Okay we are live now so let's start the session first let's divorce and after the session let me share my screen Aujbla Shaitanjim Bismillahirrahmanirrahim Alhuz saj ful la taj ful fee wa tabhu fa wak tash wa sawa wa ma zameen in fee Ayat Qaumi Takr Translation Allah Taala is the one who opened the sea for you so that ships can sail in it by His command and so that you may seek Maash by His grace so that you may be thankful and it is He who opened for you whatever is in the heavens and whatever is on the earth, all of it by His command, surely in this are signs for those who ponder Allah Azim Jazak Allah. OK everyone let's start the session. So today our main agenda is that we will read what is a model? What are the parameters of a model? Ok? What are these actual parameters? Why is it needed? Let us take a little introduction about different platforms. Ok? So today we will mainly do the same things that we saw differently in yesterday's session. So we have Saras with us. Yes sir, S over to you, you can start. We will do the question answer in the middle once and then take it at the end. Ok? So today's session we will have one hour. Assalam Walekum to all of you members. Hope you all are fine. And we also start the session today. My name is S Javed and some others might know me. I keep teaching here and today we will also cover the topic of generative vi. Even after that, you guys will have a lot of clarity about Generative VI because we have two main classes. And today is the practice class and in this practice class whatever happened in the main class will be discussed. We will not discuss anything new other than that. But we will go deep into it and do some new practices and as Sir has said we will discover new tools. So today we will take an all-around introduction to our course and explore a little bit as much as we have time. We will take the remaining question answers in half. Those who have done hand raises can now do lowering as well. Later, ask it only once. And the same is true in chat as well, do not do unnecessary chat because some people have valid questions but they cannot do it because chat is disabled, so try not to share my screen, I do it, so today my topic is to discuss Generative AI models with you guys, Sir, let me know that my screen is shared, Sir, your screen is shared, your voice, ok Sir, thank you, so today my topic is to discuss Generative AI models with you guys and the parameters, inputs and performance requirements in it, we will discover them in detail today. First of all, as you might be seeing on the screen, about the models of Generative AI. So we go to next. Let's start. So what do we have models for? And inside this we have to read two types. AI Models and Generative AI Models. We are probably also calling Generative AI as Gen AI because in its short term we will use it as generative. So first of all let us see what are AI models ? Then it will be easier for us to discuss the model of generative AI. So let's start with what is an AI model ? What does an AI model do? Simply put, what does an algorithm do? We have designed this algorithm. It learns our patterns and generates new things. It generates it from the data sets that we have given it and it generates different text for us, images as well as videos, which was the main purpose of AI models, meaning that now AI models have become outdated, right now we are using generative AI models. So the purpose of AI model was that they used to take inputs in high amount and integrate a lot of models at one place, our data was trained on it and then we used to make predictions on that data because the AI ​​models were limited in the data given to them that they could give output from within that. So now if we look at the Generative AI which is the main agent, then what is Generative AI, it generates new things, the data that we have given to it is same to same AI model but it has thinking power, it generates new things, it will never give us the output that we would have given it as an input somewhere or the other because all our Generative AI models, we have to discuss further today that we must have given it more than 1000 out of 1000 models for training, so the output that we get will be new output every time. We will never get an output that is exactly the same as the input we gave. Whereas this generative AI, this concept is that it generates new things for us. Anything can happen in it. It can be text, images, videos , graphics also because the text that we have inputted into it, all of it can be like this. Text, images, they will also be based on these things. Then if we look at it, the generative model is an excellent creating output that makes human creativity, meaning that we have designed a model of generative AI in the way a human can think, that is, we have such generative technology with which we can think like humans. Machines can think like us humans. He has been made so capable. And how does she think? How is this work being done ? It learns from the patterns we have given it and generates entirely new results. Now the AI ​​models should be understood in this way, like there is a student of class 5th or class 6th, we have given him books, so we have to take that paper in the end, the final paper, so what will be there in that paper, that whatever data we would have given him throughout the year, whatever he would have noted in the notebooks, whatever he would have read in the books, all that will be taken in the paper, but if we take the example of generative AI in easy words, like the conceptual study is done in the university, that we have covered the topic. So now we have to implement it. It can come from anywhere in the paper. We have to find the answers from what we have read. Now it is not necessary that the answers will be exactly the same as the ones we have read. If we bring our new creativity into it, we will answer the main purpose through it. This is the main concept to understand generative AI models and AI models. I explained it in a very simple way because there might be some non-tech people here, so it should be easy for them to understand. Next, if we go to the architecture of generative AI, in simple words, we have such tools, such things with the help of which any of our things is designed. So now the architecture of generative is based on the four things that are written in the menu. The rest are just a lot of small concepts that we'll look at later in the course. So first of all we have Gains, then diffusion model, then Vas, then transformers. You might have seen yesterday also that Ma'am was explaining about transformers in detail yesterday. Ma'am had a non-tech background but still Ma'am clarified it very well to you all that Gains is a generative adversarial network which creates hyper realistic images. Now the work of Gains in generative AI is that it generates a hyper realistic image. Now whose image can that be? Art, fashion, medical imaging, through competitive training. The models on which we will train it. The output input will determine the type of image it generates for us. G is working a lot in the industry these days. For example, in medical, it is mentioned here that generating hyper realistic images in medical was very difficult for humans earlier, as it used to take a whole day to draw everything. But now because of this technology, it has become very easy for us in generative AI that we can easily generate any art from any big image. It is not clear whether it was done by a human or generated by a machine. That's what Gaines does. And the kind of image we'll give Gaines. Now let's give images of cats and dogs. So you will create new images of cats and dogs. Meaning that you have given him the data set. There was a limited amount of pictures in it. Now what will Agen do in this ? It will create pictures from different angles and generate realistic images of that kind and then it will train the same model again, find and tune it, then its accuracy will keep increasing and whatever model of our generative AI we are training, when I am talking about generative AI, you guys should not get confused, whether it is JPT, Gemini, CLOT, these models which we are using daily, these are the generative VI models, we have trained and integrated the model in it, so I am calling it generative AI in this article. You guys should not be confused at all. That's why I made it clear that this arm, whether it is JBT, is the same one we are using. Consider this as one type. Generate generative AI. Then what do we have with diffusion models ? Diffusion models are what define our noise. The noise is not sound noise. This is what we define to bring clarity to our pictures. Now, as you know, whatever images we create using AI, their quality is very good. I mean that absolutely clear pixels are coming, there is no blurriness in it. So to clear this noise, we integrate the diffusion model into it. And how does it work? Its function is that it generates again and again. Meaning that we see it only once but it generates one again and again in the background. Then he checks if everything is okay? That means he destroys it again. He does it again and again. Similarly, it does iterative refinement inside it. And what is its example? The diffuse model has its doll E3 and stable diffuse. We will look at the data set we are studying now. So if I search and show you the diffusion model in it, there are thousands of diffusion models in it. Meaning that it is not necessary that if one model is made diffusion then there is no need to make the rest. Every day the environment is changing and every day new diffuse models are being deployed and used with us. Then after that VS which is Variational Auto Encoder Compressor. And what is the function of this? That it first compresses our image or data, text, then it auto generates it. First of all, to save space, to insert all the data in less space, it removes unnecessary things from inside it, but when it has to give the output back, to show it, then it again generates the destroyed pixels from inside or unnecessary binary codes from inside, it generates that too and then detects the abnormalities from inside it and clears it and shows it. That means it shows all the defects present in it. The words Drug Discovery Breakthrough will also be written inside it. So, seeing this and having trouble means that the term drug discovery breakthrough is there inside it, that we could have gone through it. Then we have the last one which is Transformers. Regarding transformers, we had read yesterday what transformers are. So, let's not go into too much detail about it. We have a complete architecture for the transformers. There are a lot of layers to how it works. First they are tokenized. Chunks are created of our data. Then the work that happens inside it is that this is our model which is inserting our text. He is taking it from us as an input. So it does not process everything together. He first divides it into small chunks. Then it processes it and gives us the output and the work for which transformers are being used nowadays is mostly in text generation. Next, if we look at the real-world impact of generative AI, what it does is that generative AI creates and creates new contexts. What does context mean? Whatever input we give to it will not come from within. Those will be new things. Like articles, poetry , code, it was not possible at all earlier that anyone could come and write this without a human. Now it has become possible for machines to read entire books and create new text, i.e. poetry or new articles. GPT Four Writes Article Poetry and Code with Near Human Fluency. That is, just like a human, it is not known whether it has been written by a human or a machine. Mid January Create stunning art from simple text. Now mid-journey, he creates stunning art from simple text. Now what this means is that the mid journey we have the model creates art from scratch. Now we have given him the text. We did not give him any image that you have to create this kind of image. But the model we trained earlier has thousands of images. He looks at those images to see what kind of image he has to generate when this text comes. He knows that if this kind of print has come to me then this kind of image has to be created. Prt here comes the word. So, there is a big concept of PRIT engineering also. If we have to excel inside this field. I will try it tonight, if I have any time left then I will let you know. Stable diffusion enabled democratized image generation. That is, the stable diffuse, the model of image ah diffusion that we have, helps in image generation. This was also mentioned in the previous slide, so I told that whatever image generation is there, it becomes easy with stable diffuse. And there are a lot of models of this on Kaggle. From where we take its API and integrate it into our code and from there we create any new thing i.e. new model or new data set. Can do. And then we can get the images created from there. The use of this in science and industry is that the people who are in chemistry here will know how difficult it is to create protein structures. So the alpha fold forms the protein structure. It can predict it and create an absolutely real protein structure. Revolutionizing Medical Research. Now, due to this, medical research will become much faster and there will be a lot of advancement, that is, the work which was to be done in many years, the Generative Eye can think and do research much faster than humans, now humans need rest to do more work but the machine can work on it 247 and because of this, very fast advancement is taking place in this field and its computational power is much more than the human brain, which we also call GPU. Synthetic Data Accelerates Autonomous Vehicle Development. What is synthetic data? Synthetic Data: The term we use is synthetic data, that is, we have some data, that is, we have a model and to train it, we need a data set. Now the data sets that we get from there are costly. Those data sets are not available to us for free. So whatever data sets we take, synthetic data generates them itself. This is the only quality of synthetic data, the meaning of this term is that it creates new data sets that we have and such data sets are such that it is not known whether it is real or fake, that is, like the data of the senses, if we have to collect the data of the entire senses, then what will this synthetic data do for us, if all is not available, some is available, then by applying probability on it and performing operation on it, it will create the data and give it to us so that we can train our model and can easily use it somewhere in real life and it does not cause any harm. Most of the data in it is nearly equal to the correct one. There are not many mistakes in it as compared to the real world. Then Robotics and Manufacturing leverage AI for precision and efficiency. Which is AI working inside robotics. This means that the machine has autonomous decision making power. Many of its machines have been helpful in industries. Next, after this slide, we will answer the questions. Having covered this, now here we are going to model parameters and the fundamentals of intelligence, the main thing we have is parameters, they are very important for any of our generative AI models. What parameters do we have? A parameter is any input we have. Any data set that we have in a form of data that is going to be input to our model can also be a data set. They may also be already trend models. There may also be untrained models. We provide all of that as an input. When any model is made, we have our parameters. So how is the foundation of intelligence, within this we have to decide that whatever we want to enable our machine with, if we are making a Large Language Model ( LM), then how capable will our model be. So to improve its accuracy, we have to integrate as many good models as possible with it and train on it. It is not necessary that if we want to get image generation done, then one or two models will be enough for us to generate many images. No, we will have to train it on a much larger data set. It will have to be fine-tuned a lot. It will have to be refined so that its accuracy is good. The modern generator models of VI that are coming contain billions and trillions of data i.e. data sets. Data sets contain millions of data stored within them. That means there is a lot of irritation. Now, the data set that we provide will also contain thousands of images. Of cats, that is, if we are getting the image generated, then if there is one in it, then there will be a lot of images of cats in it. In this way we will have to provide images of all the animals. So now the trillion of parameters that we have inside of that go away. Now what does trillion of parameters mean that now trillion of parameters are not available to us so that we can upload it from our laptop or from our system or generate our LLM from scratch. So this work is done by GPUs. GPUs do. Brick tag companies do. They make data sets available to us. This Chat GBT, the lab in which Chat GBT was made, it is mentioned here further that in the lab where it was made, a lot of computational power, GPU power is installed and it has cost up to $ billion dollars. So these parameters that we have are very important and we need a lot of parameters to create the model. Now, in the GPT5 that has come, more than 635 billion parameters are integrated. Here, 175 billion of GPT3 is written and trillions of GBT4 means 1.6 trillion. So 1.6 trillion. Now in the GBT5 which is trending, I told you that the data sets of more than 635 are passed as a parameter. So now, to pass these parameters, our local system, as I told you, will not be helpful at all. Now, like Google makes models or the companies that make these big tank companies, for them, they have big systems installed at their workplace to assess GPUs. So that's easy. The work we will be doing now in this course is that we will also train models. So we will not start from the beginning. We'll pick up models from Cagle or Hugging Face. We will pick up the model, that is, we will pick up its API from there and whatever new model we are generating, we will integrate it into it. Now we will fine tune it again. We will not generate a new one but will fine tune it. We will train him in our own way so that his output is more accurate and also meets the specifications we need. Because no matter how many creations come into the world, there is still a lack of some new design. As our news grows, we have to generate more things from our end and in the end of this course, we will conduct an Agathon and all these things will be used in all the activities that will take place. We have to generate notebook LM by going to the Colab notebook. You need to train your models. So there we will pick up the one from Hugging Face. If I show you the hugging face, sir, this is my hugging face screen being shared here ? Yes sir, your screen is being shared. Ah this is my interface right now with hugging face. What is this hugging face? Let me tell you from the start that we have models at Hogging Fuzz. There are ready-made models here and different data sets. Different ready-made systems mean that whatever kind of complete functional interface we have, we have also deployed it here. That means we also pick up train data from here. They also pick up models. Also pick up data sets. Then we train our models on the data sets. Then finally we deploy it here so that we can easily open it from here and we can also create a full interactive model. Now this is hacking face. When you open Hacking Face, you will not get this type of interface. You will have to sign in with your Gmail first. After signing in your account will be created. After the account is created, it will ask for some requirements. So after completing that, your profile will be created. The profile will be created. Then whatever new data you create. Now whatever things I have done here will be coming into my space that I have generated that. That will be the complete list there. Now if we look here, we have these models here. These different models are being shown with us. Now it is mentioned here that this is a text to speech model. This image is text to text. That means we will convert the text of the image into text. Then these are all updated models like this. Now look at this, it is not old. Updated four days ago These updates have been made. And this is the model of image to video. Similarly, data sets are also available here in huge amounts. These data sets will be shown to you on the right side. Now we will open this data set from here and train our models. These are all data sets. There will be thousands and billions of them here. And if you guys also create some new data set. Let's pick up the previous data set. By integrating something new into it, training it a little, you can come back here and upload it and reply to it with your name. There are some trees here also. Data sets are their minimum fee. We can use it with that. There are many other free resources available to us which we can use for learning. They are coming up here as well. Here, after that, if we look at the next Sir, this parameter wise has become that much. If you want to get questions and answers, let's get some right now. Two-three questions and answers. Okay sir, let us take some questions. So I call names one by one. You can unmute and ask your question. Abbas Khan Assalamu Alaikum Sir, Assalamu Alaikum Sir, my question is how do these AI models learn from data and then how do they apply this knowledge to new data which they have never seen before and Sir, these other parameters which may not be picking up its performance, how will we know whether it is actually learning or it is just modifying the training data. Now, whatever model we have generated, we will know that we have integrated the data set with it, so it must be generating it from within that. Now how will we know that we have just demanded him to generate the image and give it to us. Now the image that it will have generated will be on top of the same information or the same text that we have passed as a parameter. Now you are saying how will we know? Like GPT is now. There are thousands of data sets inside GPT. If he is not giving the correct answer on the data set we have, then we will know that we are lacking somewhere. But apart from that, he cannot provide us output from anywhere else. What was your first question? Sir, how do these models learn patterns from data and apply this knowledge to new data that they have never seen before? Now we have these ready-made models. This machine means the kind of machine learning we have done, which means they pick up the models from there. Whatever kind of parking face of Kagel, I showed it. Now when we integrate it from here, then we will write code for one purpose by combining different models, just like in a notebook, in a Collab notebook, we would have generated the code there, whether from GPT or any other app, now when we are integrating the API in it, then through that API our code will come to know that this data is coming from there. Now when we request that data below, when it gives us the output, our output will be coming from within the same data set. In this way, he will come to know that he has used this and our code will come to know that the data is coming from the API we have provided. He has to train on top of that. Now the train has to be built and those GPUs have to be utilized in it. Different computation power has to be used. Then whatever output will be there after that will be shown to us. Yes, if the course of some other meanings goes ahead then more things will become clear to you people because in the beginning it will seem a little difficult but later it will become easy for non-tech people. Sir, please take the next question. Okay Abdullah Abdullah Assalamu Alaikum ji, I wanted to ask that as you had said that population has to be controlled, how many people can use one GPT, please tell me about it, I have not talked about it, can you please clarify the question a bit, I mean, there is no limit on how many users can use it. Sir, if you have understood the bottom then please answer what Sir wants to say Abdullah, can you clear the question? Yes, you had said that the parameter is a parameter of the model, you had said this, that is, there are more than 1 trillion chats in Chapter 4, so how many, that is, it is mentioned in it that there is so much width or how many people are there, the company has contacted some company, permission has been taken from them, so many people can control it, that is, I will clarify this at some point of time, you are misunderstanding this thing, mainly this parameter, the parameter is not that there are so many people. You can say that this is the ability to understand the model. Ok? In simple words, how many variables does the model have on the basis of which it is doing learning. Recognizing patterns. Ok ? Now let's suppose we say that the old GPT was very old. There was GPT three, it had 175 billion parameters, that is, they were in billions. Ok? Now if it was in billions, then if you pay attention, were the results of GPT3 better or worse than the present GPT? The old GPT was improved, they enhanced it, added some things, the present one is better. Ok? And the current one also has higher parameters. His reason is that he has also learned more data. The old GPT tells you straight to your face that I only have data till 2021. I can't answer 2022 right now. But the current GPT can give you the result simultaneously at run time also. Meaning if you use free GPT then some benefits of paid GPT are given to you. If you tell him to search and bring me the data set. Search for a link and bring it. Is it available on the internet? He does it and brings it to you. Right Now gives you the date. Ok? But if you ask the old one to tell you today's date or bring you the data of 2022, then GPT 3 could not do this work. Right? Yes. clear? Ok. As the number of parameters increases, the capability of the model also increases. His ability to answer is increasing. Ok? This is what these simple words actually mean. Ok. A Shujaaz Assalam Walekum Sir. Sir, my question was that the way you told that the generator generates new input and output and gives it to us. My question was whether generative AI can generate something which does not exist before, like data which does not exist, can it create any new data, yes ma'am the concept of generative AI is that it will create such data, it will generate such output which does not exist before, we have creativity in it, it can output such things which were not available with us before, this is the quality that we have which distinguishes us from the AI ​​model, the special thing about generative AI is that it will generate such things which were not available before. Yes sir, next question ok, this has come in the name of the user. Kindly you rename it. Then I will take your question. Mohammad Jain Farid Assalamu Alaikum Sir, my question was that if we want to work at a low level, like if we use a free model, then we can work on our laptop which is a machine we have, or we will have to buy a GPU for any work, for which we will train the data set or find a solution or do anything, then what will we have to do for that, sir, please explain this process in a little easy way, yes sir, we have to discuss a little about this in the end also, but let me tell you that we will not generate any model from scratch, it is not our job that these big time companies like the one I have mentioned, they only do that work. We just have to pick up the data i.e. the data set and find and tune our model. Small changes have to be made inside it. He has to be trained on more information. That information will not be so large that there will be trillions of data sets. Those will be a couple of data sets. The kind of house price prediction models there will be. We generate small training data sets for machine learning. We don't have to work that much. Let me show you guys where we have to work is to work in the collab. So, we will have GPU power inside the Colab, but that GPU power will not be enough to start generating new LLMs right away. So sir, this was my aim also. But you tell us that if we go into any industry i.e. professional life, then how will we work there? I mean, then we mean big companies have given us some assessment. We will request them there and buy GPU power remote from them and then we will use it. Now I have opened this collab in front of you, so here in collab we go here and select the file, yes we have run time with GPU, sir, I will add a little bit in it. Mainly obviously we will be using these things. So let us tackle some more questions. Right now sir, whatever resources we use, we will use them for free. We will also give you a free GPU. Free RAM too. Ok? Where can you get all these resources for free or to what extent can you get and use them for free? And They Will Be More Than Enough for Us for Now. Ok ? The rest of the thing is that when you go into the industry and learn how to work, it is not necessary that you use only those resources. She can also provide you with company. You can also do it by combining two or three. You can also buy its paid resources. Obviously, if you charge someone to set up a production, you will also have to invest a little bit for it. Ok? So we can see those things at that time. Ok? clear? OK sir. Ok. Okay Manzoor Hussain Sir, let us take this last question, then we have to complete the slides also, yes yes Manzoor Hussain, if you are speaking then I am not able to hear your voice, okay sir, let us complete this and then take the remaining questions and answers. Ok sir. Ok. There is some confusion in these parameters. So let me tell you a little about the parameters and then we will move ahead. Parameters are what we have to recognize patterns. That is, whatever parameters we may have provided. Parameters We have input data. The parameter is nothing else. There are number of data sets. There are number of models which we have given to that model. Generator creates the model of the vi we are generating. So these parters are that. And what do they do to us ? We recognize patterns, which will determine and generate the output by checking all the patterns and inputs given to us in the model. So, due to this we have the drawback of increased computational demand and our fitting risk. It is also possible that if we have not trained our model properly, then it can give the same output as the input that we have given. So this is also a drawback of this, which can happen if we have not trained properly. If we fine tune it properly and get the correct data on it, then it will not come to us. Next, if we go to what do we have inputs and tokens ? Now the tokens that we have are chunks of data, whatever input we get from our data set, the model divides it into small chunks. Then it does processing one by one on the sons and in the end it combines the result and outputs it to us. Model process ination through tokens. What tokens do we have? The small units that exist can be words , phrases and characters. That means if we have the word hello, then h will create that and chunks inside hello now. will separate H further, will separate E. That means he will keep making more breakthroughs in it. Tokenization breaks complex input into manageable chunks for learning. Now its job is to learn from our input. It is not necessary that we learn from the model that we have integrated earlier. It also creates a history of the inputs we are giving. You guys know the way Chatty Gemini and Claude work. They keep creating history together. They know what kind of output your mindset wants. These give that kind of output. Input quality and relevance directly determine model performance. Now their performance or quality is determined by the model that we have given to it, that is, for it to be trained, first we have raw input in it, its process cycle runs with it, we first give it an input, the input can be in any form, I have discussed this earlier that it can be text, image, video also, whatever we have passed as an input to our model, as we have just seen on hacking face that there are some which accept images. There are some that just accept text as images and give us image output. Some like Sora AI or some like Mukhtalif video ones also take text and output images and videos to us. Now what is a video? That too is a collection of images. Fast moving so they also generate images. Now let's take the raw input. Hello World Text Generate Text Enter the system. Will enter from here. There will be a break in tokenization. There is a complete cycle inside tokenization in which chunks are created. There are more functions inside it. Then after that the processing will take place. How is the processing done ? Processing finds those patterns within it. The relationship looks at models given its input. Then it checks the input we have just given to it for the result. And after finding the pattern we have a best output which provides us. Now how does its generation happen? That is contextual output. Meaning that it is determining the context we have given it. Now that context can also be that whatever past history we have from this, that can also be a context. One context would be the one on which we have generated the data set. Now you must have noticed that everyone uses the same GPT. Now if a person searches something then his result and the result of his other friend will be different. Why would that happen? Because their past history and preferences are different. It will depend on their preferences. So now, depending on the preferences of some people, in all the competitions that are happening nowadays, in all the cathoons that we have, we have the openness to use any AI. We can use any API and we can take help from any model. Now, using the same, some come first, some come second , some fail completely. Now we should know how to use it so that we can get better output. If we want to get work done from this machine, then we should also know how to make the machine work. What are those methods and what are they? The biggest thing we have is Prime Engineering, that whatever input we have to give to it, it should be such that it does not go out of its plan or its mindset. He should directly bring the output that we want. That too with 100% accuracy. Now we are giving such input that our model is not able to understand what we are trying to say. That means we are telling it to generate an image of Apple. Now if we tell him that there should be a red apple, then along with it we give him the complete text that he should also make a plate. So our output from this and our output from creating a simple apple will be different. We have to use this ROM engineering a lot in our course. So you guys will learn this. There is no problem in that. Do n't worry at all about the non-text ones. Then how do we measure performance after that? We have four methods to measure performance. There are four to five methods. Now in these methods, first of all accuracy and predictability. What is accuracy? Accuracy is a mathematical number we have determined. It became 100, it became 99. Whatever accuracy our model has, if it is above 90, if we train it then it is well and good. If our model set is falling short then we need to do more working on it and fine tune it. Our model is not hallucinating. We also need to check if our model hallucinates. This is the concept of hallucination that we have just given it an image of a cat, so it is calling it a dog from now on. That means if he is getting confused in it then we will need to find and tune him further. So the accuracy will be much better than that. If we want to maintain accuracy properly then that accuracy will be a mathematical by-pass number. We will check that. If it has reached 99 then our model has become the best train. Then preplexity we have uncertainty. You all must be aware of the uncertainty, just like with every product, with every thing, there is a mention in plus or minus that how much variation it can make, how much plus it can go in its range. How much uncertainty is there in minus, preplexity uncertainty. Simply Human Evaluation: Now whenever we have generated a new model or any new model and trained it on our data set, then humans also check it to see how much accuracy it is giving and how good its output is. Now those humans check every output. By implementing it loosely, they give it all kinds of inputs to check what its output is like because they do not know what kind of usage the users do. What kind of input can he give him? So, it becomes necessary for humans to check it also. Then there are benchmark texts. Then there is continuous improvement. Now in the benchmark text, it is said that objective comparison of model capability means the objective of the data set that we have created, that is, we have just created a data set that generates an image. I am taking the example of a cat from the beginning so that you people do not get confused. Now if we have to generate the image of a cat, then when we get it generated, we will check whether it is aligning on the data set that we had given it or not. It should not be like that we have called it cat but it has made its ears like those of a dog. So now it is important for us to check whether it is fulfilling its complete functionality or not. Continuous improvement is necessary in everything we have, to keep it updated, to keep it up to date, just like Sir Tala has said that earlier our Child DBT used to say that we do not have the next data, but now if we search through Chart DBT, then it is also making future predictions and it has all the data up to date. So, in this way, improvements keep coming into the field and every day we keep integrating them with our models. Now GPU comes here, what is GPU? GPU is a GPU that we have created by integrating and combining thousands of CPUs. Now GPUs are not available to everyone. It became like Facebook in big workplaces. Whatever data of Facebook, that is, wherever their data is stored on big storage, their GPUs are installed there and they are functioning there. These are not available to everyone. State-of-the-art model training demands extraordinary computation. We need more power. To train GPT (GBT), we used thousands of Nvidia GPUs and its cost was $ million. Now it is not possible for any common man to do this. Only big companies can do this. High end consumers use GPUs. Will just fine tune it. The way I said they will pick up small data models. That means we will fine tune it. We will determine that for our output. Because whenever we have to generate something, we have to generate our profits and give it to some field. Profit means that we have to invest some new of our own in the company. If we are in any field like Chemistry, Maths, Physics, then when in the end we do something for our field that is beneficial for the future, then how will we make it beneficial? In the kind of AI era we are in, we will take it from the past data, train it, train it through our knowledge, after doing various integrations in it, we will deploy a new model, but for full scale training, distributed cloud infrastructure is necessary, we need big cloud companies like Google and various other companies to compensate us. They are available to us costly, costly ones are also available and free GPUs are also available to us for training. Consumer GPU fine tuning only for enterprise setup and timeline means we can only fine tune the consumer GPU that a single person is using. Enterprise users can use 10000 GPUs. Then it takes us weeks and months. Now our own system shuts down after a few hours due to the charging getting exhausted. But the GPUs are working weekly. Keep working for months. Finally, a model of them is generated. In which we have integrated the different models that I am telling you about. What is our cost and infrastructure demand? If we do anything as Sir has said, we will need some investment. Training large-scale generative AI requires substantial resources across three dimensions. Which are three dimensional? Hardware, time and cost. In hardware we will have GPUs. Storage will come. Storage like we will need RAM and GPUs will need CPU power. Then we will need time of week and month. As I told you, if we have to train it inside then that data is not small. He has a lot of insights. He has received a lot of input. Its cost is also very high. Because if those GPUs are running on electricity, then those 10000 GPUs are installed together, so their electricity bill will also come and definitely then there will be a cost of cleaning them, maintaining them and we will also have to hire people to manage them daily. So all these costs are determined by them. This tells us our total cost. Then a graph is given inside it to easily understand what demand our infrastructure has from us. Finally, if we look at the last slide, our future belongs to generative AI. So, democratization in this means that as I told you, if we create something new now, then we will make it available to the people. Now, where I was showing the hugging face, people have made their own models of the hugging face and uploaded them there, so they have planted trees along with it. That means it is not necessary that if you want to upload, you should upload it for free. You can assemble patterns there for people to buy from you. Then use it after that. Efficiency Improve Algorithms Reduce Computational Demand and Cost. The better our algorithm, the more it will reduce computational demand and our costs. The higher the efficiency, the lower our costs will be. Then responsibility: If we have to use generative AI models, we have to use them with full responsibility. What would our responsibility be like? Most of you will be students. So let me give an example to the students that the responsibility will be that if we are using any of our models then we should not do the work from there as it is cheating. We should tell him that I have studied this topic today in school, university or college. You explain to me from here that this was written in the book. But you give me some new data. Make some tests for you so that he can see you and from your input he can know how much you have understood the teacher's words. From there onwards he will start explaining. Then it will generate the test. I will make flash cards. This will lead to interactive learning. Meaning that we can make maximum 100% profit from this. But if we look at its disadvantages, we can do a lot of wrong things with it. The more we use our model, that is, our HRGBT or JMI, whatever it may be, if we use it for these purposes, it can also do wrong things. He can also do the right thing. Now our responsibility is to make all the people around us aware that we can generate 100% efficiency from this. Then we have unlimited opportunities. As I told you, to win the Hegathon, we can learn any online competition from there. Let us take him for learning. Get him to generate the code of your choice. We will also learn from it. Also, in whatever competition we have participated in, we have rules that we can use it. So when we use it, our efficiency will increase and our chances of winning will also increase. That means it gives us opportunities to get notes generated from it. Let us generate a video from this. With this we can fulfill all our work requirements for daily use. Now here I had said about prompt engineering that I will tell you a little bit in the end. So, the better our process of prompt engineering, the more refined the process, the better will be our output. Getting good output means that now as a perfect engineer, if I tell you about the opportunity in this also, do you have the opportunity that you are not a perfect engineer. You tell your Model Four GBT to act as a prop engineer and write me a prop. For prom, just tell her the scenario that I have to generate such a nice image. Now he himself will give you five or six options that you should give these prompts. Now after reading its proms you will get an idea which pro I should use so that I can get good output. You can copy and paste your print from there. Move your prompt. Paste it in again. Below her next who are giving her input in that give her that prom. Then whatever it generates will be your required output. That is, what I mean to say is that you have so many opportunities. You have so many opportunities that you have come from the source source. Who are studying maths. It has become very easy for math students to read it now. Because they can get it generated directly from here. Whatever they are using for their queries. What is its output and where is it being used in real life? Now as the image is generated, 3D images of the model will be generated with us. That means, if we want to use it responsibly, we have a lot of opportunities here. This was the end of my lecture today. Sir, please take some questions now. Rest we have Friday session for question answer. In that, all your questions and answers will be cleared. Yes sir, okay, I will take some questions. Jinan Ali Assalam Walekum Sir. Peace be upon you. Sir, I have a question that I was a patient of no use, so I was helpless and could not take two classes. So today I took these classes and I did not understand anything special. Actually, I have used Chat GBD. I have also got that assignment written from prom and have also done contact writing. So I wanted to ask whether, after learning this course, as a freelancer, I can earn something related to this course on Fiverr or now the difference is on LinkedIn. Amount Freelancer Yes, you can do it. Secondly you can also build your own startup. For a person like me, I mean if you watch the first session of Sir Anwar once, in which we showed different charts that after this course some people have even started a startup. Some people have done it as a freelancing career. Some people have done JOSA just for hobby or for learning. So from that you will get an idea about what you can do after the course. Ok? So I have another question that means I also did not understand this much. So I will be able to do this course because I only used Chat GPD and nothing else. There is no issue. Just keep with the pace. So we will take all the things together. There is no speedy thing. Right now, in this week, we are only introducing terminologies. Till today's session practice session. We will not read anything new this week. Ok ? We are revising the same things which we studied in mains. This weekend, next weekend, which will be Sunday and Monday, we will have a new session again. Ok? So we will do a practice session for that again. So right now I am studying in the main session only. They are just practicing this. So right now I should take first class and second class. Then I will understand everything. Yes sir. Things will become clear in that. Ok. Thank you very much. It will become clear. Ok. Thank you very much. Let me tell you one more thing, I have just shared it with my Lindin in the chat. So if you guys have any questions for me then connect with me on Lindin. Sarla has also done this, first of all you can connect there and ask me any question. Okay, thank you very much sir. Ok. Next question is Krish Bah. Krish Prasha Hello Salam Walekum Salam Walekum Sir Walekum Salam Sir my question is what do the transformers do and how are they different from Hugging Face, what is their difference, both of them are good, mainly Hugging Face is half thing. Transformer is not half the thing. The concept itself is amazing. Meaning they are not linked to each other. Someone Hugging Face is a platform on which you will find multiple open source free of cost models. Ok? Yes. Database data sets are also available. You can also create an application on top of it. It has multiple uses. It is available for free and you can also take that platform for a fee. I mean, what is a transformer for him now? Transformer is a mechanism or a logic whose main purpose is to extract context. I just want to find out what actually is the meaning of this entire paragraph? What is understanding? Ok? The chat you give GPT a paragraph and it starts understanding what you are actually talking about. So the logic of understanding and how he starts understanding is explained by the transformer. Ok? That is a complete machine learning architecture. For this, just look at one paper, attention is all you need. Ok? All the things about its architecture are mentioned in it. You will search online that tension is all you need. You will get the paper in advance. These things are mentioned in it. Ok? So sir means the brain of transformer as an AI is working. You could say this specifically for models of generative AI.