 Okay, going live. Okay. Hello, everyone. Welcome back to Sanjay Gupta Tech School. So today we have one more session, which is based on associate AI certification, which is launched by Salesforce. So today's topic is generative AI and predictive AI. So in last session, like you, you got to learn what is generative AI. So today, there will be a comparison. So I have Nikita with me. Welcome, Nikita on the channel. So she's... Hello, everyone. Yeah. So as you all know, like, I'm running this tech school. So she is also running her own venture. And you can see the name of that venture Silicon Heights. So she is also doing some education related stuff there. So more, she will be talking about that. So over to you. And just give a quick introduction about this and you can continue with the topic. Thank you, sir. So hi, everyone. I am Nikita as we have been doing so many sessions on artificial intelligence, specifically for associate AI programs. So we are quite familiar now. But I would like to just pin one thing, which is like, I would like to go to the LinkedIn where I am posting the content regularly. I would just take you there where, you know, we can just have a look on what I am doing, which is here. So if you can see, this is a page of artificial intelligence. Are you sharing your LinkedIn? Yeah. You need to share that. Okay, one second. Yeah. I have shared it. So this is our page. This is our LinkedIn page where you can see artificial intelligence by Silicon Heights. And I am constantly posting some content over here. Can you look down this check box? Yes, sorry. So if you all can see, we have... I have posted some content over here, which is of predictive AI. Then we had this GANs. And it started from the very basic of mathematics. And then it went on to, you know, giving the idea about what all the intricacies of data science and artificial intelligence are. So there is some content that is posted nearly every day. You all can follow it. And if you are looking forward to do the intense AI studies, then certainly this is one of the stops that you can, you know, stop on. And you can just check out the content and just read and imbibe all the important things that are going to be posted over here. So you can see it is all about maths also, the one that I taught on day one or day two. All the linear regression things are posted over here. Then we have some mathematical rules over here. So it is not as per the associate program, but it is the intense artificial intelligence program that we are going to run in future. So you all can just have a look and follow if you like, right. That's the page by Silicon Heights. And the name is artificial intelligence by Silicon Heights. This is it. Now we'll go into and we'll go and dive into the session and let's check what is there for us. Yeah, and I already shared the link in the chat. So I hope everybody will follow this page as well. So you will have lots of followers soon. Yes. Okay. So yeah, there we are. So we have already discussed about generative AI in the previous lecture, where I taught you about the intricacies of how generative AI has made certain advancements in the current scenario in the current technology. So let's just go through it. And the first slide explains us the Venn diagram approach for generative AI, where actually generative AI stands out. So you've got the artificial intelligence as a holistic diagram wherein we have got a subset of machine learning, then you have got a subset again, which is deep learning and then finally you've got a super subset which is a generative AI. Artificial intelligence is the broader field which encompasses everything which is in the development or if you have to consider something where you are transferring the skills of human mind into a model or computer, then that's what you can understand that artificial intelligence holistically is. That means that is what our target is that we are able to transmit the human intelligence into the computer or robot or any other model. Okay, this is our whole achievable thing. Then the subset that is through which AI attempts to achieve those targets is machine learning. So a subset of AI that focuses on teaching computers to learn from data and make predictions. So prediction means predictive AI, we are going to dive into it and we are going to see how predictive AI actually, what predictive AI actually is and how much of it is, you know, lately on the floor that we are going to study. Then you have a subset of machine learning that utilizes deep neural networks, very, very important because we are going to see the GANs, generative adversarial networks and the LLMs and the other technical things. So that's where the neural networks play the major role of transmitting and understanding all the content and then producing the desired output. Then you have finally got a subset of deep learning. This generative AI encompasses everything. It has to have everything in order to generate a specific output, which is in the form of text, image, music, video, code, anything that you opt for. So a subset of deep learning that specifically focuses on teaching computers to generate new and original content is the generative AI. I hope this is now pretty clear to you. If you have any doubts, you may please ask and then we can follow to the next. So I'll just have a recap of generative AI. Generative AI is a technology that takes a set of data and uses it to create something new like poetry now. Certainly, what am I supposed to generate? That generation has to be based out of something. So that training data set has to be input into the model so that I am now able to generate a text or generate a video, which is new in the sense, but it is based on the training data set. It is not as new as coming out of totally raw content, but it is based on something, however new we can say. So that's what generative AI technology is. It takes a set of data and uses it to create something new out of it. That is the difference also from the predictive AI and other traditional AIs. So unlike the traditional AI models, generative AI doesn't classify or predicts, but it creates the content of its own. So your LLMs and other models are creating the content of your own once you have input the desired training data into the models. There you go with the generative AI tools now. So as I mentioned that since you have generative AI gotten into the nerves, so we need to classify the generative AI on the basis of the functions it does. So whether it is creating audio or it is creating video or it is creating images or it is creating code, voice or music, anything that you opt for. So it has to give you these generations, these generations are based out of the trained data set. So you can see that the generative AI tools are visual, text generators, code generators, audio generators. These are all the types of generators that you can have. And then visual generators will generate image generators, video generators, design generators. So you will have different text to image, text to video, text to design, text to voice, text to music generators. And also from image to another image you can transition from. So this is what it looks like. So I mean this is how the functional classification of generative AI tools looks like. So then you have how does generative AI work. So let me explain it to you like in the previous session also and today also I'm giving a quite recap for it. That how generative AI models perform beyond the machine learning. So the first thing is to have a chunk of data that you want to train your model with. And then by utilizing multiple forms of machine learning systems, models and algorithms and neural networks, this gen AI is going to give you the outputs. So these models are typically trained on large data sets containing a wide range of information such as text, images and videos. And by analyzing these sorts of patterns, now my LLM is ready to analyze the patterns and then create its own prediction or generation, whatever it is referred to, whatever it is made for. So by analyzing patterns and relationships within the data, you can understand the underlying structure and generate new content similar in the style and context. So your contextual beauty has to be typically gotten into this model and now this is going to generate different outputs for you. So I can also say that you do not have to have billions or millions of content or data set for your LLMs. What you can just do is be domain specific, organization specific and whichever company specific you are, you can just train the model on behalf of whatever your company requirements are. You do not have to have some extra because you're not creating something universal. Something universal has already been created in the form of chat DVD. So you can restrict yourself into a domain and let's say your organization is something like FMCG. So you can just have a training or the train data that is based on your FMCG context or any other context which you think it is preferable to. So this is all about the generator AI. I just hope that, you know, watching the previous video and this one is getting you the nerve of Generative AI. Yeah, this one very helpful like with the diagrams, those who attended the previous session and if we like you are explaining things through diagrams. So like those who have attended the previous session, they can easily relate what are the functions of Generative AI and what are different components available there. So it is I think really helpful. So now we are going to hop on to one very common type of Generative AI model, which is also a very, very important part of your Associate AI program because there have been several questions on this part, which is Generative Adversarial Network. So be very much attentive while you read out of it. So you can see one common type of Generative AI model is Generative Adversarial Network, which is marked over here. Now I'm going to go on to the different slide to explain you what it actually does. Okay. So when you have got a generator over here, like there are two components of this GM, one of them is generator, one of them is the discriminator. These are the two components that you have to be very much, you know, you means you should learn that these are two components. Now what does this do is the generator is going on with the synthetic data, like you are actually wanting some synthetic output. So what are we doing is we are granting some input. We are just doing some input of raw data, which is not very real. It is synthetic. It is also very much deflected from the real data that, you know, that is the input from here. So there are two types of data. One is very much deflected from the real data. However, this is artificial and raw data, which is being fed into the generator. And then you have certain samples of real data and artificial data. Now what are we doing with the discriminator is we are going to seclude that. Okay. If the discriminator is able to figure out that, you know, this is fake data and this is real data. That means the work is yet not done. The work is going to be done only when the discriminator is not able to discriminate further between the fake data and the real data or fake images and the real images. That is when our process is finally done. So the outputs are fed into the generator until and unless the discriminator forgets or the discriminator is not able to differentiate between the fake and the real image or fake data and the real data. That's how the GANs would work. So now we will have the study of it. Before that I have made these diagrams to make you understand that GANs have got generator and discriminator as two major important components, which are generator neural networks and discriminator neural network. This is how the GANs are made up of. These are the components of GANs and how are they working? I have also put these working into the LinkedIn page where you can refer to and I have also tried to put and compile it over here also in a very concise way. Wherever, whichever way is going to be okay with you, you can do it like that. So GANs or generative adversarial networks are types of AI architecture that have revolutionized the field of synthetic data generation since their inception. Very, very important, the synthetic data generation is a VVIP thing. Because you are creating avatars, you are creating fake images, the deep fakes, whatever you are creating, the different, different pictures. You must be also been aware that there have been some reels and the picture outputs of different celebrity and how would their future look like? How will they look like when they are old? This sort of content is pretty popular on Instagram these days. So this is all out of this GANs thing. So when you play with the images, you create different images which are fake images. So you should see that this is the important technology that you are working on. So GANs consist of two distinct neural networks that engage in dynamic rivalry. One is generator and one is a discriminator. Dynamic rivalry means, so until and unless the discriminator is able to figure out the difference, that means the GAN is not properly working. As soon as the discriminator stops to differentiate between the real and the fake data, that means our GAN is now working fine and this is how the model should be looking like and working like. So the generator is tasked with generating synthetic data such as images based on random noise or some other input. Its goal is to produce data that is indistinguishable from the real data. And the discriminator, what does it do? The discriminator is trained to distinguish between the real data and fake data and its objective is to correctly classify whether a given input is real or fake. When it stops doing so, the GAN model is perfectly working fine. So here I have also deep dived into the processing of the GANs. There were, I remember some questions based on it, so that is why I have taken it into so much depth. During training, the generator aims to produce increasingly realistic data to fool the discriminator. While the discriminator aims to become more adept to distinguish between the real and the fake data, this adversarial process encourages both networks to improve their performance relatively. What do you mean by iteration? So iteration, if you must have heard about this term in two different programming languages, if you are aware of, or if you are somebody who is new, so I'll repeat that iteration means a regular continuous process until and unless a specified condition is fulfilled. I think because Sir only made this iteration process very much familiar to us when I was in the BTEC first year. So I think he would agree with this that my definition is working fine. So once trained, the generator can be used to produce new synthetic data that resembles the training data. So GANs have been widely used in various applications including image generation, image to image translation, super resolution and data augmentation. So in future, you are going to see very clear images, highly high definition images, which will have high pixels and the dimensions also will be very wide. So this is how they are going to work in the future. And this was about GANs. That's the major important study for the day, the GANs. Now if you have any problem, you all can please ask or comment whatever is good for you. And if you think that there's something that I have not explained, you can also let me know. The second architecture that we are going to see today will be large language models. Now what are large language models? These are again based on certain neural networks only because those are the, I should say the foundation of any, you know, LLMs. So chat GPT, one of the major thing that was a hype in 2022, I guess, and since then it has been hypes. So chatbot generative pre-trained transformer, there are three terms that are over here, which is one is generative. Second is pre-trained, third is transformer. Pre-trained means, as you know, it has been trained on a lot of data that was surfaced on the Google. And Sam already, who is the creator of this chat GPT, the OpenAI, he also mentioned that, you know, all the data that has been fed into for this chat GPT model has been taken from various sources, not only one, but there has been, there have been various sources for the training data that is fed to this model of chat GPT. And it's not only this model, OpenAI has got so many models. The lately developed models is Sora, which is text to image model where you can plug in the prompt and you will have a video generated out of it. So the chatbot's foundation is chat GPT large language model, LLM. A computer algorithm that processes natural language inputs and predicts the text, the next word based on what it's already seen. So here the intelligence is working, the neural network is working, which is observing certain patterns out of the language that we have put in. So whatever prompt you put in, it's not a language which is Chinese or Japanese or, you know, like you know what you are putting in. So this chat GPT has been made familiar with a lot of languages out of which English language is one of the language that it works on. So when you are putting some content in it, like, you know, create something for me, create a PPT for me, which is based on, you know, let's say conservation of animals, conservation of wildlife sanctuaries. So you can put in any sort of data. So this data is, you know, in the form of prompts. Now what are prompts? There are various, I should say that designations also related to the prompt engineering, one of the designations that is related to this prompts generation is prompt engineering, where you create certain prompts to get out most of from the AI model, whatever you are looking for. So prompts are generally the orders that you give to an LLM in order to generate some output or generate some required data that you are looking for. Okay, so some major important prompts that are going on lately, I have told you one is, you know, how do you send me a PPT or create a PPT, sorry, create a presentation for wildlife conservation, send or create a PPT that is related to the conservation of animals, which are, you know, relatively in the endangered zone. So these sort of prompts would generally work and get you the results. So next is, then it predicts the next word and the next word and so on until the answer is complete. So you can even ask it like, you know, what is artificial intelligence? It will tell you and you will ask the chat GPT like, who are you, you know, what is chat GPT if you want to learn more about it. So it is going to get you the answers from various websites collectively and then give you the desired output. So this is how it works. It works on word to word and then, you know, until and unless your paragraph is complete, your answer is complete with the rationale that it has had. So, you know, that's how LLMs would work. Some LLMs that you must have heard would be open AI is GPT-3, first in the foremost, then you have 3N4 and GPT-3.5 also. These are all the large language models that we are familiar with. Then you have Google's Lambda and LLM, that is the base for BARD. BARD is another LLM that is there. Then you have Hugging Faces, Bloom, and XLM-Roberta, which is another LLM that you can, you know, just work on and see how does it go. NVIDIA's Nemo, LLM, XLNet, these are another set of LLMs that we have. And finally, we have Meta's large language model, Meta AI, that's Lama in the short. So, these are some of the large language models that we have been served as a study or as a model to work on. So, this is all about the large language models. We might conclude a little bit to make it a holistic study for you. So, in the conclusion, what we have is an LLM is a machine learning neural network trained through... One second, yeah. So, an LLM is a machine learning neural network trained through data, input-output sets. Frequently, the text is unlabeled or uncategorized and the model is using self-supervisor, semi-supervised learning methodology. Both these terms have been explained to you. I only go to the previous lectures where you can see the supervised and supervised machine learning. I think this you covered in the very first session. Yeah. I remember. Yes. So, that's about machine learning and then information is ingested. Or content is entered into this LLM and the output is what the algorithm predicts the next word will be. So, the input can be proprietary corporate data. You can see over here. As I said, it can either be organizational data that you can feed to it and it can be fed with. Or if you want to make another GPT, chat GPT sort of thing which is worldwide for worldwide. So, that is you will have to train. Again, you will have to put in a lot of energy into the training process of your model. So, or as in case of chat GPT, whatever data it's fed and scrapped directly from the internet. The last is training LLMs to use the right data requires the use of massive expensive servers. Farms that act as supercomputers. This is very, very important like what is the actual requirement for the server of a large language model. That is stated over here, which is supercomputers. So, we started from a recap of generative AI where we discussed how gen AI works and what are the future outcomes out of it and how you can learn it. So, that's all for the presentation up till now. Now, we are going to study about the predictive artificial intelligence. And I'll also take you to the Salesforce content that is there for the generative AI and predictive AI because they have given two parts for it. And that's only content. There's no quiz out of it. So, you all can just, you know, click on the site and just have a visit. Though everything has been covered over here, but there might be a possibility that we might have left something that you can cover from there in order to have a substantial amount of data that you have to have for your associate program. Correct? Yes. So, let's read the predictive AI also and then I'll go there. Okay. The predictive artificial intelligence is another type of AI. So, fundamentally, if I have to iterate it, I will say that it will be an astrologer for the business. Now, why would I say that? It's because predictions are made by the astrologers and, you know, that's something that they think that, you know, can happen in the future. And then they restrain you to do some things in order to be a little careful about what might happen in the future. Same thing the predictive artificial intelligence does. It is going to predict the outcomes of your business. And if you see that the outcomes are not in compliance with whatever, you know, you have expected, that means either there's a problem in the model or if there's no problem in the model, that means there's a problem in the business. So, two things you can shortlist from its predictions. So, predictive AI is a computer's computer program's ability to use the statistical analysis to identify the patterns and anticipate. Now, what do we mean by anticipate is to realize if or not, you know, these behaviors would eventually come out and give us some essence out of it. So, to anticipate behaviors and forecast the future events. So, we are fundamentally wanting to know what is the future holding for us. So, your weather data that, you know, it will rain today and your data related to whether it will be snow in, you know, in the hill station or not, whether it will be delayed. So, these all are the predictions that are having a certain advancement today. Previously, they were not as act as they are today. So, this is all the essence of predictive AI. Also, if I would say any example of a person of a human being. So, let's say I go to the park and I generally see some orange clouds, orange dusty clouds and you know, which is followed by a storm. But this is just one experience. And twice and thrice, I see the same things happening constantly. So, when I go to the park again and I see that, you know, there is some dusty clouds and there's some wind that is happening. So, this has happened for twice. And then again, I go third time and then fourth time. So, I can now predict enough that whenever I see dusty clouds in a little wind, it is generally followed by the storms that are there. So, this is how I can predict from certain. Now, obviously, predictive AI takes a lot of information. What I am predicting is it just on the basis of single, what should I say, single data or single entity. This is what I'm counting that okay, red sky and a little bit of dustiness. That means I am predicting that there will be storms out there. So, I'm just counting one thing. However, predictive AI would require a lot of things to be entered and then you can have a desired output from the predictive AI. So, they are not creative. However, generative artificial intelligence, if I talk about, they are creative in nature. Now, it's not a person that I'm saying that, but the model generally is creative when we talk about the difference in the creativity. So, they are just outcome based. They would predict something and then they'll give you this desired or not so desired output. So, predictive AI makes statistical analysis faster and more accurate via ML and access to vast amounts of data. Again, everything wants data. So, data is a new world that's again justified. And then you have wireless predictions are not guaranteed to be correct. Predictive AI can help businesses prepare for the future and personalize experiences for their customers. So, this is one part of, what should I say? We have completed about 30 to 40% of this content that is there in the Salesforce AI associate program. Then we will move forward. So, the essence has been covered. And now I'll take you to the different... So, is that mean like generative and predictive AI? Is it complete or we have some more content left? It is done. It is done. Okay. So, for AI associate certification, whatever knowledge one should have. So, those things are covered in this session and in the previous session, right? Yes. Okay. So, now what's next? Should I just show what is left and what are we going to do next and a little bit about the... Yep. Sure. Yeah. Just give me a moment. Yeah. We have few viewers who regularly watch your session. So, I'm receiving the comments. So, it would be better if you can showcase what will be the next. And I think guys, I'm sharing the session tracker already with you. So, if you go to the description of this video, this live stream, you will find a session tracker. So, there we are also maintaining like what all sessions we have already done. So, you can just go through with that also. I need to change the subject. Yeah. You can stop sharing and then I think you need to share. The link, I guess. So, I'll manage from my side, but I'll just... Okay. No problem. So, maybe in the next session, you can start with... Okay. Not a 15. Yeah. So, yeah. Or maybe you can stop sharing and share then I think it will work. I had marked it over here, but I think it has... Okay. No problem. So, we can wrap the session here only. So, guys, thank you for joining the session. So, I hope you understood the generative and predictive AI differences. And I hope with the help of these two sessions, the previous one, which we did in this week and this one, you will be able to prepare these concepts for the Salesforce AI Associate Certification Exam. And next week also, we'll be having two more sessions on two different topics. So, keep following all the sessions. She's doing lots of efforts. I just know she's preparing content from everywhere, doing some research. So, if you follow all those sessions, I'm sure you will be able to know everything. Basic things, those are needed, one to know about the AI. Okay. So, do follow all the sessions and do follow the LinkedIn page. Link is already shared in the chat so that you will be updated with whatever is happening in the field of AI. Okay. So, thank you for joining the session and thank you, Nikita, for sparing some time and sharing your knowledge. You're welcome. See you in the next session. Bye, everyone.