Explore AI Literacy for IT, HR, and Legal Teams

0:00Hello and welcome to the first skill in the AI Compliance Privacy and Copyright for HR,

0:07Legal, and IT course. Before we talk about policies, privacy rules, or copyright,

0:12we need to understand what AI actually is and how it evolved. And we'll do that in this course,

0:19which is part of a series of AI mini courses. And we are here. But it is worth mentioning

0:25that technical teams can take a very similar course here in AI Compliance Privacy and Copyright

0:33for developers. And most of these mini courses include a micro crash course on prompt engineering.

0:40And this is the mini course on prompt engineering that I highly recommend that you take if you want

0:46to learn more about prompt engineering. Here we go into much more detail on prompt engineering,

0:51but it's still a mini course, which means that it's not a full course.

0:56And if you do want a full course, please let us know in the comments, because this is precisely

1:00the data that we need to hear in order to make the content that you are most interested in.

1:06And for those of you who are interested in programming, but don't have a programming

1:10background, AI vibe coding and security is a great place to start. Essentially,

1:15it's like prompt engineering, but for non programmers who want to leverage AI for

1:21programming with a special emphasis on security, because if you can't read the code, you might want

1:26to be careful. And there are steps that we can take to mitigate the dangers when vibe coding,

1:33which is incredibly fun. And if you do have a programming background, then I would suggest the

1:38AI genetic coding course. And finally, if you're building an application from scratch or you're

1:43part of a leadership team, then I would suggest the artificial intelligence for executives and

1:49leaders course. And finally, these two are AI productivity courses. The first one down here,

1:55AI productivity for professionals is a platform agnostic approach to AI productivity. But really,

2:02we're going to lean very heavily on Google in this course, because there's really just two choices

2:07when you're talking about these large platforms that have self-contained ecosystems of applications.

2:13There's Google that uses Gemini and then Workspace. And then you have this one, AI Essentials with

2:20Microsoft Copilot, meaning we're going to use the Microsoft ecosystem with Copilot and Microsoft

2:28365. So you can choose A or B. And the only other choice is to use different AI chatbots and connect

2:38different parts of these ecosystems into them. That is not a best practice for many reasons,

2:44particularly when we're talking about AI compliance, privacy and security, which you're

2:49going to learn about in this course. And on that note, what's included in this course? So what is

2:54in scope? We'll identify the difference between main types of AI, because it's not just one thing.

3:01There are many different things and have a really cool way to break this down in a non-technical way

3:08because it can get very complicated very fast. And if we think of it in basic terms, then you can

3:13build a mental model that you'll not only use for the rest of this course, but also outside of work

3:20as AI becomes more and more ubiquitous in the workplace. We'll also understand basic non-technical,

3:26that's the most important part, AI compliance, privacy and copyright. This is the non-technical

3:33course for HR, legal and non-technical IT teams. We'll also leverage AI tools using prompt

3:40engineering for basic governance. But it's also important to mention, this is not a full course

3:46on AI governance because that is a much more in-depth course. That would not be a mini course,

3:52that would be a full course. But what we're doing here is a mini course, sort of like a crash course

3:57on the most important parts of AI governance distilled into the key topics of compliance,

4:04privacy and security. And don't worry because we're not going to learn how to code. This is a

4:09non-technical course and no tutorials on how to build AI chatbots or AI agents, which is a lot of

4:15fun, but we're not doing that in this course. But first a little bit about me. My name is Jonathan

4:21Barrios and I'm excited that you're here because I love to teach data science, machine learning and

4:27deep learning. And in this course, that really means AI, but really what I call myself is an AI

4:34data scientist. It's a combination of data science and AI engineering, which includes

4:40deep understanding of both machine learning and deep learning. What you're going to learn about

4:45in this course. I've worked with many different types of AI built SAS applications, AI applications

4:52and worked on streaming platforms. And more importantly, I've been teaching for quite some

4:57time at top platforms like Treehouse, Thinkful and Chegg. And now I'm happy to be part of the

5:02CBT Nuggets team. So where can you find me? You can definitely find me on LinkedIn at Jonathan

5:08dash Barrios dash AI or on my blog website at Jonathan Barrios dot com. And you can also find

5:16me here on X at AI underscore data underscore science. And again, that's what I call myself.

5:23That's the field that I'm in because it's an emerging field and it's also my X handle. So a

5:28quick note about X. Some people love it. Some people hate it. But to be quite honest, I don't

5:34really engage in social media because I think it's a waste of time. So you might be wondering,

5:40why are you telling me to follow you on X? And here's why. The only reason I'm there is to keep

5:46up with all of the developments in the AI community. And the AI community is mostly on X.

5:53It's incredible. For example, top minds like Andrej Karpathy are continuously posting amazing

5:59things directly on X. And it's really the simplest way to get information about the AI community.

6:06Plus, you can use a built in AI grok, which is part of X AI, which is integrated into X.

6:13And that's all I do. I only follow people who talk about AI and I do not follow anybody else.

6:19That way, my feed is really amazing. It's kind of a superpower, really. So follow people like me

6:26and others that talk about AI and you'll be in a great spot. And otherwise, I highly recommend

6:32not using social media. It's not great. And instead, use that time to learn.

6:38All right. Now, here is the first skill. We're going to identify the main types of AI,

6:44because like I said, AI is not one thing. And before we talk about policies, privacy rules,

6:51or copyright, we really need to understand what AI actually is and how it evolved.

6:57The truth is, AI didn't appear recently after some new breakthrough. It's the result of decades

7:04of change in how computers learn. And instead of a technical explanation about the various types of

7:10AI, I have a very simple way to approach this. And I like to call it the path of least resistance,

7:17meaning it's the easiest way to achieve a goal. And I'm borrowing this term from physics. And

7:23while this can mean a lot of different things, it just means we're building a mental model to

7:28understand AI so that we can leverage it to better understand AI governance. And that's it.

7:34And the way that I've chosen to do this is by breaking down AI into its simplest components,

7:40meaning we'll look at the evolution through the lens of three types of software. Software 1.0,

7:47which is traditional programming. Software 2.0, where the software learns from data on its own.

7:53And software 3.0, which is the generative era. And this is what we know as chat GPT. And that's

8:01what I mean that AI is not one thing. A lot of people think it's just chat GPT, but that's just

8:07one slice of the pie. And by the end of this skill, you'll see exactly how this evolution

8:12affects today's bigger compliance questions from protecting employee data to managing bias in

8:19hiring or ensuring copyright safe content generation. So in this first skill, we're

8:26really laying the groundwork for everything that follows. We'll explore what AI really means,

8:33not from a coding standpoint, but from a business and compliance perspective, you'll learn to spot

8:39whether an AI tool your team uses is rule-based, software 1.0, data-driven, software 2.0,

8:47or generative like chat GPT, which is software 3.0. And this is a much simpler way to understand AI.

8:55And each type is associated with very different risks and responsibility. For example, when we're

9:01talking about HR, systems using data-driven AI may accidentally learn hiring bias. Legal teams

9:09using generative AI tools may face copyright or confidentiality issues. And when considering IT,

9:17these departments deploying AI chatbots may expose private information if prompts aren't handled

9:24carefully. Once you recognize the type of AI behind the tool, you'll know what kind of safeguards or

9:30policies that it needs. So that means we're going to define what AI is in plain language, explore

9:38software 1.0, 2.0, and 3.0, and understand traditional learning and generative systems.

9:45Then we'll connect it to HR, legal, and IT examples. And finally, we'll identify early

9:51compliance and privacy challenges. We'll do all of this to bridge to the next skill,

9:57on-prompt engineering. That's where we're going to apply prompting for safe and compliant outputs.

10:04Meaning, when you understand how AI learns, you can better manage how it behaves.

0:00When people hear AI, they often think of tools like ChatGPT or Copilot or other models or systems

0:08that can write or summarize text, which is also known as generative AI or Gen AI for short.

0:15But like I said earlier, AI is much broader. At its core, AI is simply software that learns

0:22from experience. Instead of following strict rules, it identifies patterns in data and uses

0:28them to make predictions or generate ideas. Even when it's generating something like text in ChatGPT

0:36or Copilot, it's predicting the next word. And this is how it generates those ideas.

0:42So let's break it down into three easy categories and also demystify what artificial intelligence is.

0:49You can really think of it as an umbrella term, which could be many things like rules,

0:54search, machine learning, deep learning, and generative AI. But we're really going to focus

1:00on the core technologies that represents AI, and that would be machine learning, which finds

1:06patterns in the data. For example, HR systems might predict which candidates could be a good

1:12fit based on past hires. And then a subset of machine learning is deep learning. Here we use

1:20layered models to understand complex data like images or speech. In short, we're using software

1:28modeled after the brain. So deep learning uses something called neural networks, which are

1:33artificial neurons that mimic human neurons. We have neurons in our brains that are connected,

1:39which means our brains are made up of neural networks. And that's how we think, talk,

1:46store memories, and so on. And that's what makes deep learning so fascinating. You can think of

1:51machine learning as algorithms and deep learning as these neural networks. For example, in IT,

1:59you could have security tools that recognize phishing emails or facial authentication

2:05on devices. We'll explore each one of these in more detail soon. And a subset of deep learning

2:11is generative AI. Here, gen AI creates new text, images, or even code based on prompts. An example

2:20would be legal teams using AI to summarize contracts or draft policy templates. The takeaway

2:27here is that gen AI is just one slice. It's not the whole pizza. And what we're doing is simplifying

2:34this because we don't need to know all of the learning styles like supervised, self-supervised,

2:40unsupervised, reinforcement learning. These are all really interesting aspects of AI, but we don't

2:46really need to know that. And when we're thinking about modalities, we can use this simple mental

2:51model to see how it's used with vision, speech, and text. NLP is just the technology, natural

2:58language processing that we use for text and also recommendation. For example, when you're watching

3:04a YouTube video at the very end, it gives you recommendations and even robotics. Here's the key

3:09concept. AI doesn't know anything. What's most likely based on its training data. If the data

3:16includes bias, private information, or copyrighted content, the AI may reproduce those in its output.

3:24And that's why compliance isn't just about setting rules. It's about understanding how the tool

3:30learns and how to use it responsibly. And when we look at AI use cases in everyday life, you can

3:38see it in your phones when you use it as a way to authenticate who you are so that it opens up your

3:44phone. That's facial recognition and that is deep learning. So it's classifying you specifically,

3:51but it can also be used for other applications, identifying individuals in a crowd, and so on.

3:57Computer vision also uses deep learning and here it's classifying a person, a car, a tree, a sign,

4:05and so forth. And this is how computer vision allows cars to drive themselves in the city streets.

4:12And when you're thinking about robotics, well, that's also deep learning because it uses

4:17computer vision to see. But how about if it speaks? It could be a combination of deep learning plus

4:24gen AI. How about these little motors that are in the joints? Maybe those use machine learning.

4:29The takeaway here is that all of these can combine. It doesn't mean that one AI type is for one thing.

4:37When you're thinking about fraud detection, machine learning is an algorithmic approach

4:42to identifying outliers, such as very large transactions in a completely different country.

4:49And that's used for fraud detection. And when you think about intelligent document processing,

4:54IDP, you can take a picture of a document and convert that to text. That would be deep learning.

5:00But when you use that document to generate new text, well, that would be gen AI. And again,

5:06we're combining these technologies to create different products. You could use deep learning

5:12by itself, machine learning by itself, or generative AI by itself, or you can combine them.

5:18And here is where we build our mental model. A much simpler way of thinking about AI and all of

5:25the technologies contained in that umbrella term, such as machine learning, deep learning, and

5:30generative AI, is to really think of it as three types of software that evolved pretty quickly.

5:37Software 1.0 is traditional programming. If you've heard about coding or programming,

5:44that's software 1.0. Software 2.0 is where machine learning and deep learning come in,

5:50because those models learn from data. We're not using rules to program them. And then finally,

5:563.0 is generative AI. These are models that generate something new from its training data.

6:03And using this mental model allows us to think more clearly about AI compliance, privacy,

6:09and security, instead of being bombarded with all of these technical terms that can be complex.

6:16While they are interesting to learn about, I think it really doesn't help to get too deep

6:22into those technologies, unless you want to build the models yourself. If you are somebody in HR,

6:28or a legal team, or you're on a non-technical IT team, it's much better to think of it in terms

6:34of software, how they behave, so that you understand how you can apply compliance,

6:41privacy, and security to your organization for your department or team. And in the next video,

6:47we'll begin by discussing software 1.0, how rules are created by hand, meaning code creates behavior.

6:55See you there.

0:00Welcome back. In this video, we're going to explore software 1.0, 2.0, and 3.0. We already

0:07know what artificial intelligence is, more or less an umbrella term for machine learning,

0:12deep learning, and generative AI. We know how they're implemented in everyday use cases,

0:17such as computer vision, robotics, and facial recognition. And to manage AI well, it helps to

0:24see how software has evolved, because every new era changed what humans are responsible for.

0:32And that's why we're doing this comparison. When you consider software 1.0, these are the

0:37rules that we wrote, meaning we wrote rules by hand. And by hand, I mean code. So programming

0:44that creates some kind of behavior. You can think of early HR systems that only followed exact

0:50filters that you created. For example, when looking at potential candidates, you might look

0:55at years of experience, and you might set a criteria that you want that to be greater or

1:00equal to three years. In this case, the privacy risk was low because nothing was learned or stored

1:06beyond what you entered. If you use this explicit rule, greater than or equal to three, it's not

1:12going to change magically, unless there's an error in the code. But once it starts to work, you'll be

1:17able to rely on that check for the years of experience. And that's why the risk was low.

1:23And when we look at software 2.0, things are a little bit different, because now systems began

1:30learning patterns from examples. For instance, analyzing past hires to predict future performance.

1:37And so you might imagine, what's the challenge here? If your data has any bias, your system is

1:42going to learn that bias. And this is an important key point. Responsibility shifted from writing the

1:48code to curating clean and fair data. So here, the code in software 1.0 is really responsible for the

1:57behavior. But here in software 2.0, we are training the model, also known as fine tuning, that creates

2:04the behavior. But really, this is data quality. If there's bias in that data, well, then your system

2:12is going to learn that bias. So the responsibility is to make sure that that data is clean and fair

2:18to begin with. And when we look at software 3.0, these are rules that are created using prompts.

2:24You can use tools like ChatGPT to summarize this confidential report or draft a job description.

2:31This might save time, but it also raises new questions. Could private data leak through

2:38prompts? And that's a great question we're going to explore in this course. Who owns the AI-generated

2:44output? How do we ensure the response is accurate and compliant? So what we're doing when we're

2:50entering prompts, we're steering the model to create the behavior. And when I think of prompts,

2:55I really think of programming in plain language. Instead of using programming for software 1.0

3:03or using data for software 2.0, we're using plain language to steer that model to create that

3:11output. It's like creating a program using plain language. And this is the framing from Andrej

3:17Karpathy, who also said English is the hottest new programming language. And again, we're talking

3:23about software 3.0. These are the chatbots like ChatGPT, Copilot, Gemini, Grok, and others. And

3:33while each stage brought huge productivity gains, they also shifted the meaning of responsible use.

3:40And that's why we have this course, because it is counterintuitive where the responsibility lies,

3:46mostly because we don't understand how these models are trained. And that's what we're going

3:51to explore in this skill. We're going to use this hamburger and this hot dog as examples to train a

3:58model. We're not going to write any code. It's going to be fun. And it's easy to understand how

4:03we're training a software 2.0 model. And if our data, which are images of the hamburger or the

4:10hot dog have bias, then the model is going to have bias. For example, if I use 100 images of hamburger

4:18and one of hot dog, guess what? It's going to favor hamburger. And that's pretty easy to

4:23understand. And it's also a lot of fun, like I said. And this is going to give you the insight

4:28that you need when you're thinking about compliance, privacy and security. You can simply think to

4:33yourself, well, is this a program? Is this traditional software programming where we

4:38enter some data? The rules won't change until we change the code. And in that case, well,

4:43the responsibility lies with the person that is doing the programming. And when we think of

4:48software 2.0, when we're using these different labeled images where the model learns from the

4:54data, the responsibility shifts to the person creating the data for the training. And to make

5:01it simpler, we can just say the person that is gathering the data and training the model is the

5:07person responsible for the fairness and the behavior of the output of that model. And when

5:13we get to software 3.0, well, that's going to be prompts. So the responsibility shifts to the person

5:20working with prompts. And that is precisely why we're going to get into prompt engineering in the

5:24next skill, because that is how you can make your outputs compliant and mitigate any privacy leaks

5:31or data leaks or security vulnerabilities. Now let's take a look at an example of software 1.0.

5:39We can say that this is pre AI, meaning computers didn't learn anything. They just followed the

5:45exact instructions written by humans. So the humans write these rules by hand, meaning they

5:52create the code. And that's what creates the behavior. Every decision had to be hard coded.

5:58If this, then that. So let's take a look at what this looks like in Google Colab. The reason that

6:04we're using this is again, you're not going to learn how to code, but this is probably the easiest

6:09way to get started with any programming example without having to do any setup. And all you need

6:16is a Gmail account. That way I can just type some code in here and then run it and you can see how

6:22it behaves. So for this short Google Colab example, we'll build a simple HR rule-based system, the

6:30kind of logic early recruiting software might have used to screen applicants. Here's how that works.

6:36So first I'm going to define some kind of criteria. So applicant years equals four, and we can easily

6:43change this number. So for right now, I'm going to click on this play button, which runs the code,

6:48and it's not going to do anything. It's just going to save the number four so that we know

6:52what the applicant years are. So in this case, the applicant years are four. Now we can do some sort

6:58of hard coded rule. If applicant years is greater than or equal to three, then I want you to print

7:07this statement. And here we can say something like candidate meets experience requirement,

7:13but we need to also handle if that is not the case. So we can use something like this else print

7:19this. And here we're going to say candidate rejected, not enough experience. All right. So

7:25let's run this code. Okay. Candidate meets experience requirement. Perfect. What happens

7:31if I run this again? I get the same exact output. If I do it again, the same and again. So it is

7:37deterministic. So here's what's happening. The system checks the applicant years of experience

7:43up here, which is four. And it does that here. If the applicant years are greater than three,

7:49print this statement. If not, then print the statement. If X, then Y there's no flexibility,

7:56no learning and no hidden logic, which is good and bad. So to recap software 1.0, everything is

8:04deterministic, meaning it's going to be the same predictable output every time. Once you have this

8:10logic set in place, unless something happens to this code or it's changed, then it's always going

8:16to be predictable in the same. That makes auditing and compliance much easier because you can trace

8:22every outcome to some kind of rule. And here's the rule print this else print this pretty clear.

8:29If the rule was biased, like filtering out applicants without a degree, every result

8:35would reflect that bias. If the laws changed or the job criteria evolved, the software wouldn't

8:42adapt until a human rewrote the code. So while software 1.0 carried low privacy risk, it offered

8:50no adaptability. Everything depended on human logic. The next video, we'll take a look at

8:56software 2.0 and get the hamburger and hot dog into this classifier so that you can understand

9:04how we train these models and how easy it is to introduce bias when you don't have quality data.

9:12See you there.

0:00Welcome back.

0:01Now that we've seen how software 1.0 followed explicit rules, let's move on to software

0:072.0.

0:08Here, the world of machine learning and deep learning.

0:13Sometimes when you hear machine learning, deep learning is the subset of machine learning,

0:17but they are sometimes also used interchangeably.

0:20But the easiest way to think about this is algorithms, some kind of sequence of steps.

0:26And in here we're using neural networks.

0:29But to keep it super simple, you can just say that all of this is machine learning.

0:33It's not super important to know the distinction between machine learning, which generally

0:38uses algorithms and deep learning, which specifically uses neural networks.

0:43But it is super interesting because chat GPT does seem to work like a human in some cases.

0:49And that comes from deep learning and the artificial neural networks, which is really

0:54software modeled after the brain or software modeled after us.

0:58So software 2.0 really is the world of machine learning, which includes deep learning.

1:04That means that we stopped hard coding logic that was software 1.0.

1:10And instead we show examples and the systems can learn its own rules from the data.

1:16And to bring this to life, we're going to use something called Teachable Machine, which

1:21is an educational application by Google.

1:24And we are going to step away from the H.R.I.T. and legal examples and take a look at the

1:30hot dog and the hamburger as training data.

1:33And in doing so, you'll have a firm understanding of how software 2.0 works.

1:39We show it these examples and then it learns to make predictions.

1:42And then we can talk about bias very easily and hallucinations, meaning when the model

1:48just starts to make stuff up.

1:50And that has to do with the examples that we're showing them.

1:54This is very important.

1:55And that's why we're going to do this example.

1:57And as I promised, it's going to be fun.

1:59All right.

2:00So here we are at Teachable Machine.

2:02It says here, train a computer to recognize your own images, sounds and poses.

2:07A fast and easy way to create machine learning models for your site, apps and more.

2:13No expertise or coding required.

2:15So let's click on get started.

2:17And here we have a choice.

2:18We can do image project, audio project or pose project.

2:22In essence, these are all the same thing.

2:25Computer vision.

2:26Here we're giving it images.

2:27So that's pretty straightforward.

2:29Here we're turning the audio or sound into a visual representation of sound.

2:35So again, images.

2:36And then finally poses.

2:38These are just sequences of images, lots of images.

2:41It takes all the images from your webcam and stitches them together to create the training

2:46data.

2:47So much larger project.

2:49So let's start here with the image project and we're going to use standard image model

2:52here.

2:53And now we have two classes, some kind of a training.

2:56And then here we preview, meaning this is where we test the prediction accuracy of the

3:01model.

3:02So I'm going to go over here and click hot dog.

3:05That's our first class.

3:06And then here, hamburger for the second class.

3:09And the way that this works is pretty straightforward.

3:11You click on the webcam and then click on crop here.

3:14And then let's move this into an area where I am not.

3:16I can move out of the way a little bit and then that way I can grab a hot dog.

3:21And there we go.

3:22There's the happiest hot dog on the Internet.

3:24So we can just hold this down and we want to get, let's say, about 100 samples.

3:29OK, we have 103 and to delete them, I'm just going to hit the trash can by hovering these

3:33images.

3:34And now we have 100 samples of hot dog.

3:36Let's do the same thing for hamburger.

3:38OK, I'm going to crop this, get out of the way, and now we can use the hamburger.

3:43So what would happen if we just did very few of these, like three?

3:48I don't know.

3:49Let's try it out.

3:50So now we're going to train the model.

3:51And this is the part where it learns from the data.

3:54It's looking at all the different patterns.

3:56So now it's ready for predictions.

3:57I'm going to click on crop again, move this out of the way so that I'm not confusing the

4:01model.

4:02And now we can hold up a hot dog and it gets that right.

4:06How about hamburger?

4:07It's a little bit confused, especially if you do something like this.

4:09Oh, it thinks it's a hot dog.

4:11Oh, that's definitely a hamburger.

4:13Now it's a hot dog, right?

4:15It's a little bit confused.

4:16And that's how we introduce bias, because we only have three samples here.

4:19So let's go back to the webcam and let's do this.

4:22Let's add some more images to fix that.

4:25OK, 101.

4:26Let me take away one of these.

4:28Perfect.

4:29But if I go over here and I try to make a prediction, it's still a little confused because

4:34we need to train the model again.

4:35All right.

4:36It's done training and now we can test it out.

4:39Hamburger 100 percent.

4:40No problem this time.

4:42And if I take it away, it thinks that the background is a hamburger.

4:45And if I get into the picture, it thinks I'm a hamburger, which I'm not.

4:48So let's click on add a class and now let's add background or we can say waiting or let's

4:54say background.

4:55And all I have to do here is crop, get out of the way and hold this down for 100 images.

5:01I can delete one of these.

5:03And if I go back into the sample, it thinks I'm a hamburger still.

5:06So I need to train again.

5:08And now we have three classes.

5:10This is pretty common because you don't want to classify the background is either a hot

5:15dog or a hamburger.

5:16But this is a way to show you how this works.

5:19Not saying background.

5:20But the minute I do this hamburger and then the minute that I do this, we now have hot

5:25dog.

5:26And if I bring a phone into the picture, it has no idea what's going on because this is

5:30not fair.

5:31It's like I don't even know what this is.

5:32This is a phone.

5:34Same thing with coffee mug.

5:36It's really confused.

5:37And all that means is that we haven't given it the training data on different samples.

5:42We only have hot dog and hamburger and now background.

5:46Because if I move out of the way, it's background.

5:49If I get into the picture, I'm a hamburger.

5:51Well, not really.

5:52And that's exactly how chat GPT in these 3.0 software models can make mistakes.

5:59They're built using software 2.0 through this process.

6:02But instead of images, we're giving them large amounts of text.

6:06So imagine we're giving it a lot of text in English, but then no other language.

6:12So these chatbots would be biased towards English.

6:15And that's why fairness is such a huge component for software 3.0 and software 2.0.

6:22Because training the model means selecting the right data and making sure that the corpus

6:29or the body of text that you're feeding into these language models is representative.

6:35All languages, all kinds of knowledge.

6:38And as you can imagine, that can be pretty difficult.

6:40All right.

6:41So let's break this down.

6:42In software 2.0, we taught a tiny classifier, meaning it was doing two classes, hamburger

6:49and hot dog.

6:50The background was just a placeholder for the educational example.

6:55But software 2.0 doesn't mean that it has to be just two.

6:58It could be many different classes.

6:59It could be 10 types of food or 100 types of food.

7:02So the main distinction between software 1.0, which are rules that are hand coded through

7:07a programming language, here we're using rules that are in weights.

7:11So weights just means how important are the different features.

7:16For example, when you look at a picture of a hot dog, it's looking at all of these lines.

7:21So it has the mustard and ketchup here, the two little eyes, the smile.

7:26These are all features.

7:27And it's asking itself, what are the most important features in this hot dog?

7:31Even when we move it upside down and all around.

7:34And the same thing goes with hamburger.

7:36Here we're looking at all of these different patterns and shapes.

7:40And it's saying what's important in order to identify each class, hot dog and hamburger.

7:47And it's not important that you understand weights so much, but you do now understand

7:51training because we did that.

7:52We fed it all the images and then it trained.

7:55And that means it collected patterns.

7:58And as you can imagine, certain patterns are going to be more important than others.

8:02So the importance are weights.

8:05So how important is a shape like this when you're trying to identify a hot dog?

8:09Very important.

8:10How about a shape like this?

8:12Not so important because we don't see that shape inside of a hot dog.

8:15How important is this shape in a hot dog?

8:18Not so important because we didn't really see that.

8:20So this one is going to get a higher weight and this is going to get a lower weight.

8:24And that's how that works.

8:25So instead of hard coding weights, the model learns patterns through training.

8:31Fine-tuning is just a specialized version of training where you train on a specific

8:35set of images.

8:36But you can really just consider both of these as training on examples that we give it.

8:42And it identifies patterns and it creates rules by assigning different weights to different

8:47patterns.

8:48And that's how it works.

8:49And we use Teachable Machine to see that just now using that image project.

8:53So instead of 20 to 50 samples, we gave it 100 samples each.

8:57But when they were imbalanced, we started to introduce bias.

9:01So just remember software 2.0 learns from examples or you can say it learns from data.

9:08And every time you train the model, it's going to change the weights depending on the examples

9:13that you give it to learn.

9:15The idea here is not to have a deep understanding of how this technology works because we really

9:20didn't go into all of the details.

9:23We're just talking about this at a high level, which is super important for prompt engineering

9:28in the next skill.

9:29This is what we're doing.

9:31We're setting us up for success in prompt engineering and we can create compliant, private

9:37and secure outputs through prompt engineering because we understand that the software that

9:44we're using, software 3.0 was created using these rules.

9:49It learns from data.

9:51And that serves as a foundation for prompt engineering.

9:54It will really start to click when we get into that next skill.

9:57And on that note, let's talk about software 3.0 in the next video.

10:02See you there.

0:00In software 1.0, we wrote the rules.

0:02In software 2.0, data taught the rules.

0:06And now in software 3.0, we don't write or train directly.

0:11We simply guide the AI with language.

0:14And that's why I like to call this plain language programming.

0:18You don't need a programming background,

0:21but you do need to know some basic prompt engineering.

0:24And if you're asking what is prompt engineering,

0:27it's plain language programming.

0:29What does that mean?

0:30We're guiding the model or the chatbot using plain language, meaning prompts.

0:37So prompts are plain language programming.

0:40And this is really important when it comes to compliance,

0:44privacy and copyright for non-technical teams.

0:48We can call this a form of AI governance,

0:50meaning you can use prompt engineering for safe,

0:54private and copyright friendly outputs for non-technical teams.

0:59And the umbrella term here is AI governance, which includes all of that.

1:03And software 3.0 is also called or referred to as generative AI

1:09because it generates outputs based on our prompts.

1:12So tools like ChatGPT, Gemini, Microsoft Copilot, Grok, Claude from Anthropic,

1:19they can write summaries, draft documents or even answer questions.

1:23But instead of coding or feeding them structured data,

1:27we communicate in natural language by prompting them.

1:31Let's see what this looks like in action with a special attention to AI governance.

1:35Let's say that we have some text here.

1:37And this is an employee feedback file that I asked the model to generate.

1:42And it did so begrudgingly because there's a lot of sensitive information here.

1:46But it also is a little funny because it's kind of pretty negative.

1:50I didn't ask it to be negative, but this is what we got from the model.

1:53And again, that has to do with its training data.

1:55So let's put this in quotes like this.

1:58And I'm going to say that this is an excerpt.

2:00And I'm going to go up here and just say summarize this

2:03and include names and performance comments.

2:07Let's see what we get.

2:08And by the way, I'm at GPT.com and I'm using my personal account here just to make a point.

2:14Normally, when you're working with an organization,

2:17you're going to be using the organization account,

2:20meaning it's either going to be Google Workspace that uses Gemini

2:24or Microsoft that uses Copilot.

2:27Very rarely do you use chat GPT and drop documents in there for an organization

2:32because compliance in that regard is very difficult to do.

2:36Once you have Google Gemini or Microsoft Copilot set up by your organization,

2:42they handle the compliance behind the scenes.

2:44The administrator is doing that,

2:46which means that you don't have to do that manually.

2:48It should be taken care of for you.

2:51But when you're using chat GPT like I am here and you're dropping in information like this,

2:56which is not real information, so we're OK, it's safe.

3:00This is a security issue and a privacy issue and a compliance nightmare.

3:04You definitely don't want to do this.

3:06And you can click on this button and connect to Google Drive or Microsoft products.

3:11But again, you don't want to use these connectors

3:13because how can you ensure compliance, privacy and security?

3:19It's really hard to do that.

3:20So let's run this and see what we get.

3:22OK, so we have a name, we have bonuses,

3:25and then we have a project name here, which might be intellectual property.

3:29And then we have all of these comments that are associated to this person.

3:34And when you combine this project name, this person,

3:38these negative comments in this amount, these can be pieced together.

3:42And while it might be difficult to do that and the chance of that is not 100 percent,

3:47it is still something that you want to avoid.

3:49And you might say, well, isn't this innocent?

3:52Well, while it might seem innocent, this request could expose personal or

3:56sensitive information because the AI doesn't know what is confidential.

4:01It just follows your words.

4:02So we can go back up here and we can add something called safety.

4:07And I'm just going to say if you find the AI, replace it with placeholders.

4:12Let's just try that very simple line.

4:15OK, with that one very simple line.

4:18Now it's employee A, but we still have project stellar.

4:21So it's not the greatest way to do this.

4:24You are protecting the output.

4:26However, this is still going into a chatbot.

4:29So prompt engineering can help you control what the AI does with language,

4:34but it doesn't anonymize data.

4:36True compliance starts before prompting.

4:39That means that when your organization is collecting this data,

4:43they should have removed that data or anonymized it or

4:47done something to the data pre-processing so that when it hits the database and it's

4:52ready for all of the employees in that organization to consume that data for their work.

4:58You don't want to have to do this in a chatbot.

5:00That is the absolute wrong way to do it because it's already gone into the model

5:05and you don't want to feed that to the model for many reasons.

5:09Even if your chatbot says that it doesn't use the text for training,

5:12you still don't want to do that because data leakage can still happen.

5:16The worst case scenario is that you are using chat GPT and it's being used for training

5:22and you're entering personal data of some individual for a company

5:27and maybe even intellectual property.

5:29That's called PII, personally identifying information like employee A's name,

5:35their performance for their job and the different bonuses that they've made,

5:39financial information.

5:40All of that should not be included in the database.

5:44And if you do have access to that information,

5:46you don't want to put it into the database before it's been redacted

5:50or anonymized or filtered and pre-processed in some way.

5:55So if you do find this,

5:58you can either do it manually before you put it into the model

6:01or reach out to some administrator or the data science department.

6:07They're the ones that would handle the pipeline

6:09that ingests data and makes sure that it's compliant.

6:13And it's not everybody's job,

6:15but when you're working with these models, you want to keep an eye out for this.

6:19It's very important.

6:20And we're going to use prompt engineering to help you identify when this is happening.

6:25And you definitely don't want to use it as an output.

6:28If you're going to send it in an email, for example,

6:30or if it's going to be an output that you're going to include in a newsletter,

6:34that's where you really want to make sure that that's not happening.

6:37But if you do identify any PII, IP or other sensitive data,

6:42you definitely want to reach out to the appropriate department or team

6:46so they can make sure that that data is handled before you use it in any of these models.

6:51What I showed you up here, safety.

6:53This is part of prompt engineering.

6:56And in the next skill, we'll explore concepts like C, R, E,

7:00context, role and expectation, as well as something called ask to and safety.

7:06This is the foundation for prompt engineering,

7:09which will allow you to communicate effectively with software 3.0 models,

7:14such as chatbots like chat GPT, Gemini, Copilot, Grok and Claude from Anthropic.

7:22If you're working with these tools, you absolutely are programming in plain language.

7:28And that really is the best way to put it, because if you talk to it like a human,

7:32because it sounds like a human, that in itself is misleading.

7:36And not only will that give you a false sense of security

7:40to give it information that you should not be giving it,

7:43but it won't give you great outputs either.

7:45So prompt engineering does two things.

7:48It really handles AI compliance, privacy and copyright,

7:52but it also gives you much better output.

7:54So your productivity is going to skyrocket once you use prompt engineering.

7:59But furthermore, you're going to create an environment of compliance,

8:03safety and privacy at your organization,

8:06which is invaluable when you're working with software 3.0.

8:11On that note, I will see you in the next part of the scale, which is the challenge.

8:16And then we'll get into prompt engineering.

8:19Until next time, I hope this has been informative and I'd like to thank you for viewing.