AI Basics for Beginners

When you start learning AI, you will come across so many different jargon that will most likely confuse you. My goal today is to teach you all these fundamental concepts in a most simple and intuitive way. Let's get started. Artificial intelligence is this overarching field in computer science where we train computers to perform tasks which we the humans are generally good at. for example, recognizing Patterns, visual capabilities, uh our voice or text capabilities and so on. Machine learning is an important subdomain within artificial intelligence. And in machine learning, there are two main branches. Statistical machine learning

and deep learning. In statistical ML you will see all these statistical algorithms such as linear regression decision tree kins etc that you use for tasks such as classification regression etc. We'll we'll talk about Those task more in detail going forward. But these are some of the algorithms that comes in the statistical ML. When you talk about deep learning, deep learning is mainly about neural networks. Okay. Whenever you are using neural networks, you are using deep learning. Within neural networks, there are different architecture such as CNN, convolutional neural network, RNN, recurrent neural network, transformer. very important

architecture folks Because transformer is the reason behind modernday generative AI and agentic AI boom and when you say genai okay or agentic AI that is nothing but an application of AI okay it is an application or a use case of AI which mainly uses this transformer architecture now you might be wondering what comes outside ML which is part of AI but not machine learning Well, there are a couple of things. For example, regular expression. You might build a System which you can call AI based system which is as good as humans or similar to human

level performance but it might just use regular expression which has nothing to do with machine learning. Regular expression is a general concept in computer science. Then you might have a rule-based system. Okay. Or robotics. I mean in robotics you will deploy a model which is machine learning train but there are other components of robotics where you don't Have to use machine learning. So all of these things help in building this AI system but all of these components are not part of machine learning. Okay. So I hope this diagram helps you clarify certain concepts. Many times

when I talk to people they think statistical ML and deep learning is same. They think ML and DL is same but they don't really know that these are overlapping concepts but there are differences between them and I hope that this diagram helps you clarify Those concepts. Let's discuss statistical ML a bit more in detail. Have you ever wondered how Google classifies emails as spam and non-spam? Well, when you look at any email, you will find some patterns such as limited offer, the email might be coming from a weird looking ID and so on. For a

moment, let's assume Google doesn't have any machine learning algorithm and they are doing this spam classification using human workforce. So, they have hired This person called a spam detector whose job is to classify email as spam and non-spam. So what they will do is they will go through some 10,000 uh past spam emails and 10,000 no spam emails. Okay, they will train this person on all this data. And when you look at all these spam and non-spam emails as a human, you will try to collect patterns. Okay, you will have all these patterns in your

brain that if email has terms such as urgency, lucrative offer, if email is Coming from suspicious email ID, then that is spam otherwise it is non-spam. So the way we can train this human on spam and non-spam emails in exactly the same way you can train a computer program as well. In computer science there are methodologies which are available using which you can train a computer program. Okay. So that's called a training part. You give it spam emails, no spam emails, you know thousands of them. And now computer Program will be able to detect when

you give a new email, it will be able to tell you whether it's a spam or no spam. It is sort of like when we look at certain patterns. For example, when I give you this particular table and when I give you a new input. Okay, so 1 2 3, this is the output. And I'll be like, okay, what is the output for input 4? Well, pause this video. You'll be able to recognize, right, that it is 16. Because when you looked at this data, You extracted certain patterns that it's a square, right? 1 square

is 1, 2 square is 4, 3 square is 9 and so on. Similarly, computer can also extract patterns. So, in machine learning, typically there are two phases. The first one is training and the second one where you give computer the real task where it will give you the prediction is called inference. So if you compare this with a traditional software program, if you Have done coding, uh you would give your input and your equation to your program. Your program is essentially your equation, right? Like you write all the code and code is some kind of

logic such as y= x² and then you give input and computer program will give output. In machine learning, what happens is you give input output. See there are three things, right? input, output and equation. So here in traditional software program you give input and Equation and it gave you output. In machine learning program you give input output and it will derive equation. It will derive the patterns you know and then it will come up with this kind of logic that this is the common pattern that I have observed and that equation you store it in

a model. So when they say machine learning model, machine learning model is nothing but it has some kind of equation, some kind of logic in it and you can save it on a Disk. So to talk in generic terms, traditional software program takes input and logic and it will produce output. Whereas machine learning during machine learning training especially, it takes input output and it will produce logic. Now that is a training part but during inference part which is once the machine learning model is trained you will give a new input to that model and it

will give you the output. So you have to remember this particular diagram this Draws an important distinction between traditional software programming and machine learning. So just to go over the definition, machine learning is a discipline in computer science where we train machines on data so that they can make predictions without explicit programming. Folks, this is very important. You don't do explicit programming. You just give input output and it will come up with the logic or the pattern that it is noticing. In Machine learning, there are two main major categories of tasks that we perform. Classification

and regression. So when you are classifying email as spam or non-spam or let's say you're giving an image and you are asking it to classify as cat and dog or let's say you have a damaged car, you're saying okay whether it's damaged or non damaged. This is classification because you have input and you're mapping it to one of the categories and you don't need to Necessarily have two output categories. You can have multiple categories. For example, in Google news, you will see that there is a news article and it will classify it as business, sports,

technology, health and so on. Okay? So, you can have multiple categories or you can have two categories. If you have two categories as an output, it's called binary classification. For multiple it's called multiclass classification. Okay. So that is classification. Now let's Talk about regression. So I live here in US and we use this real estate website Zillow. In India there is magic bricks macan.com etc. And when you look at any home's price it will show you an estimated price of that home. Okay. So this is the listed price. But if you scroll below you will

see this estimated price which they call it Z estimate. Now the way they do this is they have all past home prices. Okay, where you have home configuration, bedrooms and area And age and so on and you already have prices because those homes are already sold. So you have all that past data using that you will train a machine learning model. Okay. And let's say you save that model. Now when a new home comes, let's say when you list a new home, you're giving area, bedroom, age, etc. And that model will infer what should be

the price. So let's say the price of this home should be 925K. Okay. So this type of problem is called Regression because you don't have a fixed categories where you are mapping your input. Okay. You your output basically can be any number. The price can be 925K, 923K, 921.45K. Uh the the possibilities are so wide. In case of classification, you have discrete possibilities. You have fixed uh class labels. Okay, let's say two categories or four categories, spam, non-spam, uh etc. Whereas here it's a number. So This is called regression. Okay. So, classification and regression are

two major category of problems that we solve using machine learning. And here I have listed a few of these uh problems. Okay. Now, whenever you have training data which is input output pairs. Okay. In if you think mathematically then this input okay this input is called X and this output is called Y. So this is x and y pair right. So this is x, this is y. Here again uh this this entire thing is X and this is y. So whenever you have these kind of labels is also called label data. Okay. This category of

problems uh is called supervised machine learning. There is another category called unsupervised machine learning. See in supervised machine learning you have to prepare all this data. You have to go through all your past emails tag them as input output. So there is a labeling effort here. But what if you have huge amount of data and you don't Want to label it and you just want to do something good out of it. Well that's unsupervised learning. So let me give you an example. Let's say you have a kid who is playing with all these toys. So,

if you have a kid in your home, you know what happens when they're done playing with the toys? It's a mess. You know, on the floors, you see all these toys. Now, you can ask your kid to uh put these toys into separate buckets. So, you can have a bucket for a doll, you can have a Bucket for a car, truck, and so on. And you can instruct your kid to classify these things into three categories. Okay, that will be supervised machine learning. But what if instead of giving this fixed buckets, you give them only

two buckets and you say okay just categorize them. You're not telling the kid how exactly to categorize. You're just saying notice the patterns and somehow categorize them into two buckets. Well, the kid who is in let's Say sixth or seventh grade is has developed enough cognitive abilities that they might classify them by color. They will put all red color toys in one bucket, blue toys into another bucket. Or they might put all cars in one bucket and trucks in one bucket or all bigger toys in one bucket, smaller toys in one bucket. Okay? So this

type of problem solving is called unsupervised learning because you're not supervising your kid. You are not saying, "Hey, do this and That." They are just using their brain. and they are noticing some patterns and they categorize them as per that patterns. Unsupervised machine learning is pretty useful uh in the industry. In my own company, ATL Technologies, we use one drive to upload all kind of documents, legal documents, client requirement documents, invoices and so on. And while we have separate folders, you know, sometimes we have around 55 people. So they just put things in a Different

buckets. So to solve that, what we can do is we can use this concept of clustering. So clustering is is a technique that comes under uh unsupervised machine learning where you notice patterns. For example, in legal documents, you will have certain type of words or certain format. Okay? So you notice patterns and you try to put them in a different bucket. Okay. So that is called clustering. When I was working at Bloomberg, we used unsupervised learning For a different purpose. So we had all these earning estimates. We will first cluster them and many times we

will notice all these data points see which do not belong to any cluster and we will mark them as outliers. So unsupervised learning can also be used for outlier detection and at that time we used an algorithm called DBcan. There is another algorithm called K means. Okay. So DB scan, K means hierarchical clustering all these are unsupervised machine Learning algorithms. Whereas random forest, XG boost, decision tree, linear regression that there's an entire list. I'm just showing few of them. These are supervised machine learning algorithms and I have classified them into a regression and classification category

here. So just to go over the definition of unsupervised learning, we provide an unlabelled data set to an ML program and it learns to identify patterns and structures in data without any explicit Guidance. If you talk about tooling, we of course have Python, right? Python is the programming language of AI. Then we have pandas and numpy that we use for data cleaning and exploration. We also have mattplot lib seabborn. Then we use jupyter notebook for developing our code. Then we have libraries such as xg boost and scikit learn that we use to train machine learning

models. Okay. So these are the major concepts that you need to know when it comes to Statistical machine learning. Now let's talk about deep learning. We previously saw that you can solve variety of problems such as classifying if the transaction is fraud or not or classifying if the image is cat or dog. But if you look at these two use cases, you will notice one important difference when you look at the data set. The first use case is a structured data. The second one is unstructured data. What I mean by that is here you have

defined Rows and columns. You have a structure in your data that tondul samosa is a vendor, 120k is a amount and so on. But when it comes to image, it is just bunch of pixels. There is no formatting to it. Same thing with text and video. Okay? There are no like rows and columns. it's not in a tabular format with very well definfined structure. So when you give unstructured data to machine learning model for the training it will have problems because here uh these columns That you see they are called features okay vendor amount location

and if you're using statistical machine learning algorithms such as decision tree or xg boost and so on they perform well on structured data but when it comes to unstructured data it becomes difficult for them to recognize those patterns. Now we as a human when we look at this image we quickly identify that this is a cat. And how do we do that exactly? If you dig deep into you know How your brain works, we would detect all these different features such as round eyes, pointy ears, this whiskers and so on. The way our brain works

is nonlinear because we have all these neurons in our brain and we can in parallel recognize all these patterns and we can make sense out of it. Okay, we can connect dots and make sense out of it. In case of cat, see there are pointy ears, whiskers and so on. You might have a different image where the Location of the whiskers, eyes etc. would change. So you can't uh run a static algorithm where you say okay if the middle pixel are brown color let's say if I have an algorithm for example where I say okay

in if the pixel here are brown if the shape here is round then this is a cat I can't do that because the location of I might change the location of whisker might change so I need to develop a novel approach where I can detect these features uh Irrespective of where they are located, what is their scale and so on. And for that we need to use neural network. I will give you a very cool analogy based animated explanation of it so that you get an idea. Let's say there is a group of students who

have never seen koala in their life. Koala is a nice cute looking animal from Australia and these students are very naive and they have never seen this animal and our job is to train this group of student so that you give them An image and they will be able to tell you whether this image has koala or not. You can make this a teamwork and give individual responsibility to each of these students. Such as Mike can work on detecting only eyes. Moan can work on detecting only nose. So each of these students are working on

detecting a specific part of Koala's body from an image. The way they give their decision is by using a score of 0 to one where zero means this is definitely not Koala's eyes. 0.5 means yeah it looks like koala's eyes I'm not sure maybe 50/50 one means it is definitely koala's eyes once these students are trained and we will look into how exactly you can train them that is also very important part of neural network but that we'll cover in the later part of the video but for now just assume that these students are trained

to detect specific uh part of koala's body so Mike has become Now, expert koala eye detector. Similarly, Moan is an expert koala nose detector. When you give them an image of a lion, Mike will say the decision is 03 because the eyes they are like round eyeballs, but they don't look like Koala's eyes. So, 03. Similarly, for nose, he can just give some score. And if the score is more than.5, we'll say, "Yeah, those are Koala's eyes." or nose. You can extend this collaboration further and make Shakib an ear detection expert. And then these three

students Can go to Serena and Serena can work on deciding if the image has Koala's face or not. So once Mike, moan and sakip says that this image has koala's eyes, nose and ears, Senna can use uh their work to figure out if the image has koala's face or not. And she can use a formula like this because if you look at koala, the nose in koala is very prominent. Koala's nose are very uh different than other animals. So if there is a prominent feature then you Can give that more weightage and Serena can derive

this formula for detecting the phase and based on the scores that Mike moan and sake gives she will simply use this formula and tell you whether the image has koala's face or not. For example, in this particular case, I have given sample scores by each individual person and Sinai computes the face score to be 79. So if anything is more than.5, it means it is Koala's face. Similarly, if there is lion's image, you Will see that based on the score, the face score will be 0.15, which means it it is remotely looking like koala's face.

Here you can now use rest of the group uh where they individually for example Ji can detect legs and Chen can detect tail and they can tell their decision to Nidi whose job is to say whether uh the image has koala's body or not and finally Sarina and Nidi will go to Sergey who will give the f final answer And he will also use some kind of formula here. I'm giving more weightage to face because when we are working on face detection either humans or animals you know when we are working on detecting whether it

has humans or which animal we give more weightage to face. So now this group is doing this awesome teamwork and telling you whether the given image is koala's image or not. This is nothing but a neural network. Each individual person here are Individual neurons and they are working on a specific subtask and they pass uh the result of their subtask to the next group. So Sarina and Nidi are actually uh in terms of neural network they are a hidden layer. Mike, moan etc. they're forming the input or the first layer and Sergey is the output

or the last layer. So this is a neural network. It's a train neural network. Now the most important thing about neural network is how do you train it and how do you Detect these features. So just imagine that uh these students are not trained right now. You put all of them in a classroom and then you supply an image to this uh classroom and they will initially make a random guess. So Mike will randomly guess whether this image has koala's eyes or not. And similarly everyone will make a random guess and finally Sergey will tell

whether this is koala or not. And once Sergey has a decision he will go Out out of the classroom and there is a supervisor who is standing. The supervisor knows the correct answer. So Sergey will say hello sir this doesn't look like koala. and supervisor will be like no you're wrong this is koala. Sergey then goes back to the classroom and says hey guys this is actually koala what score did you give me? Serena and nidi will then go back to rest of the classroom and say hey Mike moan sakib this is actually koala's ears

eyes and Nose you need to be careful next time. So they use this experience of mistake. So basically as a group they made a mistake and the mistake or the error has been passed from Sergey to Serena Nidi to rest of the group. This process is called backward error propagation. So you make a mistake and then you get the feedback that you have make made a mistake and you pass this feedback to rest of the group so that they adjust their weights. You can take n number of Koala's images and repeat the same process. So

the same process is repeated where you make a random guess then Sergey goes out and tells the supervisor that this is koala or not. The supervisor will tell you if the answer is correct or not and the error feedback is passed on to rest of the group and the group keeps keeps on adjusting their weights or adjusting their brains in a way that they can finally come up with the right answer. You can be given Thousands or 10,000 such koalas images and after training this group for so many images eventually the group becomes better at

koala's face detection. Initially it will make lot of error but as the time goes it will keep on improving. Now how it improves the weights for that it uses a derivative and some mathematics. Okay. I hope through that analogy based explanation you got some understanding to summarize in artificial neural network you have All these layers you have input layer hidden layer output layer you can have any number of hidden layer okay and in terms of neuron also you can have any number I'm just showing you one specific example here and let's say you are using

this neuronet network for cat and dog uh detection then it might happen that these middle neurons will detect whether cat's ears, eyes, tails, dog ear, eyes, tails, etc. And then the second layer will do higher level pattern detection Which is if in the image you have cat's ears and eyes that means you have cat's head and so on. I'm just giving you uh an easy explanation so that you understand it easily. In real life the kind of features that neural network detects will be little different. Okay, it won't be like exactly like cat eye uh

cat ears and so on. It will be different but you can use uh this explanation to get your concept clear. So now what happens is this sort of becomes more Like a Google map. You want to travel from Delhi to Bangalore and you have all this map but you will take a specific route. So in a train neural network when you give an input image it will activate certain paths. It's like when you in a Google map when you say Delhi to Bangalore or you say New York to uh Washington DC it will activate certain

paths or routes. So same thing will happen and it will lead to a decision. These green arrows will lead to a Decision that it's a cat or it's a dog. So the definition of deep learning is that it's a machine learning technique that uses neural networks to learn from large amounts of data. Usually in deep learning you need large amount of training samples. Okay. So it learns from large amount of data mimicking the human brain's ability to recognize patterns and make decisions. And see it won't get everything right all the time. I see this uh

picture all the time in Social media where your neural network gets confused between Chihuahua and muffin. But as a human also right when you look at this image you might get confused whether it's a dog or it's a cupcake. Anyways so that is deep learning. Now you might have a question deep learning versus statistical machine learning. When should I use what? And folks, many times when people say ML, they actually mean statistical ML. Okay? So they don't always say statistical ML. Okay? So that's why I have DL versus ML. So when do you use what?

Well, there are few criterias you need to look at. Let's say you start working on any AI project and now you want to figure out I should I use statistical ML or deep learning. You have to look at bunch of parameters. Number one, features. If you have simple features, let's say structured data, tables, uh use statistical machine learning. But if you have complex features, images, videos, and so on, use Deep learning. Having said that, this doesn't mean that uh you can't use deep learning for structured data. I have seen problem statements where you have huge

amount of structured data and people still use deep learning. So you have to use these loose guidelines and then you have to experiment to see what approach gives you the best performance. The second criteria is of course we we talked about it like structured versus unstructured data and in unstructured Data you get videos, audio, text, images and so on. Then comes data volume. If you have a huge number of training samples, usually deep learning will perform better. Okay, so these are the criterias that you use to choose between statistical ML versus deep learning. Now let's

discuss different types of neural network architectures. The architectures are similar to your home architecture like you can have a wooden home, concrete home, metal glass home. Similarly, neural network can have different architectures and I have outlined here uh four major categories. Uh now let's talk about feed forward neural network. This is like you know you have this uh juice machine for your fruits where you put fruits at one end and at the other end you will get the juice. So you are feeding information forward. There is no loop nothing. information goes from input hidden to

output layer. Okay. So this is the Picture of a feed forward neural network. A recurrent neural network is sort of like creating a soup. So you put some ingredient then you test it. You get some feedback. Now you add more ingredients or you adjust the ingredients. Okay? So that is this feedback process. So that is recurrent neural network. So here although I'm showing you this one picture, this is a time dimension. So the neural network is only one but this A1 output you feed it Back. Okay? So this is a time axis folks. Okay? Just

remember this is your time axis. So you get the feedback then you again give new input you train the model you again get feedback and so on sort of like preparing a soup. And the third the most important architecture is transformer. Uh I'm not going to go over transformer architecture here uh because it requires a detailed discussion much more in depth. But this is the reason behind all the modern day geni and Agentic application. So when you're using chat GPT to write a poem about samosa, you know, it writes this nice poem and this kind

of creativity is possible because of the underlying model. It can be GPT41 mini whatever. So that model is powered by transformer architecture and GPT stands for generative pre-trained transformer. You see there's a transformer. So transformer is a generic architecture And GPT is an architecture which is derived based on transformer or it is an approach that is derived based on transformer. So this diagram summarizes different architectures. Now let's look at the tooling for deep learning. So for the frameworks there are two main frameworks. PyTorch by Meta, TensorFlow by Google. PyTorch is by the way more popular

nowadays. uh TensorFlow gives you more fine grain control but but PyTorch is more beginner friendly more Intuitive uh and other than these frameworks you need one very important thing which is GPU okay your hardware should be such that you can train a huge volume of data because remember in deep learning you have humongous volume of data let's say you have 10 million records you can't train it on CPU you need powerful GPU which you can have it locally on your computer or you can rent uh the GPUs out in the cloud. Okay, so that was

all that you need to know when It comes to deep learning. When we talk about generative AI, the first thing that comes to anyone's mind is Chat GPT. Chat GPT is a generative AI tool that can produce new content. You can use chat GPT to write your resume to plan your next vacation or even write a poem on your love for samosa. Other than text, it can also produce images where you give a description and it will create a new image for you. You can also generate Videos from a text prompt. I have access to

OpenAI Sora tool where you see all these videos. Here the prompt is a translucent octopus climbing up a mountain. This is totally artificially generated. Then you have this video where this dog is reading a book. Again, artificially generated. Look at this video. Looks so real. But once again, this is generated from the text prompt. To summarize, generative AI is a category of AI where the objective is to Generate new content. New content can be text, audio, video, anything. If you look at chat GPT behind the scene, it is using a model called GPT. And GPT

has different versions GPT3, GPT 4, GPT40 and so on. Similarly, MATA has this uh open-source model called Llama. So you can go to meta.ai and you will get similar experience as Jet GPT. The model behind the scene is Llama which is an opensource model. There are many companies producing all these models. For example, Google has Gemini. Claude is a model from Anthropic which is a company backed by Amazon. So these are all text models. You also have models that can generate images from text prompt. And here openAI chat GPT uses DALI model behind the scene.

There is another model called stable diffusion. So these are your image models. There are audio models too. For example, audio gen. If you go to this website, you can play all these sounds. So you are giving A text prompt and it will generate sound or an audio. So this is the summary of all the models. When it comes to audio models, you have audio gen music LM is by Google. And in video models, Sora, OpenAI has not disclosed the details. But you know, sometimes people refer to these models as Sora model. You might be wondering

if generative AI is all about generating new content then what is non-generative AI or traditional AI? Well, all those problems that we used to Solve in AI pre-Chat GPT era which is spam classification, image classification, home price prediction, all of these are considered to be traditional AI. We still solve these problems using statistical models and deep learning models. We have covered all these prerequisites in the theoretical foundation chapter in this course. So you can check that. So all these problems where you're not essentially generating a new text but You are doing let's say classification or

you are doing regression those are considered to be traditional AI and in today's world even in the era of geni we still use traditional AI because it is lightweight and for certain type of issues using traditional AI makes total sense. If I have to summarize the differences between these two then the purpose of traditional AI was to analyze predict classify or make decisions whereas generative AI means you're Generating new content altogether. Okay. So these are the uh task sample tasks which we already discussed. Then the uh type of output in traditional AI is usually yes

no labels you know numbers if you're doing house price prediction etc. Whereas in Genai it is creative right you can have paragraphs and sentences which are generated newly. Then the model types in traditional AI are decision tree linear regression SVM and so on even deep learning models. Whereas in GI it is LLM based GANs diffusion model etc. For training traditional AI you use supervised learning with label data. Whereas in genai you do pre-training on massive data set you know like all the text from internet and books and so on. Then if you think about humanlike

capacities traditional AI has limited capabilities whereas generative AI has high capabilities you can do the work of a poet for example using generative AI Using traditional AI you cannot do it. And then in terms of tooling for traditional AI we use tools like these XG boost, scikit learn etc. Whereas in genai you use all these tools mainly LLMs. Modernday generative AI revolution has become possible because of large language models also known as LLMs. Chat GPT is powered by LLM called GPT. Google has Gemini. Anthropic and Amazon has this model called claude. There are open-source models

as well such as Llama, mistral etc. In this chapter we are going to explore this topic of large language models further. And in this particular video I would like to start with a simple analogy based understanding of LLM so that you can build an intuition behind LLMs and then in the later lectures we will cover some of the technical aspects of it. Peter Pande has a curious parrot called Buddy. Buddy has a great mimicking ability and a sharp memory. Buddy Listens to all the conversations in Peter's home and can mimic them very accurately. Now when

he hears feeling hungry, I would like to have some biryani. For this case, the probability of him saying biryani, cherries or food is much higher than the words such as bicycle or book. But he doesn't understand the meaning of biryani or food or cherries the way humans do. All he's doing is using statistical probability along with some randomness To predict the next word or set of words purely based on the past conversations he has listened to. We can call buddy a stochastic parrot. Stoastic means a system that is characterized by randomness or probability. A language

model is somewhat like a stoastic parrot. They are computer programs that use a technology called neural networks to predict the next set of words for a sentence. Just like how buddies train on Peter's home conversations data set, you Can have a language model that is trained on for example all movie related articles from Wikipedia and it will be able to predict the next set of words for a movie related sentence. Gmail autocomplete is one of the many applications that uses a language model underneath. Now that we have some understanding of a language model, let's understand

what the heck is a large language model. Let's go back to our buddy example. Our buddy got some divine Superpower and now he can listen to Peter's neighbors conversations. Conversations that are happening in schools and universities in the town in fact not only in his town but all the towns across the world. With this extra power and knowledge, now buddy can complete the next set of words on a history subject. Give your nutrition advice or even write a poem. Like our powerful parrot buddy, large language models are trained on a huge volume of Data such

as Wikipedia articles, Google news articles, online books, and so on. If you look inside the LLM, you will find a neural network containing trillions of parameters that can capture more complex patterns and nuances in a language. Chat GPT is an application that uses LLM called GPT3 or GPT4 behind the scenes. Other examples of LLMs are palm 2 by Google and Lama by Meta. On top of statistical predictions, LLM uses another approach called reinforcement Learning with human feedback, RLHF. Let's understand this once again with buddy. One day, Peter was having a conversation with his cute little

2-year-old son. Son, don't eat too much bananas. Else I will punish you with an iron rod. Hearing this, Peter realized that Buddy has been listening to the conversations from abusive parents in his town. What he said was the effect of that. Peter then starts keeping a close eye on what Buddy is saying. For a same Question, Buddy can produce multiple answers and all Peter has to do is tell him which one is toxic and which one is not. After this training, Buddy doesn't use any toxic language. While training chat GPD, OpenAI used a similar approach

of human intervention RLHF. OpenAI used a huge workforce of humans to make Chad GPT less toxic. While LLMs are very powerful, they don't have any subjective experience, emotions or consciousness that we as humans have. LLM work purely based on the data that they have been trained on. I hope you like this short explanation which was based on analogy. Obviously the technical working of this thing is little different than analogy but this should give you a good intuition on this topic. Let's now get a highlevel overview of what is AI agent and what is agentic AI.

As per anthropic you can build two types of applications using LLM workflows and Agents. And I'm going to explain this using a reall life example of my own company at Lick Technologies where we have 50 people and we want to build this HR assistant chat board where you can ask simple policy related questions such as how many vacation days you get in a year or what is the policy on sick leave and in that case you can build an assistant by utilizing all that policy data which you have in your PDF files. Okay. Say if

you ask this question to Chat GPT, it will not be able to answer because ATL is a private company and all that data on the leave etc. is stored in these PDF files which are private to that organization. You can use something called retrieval augmented generation and you can build a custom chatbot for ATL where you can use any LLM like claude or GPT let's say and that LLM can look into that data and how does that Mechanism work well don't worry about it we are going to talk about rag in detail in the further

lectures but just assume Assume for now you have a way where you can make LLM look into that external source of knowledge and then it can answer these type of questions. This chatboard is workflow. It is not an agent. Let's uh add a little bit of more complexity where you you as an employee can log in and you can say okay how many leaves do I have this year and you can Also apply for a leave through the chat itself and it will actually perform that action. So see now the complexity is increased. Previously you

were just asking question it was giving answer. Now it can perform actions such as registering your leaves. Okay. So here you obviously have policy documents etc. But in addition you added API integration to your HR management system. Let's say your HR management system has some kind of APIs and you Hook these APIs to LLM. You make these APIs available as tools to your LLM and then LLM can specifically look into your query. So when you're logging as an employee, let's say moan sharma is your name, it knows that it needs to go into that database

and look for moan sharma. When I ask this question, how many leaves do I have left? And then when I say apply for two-day leaves, it has that context. It knows that this is moan's login and I need to apply for Leave for moan. And it will use this API which has been made available to LLM. It's like there is an LLM which is like a human brain and you are giving it a tool like a screwdriver or a phone. These are all the tools. So you can make all these tools available to LLM and

build this chatboard. This is called tool augmented chatboard. See if you increase complexity to this you can call it an AI agent but I'm not calling it an agent yet because this doesn't have Autonomy. It is performing a very simple task. Okay. So we call it tool augmented chatbot. So the two use cases that we looked at rack system and tool augmented uh system comes under workflow. But if you look at the AI applications or AI systems that you can build using LLM then broadly they are divided into workflow and agent. And so far we

have talked about is workflow. So let's look at agent. Now let's say you want your chatboard to handle even more complex Questions such as prepare for Sara's maternity leave and it will do multi-step planning for it. Let's say new intern is joining you will say onboard the new intern joining next Monday and see it will create onboarding checklist. It will schedule welcoming meeting it will send access request to IT department and so on. So if you think about this chatbot what it is doing is it is taking the question which is onboard the new intern

joining next Monday then maybe it is using a reasoning model to create a multi-step plan where the first step is you want to schedule a welcoming meeting and it has access to outlook let's say you have provided access to outlook and internally of course it is using llm because when it sends that uh welcome welcome meeting it needs to have some text so there is some generative AI there is some generation part to it and it is using this tool to send a welcome Meeting then you want to create interns profile in HR management system

so let's say this is your HRMS right HR management system and the this this will create interns profile in that system then it will create IT help desk ticket to get a Wi-Fi credentials, email, Slack access. See, so it has access to all these tools. So when you're building this kind of agentic AI system, you provide access to all these tools, knowledge, memory, etc. And then it will Use reasoning model to create a plan and you know it will perform all these steps. uh the next step can be to order laptop and ID card and

once again you have some kind of system IT system that you're using. uh so at every step you have some LLM which is doing some generation it is extracting the details and then it is accessing all these tools to perform the actions. So this is an agentic AI system and this chatbot is your AI agent. So There are a few characteristics that you will observe here. The first one is goal oriented planning. So previously you were asking a simple question. How many leaves do I have left? It will say seven. Okay. What is a polyson

leave? It will say XYZ. So it is simple Q&A reactive. You're asking question it is giving answer. Very very basic chatboard. But here it is agentic where you're saying onboard the new intern. You're not giving detailed instructions On how to onboard a new intern and then so this one the first thing is goal. Okay. So you gave it a goal and you gave it access to the tools. Okay. So these are all the tools and knowledge and memory all these tools and it will do multi-step reasoning goal oriented planning and it will do autonomous decision

making. So it is actually performing the actions. If you have a simple Q&A chatboard it is answering a question. It is not doing The actual work. Here it will actually schedule a meeting. it will actually create the IT ticket to order a laptop for the new employee and so on. Okay, so these are the characteristics of an agentic AI system and this chatboard we call it agentic chatboard. This chatboard you can call agent as well. So AI agent and agentic AI is kind of it's the same thing. Basically there is subtle difference which is AI

agent can perceive its environment make decisions And take actions to achieve specific goal. So it is one component but when you talk about agentic AI it is a system where you have one or more advanced agents. That's the only difference in agentic AI system you have one or more advanced agents. And what is agent? agent is something that can perceive environment. It can make decision and take actions. Of course, all of this is powered by LLM. Without LLM, you can't build this. Okay. So, agentic AI is a System which has one or more AI agents

rather advanced AI agents which can do complex reasoning with autonomy. Autonomy means certain kind of freedom and certain kind of agency. Agency means it has that control freedom to take an action. See when it created that email or when it created that system in IT to order a laptop for a new intern, it had that freedom. It is not asking you every time. It will just do it. Okay? And it Will reason through complex task and proactively take multi-step actions to accomplish goals without needing detailed instruction. See this is important. you are just giving it

a goal. It will figure out a plan and it will execute it. Now you can build AI agents through coding by using um uh there are a lot of frameworks such as agno, Google agent development kit. OpenAI also has uh toolkit to build agents you know OpenAI uh it has a Python package and so on. You can also build agentic system using low code or no code tools such as N8N. Here is one workflow that uh I'm showing you in N8. This is the name of the tool where when a new employee is onboarding, you

will trigger the workflow through some event through some uh recurrent schedule etc. And then you will have an AI agent. And AI agent has memory. It has chat model which is LLM. It has tool and so on. And it will let's Say in this case it is using anthropics claude tool. It has a memory in postgress. And then in terms of tool it has access to Jira. So you know in a software organization when a new employee joins you want to create their Jira account. Okay. You can have multiple tools. You can have a tool

to order a laptop. You can have a tool to order multiple things you know whatever is needed to onboard that employee and then you will check is it a manager? Yes. Then you add them to a slack channel which is specific to managers or you add them to the slack channel for individual contributors and so on. This is just a one simple example. You can have a much more complex workflow. NAT is a drag and drop tool. You don't have to do coding. You can integrate all these LLMs and build agentic AI system. So this

is one example of agentic AI system. There are other low code tools such as Zapier and so on. If you want to summarize the difference between the three chatboards that we discussed, rag chatboard is very reactive. You know, you asking question okay how many leaves uh any employee get in a year? It's a standard answer. Okay. What is a policy on sicklive? Very reactive tool. Augmented chatboard has one additional thing which is tool usage. So it is specifically looking into your employee uh records and telling you how many Leaves you have left or if you

are applying for a leave it will register that into system. Whereas when you have agent or an agentic AI chatboard, it will have reasoning, planning, proactivity, all these features. At this stage, you already know the difference between AI agent and agentic AI. Let's explore this further and also understand the difference between these two and generative AI. To summarize, AI agent has access to tools, knowledge, memory. LLM is obviously an integral component of it and it can perform autonomous task. Okay, it can perceive the environment, make decision and take actions. Agentic AI system has one or

more agents into it and it can do complex reasoning, multi-step planning and actions. Here is one example of agents system in N8N where you have an agent. See this agent, this agent that you see is a component within that agentic AI system And it has all these workflows right like you have switch you can check you have you can post URL you can run if condition and so on and then AI agent has this LLM which is Google Gemini chat model. So what is generative AI? Generative AI is this component. So generative AI is one

component of your agentic AI system. You can argue and you can say that agentic AI is one application of generative AI. That is also true. You can also say generative AI which is using LLM models like Google Gemini etc. in your agent system and that makes it generative AI a subcomponent in the agenti system. So, so you can say it either way but you understand that LLM is part of this agents system and it is associated with either generating content, summarizing it, extracting information, it's a versatile tool. It can do so many things right this

LLM can do so many things. So that is your generative AI. Okay. So This particular part wherever you are using LLM to generate something summarize something that is generative AI but when you are building this kind of system which has one or more agents into it it can do complex reasoning and multi-step actions autonomously when you give your goal to it then you're building this agentic AI system. So the purpose for generative AI is just to create new content right whatever interaction you are doing with chat GPT Asking a question it is generating answer that

is generative AI but if you use same chat GPT to perform deep research let's say you are performing deep research in chat GPT or perplexity where it is taking multiple actions it is going to internet accessing different tools and there is like a multi-turn approach there are multiple calls being made to LLM then that same thing becomes now agentic AI. So in agentic AI you autonomously reason Plan act towards goal without needing stepbystep instructions. Then the output type in case of generative AI is usually unstructured content which is like a text audio etc. In agentic

AI you are actually performing the action. It's not like you're just getting the answer you're actually performing the actions. Then autonomy in generative AI there is very less autonomy or no autonomy. It is just ask waiting for your prompt giving the answer. Whereas agentic AI has high Autonomy planning and reasoning. In Genai there is minimal planning right whereas aentici there is multi-step detail planning and tool usage in genai usually it's minimal whereas agentic there is a heavy usage of tool and proactivity wise generative AI is more reactive you ask question it it gives you answer

where aentic is more proactive you give a goal it creates a plan it performs action then it will review then it will correct review Correct and It gives you a final output. When you use chat GPT for simple Q&A, that is generative AI. By the way, chat GPT I can put here also. If you're using chat GPD or perplexity for deep research, then that is agentic AI. Okay. So, I hope you understood the difference between geni and agentic AI. And the third component which is AI agent is nothing. It's just a subcomponent within agentic. It

is one component in agentic AI. Agent AI can have one or more Agents. [Music]