Harness AI Principles Beyond Generative AI

0:00Hello and welcome to the AI for executives and leaders course my name is

0:05Jonathan Barrios, and I'm excited you're here

0:08We are living in such an exciting time

0:10Artificial intelligence is getting better and better and right now

0:14Everyone especially leaders of industry are trying to make sense of how to best

0:19use it

0:20How to scale it how to leverage it and how to make sure their teams can work

0:24most successfully

0:26Alongside it so what is this course about what's inside the scope of what we'll

0:31be learning and what is outside of it

0:33Here's what's included in the course choosing AI wisely. We'll go over how to

0:38pick the best AI tools for a

0:40Particular business task and then we'll explore automation

0:45Decisions here we'll talk in-depth about when to fully automate versus when to

0:51use a human in the loop or

0:54ITL approach finally we have scaling safely and that's the key word safely

1:00We'll start with small real-world AI pilot project examples that demonstrate

1:05security

1:06Business value and how you can scale them easily. All right, so what's not in

1:12the course

1:12What's not included and on this side? We have no coding or model building

1:17So you won't learn how to train AI models or write any code

1:22This is a non-technical course for executives and leaders and that means no AI

1:27tool tutorials

1:28No step-by-step guides on AI tools like chat GPT Gemini Claude, etc

1:35We will talk about them, but we're not going to dive deep into them and finally

1:39we have no deployments

1:40We're not going over technical setup and the focus of the course will be on

1:45high-level concepts and decision-making

1:48And that means we're going to stick pretty close to the concepts that come with

1:53implementing AI and before we get into the skill

1:56I did want to show you that this is the first course in a series of mini AI

2:02courses

2:03This one is artificial intelligence for executive and leaders the one we're now

2:07However, we do have really interesting courses that are very powerful that can

2:12work together for example prompt engineering

2:15We'll talk about different types of AI and how prompt engineering is really a

2:20form of programming

2:21That means when you're talking to chat GPT, you're literally programming in

2:26plain language like English

2:27So knowing the different techniques of AI prompt engineering will really make

2:32your models much more effective and you can mitigate

2:36Problems such as hallucination, which we'll talk about in depth in this course

2:41There's also agent decoding and this is really programming with an AI agent

2:47And if you use the Microsoft ecosystem

2:49Microsoft co-pilot is embedded into Excel word and all of the different aspects

2:55of Microsoft's ecosystem

2:57Vibe coding is programming for non-programmers leveraging AI and while it can

3:02be done

3:03Taking this course is really important because there are a lot of dangers that

3:07can come up and avoiding those are key

3:10And then here are two very important ones AI compliance

3:14Privacy and copyright this one over here is more for HR legal and IT

3:19So non-technical audiences even though IT is a technical audience

3:25This course stays at a high level and really just talked about compliance

3:29copyright and privacy while this one over here

3:32focuses on language and techniques

3:34specifically for developers and finally if you want to get your teams up to

3:39speed

3:40On AI productivity. This is a great course to get them up to speed and a little

3:45bit about me

3:45I teach data science machine learning and deep learning and I've worked with

3:50many AI tools and built SaaS websites

3:52AI applications and streaming platforms

3:56I've been teaching for a while now from treehouse to thinkful and check and now

4:00I'm happy to be here with the

4:02CBT nuggets team you could also find me here on LinkedIn or my website where

4:07post blogs and

4:09Social media at AI underscore data underscore science and here at X even though

4:15some people don't like it

4:16It's really where all of the AI community is in social media

4:20So I use it to stay up-to-date with all of the new developments and I ignore

4:25pretty much everything else

4:27And now we're ready for the first skill in our course where we're going to

4:31harness AI principles

4:33beyond generative AI and don't worry. We're going to explain what generative AI

4:38is and just a quick hint

4:40It's more than one thing. We'll start with the fundamental principles of AI

4:44itself

4:45These are the core technical principles that never change and are always action

4:50able even as trends come and go

4:52These core principles will empower you with the right tools to integrate AI

4:57into everyday processes and workflows

5:00In fact, you might be surprised to learn that AI

5:03Fundamentally is software modeled after the human brain. That's right. Hey, I

5:09was modeled after us

5:10So as we progress through this course, not only will we cover the core

5:15principles

5:16But we'll go over the different types of AI

5:18AI's for core models and how to evaluate which I approach we should be

5:25leveraging

5:25So as we progress through this course, not only will we cover these core

5:30principles

5:30But we'll go over all the types of AI

5:34AI's for core modes and how to evaluate which AI approach should be leveraged

5:40to drive business goals

5:42Before we get started, let's talk about everything

5:46We're going to cover in this first skill and this first part is the AI toolbox

5:50and the takeaway here is that it's not one tool

5:54It's many tools inside of this AI toolbox

5:57Generative AI is only one option among several and when we say generative AI

6:03We're really talking about all of these language models like chat GPT

6:07Claude

6:10Gemini and grok which uses real-time data then we'll segue into the four modes

6:16of

6:16AI because again AI is not just one thing here

6:20We're going to talk about rules classic machine learning deep learning gen AI

6:25all the good stuff and which model you should use for

6:28Which type of task, but how do you do that?

6:30And that's why we will talk about decision and safety here. We're going to use

6:35a traffic light metaphor of green

6:38Yellow and red these are going to be rules that are going to help us

6:42Understand if we need a human in a loop for anything that is sensitive or if we

6:47can just rely on AI out of the box

6:50And finally we have the challenge which is something that we do for all of the

6:55skills

6:55And what does that mean?

6:56Well, we want to make sure that we understand all of these concepts before we

7:00move on to the next skill because they all build on each

7:03other progressively

7:04Alright, so that's it for this introduction and I'll see you in the first video

7:08where we answer the question. What is AI?

7:11See you there

0:00Welcome back. In this video, we're going to kick it off by asking the question,

0:04"What is artificial intelligence?" But before we answer that question, let me

0:09ask you this question.

0:10When you hear artificial intelligence or AI, what do you picture? Do you see a

0:16chat bot generating

0:18big chunks of text? If so, that's normal. Most people do. That's because

0:23popular tools like chat

0:25GPT make it easy to assume that AI is all generative AI. So the tool generative

0:31AI is something that

0:32we're going to define. And that's part of models like chat GPT, which makes it

0:37easy to assume that

0:38all AI is generative AI. So if we ask the question, "What is AI?" we could look

0:44at this formal definition

0:46here. And it says that it's a field of computer science that focuses on

0:50building systems capable

0:52of performing tasks that typically require human cognitive abilities. And then

0:58we can talk about

0:59each one of these. We have perception, reasoning, learning, problem solving,

1:04and decision making.

1:05But today, I want to zoom out and away from generative AI, because AI isn't

1:11just one thing. It's

1:13actually that toolbox we were talking about earlier. And if we understand each

1:17one of the tools,

1:18we can pick the simplest, most effective tool to improve outcomes for any

1:24business. So instead

1:25of thinking about human cognitive abilities, we can think about the tools and

1:30better yet,

1:31outcomes. One of the outcomes we'll explore is onboarding. And we're not really

1:37thinking about

1:37building an AI tool for onboarding for our organization, but it's a great

1:42example. When I

1:43onboarded not just to one company, but many companies, there was a recurring

1:47theme. If I had a question,

1:50I would reach out to somebody. They wouldn't know. They'd have to reach out to

1:53somebody else,

1:54then give me an email reply, or I'd get a calendar invite and jump on a quick

1:58call.

1:59All of this could have been handled very easily with AI, but the trick is AI

2:05can get it wrong

2:06sometimes. So what do you do? Do you add a human in the loop? Or do you use a

2:11specialized AI model

2:13that your team will build to make sure that it doesn't say the wrong thing?

2:17These are all great

2:18questions. And we're going to dive into that in this skill. Okay, so here is

2:22generative AI. I

2:23promise we'll define this. But for now, let's look at the overall picture. So

2:29here we have our AI

2:30toolbox. And there's a lot inside of it, right? So gen AI is in the box, but it

2:35's not the whole box.

2:36So the general umbrella term is AI or artificial intelligence. So that's the

2:43first one. And then

2:44we have machine learning. And then there's deep learning, which you could think

2:48of it as a newer

2:49type of machine learning. And then inside of that, you have generative AI. And

2:54the takeaway

2:55here is that AI isn't just generative AI. This is what everyone sees, because

3:00these are things

3:01like chat GPT. But when is the last time that you saw deep learning or machine

3:06learning in action?

3:07Well, those are going to be abstracted. And you're going to see them inside the

3:11tools like chat GPT

3:13or Gemini. But they're all subsets of each other. So they're interconnected. So

3:17here's a quick

3:18disclaimer. It's not particularly important to know how all of these types of

3:24AI work

3:25deeply at a technical level. The takeaway here is more about understanding the

3:30tools and how we can

3:31use them for different business use cases. So let's take a look at some

3:36different use cases

3:37that you might see in the real world, and then we'll apply them. So let's take

3:42the first one here,

3:43facial recognition. This uses deep learning. And you can think of deep learning

3:48as a part of AI

3:49that is especially good at recognizing patterns. So instead of writing code, I

3:55'm going to use this

3:57hamburger and a hot dog here. And we're going to build our own deep learning

4:01model. And we're

4:02going to use not for facial recognition, but the next one computer vision. And

4:07for computer vision,

4:08this also uses deep learning. So jumping back to facial recognition, this works

4:14because deep

4:14learning can spot unique patterns of a person's face, like the spacing of the

4:19eyes or the shape

4:20of a nose across thousands of photos. And when we're thinking about computer

4:25vision,

4:26this is a larger umbrella. So facial recognition fits into computer vision. So

4:31it's easy to get

4:32mixed up. But here, we're thinking about self driving cars that can see road

4:37signs or medical AI.

4:39And that's why they both use deep learning, because as I said, deep learning is

4:44great at learning

4:44patterns. And that allows them to see a make sense of images and details. And

4:51the way I like to

4:51explain computer vision is that we're teaching the computer to see. We also

4:56teach it to read and

4:57write. And again, this is how software is modeled after the human brain. When

5:03we look at robotics,

5:04you could think of this as ML plus DL. So machine learning and deep learning.

5:10So why does it use

5:11machine learning robots use machine learning to learn from experience? So not

5:16patterns this time.

5:17For example, it gets better at gripping objects, walking on uneven ground and

5:23optimizing routes

5:24in a warehouse. And it uses deep learning when it needs to see with the camera

5:29or sensors. So

5:30here, it's maybe looking at this cup of coffee and the person that is going to

5:34deliver that coffee

5:35and that their hand is right here, it's calculating the distance. So robotics

5:40is one of the few areas

5:41that truly spans both machine learning and deep learning, since it needs

5:45learning and perception.

5:47Okay, so now let's look at documents. This is IDP intelligent document

5:52processing. And here,

5:54this is about teaching the computer to read forms or invoices may be contracts.

6:00And here,

6:00we're using machine learning, because it learns things like the total amount

6:05and how that total

6:06amount is usually at the bottom right. But you can also add deep learning,

6:11depending on different

6:12use cases, for example, deep learning can help when documents are scanned. And

6:17then you convert

6:18that to clean text. And finally, here we have fraud detection. This is about

6:23spotting unusual

6:24patterns in data, like let's say a sudden $10,000 charge in another country. ML

6:30is right there,

6:31because it learns what normal looks like from past transaction. And it flags

6:36what doesn't fit.

6:37It learns from historical data. So it learns from the data. So you could think

6:42of this as a

6:43fraud analyst who's seen millions of transactions and gets a gut feeling when

6:47something looks off.

6:49But we do this using machine learning instead. And takeaway here is that I

6:53haven't mentioned

6:54any use case here using generative AI. We could, but this is all deep learning

7:00and machine learning,

7:01just to show you that a lot of the different things that we interact with that

7:05use AI are not

7:07generative learning generative learning is one of the tools, such as machine

7:11learning

7:12and deep learning. Alright, so that's a good stopping point. And I'll see you

7:15in the next video

7:17on building a mental model. See you there.

0:00Welcome back. Remember when I said that we really didn't need to understand all

0:04of these

0:05different AI tools at a deep technical level? Well, here's why. My goal with

0:10this course

0:10is to give you a mental model for applying AI responsibly in business, meaning

0:17you can

0:18go to a meeting and you can use this information on the fly and name which of

0:23the four AI types

0:24is in play, connected to the framework and decide where human review needs to

0:29stay in

0:29the loop. So here is our toolbox and here we can think of this as compliance

0:34and here

0:35we have our model and here's a human in the loop and then here are the

0:39decisions frameworks

0:41that we're going to build using the red, yellow and green. And then here's the

0:45mental model

0:46so that you can use this on the fly and the takeaway of all of this is so that

0:51you don't

0:51ship poor AI products that are unsecured solutions or worst of all, leak

0:57sensitive company data.

1:00Here's how. Instead of thinking about all the different kinds of AI tools,

1:05which you

1:05will be able to explain after you become familiar with them, I think it's much

1:10easier to talk

1:11about how software has changed since AI showed up. So we have traditional

1:17software, which

1:18is software 1.0, you type something into a computer and it's not going to

1:24reason. It's

1:25not going to use AI. You can say, if this do that and it's always going to do

1:31that,

1:31it's never going to interpret or think, oh, there's another way I can answer

1:35this question.

1:36So that's the first part of the software. The second part is when AI showed up,

1:41it started

1:41to do things that are using human cognitive skills because it's software

1:47modeled after

1:48the brain and that's software 2.0. And then finally, we have generative AI,

1:54which is software

1:553.0. So it's much easier to think about AI in the context of the development of

2:02software

2:02and it's only three different kinds. So we have normal software that just

2:07follows

2:08instructions that are rules, then we have software modeled after the brain. And

2:12number

2:12three is really easy to remember because we use plain language. When you talk

2:17to chat

2:17GPT and you type in a prompt, that's software 3.0. You're literally programming

2:22without

2:23knowing it. So let's break it down. Software 1.0 is the rules that you write.

2:29And this

2:29is perfect when the outcome is clear, if X, then why this could be used for

2:35thresholds,

2:37lookups or dashboards. A good example is if the ticket is older than 24 hours,

2:43escalate

2:44because you don't want support tickets to be waiting for a long time. You want

2:48to get

2:48back to your customers. Now let's take a look at software 2.0. When it first

2:53came out,

2:53it was mostly machine learning. And machine learning is really about learning

2:58from the

2:59data. And these are things like algorithms that we can use to, for example,

3:04given past

3:05tickets, predict the priority one through five. So this is definitely software

3:102.0. But instead

3:11of thinking of it as machine learning, we just say that, well, it's learning

3:15from the data.

3:16And the same thing for software 2.0 here. So the difference is that this is

3:20early 2.0. And this is

3:22really the full version of 2.0, which is deep learning. And deep learning is

3:27where we go from

3:28algorithms to software models after the brain. For example, read the email body

3:34and detect the

3:35intent. So right now, if you think about rules and analytics, that's software 1

3:41.0. But if it

3:42starts to learn from the data in some kind of way, either through algorithms or

3:47by using

3:47software modeled after the brain, that's 2.0. Now we have software 3.0. And

3:54this is generative AI.

3:55So here we can draft language code and images from instructions. You steer it

4:01with prompts

4:02rather than changing the code. What you're doing is programming in plain

4:07language,

4:08for us that is English. An example of generative AI is summarize this thread

4:14for the next agent,

4:15or maybe draft a welcome email. So that's when we get into generative AI. And

4:20what people don't

4:21realize is that it's software written in English. And that's a very powerful

4:27statement. When you're

4:28typing into chat GPT, it's a form of programming. You're doing all of the

4:33programming things that

4:35you would do in software 2.0, but it's abstracted. It's not as direct, and you

4:40can't get as custom as

4:42you would in software 2.0. But the output that you get can be very, very

4:47similar to 2.0. So that's

4:50the power of software 3.0. And that's why prompt engineering is so important,

4:56because there are all

4:58these little details that go around software 3.0 that are paramount for

5:03security, for getting better

5:05prompts, for mitigating hallucination. That is where the model will just make

5:10things up. So if

5:11you're talking to GPT, sometimes it says things that are not correct, but it's

5:16very, very confident.

5:17So software 3.0 and prompt engineering go hand in hand. We'll talk more about

5:22that as we progress

5:24through the course. All right, so let's do a quick review. Software 1.0 are

5:29rules in analytics.

5:31It's always going to do if X, then Y. For example, if the ticket is older than

5:3724 hours escalated,

5:39it's always going to do that. The only reason that there are errors in rule

5:43programming,

5:43or software of 1.0, which is traditional programming, is when the rules change,

5:48and someone updates

5:50something, something changes. So that's why you get these errors. But if there

5:55's no errors, or

5:56nothing changes, they're going to do this literally forever. Now we have

6:01software 2.0,

6:03which is both classic ML and deep learning. So let's look at machine learning

6:08first.

6:09Here it learns from small patterns from structured data. So scores or labels.

6:15In this case,

6:16it was past ticket history. And when we're talking about deep learning, it's

6:20learning complex patterns

6:22from structured and unstructured data. All that means is things like text,

6:27images, and audio.

6:29It can read the email body and detect intent, but it can also look at an image

6:34and say that's a

6:35sunrise. And in the upcoming videos, we're going to explore each one of these.

6:39I'm going to show you

6:40some code. If X, then Y, so that you can see how it works. And that's pretty

6:45straightforward

6:46software 1.0. And then we're going to use the hot dog and the hamburger or

6:51computer vision.

6:52That's software 2.0. And then software 3.0, for that example, we're going to go

6:57to chat GPT

6:58and produce a similar output. We're going to ask it, what is this image and

7:03what is this image.

7:04And you're going to see that it's the same output that we got from software 2.0

7:09, where we use a lot

7:10of complicated software. But in 3.0, we're using English. And that's to take

7:15away. So let's finish

7:17with generative AI. And here we simply prompt the model and it just does

7:21something. So this could be

7:23text code or images. We could say summarize this thread for the next agent or

7:29draft a welcome email.

7:31So all of this is done in plain language, in our case, English. And in the next

7:35videos,

7:36we're going to go over each one of these. And I'm going to show you examples.

7:40So this is

7:41very, very clear. And this is the foundation that's not going to change. You'll

7:46be able to use this

7:47no matter what new model comes along. See you there.

0:00Welcome back. In this video we're going to explore software 1.0 and then in

0:05subsequent videos we'll go through software 2.0 and software 3.0. So this

0:10one is rules and analytics based. Let's take a look and let's consider a very

0:16simple scenario when ML or machine learning is not the right solution. So

0:22let's say and again this is just a classic probability problem and we have

0:27a six-sided die that is numbered one to six and we want to know what the

0:32probability of rolling a four is. So we would not use machine learning. The

0:38answer is the outcome is deterministic. So we can easily find out what that is

0:43which is sixteen point seven percent. So that's not a machine learning problem

0:49that is a software 1.0 problem. Meaning regular software. The software we've been

0:56using up to today before AI showed up and here's why. Machine learning is not

1:01needed because the problem is well-defined and computable. Here is

1:06exactly what we're going to do and I'm going to show you this and a couple

1:09other examples in Google Colab which is a browser based programming environment

1:14so I can show you how this works. And another reason we don't need to use

1:18machine learning is that a rule-based solution is exact and requires no

1:23learning. So this is not learning this is just doing what we tell it to do and

1:27using machine learning would add complexity and reduce accuracy. So

1:32reducing complexity is key. We want to make sure that we don't do that and

1:38that's going to be a recurring theme in this course. So when would you use machine

1:42learning? So you would use it when the patterns must be inferred from data not

1:48when the exact answers can be computed directly as we see here. So we know that

1:53software 1.0 is deterministic. If X then Y and we can call this a rule-based

2:00system. But when you have a problem or let's say an AI problem and you're

2:05trying to solve it how do you choose? You start with the simplest tool that works.

2:10If a rule-based system like software 1.0 works then you ship 1.0. If you need

2:16prediction from structured data then you might consider ML but what you're doing

2:21there is you're adding complexity. If the signal lives in unstructured content

2:26then you might move to deep learning and that you add more complexity. Maybe you

2:31need fluent text or code or summaries then you might use Gen AI. And they're

2:37all useful but you would add more and more complexity only if you need it

2:42because software 1.0 didn't work and then here software 2.0 didn't work and

2:48then you would add software 3.0. So as you go higher in the software scale

2:53you're adding more complexity. So let's try out some examples here. Let me make

2:57this a little bit bigger and this is Google Colab. So you can just start to

3:02write code. You can say 1 plus 1 and then you can run that and immediately once it

3:07connects to the cloud it'll give you that answer. But we could also say

3:11something like text. There's an answer 2. 1 plus 1 is 2. That's a rule-based

3:15system. It's math but it's the same idea. But let's explore this. So let's say we

3:20have text equals and I'm going to use an input function and what this does is

3:24ask me a question. So I'm going to say what's your name and it's going to store

3:28my name here but I wanted to do something. So I wanted to return that

3:32name. So I'm going to say print name or text in this case but I want it to be

3:36uppercase. Alright so when I do this it's going to ask me my name. I'm going to do

3:40it in all lowercase and then enter and it's going to say Jonathan in

3:45uppercase. This is a series of rules. It's not learning anything. If I do it

3:49again and I say Jonathan it'll do it again. It doesn't matter what I do. If I

3:54do uppercase it's still going to give me uppercase. So this is rule-based. But how

3:58about the six-sided die? So I'm going to say total. So total outcomes is six. We

4:03have a six-sided fair die and the desired outcome. We want to know what the

4:08probability of rolling a four is and we're going to say there's one desired

4:12outcome out of six and you can calculate the probability by dividing the desired

4:17outcome by the total outcomes. Okay so now all I have to do is just print the

4:23probability. If I run this I'm going to get sixteen point seven percent and if I

4:28run it again I'm still going to get the same answer no matter how many times

4:31that I do this because it's a deterministic rule-based answer. And I'm

4:36going to share this notebook with you so that you can try it yourself and just

4:39become familiar that this is, let me go up here, what's cool about these notebooks

4:43you can add text. So I'm going to say I'm going to double click it I'm going to say

4:46software 1.0. Okay so this is example of software 1.0. In the next video we're

4:52going to explore software 2.0 but instead of programming it we're going to use

4:57these animals here. Well a hot dog and a hamburger and we're going to use a

5:03computer vision model. And normally when I teach courses about deep learning I

5:08would write out the code on how to do that. We're not going to do that. I'm

5:11going to show you a fun way that you can build a model yourself and so you can

5:15see the classification of each one of these objects yourself. We're going to do

5:20that with Teachable Machine. See you there.

0:00Welcome back. This is the fun part.

0:02We're going to use the hot dog and the hamburger

0:05to build our own computer vision model.

0:09So it's going to classify whether or not it's seeing,

0:12so we're teaching the computer how to see,

0:14if it's a hot dog or a hamburger.

0:18Normally, there would be a lot of code,

0:21and it'd be complicated.

0:23But what we're doing is really a no-code version of this,

0:27to give you the same output,

0:29and that's why it's fun.

0:30So you can understand,

0:31okay, I have a good idea of software 2.0.

0:35Let's check it out.

0:35All right, so here is Teachable Machines.

0:37Let's just click on Get Started,

0:39and we're going to use an image project.

0:41But what's interesting to know

0:43is that you could also use audio or a pose.

0:46So really, the image project and the pose project

0:49are very similar,

0:51because it's using the webcam for both of these.

0:54You could just use images as well,

0:56but we're going to use our webcam.

0:57So they're really similar.

0:59This one, the audio project, that's very different,

1:02but they're all the same 2.0 idea.

1:06And this is all deep learning.

1:07So let's click on Image Project.

1:09We're going to use the standard image model.

1:11And here, we're just going to name this

1:13Hot Dog and Hamburger.

1:15All right, so now we need to give it samples,

1:17and this is how it learns.

1:19This is the amazing part.

1:20So let me go ahead and click on the webcam.

1:23Okay, so here we are.

1:24I'm going to get out of the frame,

1:25and I'm going to make this a little bit smaller,

1:28something like that.

1:29So I'm going to crop the image a little bit,

1:31and I'm going to get out of the way.

1:33And the way we do this is we're going to hold the image up

1:36and move it around into different perspectives,

1:39all right, to give it training data.

1:41So it's learning to identify this

1:44by looking at the training data.

1:46So let's go ahead and hold record,

1:47and let's aim for about 100 samples, maybe 200.

1:51Let's do 200.

1:52It goes pretty quickly.

1:53Let's do 300.

1:54Okay, 310.

1:55That's good enough.

1:56All right, so let's turn this off

1:59and now go to the webcam here,

2:01and I'm going to crop it again over here.

2:04There we go.

2:04Done cropping, and now we're going to use the hamburger.

2:07So this one's a little more challenging.

2:09I have to move it in different angles.

2:10So 310.

2:12We want to have the same number of samples.

2:14That way the deep learning learns better.

2:17It learns faster.

2:18So I'm going to twist it around top, bottom.

2:21There's a magnet there.

2:22It's not part of a normal hamburger,

2:24but let's go ahead and do a little bit more.

2:26There, 309.

2:27That's close enough.

2:29So this is how the deep learning model learns.

2:32It's going to look at all of these images

2:35and understand what it's looking at.

2:37So let me click on this,

2:38and now it's preparing the training data.

2:41This is stuff that we would have to do programmatically.

2:43This is most of the code, and then it starts to train.

2:46Now you can see that it's going through different epochs.

2:49This is how many iterations.

2:51It does it over and over and over,

2:53and it passes all of that information through neurons

2:56in the artificial neural network.

2:59Okay, so now, moment of truth.

3:01I'm going to crop this as well and go over here.

3:04All right, so now I'm going to show it.

3:06That's, okay, now it's saying hamburger.

3:09And actually what I'm going to do here

3:10is I'm going to go back over here at a class,

3:14and I'm going to say nothing.

3:16So when we're looking at nothing, it knows that.

3:18So I'm going to say, I'm going to go to the webcam.

3:20I'm going to crop it just like we did before.

3:22Something like this,

3:23and I'm going to do 300 samples of nothing.

3:26And that's really important because otherwise

3:29it's predicting that it's looking at a hamburger,

3:31and it's not.

3:32So really, this is the nuts and bolts of machine learning

3:36and deep learning in this case,

3:37when you're talking about preparing the data.

3:40So now let's train it again.

3:41All right, so now it's saying nothing, 100%.

3:44You can see that down here.

3:45So it's working really well.

3:46Now I'm going to take the hot dog,

3:48and you can see that it says 100% hot dog.

3:51No matter which angle that I put it in,

3:53it knows that that's a hot dog.

3:54Let's try it all the way in the back.

3:56It even knows, look at that.

3:57It's performing really well.

3:59So now it's looking at nothing.

4:01Let's try hamburger.

4:02There we go, 100%, it's a hamburger.

4:05So our model is doing pretty good now that we added nothing.

4:08And this is a great example of our neural network classifier

4:12for computer vision.

4:14Now it's classifying hot dogs and hamburgers.

4:17But if I get in there,

4:18it thinks that I'm a mixture of all three.

4:21It thinks I'm mostly a hot dog, but I'm not.

4:24So that's something to consider.

4:26When you give it some other information,

4:28it starts to hallucinate.

4:30And by hallucination, the best example I can give you

4:34is when we get into software 3.0 in the next video.

4:37When you ask ChatGPT something that it doesn't know,

4:40it's going to guess,

4:41but it's going to do so with so much confidence

4:45that it's going to throw you off or anybody, really.

4:48It does it in a way that it's like,

4:50well, of course it must be true.

4:51It's so confident.

4:53But as you saw earlier,

4:55did you see all the different probabilities

4:58of me being a hot dog or maybe a hamburger or nothing?

5:01It was confused.

5:03That's when it would guess.

5:04So that's the danger that we need to look out for.

5:07And that's why there's so much guardrails

5:09that are built into these models

5:11and so much safety that goes into this

5:14when you're trying to apply this to a business situation.

5:17All right, I'll see you in the next video

5:18when we explore software 3.0 or generative AI.

5:23See you there.

0:00Welcome back. Here we're talking about generative AI, and it's a subset of deep learning. So here

0:06we have built on deep learning. So that's what it is. So again, AI is the umbrella term,

0:12and then you would have machine learning. And then the subset of that would be deep learning.

0:16And then the subset of that would be gen AI. We don't really talk about rule based AI,

0:21because that's the older version of AI. But you do know that it exists. But for the context of

0:26this course for executives and leaders, this is really all you need to know. And more importantly,

0:32software 1.0, 2.0, 3.0, because that's how you're going to choose how to solve different business

0:38problems. So what is generative AI? So let's say you have text, some kind of historical data,

0:45videos, or unlabeled data, then you would pass that into a model. And you see where it says

0:50pre training, we just did training ourselves. So we took, in our case, video feed, and then we

0:56trained our model. And then it was able to do we could say information extraction, we can call it

1:03computer vision, but that's what it did. But we didn't use plain language. So when we're talking

1:09about adapting, we're not going to be using lines of code like we did before, or even teachable

1:15machine. Normally, when you're doing software 2.0, you're writing code, but the model learns by

1:21itself, we're not telling it what to do, like you would in software 1.0. And the final part here is

1:28software 3.0 is that we're writing English, we're just telling it what we want. And that's the

1:34adapting part, right? Once you train that model, you can just talk to it in plain language. And it

1:40gives you the output much like we did with software 2.0. Let me prove it to you. It's really

1:46interesting. Okay, so let's take software 2.0, and software 3.0 and compare it. So first, I gave it

1:52some input data, which was a hamburger, or 300 images of hamburgers, and then 300 images of hot

1:59dogs, then we trained the model. And then it made a prediction, we didn't say here's the shape of a

2:06hamburger, the difference between a hamburger and a hot dog is that one is going to be longer

2:11like this and have a bun with, you know, ketchup, mustard and things like that. Hamburgers more like

2:17round shapes with a patty and then lettuce and stuff. We didn't say any of that. It learned it

2:22by itself during training, right? Software 1.0, you would have to explain all of that, but not

2:28software 2.0. And then when it made the prediction, it gave us a probability. In our case, when we

2:34showed it a picture of this hot dog, it said that it was a 100%. That's a hot dog. So very high

2:41degree of accuracy. With software 3.0, or Gen AI, let me put machine learning and deep learning over

2:47here. But really, we were doing deep learning, which is better for more complex data. So here's

2:52how it works. We give it input data. So I'm going to give it a picture of a hot dog. And then it's

2:58just going to give me an output, it's going to give us text, human, human like text, in fact. So

3:05the main difference is that we didn't give a bunch of these images to train, we don't have to train

3:11the model. But when we tested our teachable machine, we held something up, right? What is this

3:17hamburger 100%? Same thing happens with software 3.0. If I were to give it an image like this,

3:23and I would ask it, what is it, it would reply in human like text, say that's a hamburger, and it

3:29would describe the hamburger, it's more complex. And we're not training one specific thing.

3:35Software 3.0 has been trained on a gigantic amount of data, in this case, most of the internet,

3:42and it's able to generalize. So if you give it almost anything you can imagine that you have a

3:48picture of, it would do a pretty good job of telling you what it is, right? As long as it's

3:54been trained on that data. So let's take a look. So here we have chat GPT. And that just gave it

4:00text, I told it what I wanted. And it gave me this nebula painting of basketball player, right? And

4:07here I give it a really detailed prompt. And again, this has to do with prompt engineering, if you say

4:13give me a dinosaur image, it's not going to be great. But when you're really leveraging prompt

4:18engineering, you can get a great image like this. But let's start a new chat. Okay, so here we have

4:24an image of a hot dog. This is the exact hot dog that we used for software 2.0. But we're not using

4:31it to train, we just want to know what is this. So I've asked the question, what is the probability

4:36of the class of this image? Before we gave it the class, we said these are 300 pictures of a hot dog

4:42300 pictures of nothing 300 pictures of a hamburger. Okay, so now look at this,

4:48it goes really into detail. So this is software 3.0, which is a subset of deep learning. So we're

4:54not training the models custom, this can just do almost anything that you ask of it. So now I'm

4:59going to ask for a percentage like we got it before. And here's an interesting let me make

5:04this much bigger. I don't have a classifier running yet. That's what we did in software 2.0.

5:09So I can't give you the actual probability values directly from a model, unless we set one up and

5:14run the image through it. That's what we did with TensorFlow inside of teachable machine. So saying

5:21use a pre trained image classification model, which is what we did, run the image through it,

5:27and then return the predicted top classes with their probability and percentage. Would you like

5:31me to run a plus hot dog image through one of these models and generate the probability

5:36percentages? Let's say yes. So instead of writing all the code, it'll just do it for us. But what

5:42we're doing is programming in plain language, because we're going to get the output. And that's

5:48where people don't really realize the power of software 3.0. They feel like they're talking to

5:55a human and the human is able to do stuff because they're a programmer and they can do all of this.

6:00Okay, but look what happened. Let's check it out. It looks like the classification pipeline failed

6:06because torch vision isn't fully supported in this environment. So it failed, it couldn't do it,

6:11it has limitations. So going back to this, we know that generative AI is built on top of deep

6:18learning, right? And it's trained on large scale data, meaning most of the internet, but you can

6:23adapt it to specific tasks using fine tuning or prompt engineering. So or prompting for short,

6:31so you can be very specific and leverage prompt engineering to adapt it to a specific task. And

6:39that's where you might not think that you have that power, because you feel like you're talking

6:44to a person. So when two humans are communicating, they understand the context, they understand guard

6:50rails, they understand all of these things. So you don't need to learn prompt engineering when

6:55speaking English. However, with software 3.0, prompt engineering is critical, more so when

7:03you're using it in a business. So when you're automating a task, for example, onboarding

7:09inside of your organization, you absolutely must consider prompt engineering. Furthermore,

7:16you might want to have a human in the loop. So they can look at the answer and be like,

7:20that is not appropriate. I'm so glad I caught this before we sent this to our new hire.

7:26Sometimes it can make mistakes that are catastrophic. And it's because of these reasons.

7:30So it's a general model. It's not a fine tune. So now that you know software 1.0, 2.0 and 3.0,

7:39let's review all of the different types of AI to make sure that we have this foundation down

7:44before we go on to the next skill. This is the takeaway. When we're thinking about artificial

7:50intelligence, this is the umbrella term for machine learning and its subset deep learning

7:56and its other subset generative AI. These two are software 2.0 and this is software 3.0 and

8:04software 1.0 is traditional rules. We tell it what to do. So in a way, software 1.0 is where we tell

8:13the computer what to do. Software 2.0, the computer learns on its own. After we build the

8:19model, we just give it the data and it learns from the data. Software 3.0, we're using plain

8:25language. Knowing each one of these types of software and each type of AI is crucial. There

8:32are other terms that you might hear that are not particularly important right now for our context,

8:37such as NLP or LLMs. NLP is natural language processing. That's how we teach computers to

8:45read and write. It's text-based. LLMs are large language models like ChatGPT, Gemini and Claude.

8:53All right, so that's it for this video and I'll see you in the challenge next. But just remember,

8:58there are three types of software and there are three subsets of AI and that's all we need before

9:05we get into the next skill. And the challenge is going to make sure that we have these concepts

9:09down before we move on to the next skill. All right, until next time, I hope this has

9:14been informative and I'd like to thank you for viewing.