GenAI overview and LLM history

0:00 this massive course about foundational generative AI was originally recorded live we've put it all together into this one video this course offers insights into generative models and different Frameworks investigating the production of text and visual material produced by AI the course is taught by three experienced instructors okay I think uh we can start with the session so hello everyone good afternoon to all uh this is the very first session for the uh generative AI uh from today onwards we are going to start with a community session of generative AI so uh yes in today's session we'll be talking about that what all thing we are going to discuss uh throughout the sessions and uh this session actually it will be happening for uh upcoming two weeks and uh it will be on the same time I'm going to take the session from uh 3:00 p.m.

0:46 onwards uh maybe 3: to 5 so here in this uh committee session we'll try to to discuss many more thing regarding the generative AI so we'll start from very basic and we'll go to the advanc there we'll try to develop different different type of applications as well uh first of all I will start with the theory uh so there I will uh discuss about the theoretical uh uh like stuff and all that what is generative AI what is a llm and after that I will uh like go with the open a Lang and don't worry each and everything I will discuss in a very detailed way and and uh I will show you the dashboard as well where all the lectures and all will be uploaded and apart from that I will uh show you the uh like uh where you can find out all the videos quizzes and uh assignments and all because along with the session I will give you the different different assignment different different quizzes so at least you can practice with the concepts got it guys yes or no so are you excited please do let me know in the chat if you are excited

1:58 then great so I think we going to start and uh so first of all guys what I will uh do I will give you the uh the detail introduction of the course that what all think we are going to discuss in this committee session and uh here basically this is our dashboard so let me share this link with all of you so uh don't worry my team will share the link of this dashboard in the chat so from there what you can do you can enroll it is completely free you no need to pay anything for this uh committee session and all the lectures and all the assignment and quizzes will be uploaded over here don't worry I will come to the curriculum also so here first of all let me show you the homepage uh this is the homepage guys uh this is the homepage of this dashboard and there uh uh like you can enroll in this particular dashboard and don't worry uh each and every video you will find out over the Inon YouTube channel as well so all the recorded

3:08 video and all it will be available inside the Inon YouTube channel uh definitely this video is going to be record after the session so this video will be available inside the I YouTube channel as well as uh over the dashboard so here guys this is the dashboard I think you got a link of the dashboard so there you can enroll and then uh like you can start your journey of the generative AI so here guys uh me and buy both are going to take this particular session so there we are going to discuss in depth uh about the generative AI about the llm we'll try to discuss about a various application various recent model llm models and all uh we have so many things uh to discuss about the generative AI we have planned for the two weeks but uh maybe uh more than 2 weeks uh we'll try to uh take if we are not able to cover and like all the curriculum whatever we have defined whatever we have thought uh so definitely we'll be extending the date as well but yeah we will make sure that within 2 week whatever curriculum we have defined we'll try to complete it so here guys are me and uh buy so if you

4:17 don't know about me guys so my name is Sun my name is s Savita I'm working in Aon from past three year and I have I'm having an expertise in data science uh I have explored U every aspect of the data science like machine learning deep learning and advanced deep learning like computer vision NLP I have worked with the mlops as well uh there I have designed a various applications and all got it so yeah uh you can search me about more over the LinkedIn uh there you will get uh you will you will get got my profile and you will uh like uh get each and everything in a detailed way so here guys uh what you need to do so first of all you need to enroll to this particular dashboard uh you no need to pay anything over here and if you are going to login after login what you need to do uh so uh you will be redirect redirecting to this uh particular dashboard and uh so let me show you the dashboard first of all this is the dashboard guys as of now there is no such videos and all uh definitely after the session we'll uh upload the videos and assignment and quizzes so definitely along with the sessions and all you can

5:27 practice so this thing is clear to all of you yes or no have you enroll uh through this dashboard did you get this Dashboard please do confirm in the chat I'm waiting for your reply please do it guys great so I can see uh uh many people are saying yes so definitely now uh we can discuss about the curriculums and all so whatever thing we are going to discuss throughout this committee session first of all I will give you the detail uh like uh detail introduction of the uh syllabus uh what all topics we have uh we'll try to more focus on the re recent Trends I'm not going into the uh classical uh machine learning and the Deep learning basically so I will focus more Focus basically on the open Ai langen and all so don't worry I will give you the detail introduction of the syllabus uh that whatever thing we are going to discuss inside this committee session so the first thing what you need to do guys you need to enroll inside this dashboard and uh whatever like videos and all you will get uh you will like go through and

6:46 basically you can watch it over here itself directly now let's discuss about the curriculum and all that what all thing we are going to discuss and uh for that basically I have created one PPT so let me show you that particular p PPT uh just a second so here is a PPT guys can you see this PPT yes or no please do confirm in the chat if it is visible to all of you great I think uh this uh PP is visible to all of you now guys uh we can discuss uh that uh what will be our uh like topics and all uh that what all thing basically we have to discuss throughout this committee session so here uh first of all I will start from the generative AI so there I will give you the detail overview of the generative AI that what is a generative AI uh why uh we should use a generative AI uh what all type of like application we can create and uh

8:07 each and everything each and every every theoretical stuff we try to discuss regarding the generative Ai and after that after the generative AI I will come to this uh large language model so just a second let me open my uh pen as well so I can write it down um each and everything yeah so here guys uh we can uh I can write it down as well now so here the first thing basically first I'm going to start from the generative AI there uh definitely I'll will be talking about each and everything each and every aspect of the generative AI then after that I will come to this large language model there I will try to discuss this llms large language model in a very detailed way we'll try to see the complete history of the large language model that what is a large language model what types of model we have what was the classical model and uh what is a recent model okay so each and everything we'll try to discuss regarding this llms and after that after completing this uh theoretical part theoretical of this uh stuff and all regarding the generative AI regarding this llms I will come to

9:17 this open AI open Ai and this lenen so there uh we'll try to discuss in a very detailed way that what is the open AI what is the open AI API and inside the open AI API we have a different different them right so in OPI openi itself you will find out a various model that like openi open has created a various model uh that different different version of the gpts okay so it is having some uh like a old model as well some legacies and all and some upcoming models so each and every model will try to discuss I will give you the walk through regarding those particular model and I will discuss about the python API python uh API uh that uh how you can uh use utilize those particular model by using the python getting my point and apart from that apart from the python API and all will try to discuss that uh like if we are going to be uh if we are going to use the Lenin right so how it is different from the open AI so at the first place I will give you the uh detailed differences between this open Ai and this lenen that how it is different to each other that why this Lenin is required then I will come to the Lenin and then again we'll try to create uh again basically we'll try to

10:29 Define the lenen and all uh by using the python we'll try to uh use the lenen or different different like a component of the Lenin like memory chain agents and all and yes after that I will try to create one application okay so here uh basically we'll try to create one application and with that uh definitely we'll be able to justify the knowledge whatever actually uh we are going to learn regarding this llm open lenion by using uh that by creating that particular project and after that I will uh come to the advanced part Advanced part like uh Vector databases uh we'll try to discuss about a different different Vector databases and first of all we'll try to discuss the need of the vector database that why it is required what is the meaning of the embedding uh how we can uh like uh save the embedding how we can retrieve that and how this a vector database this a vector database plays an important role whenever we are going to create any application related to this llms okay so there we'll try to discuss about the vector databases and then I will come to the some open source model

11:39 so uh first of all I will come to this uh llama uh and I will discuss about the Llama indexes what is the Llama index and we have a like couple of Open Source model which is a very very famous so we'll try to talk about those model as well like we have a llama to itself we have a falcon we have a bloom there are various model and we'll show You by using those model how you can create uh like a how you can create your end to and application uh you can solve any sort of a task just take a name don't worry I will give you the detail overview about the NLP and all that what all task does exist what all task basically we have what all task we can solve by using this llm each and everything we'll talk about and then finally we'll create one more end to endend project there we'll try to uh use the entire knowledge uh whatever we are

12:27 going to be learn uh like this Vector databases different different open source model and a length CH open a llama indexes and finally we'll try to deploy that model by using the amops concept so did you uh like this syllabus yes or no please do let me know guys please do let me know in the chat if you like the syllabus yes or no the agenda is clear to all of you if you can write it down the chat I think uh that would be great great so I can see many yes in the chat and many people are saying yes they are able to get yes don't worry we'll give you the PPT and all each and everything will be there in a uh resource section so from there you can download this PP you can download this entire thing whatever I'm uh like I will be using throughout the session yes uh so fine I got a confirmation now uh the first thing uh

13:49 many people are asking the prerequisite what will be the prerequisite if you are starting with this committee session so prerequisite wise uh if you have a basic knowledge of the Python if you have a basic knowledge of the Python if you know about the core python uh in a core python actually we have uh uh like a ifls for Loop and uh different different type of data structure and the knowledge of the database exception handling if you are uh if you know about the basic python the basics of python then definitely you can proceed with this go along with that if you have a like some basic knowledge about machine learning and deep learning so you will understand uh the concept basically whatever uh like we are going to teach you in a better manner in a better way because here I'm not going to talk about the classical uh ml or the basics of the deep learning like uh artificial NE Network CNN and all definitely I will give you the overview about the transfer learning fine tuning and all but here uh uh I won't talk about the neural network and this recurr neural network lstms and

14:56 all so uh if you have a basic understanding of machine learning and uh deep learning so definitely you will understand the concept in a very well manner otherwise basic python knowledge is fine for creating application basic python knowledge is uh like fine okay so you no need to worry about it uh whatever thing actually I need to explain you definitely I will do that uh in the class itself and uh we'll do the live implementation I'm not going to show you any uh pre-written code and all uh definitely I will write it down each and everything in front of you only got it so prerequisite is clear so prerequisite nothing just a python or uh I can write it down over here basic knowledge basic knowledge of ML and DL if you know this much then definitely uh like uh you will be able to understand each and everything in a well manner great so yes we'll talk about the RG approaches and all each and everything diffusion model is there there are some recent model in l each and everything will talk about and you will be capable so uh let's say if you are working in your company or maybe you are trying to switch into the generative

16:06 AI or maybe you are fresher in every uh case right so this community course will help you definitely if you if you are going to attend every session if you are going to learn along with me definitely you can build anything after learning all sort of a thing got it great so I think uh this uh introduction is clear to all of you now I already discussed about uh about the dashboard and all so uh I given you the walk through of the dashboard so link you can find it out inside the chat inside the chat and from there itself you can enroll now the syllabus is clear dashboard is clear each and everything is fine so I think we can start with the introduction of generative AI generative Ai and llm because from today's on uh from tomorrow's onwards I I will be like move to the Practical part and there I will be talking about the open a how to generate a open key how to use the open API and uh we'll try to understand the chat completion API functional API and uh we'll try to understand the concept of the token also that what is a token uh like how many token should I use

17:17 whenever we are giving any sort of a prompt what is a different different prompt template and all there are lots of thing which we need to understand so today's session actually it will be uh like completely introduction session and in this particular session uh we'll talk about the generative Ai and the history of the large language model so guys uh are you ready can I uh get a quick yes in the chat if you are ready then yeah definitely we'll talk about the uh like use cases of the generative and all u in today's session itself I will like give you that uh particular idea that where you can uh utilize this generative AI in a real time yes definitely this course content and all whatever you are seeing over here this one definitely it will be available over the dashboard as well so this is our dashboard we'll update it over here inside the course syllabus section so each and everything uh like uh we'll try to update in the dashboard itself here is a class timing and all and uh I will make sure that the uh the link also okay so uh we are not going to uh we are directly streaming over the YouTube so directly you can uh join through the YouTube so for that you just

18:31 need to subscribe the channel and you will get a notification in that case great so people are saying yes sir we are ready ready ready great yeah definitely Wy we'll try to discuss the applications and all and uh I will explain you all the thing and that uh specific way only don't worry we are going to build AI application by using the AI tools AI based application great so I got uh many yes in the chat now I think uh we can start with a uh with the introduction of generative Ai and the LM so guys uh first of all tell me that how many of you you are uh you have started with the generative Ai and all already means uh you have learned something at least you have learned the basics in all uh so if you can write down the chat so that would be great means you are starting from very U

19:48 like a scratch or you have some sort of idea great so many people are saying uh so some people are saying they know about the basics and some people are saying uh they're starting from the scratch don't worry so uh basically I will start from the scratch only now here guys uh you can see I have created One PP for all of you so let me uh first of all let me go through with this particular PP and later on what I will do I will again I will give you the revision by using this PP only and in between I will use my Blackboard also for explaining you some uh Concepts and all so here is my Blackboard so here I will be writing down uh the whatever basically thing I need to explain you and in between I will be using the ppds and all so first of all let me go through this particular PP and here you can see so uh I have written some sort of a name uh so in the generative AI whenever we are talking

21:06 about the generative AI or a large language model so couple of name are very famous nowadays and in in those name actually this chat GPT is a uh like very very famous so here I have written this chat GPT it's a product of the open AI as we know about this Chad GPT everyone knows about the chat GPT yes or no I think yes now if we talking about this Google bar so it's a product of the Google and we talking about this meta llm 2 so this meta llm 2 it's a product of the Facebook got it guys yes or no so yes uh nowadays actually everyone using this chat GPT Google B meta lm2 is it it's also a platform similar to this chat GPT uh where you can uh chat or where you can ask a specific question which you uh which you do in a chat GPD itself So Meta lm2 it's a a model from the Facebook site now here guys if we are talking about the generative AI or we are talking about the large language model so in our mind the first image which comes into the picture that is a chat GPT Google B and meta llm 2 yes or no tell me guys yes because of that only uh because of this chat GPT Google B and

22:20 like the other the different like whatever application you are seeing nowadays right so mid journey is one of the application or maybe Delhi uh or different different application because of that only I think you are learning this uh particular uh thing this particular course this generative AI course yes or no yes so but guys this generative AI is having their own Roots it's not all about the chat jpt Google B and some other uh application which you are seeing chat jpt is just a application of this generative AI chat GPT or this Google B is just a application of this uh like llm large language model basically we are using this large language model in a back end U like whatever application you are seeing like chat GPT and all in the back end but apart from this this generative Ai and this llm is having their own Roots so first of all what I will do I will explain you the concept of the uh like first of all I will uh start from the deep learning itself means I need to explain you few uh terms and terminology regarding this deep learning so let me uh back to the Blackboard so there I will be talking about the basics of the

23:31 deep learning So within uh 5 to 10 minutes I will be discussing the types of the neural network and all and then I will directly move to the uh like LMS and this uh genem so here guys uh you can see uh what I can do I can uh draw one box over here so this is the uh you can think this is what this is the neural uh uh basically if we are talking about the okay so first of all let me start from the deep learning itself so uh if we talking about a deep learning so uh we can uh divide this deep learning into three major segments so let me write it down over here this a deep learning so guys this deep learning actually we can divide into three major topic so the first topic actually which is called artificial neural network artificial neural network netork the second topic is called convolution neural network CNN the third one basically which is called recurrent neural network so we have a three types of the neural network

24:42 and we can divide this a deep learning into this three major section apart from this you will find out other uh like topics as well so let me write down those uh thing over here so the fourth one uh which I can write it down over here that is a uh reinforcement learning and uh the fifth one we generally talk about it uh so that is what that is a gain so this gain also it comes under this generative AI I will talk about it I will talk about this game I will like give you the glimpse of this uh generative advisal Network that what is this and how the architectures look like of this gains and why I'm saying that this gains comes into the generative AI so if we talking about thisn so let me draw the box now so if we are talking about this n so here guys see we have an

25:30 input layer inside this Ann actually what we have we have a input layer and uh you will find out the output layer and in between actually in between this input and output we have a hidden layers so just a wait now over here guys see we have a input layer and we have a output layer now in between actually you will find out a hidden layers various hidden layer so let me write it down over here input and here you'll find out the output now here in between this input and output you will find out of various hidden layers so let me write it down the hidden over here so this hidden layer actually it is nothing it's a hyper parameter so we can have as many as hid layer we can have as many as node inside the hidden layer we all know about the artificial neural network I'm assuming that thing now if we talking

26:18 about this uh CNN actually so the CNN is nothing so in the CNN uh one more thing you will find out in terms of this CNN that is what that is a convolution we always perform the convolution in terms of this CNN so here if we are talking about this enn so uh we are using the uh like structure data where we have a like different different features numeric feature or categorical feature and uh we try to solve the regression and classification related problem but whenever we are talking about this CNN so here uh the CNN actually specifically we use for the image uh related data image or video related data you can say that uh we use the CNN and all for the grid type of data okay so we use the CNN for the grid type of data and there uh like you will find out one more component that is what that is a convolution so here uh let me write it down so the component name is what component name is a convolution so in the convolution actually you will find out a various step so uh we have a various step in the convolution itself so the very first step which we perform what we do guys tell me we perform the feature X section by using a different

27:29 different filter after that what we do we perform the pooling and then we flatten the layer so there are different different like uh uh steps you will find out inside the convolution itself and after that what we do we apply the fully connected layer so that is nothing that is my Ann itself so over here I can write it down we have this convolution and we have a artificial neural network so this is my first architecture which is a like uh which is the Ann itself and and this is my second one that is what that is a CNN now if we talking about the third one which is a very very interesting that is called recurrent neural network that is called recurrent neural network so this enn we generally use for the structured data where we have a numerical column or categorical column and in the Target column like it will be a numeric or categorical one and based on that basically we are going to decide whether it will be a classification problem or a regression problem now if we talking about this CNN so already I told you if you're talking about this RNN so the name is what the full form what the full form is the recurrent neural network so this RNN actually we are using for the sequence

28:38 related data so wherever we have a sequence wherever we have a sequence so this RNN we used for the sequence related data now let me do one thing let me draw the architecture of this RNN so over here guys in the RNN what you will find out so let's say this is my uh box and here is what here is my input so this is what guys tell me this is my input now here is what here is my output so let me draw the output one more time this is what this is my output got it now here guys see uh this is my input this is my output and this is what this is my hidden layer now in the hidden layer actually you will find out one thing one concept and the concept is nothing the concept is called a feedback loop okay so whatever output I'm getting from the hidden layer actually again we

29:26 are passing that output to hidden layer until the entire time stem so that thing actually uh we learn or we learn in in the RN itself actually this RNN is nothing it's a special type of neural network and there you will find out the feedback loop feedback loop means what so whatever output we are getting from the hidden layer again we are passing the same output to the hidden layer until we are not going to complete the entire time stem that is what that is the RNN now uh you are uh we are talking about the llm so why we are uh why I'm discussing this RNN and all because this llm actually somehow it is connected to this RNN itself before starting with a llm a large language model we'll have to understand the concept of the RNN lstm attention uh like encoder decoder and then attention self attention and all so here I'm not going to discuss in a very detailed way I'm just giving you the glimpse of that that what is a like RNN what is the lstm what the Gru and then what was the sequence to sequence mapping and where this attention comes into a picture then how they have

30:34 invented the self attention then how they started the using this transfer learning and this finetuning in terms of this uh in terms of this large language model why we are calling it is a large language model why we are not calling it a model okay so each and everything we'll try to discuss now uh you all know about this uh reinforcement learning and all so in the reinforcement learning uh you will find out one agent environment regarding that particular agent you will find out a different different state getting my point and then you will find out the feedback so that actually it comes inside the reinforcement learning and that is also part of the deep learning only now if we are talking about gain so gain is uh nothing actually so in the gain again you will find out a neural network uh which we are using for generating a data and that also comes in under inside the generative Ai and we have a different different types of game so first of all tell me guys uh this uh uh like types of the neural network this is clear to all of you please do let me know in the chat if uh this thing is clear then uh I will proceed with the next topic use cases wise I will come to the use case and I will uh try to discuss a

31:45 different different use case I will come to the use case then I will tell you the applications of that and then I will come to the domains as well then in what all domains you can apply those use cases so don't worry each and everything we'll try to discuss over here yes I will directly come to the generative a itself but before that I will give you the timeline don't worry from Tomorrow onwards I'm going to be start uh I'm going to start from the uh like uh from the open ey itself uh like complete practical and all so no need to worry about it s Prasad I think you got your answer I think uh this basic introduction is clear now yes coming to the generative only don't worry yeah it's going to end to end uh we'll try to discuss end to end thing don't worry about it if you have any questions and all so

32:55 you can directly ping into the chat uh so I will reply to you don't worry okay so I think now we can proceed so guys here uh in the uh PP itself I was talking about the generative AI then I given you the uh like uh uh the types of the neural network and I just explain you the uh like the regarding the artificial neural network and the CNN and this RNN so here in the generative AI uh you'll find out that I have like included a few slides and all so let's try to understand a few uh thing from here and then again we'll go back to the uh the Blackboard and then I will try to discuss few more concept so over here uh we have seen the chat GPT like I was talking about the different different application like chat GPT Google B and metm and all now let's talk about the generative AI that what is a generative AI now here you can see the definition of the generative AI uh which I have

34:13 written over here uh that is what that a generative AI generate new data based on a training sample right so the name is uh the name is self-explanatory right so the name is explaining everything generative AI the AI which is generating something now what all thing we can generate so here if we are talking about the generative AI so you can generate images you can generate text you can generate audios you can generate videos as a output you can generate anything uh so uh this image text audio video it's nothing it's a type of the unstructured data and definitely it is possible by using the generated AI we can uh generate this type of data by using the generative AI now if we are talking about the generative AI so uh as I told you that it is having their own Roots okay so it is having their own roots and if we are going to divide this generative AI so we can divide into two segment so the first segment is called generative image model and the second segment is called generative language model and this llm actually it falls into this particular segment into this generative language model are you getting my point yes or no

35:24 I think yes so if we talking about this generative image model so I told you when I I was talking about the Deep learning uh like Ann RNN and CNN reinforcement learning and there was a gain so initially we were using the gain for generating a data so let me show you the architecture of the gain so the how the architecture of the game looks like so with that you will get some sort of idea in the game we are using this uh neural network only so let me show you the architecture of the game so let me search it over here over the Google game architecture now uh here in the image uh let me open the architecture of the game so here guys uh just see so in the gain actually we have two main components so the first component is a generator this is what this is a generator okay so uh I think this is visible to all of you this is what this is a generator and here you will find out discriminator so this generator and discriminator is nothing it's a neural network so we are passing this real data so here basically what we are going to do we are going to pass a real data and here we have a generator

36:33 which is generating some sort of a uh like synthetic data and here we have a discriminator based on that we are going to discriminate between real data and the synthetic data so this is the architecture of the game and inside this architecture you will find out we are using two main thing we are using two main component the first one is generator and the second component is a discriminator I think you're getting my point and this generator and discriminator is nothing it's a neural network got it so this is also comes under this generative AI so now let me show you this generative AI now over here I have written two points generative image model and generative language model so if you're talking about generative image model so in our previous days in our back back days actually in our old days in 2019 18 so this gain was very popular for generating a data again uh this gain is very uh like EXP uh like expensive in uh terms of computation power and all so it is very like very much expens uh like expensive in terms of like uh computation uh so over here you can see so we were using this gain uh we were using this gain for generating images and all in our back days in 2018 and in

37:44 2019 and it was very very popular and we have a different different uh variants of the Gams if you'll find out the type of the gain you will find out many types now uh recently actually you will you have find out the trend of the llm large language model now guys here uh we are uh we are talking about the game and then uh this gain basically it was the old concept it is a old concept basically and there are different different variants of the gain as well now over here if we are talking about this large language model so it become very famous from the Transformer I will come to the Transformer I will tell you the complete history of the Transformer as well now uh this image model and this language model but a recent days in a recent days what I have seen in terms of this llm and all even we can generate the images by using this llm we got those llm basically which is like that much powerful so by using those particular llm we can generate generat images as well okay so we I will show you couple of models and all uh regarding this uh like image generation and definitely you will get some sort of idea that how the uh those

38:56 part particular alms is working in terms of image generation I can give you a couple of example uh like Delhi so Delhi is a example you can uh check over the open a which is a model which is like a uh like a famous for the image generation now here uh if we are talking about this image model so actually see this image model basically it was working for image to image Generation image to image generation now this generative model actually so if we are talking about this generative model it is working Tech uh it is working in terms of text to image generation text to image generation and text to text generation so this two tasks definitely we can perform by using this llm model and this image to image generation before we were doing it by using this gain model in 2018 and in 2019 now uh as I told you that we have those powerful model in our recent Days by using those particular model definitely we can Implement image to image generation as well that is also possible so uh regarding that uh definitely I will show you couple of model so we are having four tasks here I have written it now

40:05 let me move to the next slide and let me show you that what I have so here guys you can see uh this cat is representing a genitive model where you are giving a prompt uh means where you are giving a question and uh as a response U again as a output basically you are getting a response so in terms of uh see here we are talking about generative model I'm not talking about specifically this El M okay so I told you this uh generative model actually uh you can think it's a super set this generative Ai and under this generative AI you will find out this llm and gang is also part of the generative AI getting my point I think this thing is getting clear so over here we are talking about the generative model so we are giving a input and we are getting a like output now specifically if I'm talking about

40:52 regarding this llm regarding this large language model so this input actually this is called input prompt and the output actually it is called output prompt so this cat you can imagine as a generative model or as a llm model so what we are passing as a input we are passing input prompt and we are getting as a output output prompt so this prompt term is a very very important I think you have heard about this uh prompt engineering and all that uh uh like uh prompt engineer is getting this much that much and this prompt engineer plays a very important role if uh we have to design any sort of of a prompt now uh different different types of prompt of like zero short prompt few short prompt few short learning and all we'll talk about it as I will progress with the like implementation and all in between I will give you like idea regarding each

41:39 and everything now over here guys you can see uh where this generative AI exist so if you will look into the uh look into through this particular slide so here you will find out this generative AI actually it lies inside the Deep learning getting my point so the generative AI actually it uh like reside inside the Deep learning uh initially only I have explained you that uh we have a different different types of neural network and it's a part of the deep learning only now whether we are going to generate an images by using the llm or by using the gains or whether I'm going to perform text to text generation text to image generation or image to text Generation by using the llm both lies inside this generative Ai and this generative AI is a part of the it's a part of the tell me it's a part of the deep learning now over here guys I have

42:26 written a couple of more slid so I will try to explain you uh but first of all let me give you the timeline of the llm and then I specifically I will come to the llm and all and I will be talking about this discriminative Ai and the generative AI as well so tell me guys uh this part is clear are we going good are you able to understand whatever I'm explaining to all of you so if you are getting it so please write down the chat and you can ask me the questions as well if you have any uh type of Doubt or U like if you're getting it or not getting it whatever you can ask me in the chat uh like chat section uh I will reply to your questions no reinforcement learning is not required uh uh specifically we should not go for the reinforcement learning and all yes this is a part of the uh like this Genera way is a part of the deep learning right right yes llm model used in a generative

44:06 AI correct you got it uh guys mathematical intuition so we will talk about the mathematical intuition and all but this uh more uh this course this comp session is it is more focusing on the applied side so I will create a various application in between whatever math iCal concept and all will be required I will let you know that don't worry great so I think uh people are getting it and uh they are trying to understand fine so whatever I have explained you let me explain you the like Blackboard uh and then again I will come to this PP and we'll try to uh wrap up the theoretical stuff and U then I will explain you the applications and all so over here guys see I was talking about this Uhn CNN RNN RL and G now I started from the generative AI itself so I have started from the generative Ai and I told you this generative AI is nothing you can consider it as a super set as a super set now inside this generative AI you will find out many uh like uh many uh concept many topics and all so here uh regarding the generative

45:33 AI there is uh two main thing which you will find out the first one is gain gain that is a generative aders Network the second is what llms llms large language model now we have a various task so here let me write down the task as well so the task wise so here I told you the different different task basically so the first task which I can write it down over here that is a image to image Generation image to image generation now the second task was the uh image to text uh text to text generation text to text generation text to text generation now the third task was the uh image to text Generation image to text generation and the fourth one was the uh image to image generation sorry uh text to text Generation image to text and text to image generation so let me write it down over here text to image generation text to image

46:50 generation now if we are talking about this image to image generation yes we were able to do this particular thing by using this gain we have seen the gains now we are talking about this text to text generation yes it was possible by using the lstm RNN and the uh different different by using the different different model as well but yeah this text to text generation actually nowadays you are seeing uh we are preferring this large language model for this text to text generation and you this chat GPT is a biggest application uh biggest like example for that the chat GPD which we are seeing image to text generation yes uh this is also possible by using a different different model like RNN lstm and Gru image captioning if you have if you uh if you have heard about this uh like image capturing task so that is also possible uh by using this uh like uh classical model but yeah by using this llm also we can perform it we can uh do it now if we are talking about this uh text to image generation so yes uh this type of task nowadays it is possible by using the uh llm so yes yes uh llm is able to do llm is able to perform a various amount of

48:00 task uh whether it's a homogeneous or it's a hetrogeneous now uh I was talking about the uh llm uh sorry I was talking about this generative AI so where it exists so this generative AI actually it exists uh in a u like a deep learning itself so you can think that AI is a superet machine learning is a subset deep learning again is a subset of the machine learning and this generative AI is a subset of the deep learning because as I already told you we have a different different uh like other neural network also in a uh like a deep learning and this CNN is one of them this a convolution neural network okay so I think this part is clear to all of you now let me draw the architecture uh that where this uh generative AI exists so you can think this is what this is my AI this is one uh this is the like a super set now here this is what this is my machine learning this one now uh inside that you will find out the uh deep learning and inside the Deep learning you will find out this generative AI uh so let me take a different color over here uh let me take uh this color so here you will find out the generative AI so this is what this

49:10 four circle is what this is the generative AI got it now here uh you can see why we are saying so why we are saying this is a like a subset so I think each and every explanation I have given you over here uh you can uh prefer this uh like a this particular slide that why I'm saying this generative AI is a subset of the deep learning so let me write it down over here this is what this is nothing this is a gen Ai and it's a subset of the deep learning now guys let me explain you the timeline of the uh this llm so uh now you got it that this uh llm is nothing it's a part of the generative a itself this large language model now let me talk about the complete timeline of this large language model so how it evaluate and uh uh I can like talk about the complete history of

49:58 it and uh here guys you can see that first I was talking about the RNN so as you know that uh what is the RNN tell me RNN is nothing it's a type of the neural network it's a type of the neural network so uh there basically we have a feedback loop again we can pass the information to our H layer now you will find out a different different types of RNN or some Advanced architecture in terms of this RNN itself El the second uh like thing which is a type of the RN itself that is called lstm lstm right so in the lstm actually uh if we are talking about this lstm so here you will find out the concept of the cell state so in the RNN we just have a Time stem and it is for the shortterm memory it is for the shortterm memory we cannot retain a longer sentences by using this RNN it is not not possible if our sentence is a very very huge or it's a very very long so we cannot retain that particular sentence by using this RNN but if we are talking about this lstm yes we can do it by using this

51:08 lstm so in this lstm you will find out the concept of the cell state so uh this uh lstm is nothing it is for the short-term dependency and it is for the long-term dependency also it is for the short like a memory short-term memory and is for the long-term memory as well if you look into the architecture of the lstm so you will find out along with this uh time stem so here we have the time stem U like it's a hidden State actually uh like on a different different time stem along with that you'll find out one cell state so it is going to retain it is going to retain the long-term dependency and in between in this a time stamp in this short-term memory and in this cell State you will find out a connection the connection in terms of gates so here you will find out one connection uh like uh one gate basically that is called forget gate so here I can write down the forget gate now here you will find out one more gate actually so that is called input gate here you are passing the input now here you will find out one more gate over here that is called output gate output gate okay so we have three gates inside

52:18 the lstm for sustaining a long-term dependency or for reminding a longterm uh long sentences now uh you will find out one more updated version of the lstm so this RNN is a old thing this lstm is also old thing now you will find out one more updated version of the lstm that is what that is a GRU so this Gru actually they have invented in 2014 and they had they took the inspiration from the lstm itself now inside this Gru you won't find out the concept of the cell State everything is being done by the hidden State itself and here basically in the gru we just have two gate update gate sorry reset gate and update gate and it's a uh Advan or you can say it's a updation on top of this lstm it's a updated version of the uh like lstm itself now what is the full form of the gru G and recurrent unit now over here guys see this was the three architecture which was very very famous during 2018 and 19 in our old days now here see uh one concept comes into the picture if we are talking about this RNN lstm and Gru so by using this particular architecture what we are doing so by

53:30 using this particular architecture we are going to process a sequence data yes or no we are going to process a sequence data now here one concept comes into a picture sequence to sequence mapping and for that only we are using this particular architecture so we have a different different type of uh like a mapping technique so let me write it down over here uh different different type of mapping technique uh now it is fine uh I think I'm audible to everyone now now I am audible guys please do confirm in the chat I think there were the issue from the mic side now I'm audible so please do confirm in the chat if I'm audible then and uh is there any Eco or uh what so guys are you facing any Eco in my voice now it is fine yeah it is perfect I think great fine fine fine uh it's clear great uh I think now I am audible to everyone sorry I think there was a issue from the do let me know in the chat uh from where I lost my voice so this concept is clear this one

57:48 to many or one to one one to many many to one many to many [Music] yeah so I think uh I was there RNN lstm and Gru now I think it is fine I'm audible to everyone great so I was talking about RNN lstm Gru and then U I talked about the different different mapping sequen now uh this mapping sequences actually we can Implement by using this uh RNN lstm and Gru so over here uh yes so 1 to 1 one to many many to one many to many RN and LSM and Gru this was the sequences actually I was talking about now in 2014 actually see this was the sequences by uh we can Implement by using this different different models getting my point now over here uh if we talking about this particular sequences definitely we can uh like uh per we can uh create a various uh application by using this model but here basically we are having some sort of a restriction uh as I told you the different different application like one to many many to one so many to one means uh you can think that sentiment analysis one to one to many means what one to many you can say image capturing many to many image uh sorry uh language translation so there are various application of the sequences now

59:27 see uh we are talking about the sequences uh the sequence to sequence mapping now uh we can definitely implement it by using this particular architecture so the problem we were having the problem was actually uh we cannot see let's say we are giving an input in the input actually we have a five words so whatever output we'll be getting in the output also we should have a five words so it's a fixed length input and output getting my point what I'm saying so by using this particular mapping 1 to one many to one or like many to many specifically we are talking about many to many so there was some problem there was some issue the issue was fixed length input and output so whatever number of inputs we are passing in terms of this many too many I'm talking about okay so whatever number of inputs we are passing so those many output on we can get it over here in the output itself so uh here actually one a research paper came into the picture in 2014 you can search about uh the research paper sequence to sequence learning so inside

1:00:40 that uh paper they have introduced the concept of the encoder and decoder in the encoder and decoder actually the one segment the one segment was the encoder segment segment so let me uh draw it over here so the one segment was the encoder segment and the another segment was the decoder segment this another segment was the decoder segment and in between actually in between we were having in between actually we are having the context Vector so here uh in between this encoder so we are having the encoder and we are having the decoder decoder one part was the encoder and one part was the decoder and in between we having the context Vector means whatever information was there whatever information was there from encoder to decoder we are passing through this context Vector means we are wrapping all the information in this context vector and we are passing to the decoder that actually the paper uh has been published in 2014 you can search about it you can search over the Google sequence to sequence learning so let me uh search in front of you only now over

1:01:52 here I can write it down sequence to to sequence learning research paper now uh over here guys you will find out this uh particular research paper now just try to read this paper now here in this particular paper they have clarify the issue that what was the issue with the classical mapping so that was restricted to the input and output now over here you if you will read this particular research paper so easily you can find it out the issue here itself in the like introduction itself they have mentioned they have mentioned this uh despite their flexibility and power can only be applied problem who inputs and targets can be sensibly encoded with the vector of fixed dimensionality it was just for the fixed dimensionality and basically there was we were having a limitations so for solving that particular limitation this a sequence to sequence learning paper came into the picture and there was three person Ilia sasar and orol and this there was one more person and this paper from the Google side now here guys uh let me open this uh Blackboard again so there was a context Vector but this uh encoder and

1:03:03 decoder also was not able to uh perform well for the long uh longer sentences so here in the research basically they have proved if my sentence is going uh is going uh like above from 30 to 50 words right if it is longer than 30 to 50 words so in that case it was not able to sustain the context it was not able to sustain the context if we are using this encoder decoder architecture now you will ask me sun what we were having inside the encoder and decoder so we are talking about the encoder so again here we were using the either RNN lstm or uh lstm and we were using this Gru and here also in the decoder also we are using this rnl we are using the lstm and we were using this Gru got it I think you got the problem now and you got to know about the encoder and decoder so we have started from the RNN then now we came to the lsdm Gru and then we have a different different mapping and for solving this particular issue which is related to this many to many uh like uh mapping many to many

1:04:13 sequence mapping and this uh uh this language translation is one of the example if you will search over the Google translate uh just search over there anything let's say in the in H you are saying that or whatsoever so it will generate output so this input word and output word will would be a mismatch but that was a restriction with this uh like with the classical mapping so for using this encoder and decoder architecture we can solve that particular problem now here also we are having the issue that we cannot proceed a longer sentences we cannot proceed a longer sentences so here One More Concept comes into a picture inside this context itself and that was the attention that was the attention so uh here neural translation with just a second let me search about the neural trans TR a NS relation with attention yes this was the paper and uh this was the first paper let me search

1:05:27 about the research paper yeah now guys uh this was the paper in this particular paper they have introduced the concept of the attention and just try to download it you need to download this particular paper and uh then you can see there so just a second let me show you this paper as well and this is the main uh like main papers uh basically which you will find out uh while you will be learning this deep learning and all so this paper actually this has been introduced in 2015 I think in 2015 or 16 now here they have introduced the concept of the uh attention actually so just try to read uh this particular paper at least try to read the introduction of it uh there we have uh there they have defined that uh what was a problem with the encoder and decoder and where this attention comes into the picture and what is the actual meaning of the attention they have introduced each and everything over here inside this particular paper inside this particular paper they have introduced each and everything regarding the attention see this is the architecture of the attention model and uh before going through with any blog any website or any tutorial try to uh go through

1:06:37 with the research paper and try to understand the motive of that research paper now see guys uh here I'm not going into the detail of the attention because this attention itself uh is a longer topic but I can uh like give you the glimpse of that that uh uh what what they were doing in the attention so they were mapping so let's say we have a five words in the sentence so they were trying to map each word whatever word we have a input we were trying to match each each input word with the output word means this input and output this encoder and decoder if we are talking about this decoder actually so this is having the uh information each and every information of the Hidden State whatever like in the encoder like you will find out this RNN lstm or whether it's a GRU so uh we have a hidden State actually right so uh this decoder part is having the information regarding those particular hidden State all the hidden State and because of that it was able to uh it was able to predict so whatever like sentence and whatever words or longer sentences or like U like uh the longer sentences and all which uh

1:07:46 whatever basically we were passing it was able to predict okay so this word is related to that particular sentence so what I will do I will create a like a one dedicated video on top of it there I will try to discuss uh this attention mechanism but yeah here I'm just giving you the timeline U and with that um you can clearly understand so uh here we were having the attention mechanism now guys by using this attention mechanism by using this attention mechanism in 2018 Google again Google published one research paper and the research paper name was attention about this encoder in the encoder and this decoder we were using what we were using guys tell me we were using this LSM either we are using the lstm RNN or maybe Gru now uh there also we are having the lstm maybe RNN or maybe you are having the Gru and uh if we are talking about the attention so whatever uh information we are passing from here to here so you are having the context Vector context Vector now on top of that we are having the attention layer attention layer and it was nothing

1:08:58 it was just a mapping from input words to out output word now here actually they have published one paper in 2018 and the paper name was attention all your need attention all your need now this paper actually it was a breakthrough in the NLP history this paper has been published in 2018 and here actually decoder but there is one uh there is one thing basically in terms of this encod and decoder you won't be able to find out this lstm RNN and Gru they are not using any RNN cell any lstm cell or any Gru cell so here actually they were using something else and here the what is a uh name of the research paper so they were saying that attention all your need only attention is required for generating us let's say we are passing any sort of an input means any longer input so from that particular input only attention is required for generating output now how let me show you that this Transformer architecture or

1:10:12 let me show you the attention all your need research paper so attention all your need research paper so guys this is a very uh prestigious paper in our NLP history and uh this changed the complete history of the NLP and whatever llm and all whatever you are seeing like nowadays so they have used this Transformer architecture as a base model I will come to that and there I will try to uh discuss that uh what is the encoder and decoder again I'm not going into the depth of the mathematics but yeah definitely I will try to give you some sort of a glimpse so over here uh let me zoom in first of all this paper and and here guys the paper name was attention is all your need so this was the researcher asish Nome Nikki Jacob you can uh search about these particular people and here is the abstract uh you can see and this is the introduction at least try to read the introduction try to read this particular background and the model architecture so this was the model architecture which has been introduced by the uh by the Google researcher and the architecture basically which you will find out inside this research paper I think everywhere

1:11:24 you will find out this uh this particular architecture in uh whatever NLP tutorial or if you are going to understand the attention mechanism and all so this is the architecture now in this particular architecture let's try to understand that what all things we have so see first of all we have a input okay try try to understand try to focus over here so we have a input over here then we have a input embedding so this is my first thing input and this is what this is the input embedding the third thing which we have that is a positional encoding getting my point and then after that you will find out the multi-headed attention then we have a normal uh normalization and all and then we have a feed forward Network now guys just tell me this is what this is a encoder part this is what this is a encoder part and this is what guys this

1:12:11 is this is a decoder part this is the decoder part got it getting my point so here also we have a two segment first was the encoder and the second was the decoder but here we are not using any RNN cell lstm cell or maybe Gru cell here actually we are using something else some other concept and the concept actually I think this is not a new thing for you this embedding and all uh this embedding attention already I talked about the attention that what it is mathematically it is having a like uh like a some different explanation but yeah I think you got to know the idea now here we have a feed forward neural network you know like what is a like artificial neural network what is a feed forward neural network so it's not a like a new thing for all of you and by assembling all those thing they have created one cell one architecture and

1:13:00 the name is called this uh Transformer so this architecture itself is called a Transformer what is this guys tell me this is a Transformer now here guys just see uh this uh Transformer if we are talking about this Transformer and all so let me uh tell you few things regarding this Transformer uh so first of all guys this is a uh fast compared to the classical architecture if we are talking about this RNN lstm and all so there we are passing the input based on a Time stem based on a Time stem but if we are talking about this Transformer guys so here what is the importance or what is the like plus point which we have inside the Transformer it is a faster why because we can pass the input in a parallel manner we can pass all the inputs all the tokens in a parallel manner in a parall actually we can pass the input now over here see we have a input embedding we are doing an embedding over here and then we have a positional encoding means we are arranging a sequence sequence of the sentence then we have a multi-headed attention again we are trying to uh figure out the uh meaning see let's say uh the sentence is what I am Sunny now

1:14:12 uh here it is trying to find out the relation I with M and sunny is trying to find out the relation M with uh this I and sunny it's time to find out a relation this sunny uh and this m and this I so it is trying to find out a relation with each and every word so it is doing the same thing inside the multi-headed attention then you will find out this feed forward Network neural network actually and yes uh this is what this is my encoder part as I told you this is what this is my encoder part now if you will look into the decoder side so again we have a same thing so here we have a outputed output embedding means in uh uh like whatever uh like a sequence uh in whatever like U am format I want output so that is uh this particular thing this output Ting and then again we have a like

1:14:59 multi-headed attention and we are passing this thing uh to the next one to the next layer and again we have a feed forward neural network over here on top of this you will find out the soft Max and finally we are getting a output output probability so don't worry I will try to discuss this Transformer architecture mathematically in a detailed way in some other video but as of now I'm just giving you the GL Glimpse because whatever llms we are going to discuss okay as a base architecture they are using this Transformer so guys until here everything is fine everything is clear please do let me know in the chat yes or no yes you can uh let me know in the chat uh then I will proceed with the pp and all and uh we'll try to wrap up the uh introduction of this llm and all and in tomorrow's session we'll try to talk about the open a and we'll discuss about the open API and all and a different different models of the openi any doubt anything so if you have any sort of a doubt please do let me

1:16:11 know guys please do let me know in the chat uh I will try to clarify that uh those doubt and uh so did you get a timeline timeline of the llm I will come to the llm now the specific word and uh after deep learning an NLP what is the topic uh for generative AI please give up uh so after the Deep learning and see after the Transformer actually by using this particular Transformer people has created a different different llms and all large language model now I will come to that by using the slide I will try to show you that I think this is pry much clear now let me go back to