And that's why we need to add another class down here for the background.

6:08But before we do, let's test it out.

6:10100% hamburger.

6:12Even if I do this, it's still pretty good.

6:15Look at that.

6:16Now let's do hot dog.

6:17100% hot dog.

6:19Even the back.

6:20Even in this way.

6:21Any which way that you do it, it's seeing that this is definitely a hot dog.

6:26OK, so now let's create one more class.

6:28And let's go over here to webcam.

6:30I'm going to do the same thing.

6:31Crop.

6:32Get out of the way.

6:33Move this over here.

6:34Done cropping.

6:35And I'm just going to hold this for 100 images so that it has training image or training data.

6:41But we can just delete these two here like this.

6:44And now we have equal amounts of samples.

6:47This is really important.

6:48And here I'm going to say background.

6:50All right.

6:51Let's close this.

6:52And now to change the behavior, you have to retrain the model.

6:57That's what we were saying.

6:58So let's do that.

6:59So now I'm training the model again, this time with three classes.

7:03Hot dog, hamburger, and background.

7:06Now that I'm moved out of the way, hopefully I don't have to do that cropping again.

7:11And I don't.

7:12You can see that the background is 100% accurate.

7:15If I stick the number two up, it gets totally confused.

7:18It's like, what is this?

7:19A hamburger?

7:20A hot dog?

7:21I can't tell.

7:22Oh my gosh.

7:23If I do this, then maybe it's more of a hamburger.

7:24At that point, it just doesn't know.

7:26But if you do this, wow.

7:28That's a hamburger.

7:29100%.

7:30If you do this, it's definitely a hot dog.

7:33If you do this, it's a hot dog.

7:35Nope.

7:36That's not a hot dog.

7:37So it just hallucinated.

7:39And that's one of the problems with generative AIs.

7:43If you don't have the right data, you need the data to be perfect for your task.

7:50Or it might make stuff up.

7:52And that's why we're going to get into prompt engineering in the next skill.

7:56But first, let's check out software 3.0.

7:58See you there.

0:00Welcome back.

0:01Hopefully you have a much better understanding of what AI is

0:06and what the difference is between traditional programming and models

0:10that learn from data like classifying hamburger or hot dog,

0:15and some of the drawbacks such as bias and not having

0:19good representation of each class, or more importantly,

0:23no representation of, for example, me, it thought that I was a hot dog.

0:28So these are important clues.

0:30And now let's talk about software 3.0,

0:33which is really the main focus of this skill.

0:36Now we're stepping into software 3.0,

0:39the world where most of your modern tools live here.

0:43We're not coding rules like 1.0 or even training models like we just

0:48did in software 2.0.

0:51We are steering intelligent systems in real time

0:55using language.

0:57I like to say that software 3.0 is prompting as

1:02plain language programming.

1:05And that's why I really like thinking of prompts and

1:09Gen AI as software 3.0.

1:13When you use chat GPT, Gemini or Microsoft Copilot,

1:18you're not retraining the model. You're conditioning it in the moment.

1:23It's incredibly exciting and powerful.

1:25Every prompt you write acts like a short program in English.

1:30So really they're rules in prompts,

1:34which are plain language. And what does that do?

1:37You're steering the behavior. If you want to change the behavior,

1:42change the prompt. What we're doing is setting context,

1:45assigning the role and defining what a good output looks like.

1:50We'll get into that in the next skill with C R E.

1:54Context role and expectation.

1:57When you do that, it's much like software 1.0.

2:01You're saying context equals,

2:03and then you define what the context is or role equals.

2:07And then you define the role or expectation equals,

2:10and then you define it.

2:12But instead of using numbers or data types or some kind of programming

2:16language syntax, you're using plain language.

2:21So that's what you add here. What is the context?

2:24This prompt, what is the role? What is the expectation?

2:28You do this and you're going to really steer that behavior.

2:31You're going to get really, really good at it.

2:34So prompts are runtime instructions, meaning no code,

2:38and we're not changing the weights. That's what software 2.0 does.

2:42But what we are doing is conditioning the behavior using context.

2:46So we're steering via constraints and examples.

2:50Examples can be thought of as patterns and we keep the facts as excerpts

2:55where each example teaches this style in the format. Again,

2:59this is all prompt engineering. The goal here is shorter,

3:03steadier and paste ready outputs.

3:06If you've used chat GPT and you've had to talk to it over and over and over,

3:10or it says stuff that's really wrong and confident,

3:15or it just says really encouraging things over and over. You're doing great.

3:20When you know that you're not, then you're going to love prompt engineering.

3:24So let's try to come up with an example together.

3:27Let's say you want Gemini to draft an email for your coaching business.

3:31And we are going to use the example of Thrive 365,

3:35which is a business that we're going to build.

3:38It's an exercise and nutrition startup. So we'll keep using this example.

3:43So here,

3:44how would you write an email to the new clients of Thrive

3:48365? You could say, write an email to new clients,

3:52but the output might not be great.

3:54But if you use structured clear prompts, which we call CRE,

4:00then you might say, well, this is for Thrive 365,

4:04a hybrid fitness and nutrition program. And how about role here?

4:09You might say act as a professional coach who writes friendly and

4:14confident emails. How about expectation here?

4:17You might write create a short welcome email under 120 words that

4:22thanks them for signing up and explains the next steps.

4:27When you do that, you're doing CRE, which is prompt engineering.

4:31Now the outputs fit your purpose,

4:34your brand voice and your audience.

4:37You didn't retrain the models like you would in software 2.0.

4:41You programmed the AI with words.

4:44That's the key shift from 2.0 to 3.0.

4:49In 2.0, we shape the behavior by training it with data.

4:53In 3.0, we shape the behavior with language.

4:57And this is the foundation for everything else we're going to do in this course

5:01from writing emails in Gmail to documents in Google docs,

5:05to slide decks in Google slides,

5:09and even research in notebook LM.

5:12But how do these generative AI models,

5:15also known as large language models or LLMs produce this

5:20output? As we saw with the hot dog and hamburger, they use probabilities.

5:24It's never 100% sure. I mean,

5:27it is when we were looking at the hot dog and hamburger,

5:29but the idea is that it's anywhere from zero to 100.

5:33It's not always zero and it's not always 100,

5:36meaning it's not a true and false situation.

5:41There's a wide range of possibilities.

5:43That's why it uses probabilities to make the next token.

5:48So these large language model picks the next token full stop.

5:52What is a token?

5:55So a token is a numeric representation of

6:00text, and they're not always one word for one token.

6:03There can be multiple tokens from multiple words.

6:06So I like to call them word chunks.

6:08So why are we using word chunks? Well, that's what the computers prefer.

6:12They like numbers. So we could have a word like reproduce.

6:16And since computers are so efficient, they might say, okay, re and produce.

6:20Two tokens.

6:21That way they can reuse them quickly and they're not going to be in text.

6:25They're going to be numbers. We'll get into that a little bit later.

6:28So they picked the next token from a probability distribution.

6:32And so what does that mean? Well, you could say the cat sat on the,

6:36and then you would have the top three picks.

6:39The one that has the highest probability will be picked.

6:41So it'll probably be the cat set on the hat instead of

6:46the cat set on the Lexus.

6:49One is going to be way more probable based on all of the training data on the

6:53internet that it used.

6:54Another key concept here that is super important is temperature.

6:58The higher the temperature,

7:00the more variety that you're going to get from your model.

7:03The lower the temperature, the more stability,

7:05but also could be a little bit more boring. So if you want more creativity,

7:09crank up the heat, turn that temperature up,

7:12but you're going to risk getting some weird and maybe incorrect output.

7:16So you might want to find a nice balance. Normally we cannot set this,

7:21but sometimes you can,

7:23this is just to give you an idea of how the output is produced.

7:27The temperature really matters.

7:29Different models will have different temperature settings. And finally,

7:33consistency is the same thing as reducing the degrees of freedom,

7:38meaning the number of choices.

7:40So if you give the model a specific shape, like I want labels or count or format,

7:46you're reducing the degrees of freedom, which gives you more consistency.

7:50That's really what prompt engineering is about.

7:53You can add one or a few shots. Shots are simply examples,

7:58and this is going to help you get a better output or steady the voice.

8:02So imagine a forecast that says that there's a 70% chance of rain.

8:07The model can't know for sure.

8:09It's just estimating the probability from the data.

8:12So those are called forecasting models and AI is often used for these models and

8:19large language models work the same way.

8:21They sample from a probability distribution of what's most likely to come next.

8:27What's the most likely next token.

8:29This also explains why your prop structure matters so much.

8:34When you give the model a clear framework, like return of three bullets,

8:39status change and action, you're narrowing the probability space.

8:44You're reducing the degrees of freedom,

8:47meaning you're reducing the choices that it can make.

8:51And that helps the consistency or improve the model output.

8:56That's how you turn creative text generation to consistent professional

9:00output. So where are we going with all of this? Here's the goal.

9:04We're going to become programmers, but we're going to use English.

9:07So you can think of talking as a form of programming,

9:09but we're going to change it so that it's a little bit more effective with AI.

9:15So programming with English for AI,

9:18that's who we're talking to generally large language models.

9:22So we're going to use CRE to pull that off because AI likes that structure.

9:28And we're going to add shots only if we really need to.

9:31Zero shots means no examples.

9:34Sometimes you need a little help because CRE may not be enough.

9:38So you maybe add one example, but rarely do you add a few.

9:42Anything more than two is not going to be great. We'll cover that.

9:45Then excerpts. This is important. You need to give it facts.

9:49And when you're using examples, you need to be very explicit. Hey,

9:53these are style and format only. They're not facts.

9:57So in the next skill, we're going to cover CRE and then examples.

10:01This is the basic for prompt engineering.

10:04And then we're going to apply prompt engineering by using schema length cap,

10:08tone, macros, safety boundaries for paste ready outputs.

10:13This is where we go from prompt and guess to prompt

10:17engineering,

10:18which is exactly what you want when you're thinking about AI productivity for

10:23professionals. All right. So that's it for this video.

10:25And next up we have the challenge where you're going to try to match the output

10:29that I give you. And then in the next skill,

10:32we're going to apply prompt engineering and I'll show you how I came up with my

10:37CRE to produce that output. See you there.