SDP on Gen AI & Agentic AI Day 2

Añadido: 2026-06-03

[00:02:28][music] [music] [music] Heat up here. [music] Heat. Heat.

[00:03:02][music] [music] [music] Heat [music] [music] up [music] here.

[00:03:36][music] >> [music] [music] >> Heat up [music] here.

[00:04:04]Heat [music] up [music] here.

[00:04:20][music] Heat. Heat. [music] [music] [music] [music] >> [music] [music] [music] >> Heat. Heat.

[00:05:18][music] Heat >> [music] >> up [music] here.

[00:05:39][music] >> [music] >> Heat. Heat.

[00:05:59][music] [music] >> [music] >> Heat. [music] Heat.

[00:06:23][music] >> [music] [music] >> Heat. Heat.

[00:06:44][music] >> [music] [music] >> Heat up here.

[00:07:08][music] [music] Heat.

[00:07:18][music] [music] Heat.

[00:07:35][music] >> [music] >> Heat up here. [music] [music] [music] Welcome to the dynamic learning [music] hub of XLR at Anderi East.

[00:08:14]A prime spot conveniently located at just a 5-minute walk from LIC Anderi [music] Metro.

[00:08:23]In our journey of over a decade, we have proudly trained over two lakh students.

[00:08:28][music] We've also been recognized by Deoid, Silicon India and other accredititation [music] bodies.

[00:08:42][music] Experience learning like never before in our spacious and comfortable classrooms. With highdeinition [music] projectors and state-of-the-art acistics in place, we ensure effective learning even [music] to the farthest point in the room.

[00:08:55]The hybrid mode is a gamecher as you can enjoy the benefits of [music] a light classroom from the comfort of your home.

[00:09:04]Even if you [music] happen to miss any classroom session, you can still attend the same session with the same trainer [music] at the same time at your convenience, maintaining consistency in your learning progress.

[00:09:18][music] Want to get started? Join us for a demo or career chat over a cup of coffee at our Anderi branch.

[00:09:28][music] XLR. We don't just train, we build couriers.

[00:09:48]>> [music] >> Yes. Uh so the collaboration with XLR was an excellent one. We conducted a PowerBI workshop with advanced Excel over a span of 3 days. The workshop was almost 12 hours long uh in total. Mr. Karthik who was an amazing trainer uh helped the students understand the concepts deeply. I'm sure that the students will be able to apply these concepts in their uh careers as well as their daily lives.

[00:10:20]>> The trainer Mr. Karthik was an excellent person where he uh explained in very detailed manner on how to give great insights using very complex data sets.

[00:10:43][music] >> [music] [music] [music] >> So, it was a machine learning workshop.

[00:11:05]We learned a lot from the course and it was a great experience attending the workshop. The >> the trainer was very helpful. He had a hands-on approach with his teaching and it was fabulous. This was a great experience for me. So I would absolutely love it if Acceler and its team would come again to our campus and conduct more events like this.

[00:11:27][music] [music] It was a great course and it started from basics and took me to a great extent.

[00:11:53]>> The experience of the workshop was really nice. The feedback from all the student was positive. They understood the concept. The workshop was also practical based with some questions. The experience with XLR was uh very smooth.

[00:12:07]They helped us in conducting the workshop smoothly right from the booking of the vendor and deciding the course till the distribution of certificates.

[00:12:15]In overall this workshop was a success during our technical phase atmos and I thank XLR for being a part of it.

[00:12:22][music] [music] Heat.

[00:12:48][music] [music] [music] [music] Heat.

[00:13:15]>> [music] >> Heat. [music] Heat. [music] >> [music] [music] >> Heat. Heat.

[00:13:55][music] Heat [music] up here. [music] [music] Heat [music] up here. [music] [music] Yeah. Hello everyone. Good name and welcome back to the day two of our STP on Jen and agent care. I hope things are going good and you are enjoying the session.

[00:16:57]So we'll start for the day two. Okay. So whatever the query question you have you can keep asking and if face any difficulty you can post in chat box.

[00:17:05]Yeah. So we can start today's class.

[00:17:07]Thank you.

[00:17:09]>> Yes sir. So good evening all.

[00:17:13]So good evening all.

[00:17:16]So we will start for the today's session.

[00:17:57]So we have the quick recap about yesterday's session.

[00:18:02]Um yesterday we saw for uh before starting of generative AI we saw the introduction about artificial intelligence and we saw for some real life examples of uh artificial intelligence and then types of AI like narrow AI uh general AI and uh surpass AI. So we have seen three different types of AI. So then uh we started seeing of what is generative AI and we saw for what is what are all the uh LLM models of or available in generative AI like chart GPT Google gemini and metal lama. So we know that chart GPT was uh developed by open AI.

[00:18:48]So Google Gemini was developed by Google. So Metal Lama is developed is developed by Facebook. So all the um all the LLMs are uh uh used for uh creating coding and debugging uh like uh creating reports uh uh presentations, lesson plans, uh support learning and skill development and uh create images, audio files and video files. Uh like uh some some LLM will integrate with Google services. For example, example, Google Gemini integrate with Google services like Gmail, Docs and Drive.

[00:19:44]Okay. And uh for research and summarization so etc. Okay. So next uh we saw for what is generative AI. So generative AI is nothing else. It it will create the new content such as text, images, audio, video, code and presentations based on user instructions. And so we have given some of the examples of uh generative AI and then uh we have uh seen for the generative model. Okay. So this generative model it's nothing but the user ask a question and this generative model a model process that input and generates the response. Okay, unlike the traditional system, uh it's uh that simply what the traditional system will do, it will simply retrieve the store the answers. But in case of generative model, it creates new responses based on patterns uh learned from large amounts of data. Okay, it understand the context of the question and generates humanlike answers. Okay, so here they have given lot of examples.

[00:21:06]For example, uh if you are giving the input like what is artificial intelligence, um it doesn't search for a fixed answer.

[00:21:15]Okay? Instead of giving the fixed answers, it generates a relevant response based on its training. In simple term, we can say generative model takes an input and produce a meaningful output such as what are all the meaningful output? What form it may be?

[00:21:37]It may be in the form of text and or images or code, audio or video.

[00:21:49]Okay, in very simple terms if I define what does generative model is. Okay.

[00:21:58]Generative model takes an input and produces a meaningful output such as in the form of text, images, code, audio and video or video. Okay. So we go to the next slide and here uh look into that why generative model are important.

[00:22:21]Okay. So I have given uh three reasons it is important. The first reason is understanding complex patterns uh like generative AI can analyze a huge data set and discover patterns that may be difficult for humans to identify. So this helps organizations make better decisions and predictions. Okay. And second reason is content generation.

[00:22:52]So, so here the generative models can create text, images, videos, audio and even computer code. So this significantly reduces the time and effort required to produce the content. And third reason is building powerful application.

[00:23:18]Okay. Um here many modern AI applications such as chart GPT uh virtual assistant image generators uh recommendation systems and intelligent automation tools are powered by generative models. In simple terms, generative model help machines learn from data, create new content and power innovative AI applications across industries.

[00:23:55]So the real world applications are in education, healthcare, marketing, software development, etc. Okay. So we move on to the next slide.

[00:24:14]So where this generative AIA exist? Can you see here? So artificial intelligence is a very big outer circle and next circle is the machine learning and the next circle is a deep learning and inside the deep learning you can see one more circle that is known as generative AI.

[00:24:33]Okay.

[00:24:35]The outermost circle is the artificial intelligence which is the science of creating missions that can perform task requiring human intelligence such as learning, reasoning and decision making.

[00:24:53]So inside AI is machine learning. So machine learning enables computer to learn from data and improve their performance without being explicitly programmed for every task.

[00:25:09]Within machine learning we have deep learning. So deep learning uses multi-layered neural network inspired by human brain.

[00:25:21]These networks can process large of data and identify complex patterns. Finally, inside deep learning, we have generative AI. So, generative AI is specialized in creating new content.

[00:25:38]It can generate text, images, videos, music, code, and much more based on user prompts. Therefore, generative AI is not separate from AI. Okay, it is not separate from AI. It is specialized area that evolved from deep learning and machine learning technologies. Okay.

[00:26:05]Okay. This machine learning is a subset of artificial intelligence. Deep learning is a subset of ML and generative AI is a subset of deep learning. Okay. Fine. I hope you you are clear about this diagram.

[00:26:21]Okay, we move on to the next and we have seen for the discriminative versus generative model. So here the discriminative model and generative model. Let me tell the uh first I'll tell about it discriminative model. So here in discriminative model it always identifies and classifies data. But in case of generative model, it always create new data or new content. Okay. Second thing, it learns the difference between categories.

[00:27:00]But in generative model, it learns patterns from existing data. Okay. And third, it always answers what is it? Okay. But here it always answers can I create something new and the fourth the example what I it is given here is cat or dog classification but here the example is it generates a new cat image or text or code or music okay so that's the difference between discriminative and geneneral generative model.

[00:27:44]So we move on to the next slide. One more example of discriminative versus generative model.

[00:27:53]So this slide explains the difference between discriminative models and gener models using a music example. A discriminative model receives music data as input.

[00:28:08]Its purpose is to analyze the music and determine which category it belongs to.

[00:28:16]For example, after processing the music, the model may classify it classify it as rock or classical or romantic music. Okay. So that that is the uh explanation of discriminative model and in simple terms we can say discriminative model focuses on recognizing patterns and assigning the correct label. Okay.

[00:28:54]Now come to the generative model. A generative model works differently.

[00:29:01]Instead of classifying music, it learns the patterns, rhythm, and structure of existing music.

[00:29:10]After learning these patterns, it can create entirely new music that sounds similar to what it has learned.

[00:29:23]For example, AI music tools can generate original songs, background music or melodies without directly copying existing music. Okay.

[00:29:37]The main difference is simple.

[00:29:41]Discriminative model identify and classify data. Generative model creates new data. So this is why applications such as spam detection, image recognition and disease classification use discriminative models. So whatever example I have told which comes under discriminative model once again I'll repeat it like spam detection I mean the application such as spam detection. So in Gmail we have for spam detection right? So spam detection, im image recognition and disease uh classification use discriminative models while chart GPT AIS and AI music generators use generative models.

[00:30:36]The very quick comparison between discriminative model and generative model is here the data creates data predicts uh predicts labels generates uh correct answer is is this what I can create example music generation like a music generation.

[00:31:00]So discriminative model is equal to classification.

[00:31:05]Generation model is equal to creation in one word you can say. So discriminative model is for classification. So here itself I have defined in the image you can see it here. So discriminative model means it mean for classification and a generative model is for creation or generation.

[00:31:33]So I hope you understand.

[00:31:36]Take a look on this.

[00:31:44]So take have a look on it.

[00:31:48]Just we are having the very quick recap.

[00:32:40]Okay, we move on to the next slide.

[00:32:46]Okay.

[00:32:48]So from here onwards uh uh we are starting new. Okay. Till this we have completed yesterday but from here we are starting with new.

[00:32:58]So we need to understand what is generative AI. So you just please kindly uh just go through this uh three to four lines theory here. So generative AI is a subset of deep learning and generative models are trained on huge amount of data. So you have to understand you know this point from the figure. Okay, generative AI is a subset of deep learning. So we saw here, right? So generative AI is a subset of deep learning. So we saw this point. Okay.

[00:33:32]And what about the next one? Okay. The generative models, how the generative models get trained? The generative models get trained on huge amount of data. So this you need to understand.

[00:33:46]Okay. Now I'm going to speak about the training. While the training while training the generative model we don't need to provide a labelled data that is not possible right what is a labelled data come to this figure okay so suppose this is a cat image right so I am giving uh I'm just mentioning the label this is cat so this is dog so when I am giving the input itself I'm just giving the each and every figure with the label. So this is cat image. So I'm I'm giving the cat image and I'm mentioning this is cat.

[00:34:29]I'm giving the floor image and I'm mentioning this is a floor. Okay. So I'm mentioning everything by the help of the label like that is a supervised learning. It comes under supervised learning. Okay. So that is not possible right here.

[00:34:47]One second.

[00:34:50]Okay. So that is known as labelled data.

[00:34:53]But when you are training the generative model, we have mill and millions of data will be there to train the generative model. It is very difficult to give all the data as the label data. It is very difficult. It is very very difficult.

[00:35:11]Okay. So that's why in the while in the while training the generative model sorry it is not required okay it is not required uh uh to give the labelled data you can train the generative model using the unlabelled data so unlabelled data is enough to train the generative model.

[00:35:46]Okay, it is not possible because this that's what I'm coming to say here. It is not possible when we have a huge amount of data. So it's just try to see the relationship between the distribution of the data in generative AI. We give unstructured data to the LLM model for training purpose. Now you understand this theory. I hope you understand very simple only three points you need to remember. What are all the three points you need to remember? I made it as bold. The first two point what is the first two point? The generative AI is a subset of deep learning. First point and the generative models are trained on huge amount of data. We know that because if you want to uh train the generative model so millions and millions data will be taken to train the model. Okay. So that's what I have mentioned. It is a huge of data. Okay.

[00:36:50]What is the third point in generative AI? It is we no need to provide the label data. Okay. We can give unstructured data to the LLM model for training purpose. So only three points and see here all this uh uh so here we have given some unlabelled data to train the model. Okay.

[00:37:20]So in this uh thing you can see there is a black cat image as well as the white cat image. So it will separate based on the features it will uh when you are giving passing the input to the model.

[00:37:33]So it will understand what are all the trends and patterns are there in the input. It will learn from the existing data and it will segregate it. We no need to worry about that. Okay.

[00:37:48]Just go through this slide and understand the points.

[00:37:55]Take some few few minutes to understand [snorts] this or else you note it down.

[00:39:42]Is it done?

[00:39:46]Shall we?

[00:39:49]Okay, fine.

[00:40:26]see I got one question.

[00:40:30]Okay. See you your uh you're training the mission with the un model. You're training the LLM model with the unstructured data. You are giving the prompt, right? you if you are requesting for the cat image it generates the different cat image but when on the training it may be trained with different uh uh animal images like different animal images it is not only cat dog cat rat or uh lion or tiger whatever it is so millions and of data will be get trained but in the chart GP you are giving the prompt you need the cat image. So you are describing how the cat should be. Okay. Uh so you are describing like you want in this color you you want nose should be in pink or eyes should be in uh gray color the color of the cat should be white. So you can describe anything. So it will generates the new new cat image. But while on the training, it may be uh trained with the unstructured data. For example, I am pouring different colors of balls in a room. Okay? So, all the colors of the balls are mixed. I'm just pouring it in the room.

[00:41:58]I'm asking the few students uh to segregate it according to the color. Will the student able to do it or not? Yes, he can do it. Right? How he can do do it? So he can separate all the pink color box, pink color balls and he will put it into the box. How he identifies that is pink because many color balls are there but but identifying the color it is because the student is understanding some pattern identifying the color and uh he is separating all the balls. Similarly, the machine will also identify some patterns and trends and uh it will segregate into I mean it will cluster into the groups.

[00:42:46]So okay if suppose I am just giving 10 animal pictures I mean 10 animals but each animal consist of 100 to 200 pictures. Do you understand? I'm giving 100 lion pictures and as well as I'm giving I'm having the folder inside that folder I'm having 100 tiger pictures, 100 uh dog pictures and 100 uh um lion pictures, 100 elephant pictures, uh like 100 uh um giraffe pictures like I'm having various animal pictures and a number of picture pictures I have it what I give I with this pictures I'm going to train the LL model what will happen it will understand see if the lion will be in this color so this much height so giraffe will be uh it is very tall and uh so each and every animal uh is unique they have different body structures they have different features uh different characteristics so according to that it will segregate all the images And it will different it will make it into the cluster. Okay.

[00:44:04]Clustering means a grouping. Okay. Now the machine understand the patterns. It will group all the lion pictures and it will group all the tiger pictures like this is called as clustering. So it comes under unsupervised learning. If you if you suppose familiar with the machine learning techniques then you understand what is this unstructured data how it will do it. So we know about K means clustering. Uh you have several algorithms under the unsupervised learning. Okay. If you go through that you will understand how it is happening.

[00:44:38]So similarly your LLM will also get trained with the unstructured data. I hope you understand this.

[00:44:47]Do you understand this clear?

[00:45:08]Okay. Fine.

[00:45:16]Okay. Let move on to the next slide. Let us see what is LLM. And here also you need to understand expansion of LLM is large language models or foundational machine learning models that use deep learning algorithms to process and understand natural language. These models are trained on massive amounts of text data to learn patterns and entity relationships in the language.

[00:45:51]Okay, what is that? LLM or foundational machine learning models. Okay, that use deep learning algorithms to process and understand natural language. Okay, fine.

[00:46:06]So, it is a language model which is responsible for performing task such as texttoext generation. Okay. Text to image generation. Okay. And image to text generation.

[00:46:21]Okay.

[00:46:23]Understand? Text to text generation.

[00:46:26]Text to image generation and image to text generation.

[00:46:33]You understand this?

[00:46:44]One second.

[00:47:04]Okay.

[00:47:06]Once again. One second. I'll share with one example.

[00:47:26]Okay. I show the example of texttoext generation. Okay. Textto text generation prompts. So what I'm going to do let us see one example of textto text generation.

[00:47:46]So what I can give um so prompt for content writing.

[00:47:59]Okay.

[00:48:01]So create a social media post promoting a new AI course for college students.

[00:48:35]Okay. So I have given the prompt as text. So what is the response? Text to text generation.

[00:48:43]See here.

[00:48:46]Okay. So it is giving the content for social media post on the title of AI course. Okay. For college students. Can you see it? So you can simply copy and you can post it. So you see here so it has given with the hashtag symbols with the hashtags. Okay. So this is one example for content creation. I mean text to text for content creation. So next uh text to text for next example is for email generation.

[00:49:25]Okay. So for email generation, so create your professional email inviting industry experts to your guest lecture.

[00:50:01]Okay.

[00:50:02]So this is uh textto text generation but the example is for generating I mean email generations.

[00:50:13]Email generation. Can you see here you are getting the email here.

[00:50:21]Okay.

[00:50:27]So this is on email generation and next I can uh question generation like question generation like MCQ question. So generate 10 multiple choice question on Python programming with the answers.

[00:51:13]Okay.

[00:51:20]You see here so generating the MCQ questions.

[00:51:27]So this is textto text generation.

[00:51:38]So in that slide what we have seen ex text to image generation. So next example is text to image generation.

[00:51:52]Let's see the prompt for example prompt for this.

[00:51:59]Okay. So let us see here.

[00:52:08]Let's uh text to image generation. Let's try it. So create a breathtaking sunrise over a mountain with a crystal clear lake.

[00:52:59]crystal clear uh lake reflecting the sky.

[00:53:10]So uh I need like a realistic photography style. So you can mention the style here like ultra HD.

[00:53:36]And I need like vibrant colors.

[00:53:44]Let us see how it is generating.

[00:53:59]It's taking some time to generate.

[00:54:23]See here. So I got a very beautiful picture.

[00:54:51]Can you see this? Okay. Fine.

[00:55:00]So this is one example for text to image. Okay. Fine. So similarly I give the small task for you.

[00:55:11]Okay, just try one text to image.

[00:55:16]Okay, and what uh image you are getting just put it in the chat box.

[00:55:36]So try it by yourself.

[00:55:51]So try it.

[00:55:53]Try one text to image and whatever output you are getting just give it in the chat box.

[00:56:03]You can uh try with different LLM I use for charge GPT. If you want you can uh give it in um [snorts] uh you can try it in Google Gemini or cla anything just what result you are getting put that picture in the chat box. So you can attach the image here.

[00:56:24]Any image you can try it. Not the same what what I gave. You try it by your own.

[01:00:28]I'm sorry.

[01:00:35]So I could see uh many images created by the students.

[01:00:43]Okay. Fine. Very good. Very good. Try fine.

[01:01:07]Okay.

[01:01:09]So, we understand what is LLM. Fine. And next, why LLM is so powerful? What makes LLM so powerful?

[01:01:18]So, so from this theory you can understand in case of LLM one model can be used for whole variety of task like text generation, chatbot, uh summarizer, translation, code generation and so on because chat GPT have the LLM M okay chd is a llm okay that only the ch the chart GPD can do for text generation chart GPD can act like chatbot board it can summarize it can translate it can do the code generation so that's why this LLM is so powerful okay and LLM is also a subset of deep learning it has some properties merged with generative AI okay hope you Understand? So, LLM is a subset of deep learning and it has some properties merged with generative AI that's why LLM is so powerful. So, it can do multiple task like text generation, chatbot, it can act like a uh chatbot. It can do summarization, translation, code generation and so on.

[01:02:41]So, that's why LLM is so powerful.

[01:02:45]So same thing suppose in the machine learning you are training a model for sentiment analysis. So see here so by giving this emoji symbol so it will understand it is fantastic the experience is fantastic I mean it is giving the positive response. So this emoji symbol is uh uh giving for the neutral response. The product is okay.

[01:03:09]Um okay and here this is negative response.

[01:03:15]Okay this emoji symbol gives the negative response. So this gives for positive response and this one gives for neutral response and this gives for negative response. So your u okay so I'm training the model using the machine learning uh and I'm just uh training the model for sentiment analysis. Okay. But come to here I mean for the this one.

[01:03:42]Okay. This is like language translator.

[01:03:44]So language translator means I'm speaking um in English. Uh sorry. Uh okay. If I'm going to a Spanish country uh sp uh sp uh if if I'm if I am the two is I'm going for Spanish. I don't know the Spanish language. What I'll do? Uh I use for the Google translator. So it is also one model that has to be already trained. We need two different models, right? But in LLM it can do this is the only one model. This will act this will work for sentiment analysis as well as it will work for language translator. So for example that is example is stat GPT or Google Gemini or cloud AI. So these are all LLM. Okay.

[01:04:35]Okay. So that's why this LLM is so powerful. Only one model can do multiple task.

[01:04:45]That's why it is so powerful.

[01:04:48]You are clear in it.

[01:04:51]You are clear.

[01:04:55]You're understanding this.

[01:05:01]Okay.

[01:05:48]Okay. Let us see some few milestone in large language model. So I have given only few milestones here. Sorry few LL models here. So first is Google Gemini Charg HLM T5 Lama Mist Falcon. So these are all the few milestones in LLM in industry. It it is much many more but I have listed only the few LLM models here.

[01:06:23]Okay.

[01:06:25]So next we are going to understand the generative AI end to end pipeline. Okay, end to end pipeline.

[01:06:37]Let us see what is that end to end generative AI pipeline. Okay. So generative AI pipeline is a set of steps. You have to understand here.

[01:06:50]Generative AI pipeline is a set of steps followed to build an end to end generative AI software. Okay. End to end generative AI software.

[01:07:03]Okay.

[01:07:05]For example, you can understand from this four lines. Okay. See, take any problem like break the problem down into several sub problems. Okay. then try to develop a step-by-step procedure to solve them. Since language processing is involved, we would also list all the form of text processing needed at each step. So this step-by-step processing of text is known as a pipeline. Okay.

[01:07:40]See you have we are taking the problem and we are breaking that problem into a several several I mean uh sub problems and what then what we are doing that so we are uh developing a stepbystep procedure to solve the sub problems to solve this approach problems. Since language processing is involved, we would also list all the forms of text processing needed at each step. So this step-by-step processing of text is known as pipeline.

[01:08:23]Right? So just understand from this much just have a look on it and note it Oh, I hope you understand this. Let me go to the next uh slide. So in the generative AI pipeline, so we have these steps. The first step is data accusation.

[01:11:15]So second is data preparation >> and third is feature engineering >> modeling and uh evaluation deployment monitoring and uh uh model updating. So these are all the uh AI steps.

[01:11:39]Okay, generative AI pipeline. These are all the steps uh are there.

[01:11:45]Let's let me see it one by one. So generative a pipeline first one data accusation.

[01:11:52]So so in this data accusation we collect the I mean data accusation is suppose if we are training the model we need some input. Okay. Okay. That is known as data accusation. So we need data to train the model. So we need the uh input data to train the model. Okay.

[01:12:18]So collect relevant data from multiple sources.

[01:12:22]Structured and unstructured data anything you can take. Okay. Collect.

[01:12:27]First thing is collect relevant data from multiple resources. So multiple sources like you can uh collect internet or websites or in database uh or you can create your own data anything.

[01:12:44]So you can collect relevant data from multiple sources. Second is you can take your data can be a structured data or unstructured data any data you can take it. Second thing and third ensure data quality and compliance. Okay. So every AI project starts with data. So data can come from websites, databases, documents, images.

[01:13:10]Your data can be come from videos, sensors or user interactions.

[01:13:16]For generative AI, large volume of high quality text, images, audio or code are collected. the quality of the data directly impacts the quality of AI model.

[01:13:34]So that's the first step. Okay. And second is data preparation.

[01:13:40]What is this data preparation?

[01:13:43]Only three points you need to remember.

[01:13:45]Okay. The first point is clean and organize the data. The first point is second is remove duplicates and errors. Okay. Third is format data for training. Okay. Let me explain. Raw data is rarely ready for training. We clean the data by removing inconsistencies, correcting errors, handling missing values, and [snorts] standardizing formats. In generative AI, data is often tokenized and converted into machine readable format before training begins.

[01:14:36]And third is feature engineering. So feature engineering means extract meaningful information. Okay. Secondly, transform data into useful representations.

[01:14:50]Third, improve model performance.

[01:14:55]Okay. The feature engineering involves identifying important patterns and characteristics from data. In traditional machine language, this is done manually. In modern generative AI models, deep learning automatically learns features from data, reducing the need for extensive manual feature creation.

[01:15:22]>> Screenshot.

[01:15:32]Okay. And the fourth is modeling.

[01:15:39]In modeling it's a AI architecture. It trains the model and it optimize uh parameters. Okay, this is the uh core stage. Okay, this is the core stage where the AI model learn from data for generative AI model such as transformers, large language models, GANs or diffusion models are trained.

[01:16:15]During the training, the model learns patterns.

[01:16:20]It will learn relationships and structures within the data. Okay. So you can take the key notes whatever I'm telling.

[01:16:31]Okay.

[01:16:35]You need to understand the modeling with select AI architecture.

[01:16:40]Train the model.

[01:16:42]Optimize parameters. Okay. Okay. And the next step is evaluation.

[01:16:52]In evaluation, it measures the model performance.

[01:17:01]It measures the model performance. It tests the accuracy and quality of the model built.

[01:17:10]Identify improvements.

[01:17:14]After training, I mean after training, see we are training the model, right?

[01:17:20]After the training, the model is evaluated using validation data. We access how well it generates content, answers questions or performs assigned task.

[01:17:34]Metrics such as accuracy, precision, recall and human feedback are commonly used.

[01:17:45]Okay.

[01:17:48]So next to the evaluation it is deployment.

[01:17:52]So here we make the model available to users. We prepared the model right? We evaluated. Now all have to try it. How the models the model should be see the the WhatsApp is created developed evaluated sorry so we are the end users right so we are all using the WhatsApp and similarly h we are creating the model we are training the model we are doing the evaluation. Now in this step I mean in this deployment step that model has to be available to the users.

[01:18:43]It should be integrated with applications provide it should be provided with realtime access. Once the model performs satisfactory it is deployed into production environments.

[01:18:58]Users can access the model through the applications, websites, APIs, chat bots, virtual assistant or enterprise system. Deployment transforms the model from your research project into your usable solution. Okay. So next to this monitoring and model updating.

[01:19:34]So here in this step we do track monitoring means we see now the application is in the user's hand like uh everyone started using that uh using the application using that model. Okay whatever we trained we prepared we deployed so now it is in that app is in everyone's hand. Now we need to take the feedback from the customers or from the users who are all using that app. So we need to track. So by taking the feedback we are tracking the performance continuously. So track performance continuously detect errors and drift.

[01:20:18]Retrain and improve models. So AI system require continuous monitoring after the deployment user behavior. So business requirements and data patterns can change over time.

[01:20:34]Monitoring helps identify performance issues, bias or model drift. Regular updates, finetuning and retraining ensure the model remains accurate, relevant and effective. So these are all the seven important steps of generative a pipeline.

[01:20:59]Okay.

[01:21:01]So first is da d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d d da stationation. So data preparing, feature engineering, modeling, evaluation, deployment, monitoring and model updating.

[01:21:29]I give all this uh as a short point.

[01:21:33]I'll give it now so you can uh note it down.

[01:21:59]Okay, wait a minute.

[01:22:27]So this is for data accusation.

[01:22:39]Second uh for uh data preparation and third for feature engineering.

[01:22:55]So wrote it down.

[01:23:05]Fourth for modeling and fifth one is for evaluation and sixth is for deployment.

[01:23:36]Seventh is for monitoring and model updating.

[01:23:45]Okay. So, note down this.

[01:23:51]Just have a look on this.

[01:26:01]Take the quick notes on it.

[01:28:19]It's not done then completed.

[01:28:32]Okay, please scroll down to the seventh point.

[01:28:37]Okay, we go next.

[01:29:27]Okay fine.

[01:29:30]Okay. Now we will do one small program.

[01:29:32]So program is AI jokes generator. Okay.

[01:29:37]What is that? AI joke generator.

[01:29:43]So we do it. So we use Python language for it.

[01:29:50]We use Python language.

[01:29:58]So I use Google Collab.

[01:30:28]Let me share the screen now.

[01:30:49]I hope the screen is visible for you.

[01:30:54]Okay.

[01:30:55]So, Okay. So, so what is the program we are going to do is so AI job generator.

[01:31:37]Yeah.

[01:31:39]joke generator.

[01:31:47]So we are going to use for Python language. So using Python.

[01:31:54]Okay. So next uh I I'm using for Google Collab. Okay. So you can slightly go to the Google type for Google Collab. So you can work on it. Okay.

[01:32:08]Let me uh do the code first. I'm going to import the library is random.

[01:32:19]Okay, I run this. Okay, it is working fine.

[01:32:25]Let's add the line here. I'm going to create Okay. So, jokes equal to I'm going to create it. Okay. So, I'm going to take computer related jokes and uh uh here I have taken a dictionary format like key and value. Okay.

[01:33:04]So I'm going to take computer related groups. So this is this I am going to declare in dictionary. So you know you are aware of Python dictionary.

[01:33:13]So here this is the key. What related jokes I'm going to take? First I'm going to take for computer related jokes.

[01:33:23]So this is key. Before colon it is key.

[01:33:26]After the colon it is values. Okay. So what is the value I'm going to give?

[01:33:31]Okay. I'm just uh I'm going to give it here. So, what joke I'm going to give here? So, why did the computer go to the doctor?

[01:33:46]Computer go to the doctor.

[01:33:55]Okay.

[01:34:03]So I'm giving the reasons here because sorry it should be inside the semicolon.

[01:34:27]Because had a virus.

[01:34:39]So the next uh next joke I'm giving here.

[01:34:45]Why was the computer cold?

[01:35:22]It left its window hit window open. Okay. And next one more joke.

[01:35:45]So, why did the keyboard break up?

[01:36:41]Okay. So that's it. So for one thing I have given Okay. For computers I have given. So next uh I'm going to take for uh okay so next uh I'm going to take for student.

[01:37:18]So next I'm going to take a student.

[01:37:35]So I did.

[01:38:28]so anything you can give it.

[01:38:33]Okay, I give the second and third comedy.

[01:38:41]Okay, I give the second and third committee.

[01:38:45]Okay, so next uh next is I'm sorry.

[01:39:02]So why was the math book sad? Because it has too because it has too many problems. Why did the student eat his homework? Because the teacher said it was a piece of cake.

[01:39:15]So I had the joke for computer. I had a joke for student.

[01:39:21]Okay. So next uh I have the joke for teacher.

[01:39:40]So why did the teacher wear sunglasses?

[01:39:46]Because her student was so bright.

[01:39:50]Okay. Why was the teacher uh happy?

[01:39:54]Because the class finally understood the lesson.

[01:39:59]So why did the teacher go to the beach to test the waters?

[01:40:46]Okay. So that's all. Let me run this.

[01:40:50]I have created see I imported the library and here I have created the dictionary. Okay. So dictionary name is jokes. So this is key and these are all values. So key and value. Oh, okay.

[01:41:07]The next I'm going to get the input enter the topic. I'm giving the options also here. So you want to if you want you can give the computer related jokes or uh student uh or uh teacher.

[01:42:17]So I'm giving only the three options.

[01:42:20]Okay.

[01:42:22]So dots I'm using for lower function.

[01:42:28]So whatever text you are typing I'm just converting that input to lower function uh I'm I'm see here it is like you are getting the input here the user may give computer or they he may give student or he may give teacher so he can give all the letter in capital or all the letter in small whatever any any case he can give either upper case or proper case or lower case anything whatever it is he can give anything but I'm converting that into lower. Okay.

[01:43:04]And then I'm going to use uh if condition okay if topic so here where I'm getting the input so topic in so I'm giving only three options right what is the dictionary name dictionary name is jokes so put jokes here keep the colon print chokes of the day.

[01:43:52]Okay. and then so print I'm going to use for random function random dot choice.

[01:44:11]Okay.

[01:44:13]Else topic is snow. Is it clear?

[01:44:23]You can alter this message.

[01:44:42]No jokes available for this topic just run this. So now you give any input I give student can you see here? So it is giving the jokes from the student you run for anything like uh teacher so here so why was the teacher so similar so it is working perfectly so this is known as AI job generator using Python I want you to try this code try this take the Google collab and write it open the Google collab And try this immediately.

[01:48:31]Okay, let me scroll for the remaining code.

[01:48:40]Just try it.

[01:48:43]Show me the answers. Just take the screenshot of your answers. and put it in the chat box.

[01:53:50]Okay, fine. I got um the screenshot from two students.

[01:53:57]Let's do two students again. One is from Adba and Aushi Malik.

[01:54:08]and from Tanvi Tanvi Singh I got from three students. I could see the a screenshot of them. And at last from Samiksha and from Prael sorry Prawell Prawell at the beginning itself I revised the yesterday's slide fully then only today I started the new session.

[01:54:49]Okay, only from four to five students I got the screenshots. What about others?

[01:54:57]So today we have done for AI jokes generator using Python we have done it still you are doing yes slowly I'm getting the results from everyone I got from hit Ashika And I'm got from Saga.

[01:56:07]Okay.

[01:56:12]Good.

[01:56:13]Fine. So, day two link is given here.

[01:56:24]Date to attendance link is given here.

[01:56:26]Without fail just give your attendance.

[01:57:03]Okay fine everyone have used collab very good very good Anushka have given the vary we have posted the outputs Then okay and uh rupta.

[01:57:33]Okay. Fine. Good. Ma it is happy to see all your uh outputs.

[01:57:54]content can't be viewed here.

[01:57:58]So you can't see my screen. Ma, I share the screen. Can't you screen see my screen?

[01:58:17]It is visibility. Okay. How about today's session? What are the things we have seen today? Today we saw for what is generative AI and what is LLM and we did textto text generation text to image generation using the chart GPT I hope you remember. So what makes LLM so powerful?

[01:59:00]Why LLM is so powerful? We see for sent language translations and we see for what are all the few milestones in LLM and we started seeing for uh end to end generative AI pipeline. So in that we have seen for seven steps. So first step is data data accusation.

[01:59:29]Second is data preparing preparation. Uh third is feature engineering. Fourth is modeling. Fifth is evaluation. Sixth is deployment. Seventh is monitoring and model updating.

[01:59:45]And we have done one hands-on session like uh uh hands-on session title I mean uh topic is AI jokes generator. So we have done that.

[02:00:05]So do you understand the today's session?

[02:00:08]How is your clarity level?

[02:00:11]Give your reply in the chat box.

[02:00:18]Just give the reply in the chat box.

[02:00:25]Okay. Great session. Thank you. Hit all.

[02:00:31]>> Fine.

[02:00:36]Okay. Good learning experience. Yes ma'am. Understood. Great session. Very good session, ma'am. Amazing session.

[02:00:43]Thank you so much. Lecture was amazing.

[02:00:46]Crystal clear ma'am. This very clear session ma'am. Okay.

[02:00:53]It was a good session. Very informative and useful ma'am. Okay.

[02:01:12]very informative and it's interesting.

[02:01:17]Okay ma'am we expect more programs like this every day rather than theory. Thank you for having patience while explaining everything ma'am. Thank you. Thank you Shahila.

[02:01:28]It was a great session ma'am. Thank you.

[02:01:30]Easy to understand full clarity.

[02:01:33]really very happy to see all your uh uh text in the chat box.

[02:01:40]Interesting and informative. Very useful and great. Good session ma'am. Good explanation. Okay.

[02:01:47]Fine. Fine.

[02:01:52]You're on track.

[02:01:56]Thank you. Thank you all for your valuable feedback.

[02:02:05]The session was good, easy to understand. I learn new concepts about generative AI and LM. Thank you.

[02:02:13]I can't show out because Okay, no problem. No problem.

[02:02:26]Very understandable providing.

[02:02:30]Thank you Vishnavi.

[02:02:37]Okay fine. So daily go through the topics ma. So the past two day sessions let's okay some of them are requested for more practical sessions. Okay we do that we do that one by one but understanding the concepts is very important.

[02:02:58]Very informative session.

[02:03:05]All of you without fail fill the day to attendance.

[02:03:10]The uh link is given in the chat box.

[02:03:14]Without fail just fill the attendance. So then only your attendance is recorded.

[02:03:22]So without fail fill the attendance the attendance will link day to attendance link will not available for tomorrow. It is only for today.

[02:03:44]If there is any doubts you can ask me here.

[02:04:11]Okay. If the jokes are already defined then how it can be termed as artificial intelligence. See we are training the mission with uh okay uh millions of data. Okay. We are uh training the mission with millions and millions of data. Okay.

[02:04:29]it will automatically create. So here you know I just uh I have created a small thing even I have not used the model using the Python I just created the small thing like dictionary concept and if condition I did the AI jobs generator but if you are training the model with millions and millions of data it will automatically create for example in the chat GPT you are typing something automatically even you take the example of WhatsApp you are typing something automatically the suggestion word will be there when you're typing itself the suggestion words are available right so okay that is a small uh scale model it is so we are speaking about LLM so it is a large scale model so WhatsApp suggestions um word or small scale model so now we are in LLM so when you are typing in the chart GPT um it it contains 1,000 tokens suppose if you are using for uh paid charg so you can get for 4,000 tokenization so I will explain the concept of tokenization in the further classes then you can understand how you uh it is doing everything I mean how it is creating the new content everything you understand the concept of tokenization to understand that.

[02:06:02]Okay. So here I have not used any model in the code itself. You can see I have not used any model. I have used um a jokes generated using the Python code.

[02:06:13]Okay. [snorts] But once you train the model with uh millions of structured or unstructured data, it will automatically create.

[02:06:24]Okay. Whatever you are asking, it will automatically create. It will understand the pattern and it will create for you.

[02:06:31]I hope you understand Adba.

[02:07:02]Okay fine. So thank you all for joining the session. Thank you all.

[02:07:08]Amar sir.

[02:07:11]>> Uh yes ma'am. Thank you.

[02:07:13]>> Yes sir.

[02:07:14]>> So dear students I hope everybody marked the attendance and share the feedback about the session. Thank you so much for joining on the YouTube and uh we'll again see you sharp at 7:00 tomorrow.

[02:07:27]Okay.