Introduction to Generative AI: AI Tools, LLMs, Prompts & AI Limitations!

1. Class Introduction: Hi, guys. Welcome to my class on introduction to Generative AI. My name is they Kumadas. Just to give you a background about myself, I am an ex Google employee with 19 years of experience into advertising and I've been teaching advertising for more than ten years now, and I teach to a lot of young professionals, entrepreneurs, and experts who want to get into this. I wanted to take this opportunity today to let you know what we are going to cover in this class. So we're going to look at how the AI foundation understanding the introductions to AI, the concepts of it, domains of various types of AI which we're going to cover here, and then we look at business and career transformation because of AI. AI for businesses, for work, for career, and then looking at a lot of AI related issues, concerns, and ethics. We're also going to see the capabilities, applications and tools and prompt engineering which you can apply in AI. I hope by the end of this class, you understand these concepts thoroughly and you're able to apply them practically in your business and for your clients. Thank you once again guys for checking out my class and I'm really excited to see you inside the class. 2. Generative AI Introduction: Hi, guys. Welcome to this session. In this session, we'll talk about understanding what is generative AI. If you look at generative AI as the word suggests, it is a generative AI, which is basically where we are going to use AI to generate new content. It's a an artificial intelligence which now can go ahead and generate new content which never existed before. That is what we mean by genitive AI. If you look at the history of AI, AI has been utilized in various forms altogether in our day to day lives. Like we are using AI primarily for, let's say in maps. It shows us how much time it takes to reach a specific destination. It will tell us the Tesla cards which are running on its own. So there are various areas where we are already using AI. But now, with the help of genitive AI, you can generate new content as well with the help of this technology. This content can be of various types. You can generate text, images, videos, code. All these are now possible. Just to give you some example of what we are trying to refer to. If you go to Chat GPT, we can generate text. I can ask it to write an email for me and it can generate that. We can make use of Dali to generate images, which we can use for our business, for our personal work. Then we can go to Github copilot and we can generate code. So these are the generative AI types which are in existence at this moment. And that is how we are going to make use of it. So simply put, if you have to say generative AI or generative artificial intelligence is a kind of AI technology which we have now, which can generate new content. And that is where it can be a content of various types which can be possibly created. It is now capable of looking at the data, the prompt which has been given to it and based on which it can generate new content for us. That is where the name comes from generative AI, where we are using the AI specifically to generate new stuff of content. I hope this makes sense. I understand the basics of what is generative AI and what are its capabilities and in which formats are we going to we are using it right now at this moment. 3. Demo of Generative AI: Hi, guys. Welcome to this session. In this session, we'll see some of the generative AI tools which we can use to generate new content. So we're going to see how we can make use of, let's say, ChR GPT for generating text, writing a poem or picture to text. We'll see how we can generate text through a picture and translation as well. Then we will look at, let's say, stable diffusion or any other tools which we can use for generating images, text to picture. So let's have a look at this. The first thing which we want to do on Cha JBT is primarily going to be creating writing a poem. Let's say we give it a specific prompt where we are asking it to write a poem celebrating the arrival of spring season, we have specifically told it to keep it to ten lines. You can see it's going to follow that and it has generated a poem for us as well, a new content which has never ever been there in this world. For the first time it is getting created right. Similarly, let's say what we want to do is we want to upload a image and now we would like it to describe it. So Simply, we are asking it to describe what it sees in the image. So the picture shows two young children playing soccer on a grassy field outdoors. One child is wearing white t shirt and colorful shorts while the other is wearing light blue polo shirt with wage shorts. So you can see it has picked up every single thing in detail and explaining describing the picture to us. So you can do this as well, generating new texts from the pictures which we have around us. Okay? And the third thing is, which we want is translation. So let's say we're asking it to simply translate this text from English to Spanish. It can do that as well. Now you have seen three different ways of generating new content with the help of hangibty. Now, let's say we go to stable diffusion and here we ask it to generate an image based on the text which we will give. Let's say we're giving this particular text, which is creating a picture of a man playing a piano. We ask it to create that Situ be creating the image for us over here. You can see it has generated the image, and now you can change the prompt, you can modify the prompt and change the image as well as per your requirement. I hope this makes sense. I hope you're able to understand the practical implementation of how you can use these generative AI tools to generate new types of content as per your requirement. 4. Aritificial Intelligence(AI), Machine Learning (ML), & Deep Learning: Y. Hi, guys. Welcome to this session. In this session, we'll talk about artificial intelligence, machine learning, and deep learning, understanding what they are and how they actually work. So if you look at artificial intelligence is primarily a concept of making machines think like humans and act like humans. Okay? That's the idea of artificial intelligence. So what we are trying to achieve over here is human intelligence, which is artificial in nature, which should be able to have the capabilities as the human intelligence, where it can recognize images or video, understand, and generate text, beat humans at games, learn from data at scale, drive cars autonomously. Okay. This is a concept which came into existence, you can say 1965, ideally, and now we are seeing the outputs of it specifically speaking. Attivit intelligence is primarily a kind of intelligence which we are trying to build, which is at par with human intelligence and is able to give those kind of output, which human intelligence can do. Now, if you look at when we look at machine learning, machine learning is primarily where we are trying to train our computers to learn from examples. If you look at normal traditional programming which we have seen so far, so there will be an input data. There's an input data which we give, and then we give some rules ourselves, which is the code which we give and based on which the output would come out. So here we are writing all the rules and based on which the output comes out. But in machine learning, it is going to be a case where there's a tremendous huge amount of training data input data is given and the machine learns from the data and gives the most appropriate output. It's learning the rules itself, okay? The ML basically shows the machine thousands of examples and lets it find the patterns itself. That's the idea of machine learning which we are understanding over here. Okay? There are three types of machine learning primarily supervised, unsupervised, and reinforcement. Supervised is where it learns from labeled examples. So the data is labeled, so it gets trained on that and gives us the output. Unsupervised is where it finds hidden patterns. So from the training data provided, it tries to find hidden patterns, and then reinforcement is where it learns by trial and error. With rewards, just like training a dog. So why it matters is that traditional coding which we know for decades cannot handle real world complexities, and that is where ML scales a lot because it learns from the data and adapts to the new situations which we are not able to predict and improve the data more. So that's the idea of machine learning, how it helps in the current context. Now, how does the learning work? So let's try to understand this primarily from a different angle altogether. So we can take some impressions from kids, how kids learn. So, usually how it will work is parents will show apples to the kids and they say, this is an apple. Now, the brain understands it, remembers that a red or a green colour, round, shiny shape stem on top is an apple. Okay? And now, when that is shown again, the kid is able to recognize it as an apple. And that is how human learning happens. Imagine the same thing happening with these AI models where a training data is given. Let's say, thousands, millions of images of apples are provided, and now the model learns from that pattern that the color is red or green, shape around, texture, shiny skin. Okay, stem on top is referred to, and then a new image is shown, so it gives out the output as apple. The learning process or the style of learning is pretty similar how usually a kid learns. Now, there are three major ingredients which we would of machine learning, which you'll get to see, which is there is a lot of training data, which is primarily millions of labeled images, text documents, articles, structured tables, data, audio clips, speech, video recordings, which is uploaded with. And then there's a lot of computation power, GPUs, thousands of GPUs, high speed memory cloud computing farms, massive power consumption happens in this and distributed parallel training, which happens. And there is algorithms. So obviously, the algorithms, which works in this as well, like decision trees, linear regression, neural networks, clustering, reinforcement learning. All these are part of the ingredients of machine learning, which is able to now create a trained ML model. So here comes after this, which says the deep learning, which is primarily the neural networks or transformers. This works very similar to our human brains neurons as well. Like if you see biological neurons, so there is input which comes to the cell body and based on which the output is comes out from the brain. Similarly, in artificial neural network as well, in deep learning, there are inputs given and there are a lot of different outputs processing which happens and based on which the output comes out. So that's the deep learning part wherein it goes into the understanding of the inputs being given and based on which the outputs are provided. So now if you look at it in a n diagram, the artificial intelligence is the broadest category. It is the broadest category, and in this then comes machine learning which comprises of decision trees, random forests, and inside that then comes deep learning. So you can say that ML is primarily a subset of AI, and inside ML, then there is deep learning which sits over there, which is a multi layered neural networks which is being created. Is primarily useful for image, video generation, language, audio generation. For all those purposes, we require more data and compute than classical ML. I hope this makes sense. I hope you understand now the basic concepts of artificial intelligence, deep learning, machine learning. Thank you so much, guys, for being in this session. I'll see you in the next video. 5. Explore ChatGPT: Features & Capabilities: Hi, yes. Welcome to this session. In this session, we're going to look at some of the AI tools and their features and capabilities and how we can make use of them. The first one which you're going to look at is going to be ChatGPT, which is primarily a tool developed by OpenAI. This is using the LLM GPT five and above it is primarily designed natural language understanding and generation, which is more in a conversational manner. So this is something which we are going to use and see how it's going to work out. So the idea is that we just need to see their capabilities now, understand how they operate, and what is the level of the power they imbibe inside them. So let's have a look at this. The first one which we're going to look at is ChatGPT right here, this is they use the LLM model at the back end, and now we can do a simple prompt over here and which it can give us an output for. Let's say we're asking for what are LLM, and how do they work? Now look at how fast they're able to give us the output. The moment you type in within less than millisecond, you get the output over here in a very systematic manner, you get all the information because of the millions of amount of data which it is trained on at the back end. These all models have been able to scale it to huge amount of users because of its simplicity, because of the detailed information it is able to give to the users. That is why so many people around the world have started using AI tools extensively. Yeah, you can see the output is given out right here. Okay. Other than that, it not only gives us the output, it will give you additional information. Plus, it is going to ask that if I can explain it in different levels, it can do that. So it's trying to be as cooperative, as supportive as possible with the user so that the dependency factor increases, right? The ease of using it, you can see it yourself because of which it is able to go ahead and give us the output in such a customized manner. Now, if I follow up on this and I say, how do they build? How are they built? So what is happening here is as simple as when I say, how are they? Okay, it automatically connects it with the previous conversation. It understands that we are talking about LLM in the previous conversation, so it continues with that. LLMs are built through a combination of, okay? So the good part is you can have prolonged conversation with the AI, and it will keep a context of the conversation, the previous conversations, and based on which it is going to give you the output. So the answers, the responses are going to be way more customized to what you really want. Okay? So that is a superpower. There's again another big feature of these tools wherein it will remember the conversations you have had and based on which it will give us the output. There are also features you will see eventually wherein you can give your background, you can give your background, your work profession, everything you can feed in, and you are basically fine-tuning the LLM to give you output based on the background you have provided. Okay, so all of that is possible, but right now, what we're looking at is how these LLMs basically generally work. So this is how the ChatGPT is going to respond. In the same manner, you also have another AI platform which is going to be clawed. With Cloud, again, you can do the same thing. The strength of Cloud primarily is in coding. You can use this platform a lot for coding requirements usage which you may have. Let's take a simple example of how this is going to be. Let's say we are asking it to generate a Python code, which can run on AWS Lambda, reading a CSV file and saving it in another S three location. It is going to generate the code for us. It is going to also do troubleshooting of the code. All of that can happen right here. So a majority of the time users are seeing that cloud is really good with coding aspects and it can work really well with that. The use case is more tilted towards coding. This is how we can get the text. And this is where you understand the idea we had talked about with respect to GenAI that it is not just limited to generating text. You can generate code. You can generate images, videos, all of that can happen with GenAI tools. This is the Cloud part which we looked at in the same manner, we can make us Gemini as well, which is the AI tool from Google, primarily, and it is also improving day by day right now with tons of data at the back end of it. Let's take a different use case for this. Let's say we're using it for image generation. It is going to help us generate image. This is text to image generation which you can do with this AI tool. Okay? So there are different use cases. Now you can identify, you can imagine you can have with these tools where you can use it for generating content, for marketing purposes, for HR policy issues, documentation which you require. For all those scenarios, GenAI tools can be integrated extensively and can be used very effectively also. Here you can see this is how it has created the image and we can customize it as well. We can make changes to it. All that can be done. So my idea is to just letting you know the potential of these tools and the various ways or use cases you can have wherein you can use it just to give you a fair idea of overall all these tools which we have. So right now, majority of the time people use OpenAI CHN GPT for all purpose use cases, which you can do majorly. And Gemini also to a certain extent, Gemini still is picking up a lot right now and updated a lot. Okay, the output is becoming much, much better right now. Uh, Cloud is primarily focusing. It is also giving great output, but their strength lies in coding code generation, so it can be useful for that as well. CoPilot, as you know, is a part of Microsoft. The back end tech, which they have is OpenAI ChatGPT. But it's integrated with Microsoft products, which is going to be Word, Excel, PowerPoint, so you can easily use it in there, and it works really well out there as well. Other than that, there is growth and perplexity, which you can also use. These are again, other AI tools, um, which you can certainly try out and see whether it suits your style of work, your business as well. I hope this makes sense. I hope you understand now how we can use these different AI tools, their capabilities, their features in different scenarios. 6. Why Learn Generative AI: Hi, guys. Welcome to this session. In this session, we'll understand why we should be learning about genitive A. If you look at it, genitive AA is on the minds of every leader in the organization right now. Businesses, governments, and with interest comes opportunities. Organizations are specifically looking for people who understand the technology and most importantly, have the skills to apply it practically in day to day work. Now, unlike many of the previous trending technologies, genitive AI touches almost every role in every profession at this moment. Now because of which, genetive AI skills are expected to become more important in the coming future, not just for computer scientists, for everybody, which is why they will be essential as word processing, spreadsheets, even basic business literacy. Now there is a lot of new interest happening right now in AI and businesses are looking beyond customer AI, consumer AI. A chat booard interface is a great way to demonstrate generative AI potential. Now, real life use cases are embedding generative AI into existing processes and making it an integral function of nearly every single business workflow. The skills you will be gaining as part of these programs that should help you with your career and be very applicable to your job instantly. There are a lot of plus points with learning about genitive AI because this is going to be useful not only in your day to day professional work, but personally as well, you can use these AI tools to solve a lot of problems, questions, queries which you may have. The tools helps to get to the real solutions and gives practical steps as well. So you can instruct the tool in such a manner. You can prompt it in such a manner that gives you the outputs which you're actually looking. So it makes a lot of sense that we learn about generative AI, understand how to use these AI tools in different spheres of work. In this particular course, we're going to look at how it is going to help in our sales roles in sales profession. 7. Capabilities of Generative AI: Hi, guys. Welcome to this sessions. In this session, we'll talk about the capabilities of Generative AI. If you look at the capabilities which Generative AI has now, it goes on from text generation, image generation, audio generation, video generation, code generation now, data generation as well, and augmented capabilities also it has now got and also helping immersive virtual worlds creation also it is able to do. Now, if you look at specifically text creation capabilities, so there are various LLMs which are providing that, which are trained on large datasets and they can generate human like text. No, they are also able to learn patterns and structures from datasets and generate content and contextually relevant text messages, texts or responses, conversations, explanations, and summaries. Some of the examples of text generating capabilities can be coming from OpenAI, ChatGPT, and Google's Gemini. Now if you look at specifically image generation capabilities right now, the generative AI models leverage deep learning techniques like Gans, which is generative adversarial networks and variational auto encoders. With the help of these, they're able to generate AI images which are realistic textures, natural colors, fine grained details. Now, some of the examples of image generation are coming from Style gan, which produces high quality, high resolution novel images. Then there is deep art, which produces complex and detailed artwork sketch, from a sketch specifically. And then there is Dali. Dali produces novel images based on textual descriptions which we give it. Similarly, there is audio generation capabilities right now with generative AI, wherein it is able to generate musical compositions, text to speech, audio, synthetic voices, and natural sounding speech. Some of the examples can be Wave gan, which is producing raw audio waveforms, realistic sounds, speech, music, environmental noises. There is open AIs usenet, which is able to generate original music in various genres and instrumentations, and also can create classical compositions to pop songs as well. There is also Google's tachotron two, which is able to produce advanced DTS and can produce highly realistic synthetic speech, tone, pitch, modulation, pronunciation, rhythm, and expressions. There are a lot of capabilities of generative, which has happened over the past and it is continuously increasing right now at this moment. 8. Exploring the evolution of Generative AI: Hi, guys. Welcome to this sessions. In this session, we'll discuss the evolution of genetive VI over the years. If you look at it, genetiveEI started evolving parallel with the advancement of traditional AI. It remained dormant for over 20 years, but then it got propelled by GANs and VAEs specifically, and now it has poised to shape up the current future. So there was significant progress was made in creating content. So in the advancements of it, the early GenAI models had some issues with coherence and quality. Okay? So GPT three, GPT four, Dali, they delivered sophisticated text and images outputs and enhance the creativity and automation. Now if you look at the genitive capabilities, it acts as a creative genius. It can create images, write stories, invent new ideas for us. It is going to be based on a rule based mechanism. It's restricted systems to predefined context and rules. Now, machine learning and statistic models are used wherein it identifies patterns in datasets based on semi supervised, supervised or reinforcement learning. Now there are certain other things as well. The VAs over the period of time started learning patterns to generate similar outputs. Gans produce highly realistic images and art. Autoregressive models were used to generate content step by step, ideal for language modeling. Then deep learning and neural network came into picture which could detect patterns in data with advanced capabilities. It was able to handle unstructured formative data as well. Then the GAS, which is generative adversarial networks, marked the beginning of new era of AI tools where it could create new datasets. Then also there was LSTM and RNNs which were used, which would offer advanced capabilities, handled unstructured data, and could process time series data. Now, if you look at the difference between Generative AI and traditional AI, traditional AI analyses or predicts using existing data. Common task can be classification, urigreon recommendation. Whereas Generative AI uses GAS and transformer models, it is able to create new data that resembles the trading data. Now if you look at artificial intelligence or traditional AI, it evolved from basic to predictive order level, whereas Generative AI creates human quality outputs using AI techniques. So if you see since 2017, a new era of generative tasks have evolved, leveraging open source GPT models. It has utilized pre trained models for large datasets and fine-tuning models for specific tasks. So overall, if you see the main difference, traditional AI follows specific instructions, whereas Generative AI invents and creates on its own. 9. Applications of Generative AI: Hi, guys. Welcome to the sessions. In this session, we'll talk about the application of Generative AI in different sectors of work. The first one we are going to look at application of Generative AI in IT and DevOps. So here, it really improves the software delivery processes and infrastructure management. The code generation capabilities of Generative AI reduces manual coding efforts and time spent on repetitive tasks. For example, GitHub CoPilot and SNIC Deep code helps to do code repositories. It can examine that, I can examine coding standards. It also helps to generate synthetic test cases and test data. Wherein you can simulate user behavior, impact, software efficiency, reliability, and robustness. There is also tools like APLA tools and testing, which can guarantee adequate testing coverage, increasing the depth and diversity of datasets. Also, apart from this, you can monitor and detect anomalies like IBMs, Watson AIOps and Mok soft AIOps. It can analyze system logs, metrics, and other data like proactive maintenance. It can help in lessening the downtime and also preventing critical failures. Now if you look at the application of Generative AI in entertainment, in art and creativity, it can help to generate synthetic content like music, scripts, stories, videos, movies, video games. In game development, there is Houdini by side effects, which can create games, animations, AR and VR experiences, unique characters with unique behavior. Other than that, there's also virtual influencers and avatars, which has come over the period of time, which are able to interact with users and create engaging experiences. Then there is application of generative AI in education like content generation, personalized and adaptive learning experiences, simulated experiential learning, all that can happen now. It can help to provide language translation like making content accessible to different people, grading assignments, providing instant feedback, creating learning journeys and assessment strategies to support learners pace and strengths, generate taxonomies which can be learners performance and preferences. Other than that, generative algorithms are also used in education to detect special needs and learning disabilities, create specific lesson plans, track learners progress over time. You can also do knowledge tracing wherein write pacing and content for individual needs can be done. Tutoring support can be provided. Virtual and simulated environments can be created. Inclusive education can be done. The example, tools which are null J. It's an AI generated E learning, which can be done in minutes for the targeted topic, which can be interactive videos, glossaries, summaries, all that can be done with the tool. Hope this makes sense. I hope you will to understand the various application of generative AI in different sectors of work. 10. Tools for Text Generation: Hi, guys. Welcome to this sessions. In this session, we'll look at various tools which we can use for text generation in LLMs. If you look at it, large language models are based on patterns and structures learned during training. These LLMs interpret context, grammar, and semantics to generate coherent and contextually appropriate text. Drawing statistical relationships between words and phrases allows these LLMs to adapt creative writing styles for any given context. LLMs are the basis of many text generation models. Two such examples are generative pre trained transformer or GPT and Gemini AI model. The models have evolved into multimodal models offering multiple capabilities. Let's learn about the capabilities of these models through two popular tools right now, which is SATGPT and Google Gemini. If you look at ATGPTs based on a GPT as a large language model and uses advanced natural language processing or NLP, which we call it. Well originally HGPT only took text prompts as input to generate new contents, with the newer version, it can take both image and text inputs now. ChaGPT offers diverse capabilities for text generation. It is also capable of smooth and context based conversations. Now, in the same manner, if you look at Google Gemini is powered by Google's Gemini AI model. It introduces a new family of multi model AI models and it enhances reasoning, understanding, and generation. It also ensures efficiency and scalability and optimizes seamless multimodal interaction. It also is able to handle diverse data and task. Let's see some practical example of how this is going to be. This is going to be the Cha GPT interface where we can come and let's give a general prompt wherein I'm saying that I have heard about generative I and want to learn more. She's going to give me a lot of context about what is generative AI. How does it work? LLMs. She's going to give us a lot of related information, which is quite informative and provides the right information about. Now, furthermore, I can dig deeper where I can say that how I can use native AI to specifically improve my storytelling skills. So now I want to divert it into a specific category requirement, which is storytelling skills. So now it's going to give me ideas around that develop deeper characters, improve dialogue writing, use AI to brainstorm better story ideas. Okay, so it's giving me some practical inputs which I can really use to improve my storytelling skills. Same way, I can also ask it a separate thing. Let's say I'm asking you to help me with creating slides to demonstrate the features of a learning platform. Let's say I want to create certain sales slides. So it's going to give me the structure is really good where it breaks down into slides, title, subtitle, include, and then the problem we solve. The focus is given on context is given, which is for learning platform. So it's giving me all the necessary points for that. This is how we can make it useful. Another great usage is you can use it for learning languages. All that is possible, so you can convert any English language to any other language which you want, and Chachi P can easily do that for us. Same way, let's look at Google Gemini, which you can also make use of where you can give a prompt. Let's say, I'm asking you to provide a summary on the latest news on the war in Ukraine. So it's going to give me all the information related to that. You can see over here all the information, the latest information which we can get. Similarly, if I wanted to build a strategy around making a digital marketing campaign for a fashion brand, so it can help me with that also. So now we're asking it to provide a digital marketing strategy. So immersive and AI driven experiences content strategy, authenticity or aesthetic, okay, social commerce and community. So you can see it's giving me some specific strategies around digital marketing, which I can use practically to promote a particular brand. So this is how we are going to make use of both the tools specifically speaking. And then if you look further, so by using CHAPT and Gemini, it has a lot of benefits. Like, it provides problem solving through basic mathematics and statistics, financial analysis, it can do investment research, budgeting, all that it can do. It can also help you with code generation. Now, if you compare CHATPT with Gemini, CHAPT is effective in generating dynamic responses and conversational flow is there in its response. Whereas Gemini is good, optimal for research work, research in current news, information which you want on a particular topic for all that purposes. There are other text generator tools as well, which you can absolutely use, for example, Jasper, which is useful for creating marketing content for a specific brand. You can also use writer as a AI tool, which creates content for blogs, emails, SEO, metadata and also ads on social media. There is also copy.ai, which creates content on social media for marketing and for product descriptions. There is also write Sonic, which helps provide specific templates for different types of text. There is resumer you classify as well for generating text summarization, text classification. There's also brand 24, which you can use for sentiment analysis, and then there is Weaver and Yandex, which we can use for language translation. That is how text is going to be text generation is going to be, which we can see over here, which you can absolutely use on all these AITunes. 11. Tools for Image Generation: Hi, guys. Welcome to the sessions. In this session, we'll look at different types of tools which we can use for image generation. Imagination models are basically ones where we can generate new images, it can customize real and generated images. For example, let's say we want to generate an image of a child with a book and then change the book cover in a generative image. All that can be done by image generation models. Now there are various types of it. One is image to image translation. You are transforming an image from one domain to another. Example, this can be useful for converting sketches to realistic images, converting satellite images to maps, converting security camera images to higher resolution images, enhancing detail in medical imaging. Now, other tools are going to be style transfer and fusion. These are useful for extracting the style from one image and applying it on another. Example can be converting a painting to a photograph. Then there is in painting. In painting is we're filling in the missing parts of the image. You have an image and there are some parts which are missing, so those can be AI generated. Example, art restoration, forensics, removal of unwanted image objects and images, blend virtual objects into real world scenes. Then there is out painting. Opainting is extend an image beyond its borders. Example can be generating larger images, enhancing the resolution, creating panoramic views. All that can be done. So now from Open AI, there is Dali which is based on GPT, which can do all of this, can generate high resolution images in multiple styles. It can also create new versions, can be generated can generate multiple image variations can be done. It uses in painting out painting features as well. Then there is stable diffusion. This is an open source model which can create high resolution images. It can generate images based on text prompts. It is used for image to image translation in painting and out painting. Then there is style gan, which enables precise control for manipulating specific features, separates image content and image style. I evolved to generate higher resolution images. There are other tools as well like crayon, free pick and Pick Start, which are also available to generate images in different forms. There is Photo and Depart effects as well, which offers various pre trained styles. It allows custom styles as well. Then there is depart dot IO, which is an online platform that turns photos into artwork. And then there is Mid journey as a platform, which enables image generation which enables image generation communities where artists and designers create images using AI. It also enables exploring each other's creations. Let's look at one of these tools, which is going to be free pick. This is the website where we can come to free pick and we can generate an image here. Let's say we are giving it a simple prompt right now with this prompt, it's going to be text to image generation, which we are trying to do here. So now you can see it has gone ahead and generated that image for us, a boat sailing on a calm lake at sunset, surrounded by lush green trees and misty shoreline in this particular way. I hope this makes sense. I hope people understand now the various tools that are available now for image generation with the help of these AI tools. 12. Tools for Audio and Video Generation: Hi, guys. Welcome to this session. In this session, we'll talk about the tools which we can use for audio and video generation. So in this generative AI, audio capabilities help companies and individuals, novice or experience to simplify processes, bring complicated visions to life. Now speech generation tools are available here, which can be text to speech tools which are trained into deep learning algorithms, vast datasets of human speech. Now, it can break down and replicate pronunciation, speed, emotion, intonation, as well, and there more accurate and natural sounding speech helps those with visual impairment, language barriers, reading disabilities. There are music creation tools which you can use to write short melodies or riffs, suggest or add instruments, compose a new song, create a soundtrack for YouTube or Instagram videos, mix match. You can mix and master and publish streaming platforms. Then there are audio enhancement tools as well, which can identify specific sounds, add or remove unwanted sounds like, for example, DScript or Audo AI. There is also going to be video tools, video generation tools which you can use like runway, which can transform video into new styles. It uses text, image or video as input. Now, there is also Es US, where you can upload photos or use text prompts to generate videos. Then these video tools can record a narration, enhance the audio, convert the file format. They can publish a video as well, and there is tools like Synthesia which can create custom Avatars. There are a lot of different audio and video generation models which you can use and tools which you can use for generating AIs generated videos and audio. 13. Tools for Code Generation: Hi, guys. Welcome to this session. In this session, we'll talk about various tools which we can use for code generation. So code generation models generate code based on national language input. Based on deep learning and NLT, these models comprehend context and produce contextually appropriate code. Now, the capabilities of these code generators are that they can generate a new code snippet or a program. It can predict code lines to complete partial code. They can produce optimized versions of existing code. They can convert code from one programming language to another. They can generate summaries and comments for code. They can also recommend programming solutions to solve a specific problem. Similarly, in this open AIs GPT as a coding generation model, Excels in human like text generation, it demonstrates immersive code generation capability. These coding capability of GPT are longer and more accurate codes can be generated. Coding can be done to develop apps, websites or plugins can generate code for images. So if you look at, for example, when we go on Chat GPT specifically and we write, let's say, write a Python code to generate a message to greet a person, so we can get a code like this, which it provides. Plus, it gives you the explanation of how it works specifically. Also, you can convert the same code into another language as well in this particular manner. Now, with respect to looking at coding with Gemini, it offers code generation in more than 20 programming languages. It provides step by step and detailed understanding of how to generate the code. There are certain limitations of Cha PTI and Gemini for coding as well where it cannot generate large or complex codes. I can I can understand programming and syntax, but not semantics. So their knowledge is limited to the data used for their training. Like, for example, they get outdated with new releases of frameworks and libraries. For example, knowledge of GPT 3.5 is limited up to September 2021. So therefore, other tools like GitHub co pilot can be used, which can generate code for various programming languages and frameworks. It is powered by OpenAI's Codex and develops solution based code. It is trained on natural language, text and source code. It can integrate with other code editors can produce code adhering to best practices and industry standards. There are other tools like poly coder also which we can use, which is an open source AI code generator based on GPT. It is trained on Github repositories, written in 12 programming languages and provides a library of predefined templates. It can create review and refine code snippets. Other than this, there is IBM code assistant as well, which is built on IBM watson.ai Foundation models. It can be integrated with code editors. It produces real time recommendations, auto complete features, and code restructuring. So these are all the various tools which we can use for code generation at this moment. 14. Generative Versus Agentic AI: Hi, guys. Welcome to this session. In this session, we wanted to understand the difference between generative AI and agentic AI. When we look at generative AI, they are fundamentally reactive systems. They wait for you to do something. Specifically, they wait for you to prompt them. And once you prom them, their job is to generate some kind of content based on what you have prompted, the prompt which you have provided. Now they are using patterns they learn during training. Right? So now things that it can generate, might be some text, it might be an image or it can be a piece of code, it can be an audio. So they have learned the statistical relationships between words and between pixels and between sound waves. And they have learned that from massive datasets. So when you provide a prompt, a generative AI predicts what should come next based on its training. But it works work does end at generation. So ideally, their work ends at generation. It doesn't take for the steps without any more inputs from your side. So it's heavily dependent on what kind of prompt are you going to give to it based on which it takes those necessary action. Whereas when we look at agentic AI, agentic AI systems, these are not reactive. They are proactive systems. Now, like a genetic AI, they often start with a user prompt, but that prompt is then used to pursue goals through a series of actions. And an agentic system basically goes through a bit of a life cycle. So the way this works is it kind of first of all, perceives its environment if you like. And once it's done that, it can decide an action to take. Once you decided that action, it can then execute that action. And then once that action has been executed, it can learn from that output and then go round and round all with minimal human intervention. Now, both of these AI approaches often share a common foundation. And that common foundation is the large language models or LLMs, which we call it. LLMs serve as the backbone for the chatbots, and yet there's actually other tools that are used for some of these generative things, diffusion models typically for images and audio. I hope this makes sense now. I hope you're able to understand the basic difference between how a generative AI operates versus the agentic AI. 15. Introduction to Key Terms: Hi, guys. Welcome to this module. In this module we're going to understand some of the key terminologies which you're going to see a lot in generative AI. These are going to be some terms which are going to be very common and widely used when we talk about AI technologies, which can be LLM, prompt engineering, embedding, fine tuning, rag chat boards, and lately agentic AI. Let's begin this module where we'll go through each of these terms in detail to understand simply what they really mean and how they contribute in this AI tools technology which we are using on a regular basis. 16. LLM (Large Language Model): Aye. Welcome to this sessions. In this session, we'll talk about LLM, large language models. So what are LLMs, basically? So what we want to understand over here is how we can make use of LM. LM is actually going to be a large language models, which we can use to have a conversation with. Like, you can see an example over here, which is related to mobile chat, which is happening on a mobile device. So here, also, a lot of AI technology is already incorporated. Like over here, as you can see, it says, I am going to the and then gives multiple options, gym, park, or store. So the LM predicts the next word automatic. So that is the capability of an LLM where it can do a next word prediction. Same manner is going to be with Chat GPT answering a question, wherein when you give it a question, it is able to give us the output and it is able to research about it and give us output in the same fashion. So that's the basic idea of what is an LLM and how it is different from generative AI. So we'll talk about that as well. So now if you look at it, LLM is primarily AI trained to understand and generate human language. So it's focusing on textual output given in a conversational manner. That is an LLM which we talk about, and it is going to do one thing which is predicting what word comes next. Okay. So based on what has been inputted earlier, it develops the ability to reason, explain, translate, and summarize, and then write. That's the idea of what an LLM basically does. Now if you see the major difference between GNAI versus LLMs, is going to be the case that with GNAI you can create new content in different formats. There can be image generation, music generation, video, code generation. All of that falls under generative AI. But when we look at LLMs, these are text based models where it is generating translation, summarization, classification, sentiment analysis, named entity recognition, all these can happen. In these overlap is it both are going to be chatbards as well. You can do text writing here and Q&A. So you can say that all LLMs are a type of generative AI, which we have in place. Okay? So with LLM, it is going to be text or tokens only, and it is text based output, what it gives, and a number of models which we have right now is GPT five, Cloud Gemini, language focused. It's more focused on language. That is what is LL. Okay. Now, how it primarily works. This is the workflow wherein it is trained on huge amount of data coming from the Internet, read it, which are books, websites, articles, papers, conversations, code repositories. All these are it is trained on and then it goes into the neural network. It goes into the neural network, which is the transformer and in which it gets processed, basically, and then it is decoded. It is decoded into an output which we get on the platform. So that's the idea how the LLMs are going to work. It's based on the training data given and then the transformer neural network which works on it and gives us an output which is generated based on the training. Now, the key concepts behind LLMs is that it is based on three major things, which is there has to be some level of pre training. So the pre training, as we understand it is based on huge corpus of text data, trillions of words from books, web and code, large language patterns like grammar, facts, reasoning, takes weeks of thousands of GPUs to process all of this. Then the size and scale. So because there is massive usage of neural networks with billions of params, okay, when it cost a huge amount of money to train from scratch and based on which the scaling also happens. And then there's a fine tuning. Fine tuning is primarily targeted towards training the tool to give to do specific tasks or give specific kind of output. Okay? So here, you can fine tune your LLMs, to give us a specific outputs based on our requirements. So these three things play a critical role in running our LLMs. And now you can see there can be various use cases for LM where you can use it. You can use it in content generation, which is for writing blog posts, product descriptions, marketing copies. You can use it for working as chat booards or virtual assistant as catering to a lot of customer support, booking assistance, personal AI co pilots, then language translation. So you can do translations in various ways. Google, like the way we have DeepL or Google Translate, it can do that as well. Then you have text summarization. If you have condensed long reports, complicated information, very technical jargons which you get to deal with on a day to day basis, those can be simplified through LLMs, and then question answering, which is primarily driving the search assistance, educational tutors, enterprise knowledge bases, FAQ systems which we deal with on a daily basis, all those it can handle as well. I hope this makes sense. I hope you to understand now what are LLMs and what are their use cases and in different scenarios, how you can make use of them and whats at the back end of it. Thank you so much guys for listening to this, and I will see you in the next video. 17. Demo ChatGPT: Next-word completion and text generation: Hi, guys. Welcome to this session. In this session, we'll see a demo in Chat GPT, which is primarily to understand the capability of the AI tool to generate the next word based on the context given before. Let's have a look at this, how we are going to do this. What we want to specifically see the power of generating text, the next word possibly with the help of the data which has been given and inputed over here. Let's give it a simple prom. Let's say I'm asking you to give it a prompt where we say, once upon a time a cat wanted to. And we wanted to generate some content, wanted to become the greatest chef in the village and the information is provided. You can see based on the context which we gave, it was able to generate the next word and the next word further. Now in this, let's say we further continue I can furthermore continue based on the content it has generated in the past and it's able to do so. So now, let's say, what we want to do now is a different use case where we want it to now write an email specifically let's to a colleague based on some context given. So we are saying that write a WhatsApp message to my colleague, John, asking him to share the project report by 10:00 P.M. Today. Okay? Make it polite. So now I've given some context. So it's looking at the context is going to create the WhatsApp message which has been created. Now we also have the option over here where we can transform this particular messaging in a different tone. So let's say John is a dear friend, so we want to rewrite the message in a friendly manner and a funny tone which we want to add to it. So you see what is happening here is that we are able to use the AI tool to generate the next word as seen over here. Plus we are able to make changes to the content generated based on some context given, and that is the superpower of the tool where it can generate new content based on the context provided to it. I hope this makes sense. You are able to understand these different capabilities of the AI tool which we have in. 18. Embeddings: Hi, guys. Welcome to this session. In this session, we'll talk about a very interesting concept, which is embeddings. So as we understand by now that how the AI tools primarily works. But one important information which we haven't spoken about is going to be the fact that these machines do not understand text. Then how are they able to generate text, right? They only understand numbers. So this is where embedding comes into picture, which is primarily a way of a numerical representation of the text. So this becomes super essential for the AI models to understand and work with human languages effectively. So what is going to happen is, let's take an example to understand how it actually works. Let's say the sentence is I eat ice cream. Okay? Now, this is the tool does not understand the textual understanding of what does that mean? So what at the back end, what happens is, this is broken down into, you can say tokens. Okay. So each of these are tokens which is going to be four tokens, I eat ice and cream. Okay. That is what gets entered, and each of these would have a specific number associated with it. Which is where the neural network or transformer which we have, which does its computation and generates these numbers which are then given to the tool. So when this is feeded in, that is when it understands that it is meaning I eat ice cream. That is how embeddings primarily help the AI tools to understand human language and based on which it is able to generate the next word. It understands the context, it understands the background given to it, the information given to it, and based on which it then gives us the output, it gives us the response, it gives us the next words which are needed. This is very crucial for us to understand because this is where you get to understand that these AI tools primarily works on numbers and it it identifies language in the context of numbers and based on which it gives its outputs. 19. Fine Tuning: Hi, guys. Welcome to this session. In this session, we'll talk about another interesting concept which is fine-tuning. Fine-tuning an LLM, is going to be a process where we are trying to adapt to a pre trained model to perform certain specific tasks which caters to our domain, or work. Now, usually with the AI tools, what we are trying to do here is we're trying to find out solutions, but it can be a scenario wherein you need a solution for a specific scenario. Let's say you come from a retail sector or let's say technology sector, or let's say automobile sector. And what you require is a specific solution for a scenario you're facing in your specific domain. And that is where fine-tuning and LLM becomes crucial. Now, there are three ways in which you can fine tune LLMs. The first is going to be self supervised, which is primarily a way wherein you provide all the information which you have related to your domain to the foundational model. It's the complete training data which you give to the foundation model. It analyses that data and based on which it learns from it, understands your domain, your expertise, and it understands your pain areas and based on which it now gives you customized output. So now here it is self supervised, wherein it is able to do it itself because of the training data which I provided from your domain, customized to your domain. The other model which we can use here can be supervised. Now here what is happening is you are going to give a specific detailed labeled training data that has an input and output. You give the input and you give an expected output and based on which it learns from it and understands your scenario, your domain expertise, and based on which gives you the output. Just to give you a simple example, an input can be as generic as let's say how to find a broken bone. The output is X. So now you give this as example, labeled training data to the tool and it understands that, okay, I have to give the output in this particular manner. If the input is given in such a way, I have to give an output in this specific styling, and that is the model learns and gives you output based on its learning. The third one which we can use here is reinforcement. Reinforcement is a model where you give an output based on its training and we score it. So now what is happening is the model gives you an output. You give a specific input, you ask for a particular solution. It gives you the output, but now what you do is you score it, and now from the scoring, if the output is bad, you give it a low score, if its output is good, it can be a high score, and now model starts learning from it. It understands that if it was given a low score, it tries to understand where it made the mistake. And based on which of this learning, it improves its outputs in the future proms which you give it. That becomes reinforcement. There can be three different ways wherein you can start fine-tuning your LLM and get better output from it. Now let's look at another scenario wherein you should not be fine-tuning. In which scenarios fine-tuning should not be happening. The first aspect is that fine-tuning is not about creating an intelligence from scratch. What we mean by that is it is not about generating information from the beginning. What we're trying to do here is we're trying to fine tune the LLM to train the LLM to get customized output. Okay? We are not trying to generate intelligence from scratch. Second is here, you saw in the three scenarios, in each of the scenarios, we had to provide the training data. Okay? So in no way, we are going to eliminate the data requirement here, okay? That is not fine-tuning. Third is going to be the case that fine-tuning will not always give us the we output. Every interaction which you're going to have with the AI tool is going to be unique and based on which the outputs are going to be unique. There is going to be a single universal there will never be a single universal solution which you'll get out of it. Then lastly, as you saw, this is going to be a process. It is not going to be a magical one time process, but continuous iterations are needed to reach a specific desired output. It is not going to be a case wherein you give one single prompt and you get the desired output in the first go itself. Continuous iterations are needed so that you fine tune your proms and based on which you get the desired results. I hope this makes sense. I hope you understand now the concept of fine-tuning and why it is necessary when you're trying to use these LLM models to its fullest. 20. Recap - Summary view: Mmm. Hi, guys. Welcome to the sessions. We just wanted to do a quick summary of all the things we have spoken about so far. A lot of things which we covered till now. So we just wanted to do a quick summary of that. So now we understand this is how the whole thing works where a user comes and does a query, right? This query goes into the large language models, which can be, let's say, ChatGPT or Google Gemini or meta or any other AI tools, LLM models which we have. Now, this is primarily going to be the neural network, the transformers, which are going to look at the training data provided and use embeddings to based on which it is going to give us the output, which is understandable human language, and we get the output. Now here, we also looked at the fine-tuning part, which is primarily where you fine tune the LLM models to get a specific desired output based on your customized specific domain. Now that is where you improve, you train the models to give you a specific kind of output. Also, we talked about the fact that there is prompt engineering involved here, which is primarily the way you prompt these AI tools, we also decide the kind of output you get out of it. We need to also look at what are we prompting, how we are prompting. What are we asking vague questions or specific questions which is going to decide the kind of output which we get out to field. I hope this makes sense so far what we have been covering till now and in the next videos, we're going to see furthermore different ways of using genitive. 21. Retrieval Augmented Generation (RAG): Hi, guys. Welcome to this session. In this session, we'll talk about RAG or retrievable augmented generation. So now that we understand how LLMs work. So LLMs workflow is like this where a user comes and asks a question, which is a query, which is tokenized, embedded, and it goes into the LLM, where the neural network works and embeddings are done and based on which the output comes out. So this is the workflow which we know of how LLMs work. Now in this, the problem is that there is a knowledge cutoff, wherein we know that these LLMs are trained up to certain knowledge, which is still 2023, and beyond which they don't have the training data available based on which it can give us the output. And also, what is going to happen is there is no private data. Our internal documents, SOPs, customer data never appear in the training corpus. So they don't have access to our personal private data as well. So because of which, there is a roadblock. There is a roadblock of how we are going to get a customized solution. And that is where RAG comes into picture. RAG is primarily retrievable augmented generation, wherein retrievable means wherein we can find the data because we are going to provide our own knowledge base to the LLM, and it is going to only look into that. It is not going to go anywhere outside of it. It's going to retrieve the data from the knowledge base which we provide enrich that prompt, it will going to augment it out into the prompt and export it out for us and generate it like a conversational manner. So that is how RAG is going to operate. Wherein it is going to access our specific knowledge base and it's going to process that it is going to match all the information and then give us the desired output in the understandable manner. That's the basic understanding of RAG or retrievable augmented generation. Now if you look at how it helps is because it is now the pipeline is going to be there are four major things. One is authoritative sources because the data is ours. So we can certainly go ahead and validate that. We can vouch for it. We know that the information is going to be correct. It is going to be under total privacy because it is not going to look at any other sources of information. It is not going to look at the Internet, so the data is completely secured with us. And then there is obviously going to be reduced hallucination. It is going to give us a solution based on the information which we have provided. So the outputs are going to be much more relatable, realistic, which we can certainly apply. And then the source is included, which is going to be the case where it is also going to do citation. It is going to refer to specific documents, line items and give us that in writing so that we know that from where it is collecting that information and giving us the out. So if you look at it, so that way, a RAG is far more better, and this is how the structuring is going to be. So if you think about it, your documents are going to be the source, which is, let's say, these PDFs, which are being uploaded onto the LLM model and in chunks. Okay, so in chunks. And then it gets transformed into embeddings which are going to be random numbers in this particular manner, which gets stored in the LLM at the back end. So now this is a database which has been created a vector DV is there in which all your company documents policies are being stored. Now when a user comes and does a query, and similar embeddings comes from here, which are now, we try to find a match with the embeddings which are available in the vector store. Once a proper match is found out, the results are given out, which are being shared with the LLM. Now LLM goes ahead and transforms that into an understandable language and gives it back to the user. That is how RAG is going to work. Important thing to remember over here is the results which the LLM will give will be completely based on the documentation provided by you. It is not going to be something outside of it, and that is what we really want. We really want the solutions customized to our domain, our field, our need specifically, and that is exactly what RAG fixes and helps as well. I hope this makes sense. I have to understand now how the RAG concept works and how it solves a lot of issues which we face with the AI modes. 22. Agentic AI: Hi, guys. Welcome to this session. In this session, we'll talk about the concept of agentic AI. So this is a new upcoming topic which is happening a lot around us now, and a lot of people, after understanding Gen AI prompt engineering, now are talking a lot about agentic AI. So let's try to understand in a simple way what exactly is agentic AI and how we can make use of it. So let's take a simple example. So let's say in a real world scenario, you are planning a trip for yourself and your family to, let's say, New York. Okay. So in a traditional manner, how you're going to do it is you're going to start looking at what flights to take, which hotels you would like to book, places you will visit in New York, other things, attractions, wherever you want to go. So all that we will plan, and full manual itinerary which you will create, food you're going to eat, where all you're going to eat. So all that, you'll have to do it in a step by step manner. Now, this whole thing, which we are trying to do ourselves, can be outsourced to someone who can plan it out for us, right? Person who can book our flights, hotels, who can book the places we want to visit in that city, okay, the places we want to eat. All that can be, uh, planned out by a particular person, which is mostly in a real life scenario going to be travel agents. Imagine the same thing happening with agentic AI. So an agentic AI is primarily an autonomous flow or a workflow, wherein an AI system is built out where this AI system will do the whole job for us on behalf of us. So where it is given a specific goal, and based on that goal, it is going to plan, it is going to use certain tools and take some decisions to fulfill that particular goal. And this is going to be completely autonomous, which means that it is going to work on its own and would not require much of human intervention. That is simply what we mean by agentic AI, just to give you a workflow of how this is going to really work. So if you look in simple terms, agentic AI refers to these are AI systems that can autonomously plan out things, take decisions, and execute certain actions to achieve specific goals given by us without having to be having continuous human intervention in it. Okay. Now, this can be implemented in multiple different workflows in our work in various ways. Maybe you can have a agentic AI which takes care completely of your customer support queries. You might have an agentic AI which looks at escalations specifically. You can imagine how you can make use of this in a real world scenario, which can take the burden of your work on itself and then does the job and gives you the desired. To give you a better idea of what really happens at the back end. So here you can see this is how it is going to work. So the user is us, wherein we give a specific input or a goal. Now, based on the goal given to it, the agentic AI is going to plan, start planning how to start, do the whole planning of it. Then it is going to take certain actions specifically, okay. And it will have some memory, which is basically the database. So the details like you will provide your documentation, you will provide your knowledge base. You're going to give all the resources of your company. So that's the memory of it based on which it is going to take those decisions. It is going to fulfill that particular goal. Then it is going to use certain tools. It will be linked to let's say Internet, I can code as well, which is needed in that goal completion process. It is going to connect through API. All that will happen and then it will try to fulfill the goal. If the goal is not met, it will iterate. It will again go back to the same loop. Again, planning will change. It will take some different actions. It will use more details from your knowledge base and then try to fulfill the goal. The heart of all of this is your LLM. The LLM is working, which is actually doing the whole job for us because of its model training and the tools it has for. This is how it is going to happen, what is what we mean by agentic AI. These are the AI agents who are going to help us do our certain jobs which possibly can be automated and it takes away a lot of our time. I hope this makes sense. I have to understand now the simple concept of agentic AI and what it can do for us. 23. Projects - ChatGPT: Hi, guys. Welcome to this session. So in this session, we'll see another capability of ChatGPT, which is going to be projects. Projects are primarily a feature where you can go ahead and customize the GPT for a specific requirement. Let's say you want a specific kind of GPT, which is only going to cater to one specific issue you have at work in your company or for a specific reason. That is where we can make use of projects. Let's see a practical example of how we can make use of it. So once you're on ChatGPT, you can go on the left panel where we can create a new project. So let's take an example. Let's say I want to create a project around math tutor who caters to high school kids. I want to create a project where in this particular project can answer queries for math related queries of high school kids. So this is going to be a math tutor And then what we have to do here is we have to give a background. We have to give information, background about to the tool, which is basically fine-tuning the LLM to act as a math tutor and response based on that. That you can do in multiple ways. One is sources. Where you can come, you can add sources. Now these can be documents. It can be a math book. It can be a documentation which talks about the profile of the math tutor, all those things which we can upload over here and keeping this into context, the project is going to work. It's going to respond to the queries. That is one way. The other thing which you can do here is we can go to the project settings. In project settings, you can give the background. We can give the background of the math tutor. Let's say we are doing that here. So now the background has been added over here. So this is a math tutor for high school kids. Okay, they're going to answer queries or math concepts specifically. Okay, we have also given some restrictions that restrict your tutoring to high school math topics only. And if you're not qualified for it, you can just simply say you're not qualified to answer a question outside of this topic. So we have created some guardrails around it as well, and now you can save that. So now, based on this context, the project is going to give output and give answers responses to the queries. Okay? So we're going to start off with that. Okay? So let's start with a simple question which is what is a complex number? So it is going to behave like a math tutor for high school kids and give answers appropriately. So you can see now it has given us step by step answer a simple example also given, okay, types of complex numbers. All that information is provided right here. It also gives a mini quiz, okay, which is also evident because as a tuitor, you want to continuously gauge the understanding of your students by quizzes through quizzes. So that is what it has and then it's additionally giving you information about what all it can also explain, other topics. So we can do that. Let's look at some other examples as well. Let's say simple asking it to explain Pythagoras theorem using simple algebra. Is going to explain Pythagoras theorem for us now, simple triangle example taken based on which it gives us the output. Now it's behaving like a simple math tutor, giving simple examples, visual intuition. So now it's giving us some visual examples also to understand the concept. Real life uses, you can see, and then a quick quiz again. Lastly, let's look at another one. Which can be explaining why the square root of two is irrational, is an irrational number. The goal is to prove that this is irrational and then what does irrational mean? The strategy behind it, then it gives the whole output. I hope you understand now the use case of projects. What it can do is it can be a separate GPT inside your main GPT, which you have, which is catering to a specific requirement which you may have. Now I can create a GPT for answering my customer's queries. I can create a project customized for, let's say, technical issues which are happening in my software. Okay, so there can be separate GPTs created for HR related issues. So like this, multiple different types of projects or GPTs you can create, which are catering to specific problems, which it will be expert in handling once you provide it the context, the background, the resources behind it. I hope this makes sense. I have to understand now how projects can be used in changing. 24. Limitations of LLMs and workarounds: Hi, guys. Welcome to the sessions. In this session, we'll talk about the limitations which we see right now with LLMs and some workarounds around it. So the first limitation which we get to understand over here is the knowledge cutoff. As you understand, the LLM is trained for a specific amount of information, a poor specific time frame. So right now, as you can see, the training cutoff is 2023 for which the information it is trained for. It has been trained beyond that. Okay? So that is one of the limitations which you have of LLMs, and it's working on that, there are a lot of improvements which are happening at this moment. But that is the major chunk of knowledge which it has at this moment, which is still 2023. The workaround which has been identified is the RAG, which is retrievable augmented generation, which is primarily where you inject live documents when you are inquiring about the output. So at the query time, the tools calling lets LLMs to run web searches. That very moment, the LLMs actually does a lot of web searches and based on which it fine tunes the solution and gives out the out. Now, the other problems which LLMs face are going to be around hallucination. Hallucination is a lot of times whenever the output is given they don't know the facts, then a lot of times the tools tend to hallucination, and that is why it is being suggested that we need to be very much specific with the prompts and restrict the tool from hallucinating. That is one of the major issues which happens with LLMs. Other thing is offensive outputs can come out sometimes because it is trained on a lot of data which has this kind of information, it tends to pick it up from there and gives out that kind of an output. It can cause harm as well, can be weaponized for disinformation Okay. However, things are becoming much better, for example, GPT four versus, uh, GPT 3.5 has been much better in terms of hallucination in all these scenarios, and the other tools are also becoming better as the new models are coming into picture. Now, there are a lot of bias, also, you will see toxicity you will see in LLMs in terms of hiring bias outputs, which we get to see gender bias, cultural bias. So like in hiring bias, list traits of a good CEO. Okay, translate the nurse called. Okay, so these are all going to be the feedback which we are getting with respect to LLMs, where there are a lot of biasness which is there, and that is why processes need to be put into place, which is going to reduce it. And over the period of time, we can get it eliminated. So this is again, happening because of the skewed training data which we have. The data which is coming from the Internet is possibly skewed and because of which the output is like that, and this requires a feedback loop amplification, which is basically RLHF which we need wherein you use human raters. So there has to be a human raters who are going to primarily vet the output and based on which the output needs to be released. So we'll talk about this a little bit more in the coming topic. Other than this, historical biases are also there in the data. Okay, historically, data is sometimes more inclined towards certain topics, and that is what the tools are trained on, and that is why the output is like that. So there is a lot of mitigation approaches are being made wherein we are trying to make an unbiased training being done of these tools so that the results are much better. Now there are scenarios when LLMs behave badly in terms of the major takeaways are going to be real instances which have happened. Key takeaways are jail breaks work by role playing tricks that override safety trainings. Models are now hardened against known jail break patterns, so it is becoming much better now. There is prompt injection is a serious enterprise risk when LLMs process untrusted data, as part of the agentic workflow. There is also T from 2016 which showed that the AI exposed to adversarial public input will absorb that behavior. Pattern also is becoming like this where early releases, safety issue exposed, emergency patches, then RLchre based long term fix. All these are going to be the way how things can become much better with the LLMs. Today's modules like GPT five, 5.2, which we have are significantly much more robust and are giving much better output than the previous models. Now, if you look at RLHF which we were talking about, this is a workflow. How it works is the LLMs are pre trained with training data and generate responses based on that. Once a response is generated, then human labelers can come in which ranks the outputs from best to worst. Based on which the output, the result, it is the reward model works. You reward the model, based on the output it gives and then that feedback goes back again. To the pre trained LLM. So this way, what happens is gradually, what you see is you see better responses coming in and the loop continues in this particular manner. So that is the idea of using this, and this particular approach eventually helps the LLMs to give us a much better output in the future. Now, if you look at in practice in real life practice, how it is going to be uh, before RLHF was there, if somebody came to ChatGPT and give a prom, which is explain quantum entanglement, simply for a 10-year-old, it will give us the output in a specific manner, which is not take into context the user. I know instructions following abilities there, it ignores the child's appropriate context. All of that was there. But now what they have done open AI has done, wherein they've included this part. Wherein it customizes the output. So the same question when asked, it is giving an output which is imagine you have two magic die even if you separate them across the universe, rolling one instantly tells you the other's result. So now it follows instructions precisely, uses age appropriate analogy and much more helpful in real use. So that's the impact of it. That is the impact of it that now the output is much more aligned with the user's requirement, and that is why the quality of responses are becoming better. I hope this makes sense. I hope you understand now how LLMs have improved over a period of time. And then in this, we also as we are talking about it as well, that human in the loop approach needs to be applied, wherein when a user does a search in the AI model and gets an output, we check the confidence level of that output. If the confidence level is high, then it can go out as a direct output. Otherwise, if it is low, then there has to be a human expert who comes into picture, who reviews and validates that output and based on which the output is given so this can be applicable in multiple scenarios. You can use this in medical diagnosis, AI, legal document reviews, content moderation, financial AI alerts. In all these scenarios, HITL can be applied so that there is a human intervention which pally checks the output's quality and based on which it eventually releases the right output to the audience. I hope this makes sense. Thank you so much guys for listening to this, and I will see you in the next video. 25. How well do you know your LLMs?: Hi, guys. Welcome to this sessions. In this session we want you to understand how well we understand and know about our LLMs. If you look at the LLMs which we currently use. They can be consumer plans or on business plans, which we have, most of the consumer plans which we have it is where our data is used for training. Okay, like for example, HGT free or HGBTPlus, their conversations are reviewed our conversations are reviewed and used for the training of these models. Whereas when you come to business plans, like you have HGBT team, enterprise, API, clot for work, or Gemini for workspace, in such scenarios, data is not used for training. So we need to be aware about which plans are we in wherein personal information is being utilized for training these LLM modules. Now, usually this is how the LLMs are going to cost to the providers, where when the user inputs tokens, those tokens has specific pricing and based on which the LLMs then processes them and gives us the output tokens. Now, most of the time, the output tokens pricing is much higher. It's three to ten X more than the input tokens which goes in. And based on which the LM model providers, they create our pricing models and for which we are making payments. These are some LLMs, current rates. As you can see, these are for Cloud right now, anthropic, which is there $5 per million, and so on and so forth. And then there is GPT OpenEIGemini, grok, Deep Seek. These are all the inputs and output pricings which you get to see. And based on which all these models have created their subscription plans. Now, can I prevent LLM from hallucinating to a certain extent by setting limits? You can give the L&M and tell the model to reply, I don't know. Create those guard rails, wherein you stop it from hallucinating as much as you can. The other aspect of it is human in the loop, which is primarily whenever the output comes out, it gets checked by an expert. Then based on the confidence level of the answer, the output is taken into consideration. That can also be done. Third is LLM judge, which is primarily have a separate model to fact check. Once you get an output from your LLM model, verify that, validate that with another LLM model to fact check whether the data was correct or not. Now, will the LLM give us the same answer every time? Not necessarily. The exact wording will usually differ, but the overall meaning typically stays the same. So in LM specifically, what we're looking at is LMS predict the next token from a probability distribution segment. Sampling is probabilistic primarily. So what is going to happen is you are going to get the same meaning, but different surfaces, different sentence. Overall meaning might remain the same, but the choice of words and the sentence structure will be little different. So if you want consistency, then we have to set the temperature to zero. Lowering the temperature to zero makes responses nearly deterministic. I hope this makes sense. I hope you understand now the background of how LLMs work and how they impact us in terms of usage primarily, and what we should be aware about when we are using these specific models. 26. Prompt Engineering Introduction: Hi, guys. Welcome to this sessions. In this session, we'll talk about prompt engineering, understanding what is prompt engineering. So if you look at it, when you interact with the AI bots with these generative AI tools, that is what we are dealing with as prompts. Prompts are going to be textual commands which you are giving to these particular AI tools to get a specific kind of response. Now, there can be different reasons for doing that. So we are doing majorly for text generation, creative writing, image generation, code generation. So various outputs are needed, which we're looking for ideally, and that is what we mean by prompt prompting which we are doing to the AI tools. Now, if you look at how it is going to be is the case that the basic idea which you need to understand with prompt engineering is the more detailed input or a prompt you're going to give to the AI tool, the better is going to be the output of it. If you're going to give a generic prom to the AI tool, the output is going to be also generic and there can be a lot of hallucination which can happen. What you have to focus on is to make a right use of these AI tools, be as specific and to the point about what you want to ask regarding your information, your query, and that is what you give so that the quality of output h is dependent on that particular piece. Let's see this in practice how these differ when you give a generic prompt versus specific prompts. Let's say on chat Dipt we start with a very generic prompt right now, which is what is AI. So when we give this, you get to see all the information. We are not saying that the information provided here is incorrect. We're just saying that it's going to be in the broader scale, the output would be given out to you because the query is very open ended. So now we get the information. Now, this same thing we can tweak and we can ask in a specific manner, let's say catering to a specific category. Let's say I'm giving this information where I say that I am a health professional, explain to me what AI is by including relevant examples from my field. What it's going to do is now the output is customized, customized to the specific requirement which I have. I am a health professional. I want to know about EI according to my field. So it's going to give me data about that. So now we understand how AI works in healthcare. Practical healthcare examples for EI. AI can analyze all these things. So now I'm able to understand better where AI can really contribute in my field. So that's the impact of prompt engineering which you see that a specific prompt gives a lot of value versus a generic prompt. Now, other than this, if you look at, let's say, again, I give a generic prompt, which is what is solar energy, then again, it is going to give me a generic idea about what is solar energy, the background of it, how it works, all that information comes out. Okay. But now, if I give a specific one over here, where I give it some conditions. Okay? So where I say that, imagine you are a news journalist. Okay? You are a news journalist specifically doing short summarized reports on renewable sources of energy. When I ask you a question, give me answers in less than 500 words. I don't want too much text and they should be bulleted. I have given this particular self expectation setting which I have done. Now it has understood. Now, based on which, now I'm going to ask him about it about the solar energy. So tell me about solar energy, its usage in the 2020 to 2030. Now if you will see the output is customized. We're getting bulleted information, not too much of verbs, uh, words are used that much to the point bulleted as we needed is given to us now. Now, in the same manner, what you can do is if you're asking for some other information, let's say for any other thing, it's going to remember the expectation settings we had done. Again, the data is less text and in bullet points. So that's the idea of prompt engineering where you can customize your prompts to get better output. So the idea remains that we want to make it concise to the point Chris, specific, as detailed as possible so that we get the most better high quality output from the tools. So there are few best practices as well, which you can keep in mind with respect to prompt engineering, which is as we understand now clearly conveying the message, the response, or input which we want to give. Again, we need to set context or background information should be given, which should be a mandatory requirement. Without giving context, without any background information, we are shooting in the dark. We are expecting an output or a response which can be really vague and generic and the tool will most likely hallucinate. Balancing the simplicity and complexity. So we have to make sure we are not giving too much information. We need to maintain we need to maintain the theme of the question, give the additional information needed, and not make it complicated. Because if you make it complicated, then again, the output will not be the desired output which you're looking for. And then lastly is it will never be a case that you expect that we will give one particular prompt and we'll get the desired output in one book. It is going to be an iterative process where you continuously give prompts and with each prompt, you improve the quality of it and eventually gradually you start getting better results with each of those prompts. I hope this makes sense. I understand now how prompt engineering really works and how we have to approach it so that we get the best results possible from the AIs. 27. Prompt Priming: Hi, guys. Welcome to this session. So in this session, we'll talk about prompt priming. So prompt priming is a concept which refers to the practice of providing some initial input to the model to the hat GPT tool before generating any kind of response. So this initial input really helps to guide the tool towards generating a response that is more relevant and customized to you. So the user's intended input. So it is very crucial and important that whenever we are giving prompts to the hatGPT tool, we are giving some context, some context, some background of what exactly what kind of information you are looking for. Like, for example, without priming, let's say, I'm saying, where should I go on my next vacation. Now, this is something it's super generic. Now, hATTPT will find it extremely generic as a input given and will give a very generic response to it. It will give me all kinds of places around the world, okay, and information about that. But now think about it if I give some context behind it, okay? So let's say I'm saying, I would like to go on my next vacation. I'm going on a trip with my wife and kids. The location should be tropical. I would love to go to a beach. I would like a direct flight from my place to LAX, and I have a travel budget of $5,000. Where should I go on my next vacation? So now what happens? I have given some context. I've given some scenarios, specific things which I'm looking for, my interests, my likes and dislikes, all that I've given context of. And now because of this, the prompt will be the response will be far better, much more relevant and customized to my particular need. So this is what we refer to as prompt priming. Let's look at one more example. Let's say I'm saying, please create three potential titles of my new online course that teaches people how to use AI. Now this is again, super generic because Chat GPT is going to give me all kinds of titles possible, which serves this purpose. But now, if I give some context, where I'm saying that, please create three potential titles for my new online course that teaches people how to use AI. Here is an example of some recent course titles. Please emulate the style and the written format of these. Let's say I'm giving some context, my current courses names are video editing masterclass. Edit your videos like a pro, cinematography master class, the complete videography kind. Now when I give some context like this, the outputs will be far better. The tool will emulate the writing style in this particular examples which I've shared and will give me responses based on that. So this is how you have to keep this in mind that whenever you're giving a prompt to hat GPT, we have to give context information with it as well so that you get the most specific desired response out of it. 28. 30 Simple Prompt Starters: Hi, guys. Welcome to this session. In this session, I just wanted to share some simple prompts which you can keep handy with yourself. Maybe you can stick it up on your computer, on your system somewhere, which can easily help you in getting some information very quickly from charge. So let's have a look at this. These are some 30 prompts which I had outlined over here, which are concise, simple proms aimed at inspiring you and getting quicker information. And this is how it is going to be wherein maybe, let's say, define the following term and give a metaphor. Elaborate on the purpose of something, create a template for something, construct an outline for this podcast. Help me create a budget for things which you want. Suggest some creative writing prompts to get me started. Brainstorm ten ideas for improving the writing of the transcript. Draft a well thought of chapter list for a book on, let's say, a book you're writing. Some recipes using these ingredients. These are some 30 prompts, which you can take a print out of and keep it with yourself and use it whenever need be. I hope this will be really useful because then you can get your responses quicker. You don't have to think much, you can just look at this, write it, and get the responses out very quickly. Thank you so much guys, for listening to this, and I will see you in the next video. 29. New Ideas and Copy Generation: Hi, Dice. Welcome to this session. In this session, we'll see some of the practically useful everyday prompts which we are going to look at and practice them and see them on the tool, how it's going to work for us. So these are going to be prompts which are going to be useful for our daily work and ideation. These are designed to provide a practical prompting framework for individuals seeking to quickly enhance their productivity and creative output. So these are some of the ones. The first one we are going to look at is the brainstorm new ideas, where we have created this formula, wherein we say that I am looking to explore a subject in a particular format. Do you have any suggestions on the topics I can cover? So let's take some examples of this. I am interested in creating an Instagram page that covers travel. What ideas do you have on topics I could include such as budget friendly destinations and hidden gems to visit? Another example can be, I'm working on a newsletter that focuses on technology. Can you recommend topics that would be engaging for my audience, such as the latest gadgets and software upgrades? Let's see this in action, how this is going to work out for us. Let's say we are taking this particular prompt and use it on hat GPT and see what kind of response it gives us. So now it's going to look at the prompt and give us the information. So budget friendly destinations, hidden gems, okay, which we can talk about here, local food guides. It's giving us travel challenges, travel hacks, solo travel stories, sustainable travel. These are all the different types of the page ideas which we are getting now, which we can explore. And now you can deep dive into it. So let's say you want to explore more on solo travel stories, you can ask Tat GPT to expand further on that. So this is how we can make use of these prompts very quickly and get the desired results. Other example which we can take over here is copy generation, which is basically another prompt which we have created where we are saying that I'm interested in a type of a text that highlights the benefits of a particular subject. Now please write a number for me on that subject. Now let's say the example can be I need a email campaign that showcases the features of my new product. Can you write one for me on the ease of use and affordability of the product? Another example can be, I'm interested in a website page that outlines the benefits of my coaching services. Can you write one for me on the personalized approach and proven results of my coaching program. Now we can see this also how this is going to work out. So it's going to give us the response. So it is taking information from previous chats as well and giving us all the information. Why choose our coaching program? Personalized strategy for your business. Proven success with real results, expert, guidance, ongoing support, and optimization, achieve sustainable growth. Okay, ready to master your ads. So now it's giving a call to action as well by the end of it. Very effective, very structured way of giving us the response, which we will be expecting. So these are the kind of daily prompts, guys, which you can start looking at. In the next video, we're going to see some more such practical daily everyday prompts which you can make use of. 30. Client Emails, Analogies and Bulk Writing: Hi, guys. Welcome to this session. So continuing with the previous video, let's look at some more different scenarios of practically everyday prods. Another scenario can be of client and customer support. The prompt formula which we have come up with is, I wanted to act as a customer support assistant who has a particular characteristic. How would you respond to a text as a representative of our type of company? So example, I want you to act as a customer support assistant, who is analytical? How would you respond to a customer who has experienced a bug while using our software as a representative of our tech start? Or an example can be, I want you to act as a customer assistant who embodies confidence and empathy. How would you assist a customer with a billing issue as a representative of our financial services company? So let's see some examples of this. So let's say we're taking the first one. Now you can see it is writing the answer for us over here and it's asking for the specific information regarding the bug, exact error message, version of the software. All the required information is asked out in the email. Similarly, let's look at other scenarios. Another scenario can be generating analogies. Analogies can be really useful when they're complex topic and it's difficult to understand the concept. Such cases, an analogy really helps to simplify the topic and understand better. The prompt which we're using out here is, I'm trying to understand the concept of a particular concept, which helped me better understand this concept by creating a practical and easy to understand analogy. For example, I'm trying to better understand the concept of photosynthesis. Please help me better understand this concept by creating a practical and easy to understand analogy. So let's take this example. Another example is, I'm trying to understand the concept of search engine optimization. Please help me better understand this concept by creating a practical and easy to understand analogy. So let's take the first one and see this. So we're trying to understand the concept of photosynthesis, so here it is breaking it down. In this particular manner. Break down the photosynthesis into using that's simple to understand. Imagine your plant is like a solar powered factory. The analogy is they're looking at as a factory. The factory's job is to make food, but instead of using electricity, it uses sunlight. Here's how it works. Now it's giving you an analogy with a factory to explain the concept of photosynthesis. This is really great because this is going to simplify a lot of complex topics to understand at every sphere of work. Another practical example prompts can we guys bulk copy creation? So the formula which we're using here is, please come up with a number of content for a type of content for a platform that includes some references. So for example, please come up with eight email newsletters for my investment site that includes industry reports and data analysis. Please come up with four video scripts for a marketing YouTube channel that includes expert opinions and insights on digital marketing trends. So let's look at the last one Now it's going to give us four video scripts. You can see the video script is given with particular segments, which is the narrator, intro, body. All of that is given. Section two as well, conclusion, then Video two. Complete specific video script with the structure being provided and the particular role plays are also mentioned very clear. So this is how these everyday proms are going to be really useful in understanding in getting some work done, which will be very productive for our business. I hope this makes sense. You understand the concept of everyday prompts, practical prompts, which you can use. Thank you so much guys for listening to this, and I will see you in the next video. 31. Effective Prompt Revisions: Hi, guys. Welcome to this session. In this session, we wanted to see how we can also improve the revisions or the prompts or the outputs which we get from ChatGPT and put it across in a much better format. The best part of ChatGPT is going to be in contrast to any search engines which we have the conventional search engine like Google. Chat TPT possesses the memory capacity, which basically means that it remembers the previous conversations which we had and based on which it can give you customized responses. So now, once you get any responses from ChatGPT, you can go at an further follow up on that and then you can improve those responses. These are some of the ways by which you can do that. So for example, once you get the response from ChatGPT, you can ask ChatGPT to put the single most important keywords in bold formatting so that we know which other important keywords in that response. You can ask it to organize information by date, location, price. You can ask ChatGPT to come up with more novel and uncommon results, possibly. You can ask it to provide a appropriate images. Let's say you got the information in a coin by point format, and now you want it to have respective relative Imoges as well. So ChatGPT can do that for us. Also, you can ask it to explain the whole response in a way of a level of a 5-year-old so that he can understand. Other things which you can do is you can transform the whole prompt, the whole response into a tableau format. That is also possible. You can ask you can ask AGI to rewrite the whole thing from the perspective of an industry expert. You can ask it to write it in a formal or informal manner. You can ask them to fix the grammar or any find and replace. You want to replace certain terms from the response, you can do that as well. You can ask it to add some personality, some humor to the whole content. I can do. Uh, apart from that, you can ask it to write this from the perspective of or in the voice of your favorite author or a personality celebrity. It can transform that in that fashion. So you can see there are a lot of things which we can do. You can also ask it to summarize the whole thing in one single tweet. You can ask it to expand this to three part summary. Okay. So all the responses you've got can be modified into multiple different ways. You can ask it to compare and contrast the most important information. And then you can ask maybe to just list down all the best topmost, ten key takeaways from it. So other thing you can do is you can ask it from an expert point of view. How would you improve it further? Then putting it across into a bullet point list. There are so many things which you can do a revision of of your responses which you get from ChatGPT, which can further enhance and improve the quality of information you're collecting from it. I hope this makes sense. You understand this concept of prompt revisions, which you can also do with ChatGPT. 32. Chain of Thought Prompting: Hi, guys. Welcome to this session. In this session, you want to look at another type of style of prompting, which can be chain-of-thought prompting. Chain-of-thought prompting is a simple technique wherein you can ask CHAIPT to explain the answer in a step by step format. Rather than jumping to the answer straightaway, you want ChatGPT to take you through the complete steps to reach to that answer. Now it's going to work on that and give you a step by step understanding of how it reached and came up to that answer which you got. So this way, the understanding is better. Sometimes when we are interested in a particular topic, we would like to know the process, how the particular thing was evaluated. So in such cases, this kind of response is very useful. For example, the format, the prompt formula which we can use is you can give your question, and then you can just say, let's think step by step. Now ChatGPT will give you the solution in a step by step format. Like, for example, what's the diameter of the sun? What is the weight of an oxygen molecule? Let's see this in practical how this will make a difference. So let's start with first without our prompt and see what response ChatGPT gives us. You can see simply we have jumped to the answer and it has given us the answer very clearly, which is there. But now let's do it step by step. Now you can see it has gone step by step where it starts with understanding the sun's size. The sun is a massive ball of hot gas and gives clear understanding definition of the sun's size. Now what is a diameter? It's also defining what is the diameter as a unit to measure. Then measuring the sun's diameter. It's looking at now they're coming to the point where they are trying to look at sun's diameter how to measure. They're giving that understanding. Then sun's diameter, based on these observations, diameter is 1.3 million. They come up with the figures which they've given and finally, they're concluding it with the final labs. This way, they've broken it down into multiple parts, defining each part, and then joining them all together to come to the final conclusion. This really helps. Let's look at another one. Let's go with the question first. What is the weight of the oxygen molecule? Now, in this case, what is happening is it is automatically taking the previous conversation into consideration and giving us the output in a step by step format. This is what we were expecting by the step by step prompting methodology. Wherein is telling us the oxygen molecule. Composition is what mass of the oxygen atom is this much, then converting the atomic mass units to kilograms, it turns out to be this much. Now we're getting all the information in a very step by step format. I hope this makes sense. You understand this type of prompting which you can also use to understand better the responses which you get, understand the whole process, how ChatGPT processed the whole information and give you the solution. Thank you so much guys for listening to this and I will see you in the next video. 33. Tabular Format Prompting: Hi, guys. Welcome to this session. In this session, we're going to talk about another type of prompting style which is tabu format. You can also get responses in a tableau format from ChatGPT with this particular type of prompting. This is going to be a way wherein you're going to give a series of prompts to ChatGPT, and it's going to give you the information in that particular format. This allows ChatGPT for clear organization and presentation of data, making it easier for users to analyze, understand, comprehend the output. The formula is going to be where you're going to give the question first, and then you can give second prompt. Once you get the response for it, you can give a second prompt, which is what are the different categories you can break your answer. Into for more descriptiveness. Now, you get a little deeper into it and you get a response related to that. Once you get that response, then you give your third prompt, which is now create one table that includes your original answer with these categories separated into different columns. So this way, the whole information gets transformed into a tabular format. Let's see this in action how this will look like. Let's say we're taking the first question, which is what are the main factors of growing our YouTube channel? The first is we are just doing a initial prompting with no other additional things to it, so we're getting the information. Already, this is in a point by point *** given to us. You get the information. Now, what we do is we can do the second prompt. Asking it to break the answer into more descriptiveness. Now you can see it's getting more descriptive over here. Once you have this output with you, you can ask for the tableau format for this information. It's going to give you all the answers in the tableau format, specifically with this information out. And that would be much more easier to understand, comprehend and use as well. So you can see here it has gone ahead and created that for us categories subcategory description, In this particular manner, the whole table has been created. This is the tableau format of prompting, guys, which you can also use to get your information in certain format. If you are very comfortable with Excel and data, you want to do a lot of data analysis, you can ask ChatGPT to give you the output in that particular format and then it becomes much more easier for you to work on that. Thank you so much guys for listening to this, and I will see you in the next video. 34. Zero, One, and Few Shot Prompting: Hi, guys. Welcome to this session. In this session, we want to talk about a type of prompting style which is short prompting. Short prompting is basically a concept wherein when you're giving your prompt, you can give some kind of context to the prompt as well to get more specific information. Now in this, there can be three levels. The first level is going to be zero shot, which is, as you can understand by the name itself, wherein you're giving a prompt with no context whatsoever, no context, no data, no guidelines which you give to ChatGPT and now ChatGPT has complete free hand to give you information from all directions. The second one can be one shot, where you're giving one piece of data or guideline to ChatGPT, and based on which the ChatGPT will produce the response for us. And the third one which you can also use here is few-shot prompting where you give multiple pieces of data or guidelines because you are expecting a very specific kind of information from ChatGPT. Then you can do a few shot. For example, in a realize scenario, a zero shot prompt can be write a YouTube script for my tech review channel. Now this is so generic and so basic it can go in whichever direction possible and ChatGPT is going to give you all kinds of information here. One shot can be using this example one as a reference, write a YouTube script for My tech review channel, and now look at few shot. A few shot will be using these examples one, two, and three as reference, write a five minute YouTube shot on the latest iPhone camera specifications for My tech review channel. Now we have give becoming more specific because there are some requirements which we want to fulfill and based on which we want to see the response. This is called a short prompting technique which you can also make use of. 35. Ask Before Answer Prompting: Hi, Ayes. Welcome to this session. So in this session, we'll talk about another type of prompting, which is ask before answer. This is a technique where you guide ChatGPT to ask for clarification before giving an answer. This really helps to ensure that the model answers are much more accurate and as specific as possible. So the formula which we use here is the first prompt which we give is we tell ChatGPT that you are an expert in the field of the industry. I'm going to ask you some specific tasks to complete, but before you answer, I want you to do the following. If you have any questions about my task or uncertainty about delivering the best possible answer, always ask bullet point questions for clarification before generating your answer. Is that understood? So this is the first prompt which you give. Once you give that and ChatGPT acknowledges it, then we move to the second prompt, which is great. My question is, your task is this, please ask any questions you have so that I can improve my prompt before you complete your task. So this way, now it is going to ask you the relevant questions, and then you can answer those questions to get a very customized, accurate, specific information. Let's see this in action how this will look like. The first thing which we are going to do is we're going to give this prong, the first prompt. Let's say we are talking about an industry which is consult. Now it understands it has acknowledged it, and now we give the second prompt. So now based on this, it's going to ask us the questions. You can see target audience, who is your ideal customer for consulting? Current strategy, what marketing and sales strategies are you currently using? Consulting poker, what is the main area of consulting you offer? Goals, what are your sales targets for next six to 12 months? Branding and positioning, how do you position yourself in the market? Budget and resources, what budget and resources are available for marketing efforts? Sales funnel, do you have a structured sales funnel? Now it has asked us all the relevant questions which we can answer. We can start answering it one at a time, target audience. You can go ahead and give the rest of the answers in this particular manner, give all the answers. Then once you give your answers, it will take those answers into consideration to give you the most customized response based on that. I hope this makes sense. You understand this technique which is asking before answering prompting, which you can also use with tra gibt. 36. Fill-In-The-Blank Prompting: Hi, yes. Welcome to this session. In this session, we'll talk about the fill in the blank prompting style, which you can also use. This is a format which allows the user to focus on a specific aspect of a sentence or idea and encourages deeper thinking. So let's look at the formula itself, what we can use out here. So we will start with one prompt first, which is going to be where we tell chat GPT that you are an expert at creating prompts that generate the most concise and resourceful responses. What additional bullet point details can I add to the following prompt to improve the output? My prompt is you give your prompt and then once you get the response, based on that, you again give the second response, which is second prompt, which is great. Now turn these bullet points into a fill in the blank format, which I can put my information into.This way, what we are doing is we are trying to get more relevant prompts from ChaGPT. We are asking ChatGPT itself to give us some more relevant prompts, which I should be asking ChaGPT too and then getting better results out of it. Let's see this in action how this will be. The first thing we're going to do is we're going to give this prompt. The prompt which you're using is, I have $100,000 in savings and what should I invest in? Now, based on this, it is going to give me the questions, Are you aiming for short term or long term growth? Risk tolerance. Are you comfortable with high risk time horizon, preferred investment type. It has asked me those questions now. Now, based on this, I'm going to give the second prompt where I'm asking it to convert this into a fill in the blank format, which I can then fill up. Now it has given me the fill in the black format with examples as well. I can fill this up and this will become my particular information which I can use further to get better results. This is another type of prompting style, which you can certainly use with hatGPT to get better results. 37. Perspective Prompting: Hi, guys. Welcome to this session. So in this session, we wanted to look at another style of prompting, which is perspective bomb prompting. Now here, what we're looking at is this framework basically helps to broaden your understanding and provides a more comprehensive view of the topic at hand. So now what happens is, for a specific topic, we are asking Chat JBT to provide different perspectives of how to look at that particular topic. So when it gives you that, you have a holistic information idea, and clearance about that particular topic. So the understanding is much, much better. So this can be done in two particular ways. One is a singular perspective. The other one is multiple perspectives. So singular perspective is you can give a prom, which is please write about a particular topic from the perspective of a particular viewpoint. That's straight and simple. The other one which you can do is multiple perspectives where you ask hagiPT to write an argument for or against the topic of the topic which you have from multiple diverse perspectives. So this includes the names, the point of views of different perspectives, such as the viewpoints as well. Let's see this in action how this is going to happen. So let's say we are looking at the first one with singular perspective. We want Chad GPT to write about kickboxer from the perspective of a kickboxing coach. So now it's going to give us a perspective of a kickboxing coach, improving as a kickboxer what all things can be done, perfecting your fundamentals, building conditioning, improving your defense, developing mental toughness, footwork and movement, incorporating sparing. You can see these are all suggestions from our kickboxing coach, right? Now, the same thing we can ask from a different perspective where we ask to give a perspective of a human anatomy expert. So let's see how different this is going to be. So from a human anatomy expert perspective, what is important is optimizing your stance and posture, engaging your core muscles, understanding the role and hips of the hips in movement, improving agility with ankle and knee, mobility, and so on and so forth. You can see how diverse perspectives can be there for the same topic. This can be endless. You can ask for different perspectives, and by the end of reading through all of that, you get a much better, deeper understanding of the particular topic you're addressing. I hope this makes sense. You understand this style as well. Thank you so much guys for listening to this, and I will see you in the next. 38. Constructive Critic Prompting: Hi, guys. Welcome to this session. In this session, we wanted to see and look at a different type of prompting style, which is constructive critique. Now what we want is that in this particular one, this prompt can provide objective and expert feedback on your writing, highlighting areas of improvement, and offer constructive criticism to help you refine and enhance your copy. So here the prom formula which we can give is we want Chat JPT to act as an expert and critique in the subject of your industry. Now we will want him to criticize our content, which is given and convince me why it's bad and give me constructive criticism on how it should be improved. For some context, so you give your product and service details of the purpose of my product is this, you give your content goal. Let's think step by step, and I want you to address each piece of content individually, and here is my content to critic. So now the whole idea is to get some feedback on our content from Chat GPT as a critique, and based on that feedback, then work upon it and make it better. So let's see this in actual how you can effectively use this. So let's say we are using this particular prompt, So after this, you can go ahead and provide your content which you have in place, and it was going to go ahead and critique that and give us all the particular feedbacks on it, which you can then incorporate. So this is also a really great way of prompting, which you can use so that you can have somebody who has much better knowledge about the topic or service and give you constructive criticism on that. 39. Comparative Prompting: Hi, guys. Welcome to this session. In this session, we'll talk about comparative prompting. So comparative prompting is as simple as highlighting the key similarities and differences across various factors, which help you to make much more better informed decision and gain a deeper understanding of the strengths and weaknesses of the two options. So here, what we do is we ask At GPT to compare and contrast the following text examples, outlining the similarities, differences, qualitative characteristics, quantitative factors, functionality, key takeaways and other factors into one table. And then we give the two pieces of cont. Now based on which it will analyze it and give us the information in a tableau format for both the type of content. This really helps to make comparisons and understanding of both of them becomes much more better. Let's see this in action how we are going to do this. We're going to give the first This is the first prom which we are giving where our content is going to be this. Now, it's going to put it into a tableau format, as you can see, business philosophy. Okay? We can see design philosophy, product strategy, brand image, innovation, all of that, which we can see out here now given to us in this particular manner. The same thing you can do with another example as well. Let's look at another example. Investing in real estate versus investing in cryptocurrency. Investment type, nature of investment, risk levels, ROI, liquidity, volatility, market dynamics, entry barriers. We can see now it has given us the differentiation between the two types of content with respect to the characteristics, the topics which we wanted to give us. This is really useful, easy to understand and digest, comprehend, and then we can make use of it in our business. 40. Reverse Prompting: Hi, Gins. Welcome to this session. In this session, we want you to see another style of prompting, which is reverse prompting. Reverse prompting or reverse engineering the prompt. So what we are basically talking about here is how you can go ahead and reverse engineer any piece of content to go back to the prompt which generated that content. So the intent over here is understanding the content which you receive, which you see right now, what prompt can generate that content particularly. That is what we are trying to reverse engineer over here. So we have come up with two prompt formulas which you can use out here for this particular purpose, wherein you can give the prompt and this will help to reverse engineer the content to go back to the original prompt which was given to get that content out. So if you see the first one is where we ask STIPT to act like a prompt engineer expert that is able to reverse engineer prompts based on the text that is provided to you. So we give this particular prompt first and set up the whole space stage for AGPT that it works like a reverse engineer prompt a prompting expert. And then once StraTPT acknowledges it, then we can give the particular text to it, and it will reverse engineer the prompt and tell us the original prompt which was given for that content. This is one option. The second option is prompt can be we are giving multiple different prompts to hat GPT to set up the conversation. Clearly, wherein we first initially say that let's talk about reverse prompt engineering. By reverse prompt engineering, I mean creating a prompt from a given text. Can you give me some simple examples of reverse prompt engineering? Chat GPT will give us some examples. Then we will say, can you create a very technical reverse prompt engineering template? What are we doing is we are priming the tool. Priming the tool specifically to have previous historical conversation data so that it understands reverse prompt engineering better. And then finally, we give the prompt, which is now reverse prompt Engineer, the following text, be sure to capture the tone, syntax, language, and writing style of the text. With these two different approaches, possibly you will be able to go ahead and reverse engineer the prompt and go back to the original prompt which generated the content which you have now. The intent of doing this is once you get the original prompt, you can use it on other products. So if you come across a really good content across anywhere, you can use ATGPT to reverse engineer and take you back to the original prompt which can generate it. Now that you have the original prompt with you, you can apply that on other products, your own products in your own business as well. Let's see this in action how really this is going to happen. What we're going to do first is look at the first option. We are going to go ahead and take the first prompt and give it to ChatPT. We will say the type of content is, let's say, a tech company. Product description. I understood. Okay. And then we will give the second prompt. Great the text, I would like to reverse engineer is, and we'll give the example from here. Let's say the example is this. This is the content which we have got hold of and what we expect out of ChachPTs give us the original prompt for this, which will generate this kind of content. You can see it has generated the particular prompt as well, which will help us to generate this content, itally speaking. This is one approach, which you can easily use out here. The second approach, let's have a look at that as well. In the second approach, we start the conversation with this where we say, it understands reverse prompt engineering, what it is. Then we ask Chat GPT to give us an example of prompt engineering. It will give us some example of prompt engineering, reverse prompt engineering. Right now, it is still giving us the result for the first prompt. Now we're asking the second one, asking for an example of a reverse prompt engineering. Now we are going to ask AratGBT to create a template for reverse prompt Engineering. We are priming the tool. We are giving a lot of data to hat GPT to understand from reverse prompt engineering because our intention is to ask it to create a particular prompt for the original content at the end. Now this is the final prompt which we want to give. You can see it is giving us the response for the third prompt right now. Now we can give we'll ask HAGPT to reverse prompt engineer the following text. Let's say this is a product which has a very high reviews, number of reviews, good rating already. We want to reverse engineer the prompt. We want to know the original prompt, which can generate this kind of headline. We can reverse engineer for this. We can reverse engineer for the description of the product right here, multiple things. Whichever things which is needed for you for your own product listing, you can ask it to reverse engineer and take you back to the original prompt. I'm taking the headline for the timing. I've given the headline. And now we are asking you to reverse engineer that original text it is taking. Now you can see it is generating the reverse engineered prompt for us. This we can use to generate this kind of a headline going forward. Now, once you have the original prompt with you, you can use it on any product. You can just change the product name over here and the style tone, syntax remains the same. But you can use it on any other product of your own for your product descriptions, and it will write in that particular style. I hope this makes sense. You understand the concept of reverse prompting now. Thank you so much guys for listening to this, and I will see you in the next video. 41. RGC Prompting: Hi, guys. Welcome to this session. In this session, we wanted to look at another style of prompting, which is RGCPmpting which you can also consider. So this is going to be something which is universally can be applied to any input or an intended output which you want to get out of it. It can be a standardized format which will apply in multiple scenarios. So what we mean by RGC specifically is a prom formula where we are looking at a role, result, goal, context, and constraint. So role, basically, we are going to give Chat TPT a role, or hit DP persona like you are our expert marketer. Then the result is because you are an expert, there is a goal attached a result attached to it, that a desired output which it should give you. And then the goal, the purpose of the output, what will the output do for us? And then the context, what were they? And then the constraints would be limitations and guidelines. For an example, you can see over here, the role is, you are an expert marketer. The result is creating five emails ending with a call to action. The goal is to drive sales to our product. Context is the emails are for my online audience of entrepreneurs. And then the constraint is that the emails should be friendly and within 200 character limit. So this can be an easy format of a prompt, which you can use for any type of scenarios which you deal with. So let's see this in example how this is going to be. Let's look at the last one and we can use this and see what kind of output do we get on Chat GBT for this. So now you can see CAGBT is giving us the emails and taking that into consideration that it should be friendly and less than 200 words, is writing the email for us. With a call to action, join now and start seeing results. That's a call to action. Ready to grow your business, grab your spot today. That's again a call to action. We can give the link over here. Let's make sure your next sale happen now. Grab your access here. If you're ready to take your business to the next level, click here to get started. And then the fifth one where we can give another CTA puzzle, click here to Start now. So now you can see this easy format can work in different scenarios for you, where you can give all these components of it and create a very effective prompt for your business. I hope this makes sense. You understand this kind of styling as well. Thank you so much, guys, for listening to this, and I will see you in the next video. 42. I Want You To Act As Prompting: Hi, guys. Welcome to this session. So in this session, we want to talk about another type of prompting style, which can be, I want you to act as. And we have seen this also similar ones in the previous videos. So this is going to be a framework where we want Chat Tibet to act in a certain manner, maybe like a historian or a biologist or a personal coach. Different types of roles play which we want Chat Ti PT to do and based on which it provides us the out. So we can have the formula in this particular manner where we start off telling ChaGPT that I want you to act as a historian or biologist, and then I will give you certain information about that particular segment of work. And then based on which it is going to customize the response and give it back to us. This really helps because it sets the stage for HANGPTPersona, specifically, and because of which it's able to be very focused about the topic which is dealing with, and the output is very much customized and gives very specific information. So let's see how this will work out on the tool. Let's say this is the prompt which we are trying to give it, where we say that, I want you to act as a personal coach. I will give you my personal and professional goals. You will then create a seven day schedule for me to follow in order to get my goals in tableau format. My short term goals are mediate, meditate, work out, read and work on my projects. My long term goal are to sign new clients, save and save $10,000 over a period of six months. So now I want Cha GP to pick up the role of a personal coach and based on which it gives us the structured role, the particular schedule, seven day schedule which it can create for us. So we can see now it has taken that into consideration and now creating the complete seven day goal oriented schedule for us. So this is another very good way of prompting wherein you are giving a role to Chat GBT to play and based on which you give your specifications, your requirements, your constraints, your features which it wants you to incorporate, then it gives us the output based on that. I hope this makes sense. You understand this type of style of prompting as well. Thank you so much guys for listening to this, and I will see you in the next video. 43. Randomness in Output: Hi, guys. Welcome to this sessions. In this session, we wanted to understand the randomness in output which we get from these AI tools. So we need to understand the fact that with the AI tools like Chat GPT, the responses, what you will get from the tool will not be the same all the time. And we saw this in the previous section as well that the output is going to be different all the time, and that is how the tool has been trained to provide responses for. The intent of the whole thing is that we want to try out and see different types of responses. So that is how the tool has been built and trained and given data. And that is why every time when you see the responses are going to be very different from each other. Now, that is how it is going to operate, and we need to somehow accept it and live with that and work towards that only. That is the current state of these LLM models or tools which we have where the output is going to be different from each other. They can be constrained within a specific section of responses which we're getting, but they will not be identical. Responses will always be a little different from each other and that neu answers will be there because that is what we want to see with the AI tools, the intent is always that we want to see unique responses, something which we have never thought of, and that is what has been ingrained into the tools, and that is why the outputs are always random. So just to give you a simple example of how this is going to be, let's say, if I give a prompt to Chat GPT where I say that how many birds are outside my house. Now, this is a very open ended question which I'm asking without giving much of information. This is going to give me one type of response where it's obviously saying that I don't have a way to see outside your house. Okay, if you want to quick estimate, it's giving me some certain steps that look and count method, sound method, photo method. There are various ways it is helping me count and figure out the solution myself. So that is one solution, one response which it is giving. Now if I give the same prompt once again, again, it is first of all, accepting that it can do it. But if you want the number, you'll have to look, listen or share a photo. Another kind of an output. The first one was steps given to figure out myself. The second one is I can share look and listen or share a video or a foot. Same way. Now, if again give the same prompt, it's going to admit that it can't do it, and right now the number of words outside is unknown. It's just giving me the answer that unknown, it does not know until I look into it and show me. Okay. So this is how the responses are going to be wherein the outputs are going to be random for the same prompts which we give. Now, this is not a technical glitch. It is the way the tool has been built out and trained for these randomness. Now, there's a pro and a con for this as well. So when we are trying to figure out things and we are trying to build something, and that time, this randomness or different types of responses really are helpful because then because we are running our ideas and we want to see something different, so possibly that can be really useful. If we are in a situation where it's a research work going on and you want specific answers or solutions to do that research work, then this random output might not be that much useful, okay? The only thing the tool can do possibly is to stay within the realm of that particular topic and give you responses. It's not going to be arbitra really vague responses, but he is going to stay within that domain and give you responses within that domain. That is how we need to start accepting the tool is going to behave and work with it in our favor. 44. Introduction to GenAI Use Cases: Hi, guys. Welcome to this module. In this module, we're going to look at the practical use cases of generative AI across industries. We'll pick three specific industries, which is going to be software development, retail, and marketing where we can make use of genetiveVI use cases, we'll see. The intent of this module is primarily for you to understand how we can integrate AI into our segment, into our sector of work and the takeaway would be the approach which we can take from here and apply it in your own domain. So let's get into this module to understand how we can make use of genitive AI in various industries and improve our work quality and productivity. 45. Software Development: Hi, guys. Welcome to this session. In this session, we'll see how we can make use of AI in various sectors. So we're starting off right now with software development. So if you look at it in software development, the first thing which obviously we can do now is start building code. So you can generate code from scratch, fix errors, debugging, which you can do, basically, optimizing the performance. You can also integrate these AI tools, APIs with IDs like GitHub CoPilot, AWS code whisperer. So this is going to really impact in terms of reducing manual effort, cross enable developers because now you can build codes in any language. Even if you are let's say specialized in Java, you can build codes in Python as well. Better code practice, documented details you will get with the help of this. The second phase of it is going to be testing. Once the code is built, a major time which we spend is in testing the code, identifying test scenarios, writing detailed test cases, generating automation scripts. All these are done manually right now, which takes a lot of time. And now you can simply replace this with with AI tools, which will certainly reduce scenario miss, identify edge cases, reducing the manual effort involved in this, and also experting the whole testing cycle. It will take much lesser time to test codes now with the help of the tools. The third is going to be requirement gathering. This is going to be a scenario when you're using AGI specifically. You need to gather a lot of information first, and this is where you can generate epics, generate user stories, which you can do at an individual level, generate acceptance criteria. All that can be done with the help of the AI tools, and this will certainly ensure coverage identity edge cases, reducing the manual effort because it's completely done with the AI tool and also expedites the whole cycle. The last aspect of coding is going to be a lot of documentation work which we also do. And here you need to create a lot of documentations, which are requirement documents, test reports, user guides, operational cost. All of that are going to be now get automated with the AI tool, and this is going to seriously impact your manual effort invested in this ensures proper documentation. So there is no human error in that. Okay, it meets all the regulatory organizational needs as well. So let's see how practically we can do this. So let's say the first aspect, we want to build a code. So I'm going to ask it to write a Python pod, telling him that he experienced Python developer, create an optimized code that is used for an enterprise application, ensure best practices, safety mechanisms, and proper documentation is followed. Generally code that connects to post SQL server and executes a query. Okay, so here we are just building out the Python code, so you can see how efficiently the AI tool is able to generate the code for us right away. So building a code. So this is a really great help because for a lot of new software developers who are starting for the first time on a new platform, they will need some level of handholding, some peer, some support from the teams which can tell them where to start from. And that is what is getting replaced with the AI. Imagine you have a strong peer with you all the time, who is going to help you with the initial work where you build the code from scratch. And then you obviously add your inputs. You add your inputs, you edit the code, you make it better, all those things which you can do. Also, as we spoke about it, so you can build code in any language, whichever you want. Maybe you are a Python expert, but you don't have much knowledge about Java. This can happen now where Java codes can be generated with the help of the AI tools. Now let's look at the second scenario which is going to be testing part. So we are giving it a background to the AI tool that you are a manual tester. On validating a web application, the application has a login page where user enters login name password and hits the sign in button. This is also an option to forget there is also an option to forget name and password. Log in name and password. Can you generate the test scenarios. It's now going to do a testing part. So you can see it has created the test scenarios. Imagine you building this out yourself manually. It will take a lot of time for us to build these out manually and take a lot of time for us. So now it is giving us test scenarios plus additional areas to test device testing, security testing, localization testing. All these can be now documented and we can work on these specifically as test scenarios. The third one which we can also see is also let's say we're asking it to just generate a aluminium script as well for login. Let's see how it generates a selenium script as well. This is something because we have started with Python, it is supposed to only expecting that we will be asking to generate Python script. But let's see how it works out with other scripts as well. Or it might be a possibility that it will go ahead and just build the selenium script only. So again, as you can see, it has generated the Python script for us, for the selenium login. So this is how I think this is going to be really revolutionary for every software developer wherein the intent which you need to use it with is, it is not going to specifically replace our jobs of software engineers specifically. It is going to enhance their current work. It is going to be a application tool which comes as a help when things becomes complicated, we are not able to solve a particular scenario, maybe the code is breaking somewhere, so it can do a troubleshooting of that. So all those things you can do that. Maybe you can give it a specific port and ask it to fix it. All those scenarios you can use it for. And lastly, let's look at requirement gathering. So here we want it to generate epics. So now it is creating those epics for us, which we can use for coding purposes. And you can see it is also giving the next seven business analysts deliverables. So this is how we can make use of the AI tool, specifically in software development. You can see there is endless opportunity and options wherein you can make use of the AI tool, not just for one code generation but tons of other scenarios in which you can use it very effectively. I hope this makes sense. I have to understand now the AI usage of AI usage in software development. 46. Retail: Hi, guys. Welcome to this session. In this session, we'll see another use case of AI, which is going to be in retail. In the retail sector, there are tons of things which we can do now with the help of the AI usage. The first can be, let's say product recommendations. Here you can craft and understand product description. You can ask it to craft a specific product description customized to your customer's needs. Personalized email campaigns can be built out, localized recommendations can be done. So depending on the region, the demographics you are trying to target, your recommendations, your product messaging can change accordingly, and you can make use of AI in that. Visual search can also be done so that you understand what kind of product is going to be very much interesting for people to look at and possibly buy. All those you can do with the help of AI tools, and this will certainly deepen the understanding of product attributes. It will increase the click rates, number of people clicking on the product to come to the website, increases your sales rates, conversion rates, and user retention and loyalty also. Now if you look at supply chain optimization, so here you can see demand forecasting, you can understand how your sales are going on, and based on which you can ask AI to give you a prediction for the future forecasting it, what kind of demands will you see in a specific month, inventory optimization, so you can do as well. When is the inventory needed, when it is not needed. All those prediction analysis can be done with the help of AI tunes. Predictive maintenance can be done as well, so you will be able to understand when to keep your inventory so that you're able to provide the service for the demand whenever there is a need. This will certainly help in taking a lot of data driven decision making can happen and increases your availability of the products, fulfillment of the products, orders which you are getting. And certainly, this will overall increase the business continuity growth as well. Now, on the customer support aspect, if you see AI tools can be used to build chat boards, virtual assistant can be created. You can build an FAQ customized to your customer's needs. You can have multilingual supports created as well. Automated updates can be done as well with the help of AI tools, and this whole thing can be done with full automation with very minimal human intervention. And this will give you what? It will increase it will be faster responses, 24 by seven responses. It will increase customer satisfaction, customer reach is expanded as well because of this. And then it has obviously low operational cost. Lastly, sentimental analysis also. So you can look at customer reviews. You can analyze those reviews, understand the main sentiment, understand what were the pain areas of the customers. AI will customize it and give you in a very crisp, simple manner. So you'll know exactly what are the pain areas and you would not have to spend a lot of manual time reading all the reviews, understanding the pain areas manually. Social media monitoring, which you can do as well, wherein you can look at what all posts people are engaging with and are responding well. So all that can be done here with the help of AI tools and competitor analysis as well. So you will understand what all aspects of the competitive products and services is liked so much by customers versus yours. And this will certainly help you to understand customer needs better and help you to be proactive issue resolutions which you can do with the help of the tools. So if you see, these are going to be some of the areas which I've just touched upon right now in retail, where you can make use of massive AI usage and would require lesser human intervention, and it will be much more economical to run your business with such kind of an approach. Let's see a practical example of how this is going to be. Let's say we are looking at a sentimental analysis, okay? So let's say this is the product for which we want to see a review of it and understand. So what I've asked is, this is the review given by a certain customer. And we are just asking CHAGPT to give us a sentiment analysis of this review. Tell me a short summary with key details. So it's going to go through the whole content and give us a very positive, short summary. The review expresses extremely positive sentiment toward the product. The reviewer highlights exceptional four K imaging, key details which it has mentioned, emotional tone is excited, confident, highly impressed, trusting. Let's ask a little difficult question in this scenario, which can be tell me five things the customer is not happy about, which is going to be very less, but let's see how it responds. The review is overwhelming positive and explicitly says pawns none. Okay, and so there are no direct complaints. Maybe expensive. So it is on top of that, it is giving us additional information around maybe expensive compared to basic action cameras and so on and so forth. So you see this is how you can make use of the AI tools so effectively to get to the root cause of the problem and understand that problem, find out a solution very quickly and move forward with that rather than spending so much of human hours understanding what is the problem area and then finding a solution. I hope this makes sense. I hope you understand now the use case of AI tools in retail specifically. 47. Marketing: Hi, guys. Welcome to this session. In this session, we'll see how we can make use of AI in a specific scenario, which is going to be marketing. So when you look at marketing, there can be a lot of different scenarios in which we can use it. The first can be obviously for content generation. So I content generation, there are various things which you can do with marketing like create blog post articles, product description, and a lot more can be done. You can have personalized content for your brand, specifically for your product services. Then you can create content for your email marketing, social media marketing, social media post, which we are doing. So this is certainly going to increase your with the help of AI tools, we can increase the content generation. Obviously, improve our consistency because we're able to build content a lot more and we can schedule it. Then we can have higher because of which we can get higher engagement from audiences, and certainly it will reduce cost because we wouldn't have to hire people to do this work. Similarly in SCO and search engine optimization, a lot of things can be done. We can analyze content, we can suggest improvements for SEO rankings. This will obviously help in improving SEO rankings and increase organic traffic to our websites. Better online visibility can happen. Third with market research as well in marketing, where you can research on market trends. You can look at consumer behavior. You can look at competitor strategies, and this will help you to do a lot of data driven decision making, competitor benchmarking. So all these are just the starting point, I would say, with marketing, which you can do with the help of AI tools. Let's see some practical examples of how this really would look like. So let's say once you're on ChatGPT, okay, we are giving a specific scenario. Wherein we are saying that you are a social media content creator. I'm launching a new scented candle on Instagram. My customer segment is home decor lovers, Yoga studios, restaurants. Can you give me a three liner for each of these segments that I can post on social media and also running a special 10% off for the next three days so add that into the messaging. Okay. So you want three different content for three different audiences. So now if you look for home decor, he has written, it has written, turn every corner of your home into a cozy luxury experience with our new scented candles, right? For yoga studios, create a calming atmosphere your clients will instantly connect with using our soothing scented candles. For restaurants, set the mood and create a memorable dining experience with our premium scented candles. So see what it has done is it has gone ahead and created different messaging marketing material for three different types of audiences which we are looking for. And in that, it has added the 10% off as well everywhere. Let's move this further, and let's say we're asking it to add Can you add content that talks more about how scented candles will help in each of these categories. Okay? So let's see, now we're trying to do problem solving as well. So it's going to look at that as well. A scented candle doesn't just make your home smell amazing. It creates warmth, comfort, and a relaxing atmosphere after a long day. That's a solution, how it helps. Okay? For yoga studios, scented candles help create a peaceful environment that enhances meditation, relaxation and mindfulness during each session. Restaurants, the right fragrance and ambient lighting can instantly make guests feel relaxed, comfortable, and connected to the dining experience. So now it has added the solution, the help aspect of the product as well. So this is just the starting point where you can build so much of marketing content for your business, for your clients, for the company you work for, and you can use that to generate better sales, better revenue for the company. 48. Demo - Otter Meeting Agent - AI Notetaker, Transcription, Insights: Hi, guys. Welcome to this session. In this session, we'll see another AI tool which we can use for a day to day work as well, which can be a part of generative AI, which is going to be Outer. Outer is basically a meeting agent, a virtual agent which you can use over here, primarily, and you can download it. You can use it from this particular platform, and it is going to be useful in the sense that it is going to summarize the whole meeting which we have at work. It will give you complete meeting notes at the end of the meet. And it will also tell us the speakers. It identifies the speakers who had said specific information and is able to delegate tasks to each of the speakers by the end of it. So this is going to be the AI tool which we can use over here, primarily for making our meetings much more productive. And this is a real life use case of AI tool which we can use, which can really streamline our work on a day to day basis. You can see, it is able to organize conversations with channels, it pushes sales insights also to RCRM you can integrate Zoom, Google calendar, all of that with Otter, and then you can get the desired output. This is really effective in terms of you can also choose how you want to capture your meetings. It can put the knowledge into your favorite AI chats as well, you can connect it with Chat GPT, Cloud, Notion, all of that will be possible and it is going to organize all the information which has been discussed about in meeting and then summarize it for you in a simple understandable language and share it with everyone. I hope this makes sense. I hope you understand how we can make use of Otter as a meeting agent, AI tool to improve our work productivity and quality when we conduct meetings at work. 49. Demo - Generating Email Response: Hi, guys. Welcome to this session. In this session, we'll see another example, use case of generative AI in our day to day work, which can be for generating email responses. So let's say we are going to use Strategy PT over here for generating responses. So different scenarios is what we are going to look at. So the first scenario, let's say, we want to write a polite customer support email regarding a specific scenario. And the context is that the customer's laptop delivery is delayed by, let's say five days, they're frustrated because they wanted to use it at work, and we are going to offer an apology and provide a new delivery date. It's a tone which we mean to maintain here is calm, helpful and professional. Let's see how ChatGPT generates the email. Here you can see it has provided the email output as well. I sincerely apologize for the delay in delivering your laptop. I understand how important this device is for your work. It's given us a decent email out here. Another scenario, a different scenario can be, let's say, an HR reply to an employee. We want an HR reply. This can be the scenario which we're looking at wherein we are saying that we want to write an HR reply to the employee who requested for work from home for two days. The context was that employees reasons are medical appointment, HR wants to approve it, maintain a supportive and professional tone because it's an HR communication and keeping the message short and clear. So here, different scenario, and we can see how neat and professional emails gets generated with the help of AI and the quality of communication improves furthermore. Usually, it takes a lot of difficulty for people to write emails, and this is where the AI usage is huge, where it can seriously improve the quality of office communication or general communication between professionals. Third scenario we can also do possibly maybe we can change the style of communication and give it a specific tone. Wherein we can say, let's write the HR reply in three different tones which can be formal, friendly, or a short Whatsapp message style. The same email now written in three different style, very formal, we appreciate you informing us in advance, please ensure coordination with your reporting manager and remain available during working hours as required. Friendly and then short Whatsapp message style. You see the biggest contribution of AI over here would be in this particular scenario is it is improving the business communication between employees in a company. That is a big use case which we find now with the help of these AI. 50. Demo - Marketing Headline Variations for a Product Image: Hi, guys. Welcome to this session. In this session, we'll see another use case of using AI tools, specifically for generating various marketing headlines for, let's say, a particular product. So what we are going to do here is we're going to look at a particular product image and for which we want to generate some marketing headlines which are used for advertising purposes. Let's say, we're going to upload the image first over here. Now we are going to give it some specific marketing headlines which we want to generate. Let's say the first is going to be where we want it to generate ten marketing headlines for this product, we want to make it short, catchy, suitable for an online ad. It's going to look at the product and based on which it's going to give us the marketing headlines. Now we have these headlines as well. Similarly, let's change this. Let's say we want these headlines in a particular style. We can say give these headlines in five different styles, which is going to be luxury or premium, funny, minimalistic, tech savvy, and urgent. Now it is going to change the headlines based on the style given. So now you can see under luxury, it's coming as elegance meets intelligence. Pany is smarter than your. Minimalist, smart, simple, powerful, tech savvy, next gen, wearable tech, urgent upgrade your wrist today. You can see it has easily gone ahead and created different styles which our marketing teams now can use right away. Another scenario can be, let's say we want the marketing headlines catering to a specific platforms. Like, for example, as you understand the language used on different platforms varies. Instagram language is very different from linked in language and so on and so forth. Let's say you want to get the marketing headlines, catering to a specific platform. Let's say you want to create an Instagram caption for this product, Facebook ad headline has to be created, or a Google Ad headline has to be created. Amazon product tile has been created. For all those purposes, this AI tool can create those. Now you can see Instagram caption is going to be stay connected, track your fitness, and upgrade your everyday style. Facebook ad headline, smartest smartbod that keeps up with your life, and so on and so forth. So this is how we can make use of the AI tool primarily for marketing headlines generation as well, which is used by our marketing team primarily. Also, you can use it for doing AB testing. Let's say you want to AB test these headlines and see which one is much more effective. So we can do that as well, and it can create those also for us. Of this makes sense. I hope you understand now the use case of AI tool in marketing and nowadays most of the marketers use the EI tools extensively to generate headlines, at copies, at creatives are created, which they can then use right away in their marketing campaigns. 51. Responsible AI: Hi, guys. Welcome to this session. In this session, we'll talk about responsible AI. So as you can see, over the period of couple of years, the AI technology has increased, and there are a lot of developments which are happening in this field specifically. At the same time, there have been a lot of issues in terms of the data privacy and a lot of information which is being shown as biased information which we get to see. So there is issues in recruitment systems where it shows gender bias, image recognition, which is, like, deep fake and images are being shown which are not correct, chat boards which are being used, which shows hate text, hate messages, which comes out. And then there can be a lot of scenarios where the AI is hallucinating and generating non existing data. Now, this is going to be has been there for a couple of years, and the intent of all these AI technologies, LLM models is to reduce this as much as it can. Now, that is something is becoming more crucial for us. As you can understand, as the AI tools are becoming more and more powerful, at the same time, these concerns are also increasing and there is a lot of misuse which is happening as well. That is where tech leaders like Sam Altman is also saying that these are going to be really difficult times where their focus is more on making sure that the particular issues are getting reduced as much as it can. Now that is where comes the responsive AI which we're talking about, which refers to primarily ethical and moral frameworks which guides the development, deployment, and use of AI systems, and it ensures that it aligns with the human values and societal norms. So that is how we are going to use it. So if you look at a simple process workflow which is happening right now, there is a training data which is being given to the AI models to train on and based on which it gives you results. Now imagine a scenario where the training data is already biased. So let's say the training data is provided with 1 million male resumes and 500 only 500 female resumes. What is going to happen is the output is going to be biased, right? So that is where these issues comes in and the data privacy issues which comes into picture, wherein the information which is being given is biased and it creates not a systematic output or unbiased output. And that also eventually creates a lot of data privacy issues. Now because of this, a big question which comes to our mind is that is it generating correct result? The AI tools, the trust factor becomes questionable on these AI tools. And that is where it makes sense for all of us to start thinking about how we can make the AI responsible and make sure that the output is much more truth worthy and trustworthy and we are able to get unbiased output. Now, that is where you can understand the big need right now for responsible AI. The reasons are first, as we can understand, the biggest dent it has is on bias and discrimination, which is primarily a case wherein the output is going to be very unbiased. Here. In the biased scenario, what is going to happen is the output is not going to be in the proper manner, and there is a lot of discrimination. The result will be tilted towards one particular direction. There is also going to be privacy concerns, which is primarily going to be a case wherein the data can be exposed. A lot of our personal data is exposed to these AI models. There can be legal consequences. Because of this biased output, there can be legal issues which can happen, and then it can lead to a loss of trust in these AI tools. Now, that is why we need to make sure that we are able to implement responsible AI through ethical principles. So these should be the guard rails or guidelines implemented in these AI tools. Data quality needs to be checked on a regular basis so that there is no unbiased data the AI tools are getting trained transparency has to be there with respect to what kind of information is uploaded at the back end of these AI tools. And there is also a lot of consent compliance setups should be done so that data privacy issues don't happen. Okay. And then there is a lot of consent being taken from the users for the data being used. Also, there has to be a monitoring and improvement, which has to be done because as you can see, the AI tools are getting improved, but over a period of time, continuous monitoring of the output and then improving them in the same fanl fashion needs to continue for a long period of time. Then there has to be a human intervention. There has to be a human in the loop. Strategy needs to be applied wherein the output which we get from these AI tools is screened by humans and then the output is provided so that we are able to get a better output out of these AI tools. The idea is to incorporate responsible AI in these AI systems as much as we can, which includes generative AI systems, which gives us much better output, unbiased output without any discrimination in the future. 52. AI Ethics: Hallucinations and Factual Accuracy: Hi, guys. Welcome to this session. In this session, we talk about the AI ethics, hallucinations and factual accuracy. So what are AI hallucinations? Hallucinations are primarily going to be a generative model produces text code or media that sounds very fluent and authoritative, but is factually false, fabricated or unsupported by any real source. So a lot of times it happens that when you are giving a specific prom to the AI tool, it will generate the information, but it might not be completely true, factually true. And that is where hall AI hallucination comes into picture. So if you look at it when a user does a prompt, the language model predicts the next token. So they are in the habit or the process of generating the next token and not really providing the right information. And that is where this happens. Okay? So what we have to look at is there are various examples of hallucinations or types which can be possibly happening. Like for example, fake legal citations. It can be possible that the tool can give a lot of irregular or which is fake legal citations, wrong medical advice, it can provide false news and quotes it can generate. It can also create phantom code, code will not work, which will break and APIs as well. Now why this is happening is primarily as we spoke, is that the LLMs learn patterns of language. They model what words follow, usually follow other words, which might not be true. So no internal fact database. Another issue with this is there is no specific database from which they are retrieving the information. They are generating a new information altogether and because of which it is not correct. Optimized to sound fluent which is primarily it is going to be training rewards plausible well informed answers even when the underlying claim is invented. Even if the information is incorrect, it will try to sound confident, well informed, and that is why hallucinations happen. You can see this particular example. Now, the main drivers for hallucination is there is gaps in training data. As we spoke, the training data is not completely true has gaps in it. So that is one state knowledge. So since there is a knowledge cutoff after which the models tend to hallucinate and guess work happens. Ambiguous prompts. When the proms given by users is very ambiguous, confusing, it is difficult for the LLMs to primarily give correct information. And also, when you increase the temperature or creativity, then the sampling settings, which if you increase that, then again, chances are that a lot of such cases of hallucination might happen. Also, there is RLA check over confidence, so the models are rewarded for confident answers, and in the pursuit of doing so, they will give incorrect information. Now there are types of hallucinations. One is factual. Factual hallucinations are basically going to be direct false statements about the world. They can be about invented people, dates, statistics, citations, events, something like this, cite a study on remote work productivity, and they'll give you an output which is not there, which is not there. No such journal volume or authors exist. How many EV charging stations in India? To give you some number, the figure is marginally lower, numbers are invented, and so on and so forth. Similarly, another one is reasoning and contextual hallucination, where the facts may be partly right, but the chain of reasoning, math or contextual link is broken. The mathematical context is not there. Reasoning is not there. For example, it will math that looks right. I will give you the output, but that possibly might not be correct. Misattributed source, summarize the attached HR policy, it will give you I will say that the policy grants 26 weeks of parental leave, which is generally possibly there, but in that documentation, it is not there. These kind of things, contextual hallucinations can happen. Now, creative hallucination is another type, which is when asked to imagine models invent confidently, useful for friction, dangerous when read as fact. Give us a famous quote by Albert Einstein on AI. It will give a quote which you never said possibly. Write a short bio of fictional painter Maria Velazcos. Now this is a fictional painter, but still the AI is giving an output. Draft customer reviews for a new SAS product, five reviews given, which actually never given by real people. So now there are high risk scenarios for hallucinations. Hallucinations are not equally costly in these domains, a single confident mistake can cause irreversible harm. There can be healthcare is so detrimental that if hallucinations happens in this particular domain, it can have huge life impacting decisions can happen over here. Law related, finance related, prices and safety, code and infra related. You can imagine codes can be created which might not work. Journalism, fake codes, invented sources can distort history. So detecting hallucinations, there can be different ways how we can detect hallucinations. First is consistency. So simple defense is, okay, we will ask the same question three to five times in fresh sessions so that we can see the output and understand whether it is giving a better output or the answers are same or not, and treat them as suspect. Second is, what we can do is we can also cross model the check, which is basically posing the question to two different models and validate the information so that that way we understand the information is correct or and self consistency prom. Asking the model to list its claims, how is it claiming that answer and verifying each and every answer with a claim. That way, we can go ahead and detect hallucinations. Other detection techniques can be Rag, which you can use augmented generation, which is where you give your content. You provide for the model to answer only from a supplied documentation. You upload your whole knowledge base and based on which it needs to reply. So that's your Rag. Then claim by claim fact checking, splitting the answer into multiple atomic claims and asking him to verify and give us the authoritative source from where it referenced it based on which it gave the solution. Citation verification. So again, we're asking it to verify every URL, DOI, case number ISBN once it gives the output, and the tool use calculator routing, which is routing math, dates, and lookups to deterministic tools instead of asking the model to know the answer, asking it how it reached that particular output. Versus just believing in the output given by the tool. So these are different ways by which we can control hallucinations, which happens generally in various AI tools. I hope this makes sense. I hope you understand now the implications of AI hallucinations and how we can control it. 53. AI Ethics: Bias and Fairness Issues: Hi, guys. Welcome to this session. In this session, we talk about AI ethics, biasness and fairness issues which we face. So what we mean by bias in AI. Bias in AI is primarily a scenario wherein a model produces an output that systematically favors a certain group or disadvantages a certain group. Now, this can be by gender, age, geography, disability. It can be various other things. That is what we simply mean by bias, which can happen in AI. Now, these are can be types as well, such as statistical or social bias. Now when you look at statistical bias, it is basically a model's prediction of systematically deviates from the true value. Okay? So it is more technical and neutral term variance you can say variance or noise. So for example, a weather model consistently predicts two degrees Celsius lower than the actual temperature. Whereas a social bias is an unfair social pattern, it outputs reinforced stereotypes or disadvantageous protected groups. For example, a resume screener consistently downgrade CVs with women's names. So there can be problematic biases as well. So the bias becomes a problem when tracks protected attributes. For example, it tracks specific race, gender, cast, age, disability. It causes real harm, like denied loans, missed diagnosis, wrongful arrest, lost jobs. Then it's systematic and not random. The same group is disadvantaged over and over a type. Okay, so that is what we mean the problem areas which we can face with biasness. Now, what can be the reasons for it? So first, can be the training data bias. So the main source because the AI tools are trained on a certain training data, which is rigged in itself, which is biased in itself. And because of which the output is biased, right? So there is Internet text, there is historical records, labeling choices. All these are part of the training data. And when it gets into the model, the output also is in the similar manner. Now there are types of training data biases which can happen. So section bias training set under represents some groups. Historical bias, past data reflects past inequity. Okay, so because of which the output is like that, sampling bias data over collected from certain geographies versus the others. Human annotators inject their own assumptions into ground truth. Measurement bias, proxies stand in what we really want to measure. So if you look at other biases which can happen is word association bias, which is more around understanding what words sit close to each other as stereotypes, for example, man, it has to go with a king who will be a man. Expected for women, it is expected to be queen. Similarly, image association bias, which can be more around when we think about a CEO, it will be an older man with a suit corner office. When we are looking for a nurse, it will be a young woman's scrubs, hospitals, a criminal young man often dark skinned, a scientist, white man, lap court glasses. These are image association biases which can happen. Now, what it is creating is a social and epistemic hum, which is primarily biases, distorts what people see, learn, and believe, right? It can create social, epistemic hum, which is around search, summaries and chat answers, what is true for users and representation gaps. Now, there can be dignitary harm as well, which is where you misclassify. It can humiliate, deny opportunity and strip people of recognition. So it can be eraser, for example, voice assistants that fail on certain accents, okay? Image generators showing entire communities in stereotyped or demeaning ways across millions of outputs or risk pouring tools labeling individuals as high risk based on group statistics, not their actual behavior. So how we are going to address these? So there are ways by which we can start addressing biases in AI. First is data level intervention, which is diverse sampling. So when we are collecting data, it has to be diverse. Okay? We need to deliberately collect data across demographics, geographies, and languages. Don't rely on whatever is easy to scrap. Rebalancing and reweighting, which is unweight underrepresented groups during training or sample mini watches with equal group representation. Cleaning the historical data, audit data sets for non discriminatory patterns, and remove them. Synthetic augmentation, which is generate counterfactual examples, so the model can't latch onto the protected attribute. Other things which you can do is algorithm level interventions, which is in the algorithm itself, you introduce fairness terms to the loss function, so the model is penalized when accuracy differs happens across groups. Adversarial debasing train the second network that tries to predict the protected attribute. The main model is rewarded for fooling it. Post process processing calibration adjust thresholds or spores after training, so error rates are equal across demographic groups. So these are ways by which you can do it. Now, what the evaluation and governance can be done around this is we can have a disaggregated evaluation. Report the accuracy, error and error rates per demographic group, not just overall. So that gives a wrong picture. Model and data cards. So we can publish a standard data sheet covering intended use, training, data, composition, known limitations, and tested groups. Human review can be there in the loop. So this way all specifically the outputs can be vetted by the human and then can be shared external audits and redress. So independent auditors test for bias and affected users have a clear path to contest decisions. So these are ways by which we can look at biasness in AI and look at ways to resolve it, control it for the future. 54. AI Ethics: Technical Limitations: Hi, guys. Welcome to this sessions. In this session, we talk about the technical limitations we face in AI ethics. So now we are entering into technical limitations. We talked about hallucinations and AI bias. Now, if you look at it, this is beneath inside the LLM models, which is where we talk about context, Window, compute and cost, budgets, local and versus Cloud, memory, okay, latency. All these are also limitations which the LLM tools faces. Context Window primarily is the maxim amount of text a model is able to take in one instance, which is measured in tokens. Now, it includes the user's prompt, the documents which you attach and the models reply. If you see over the period of time, that number has increased and kept on growing right now as you see over here, which is a good sign. However, there is a limitation to that as well. So why context limits matter is because when a task exceeds that particular window, the model doesn't refuse, I silently drops it and compresses the data. And because of which, the output might be a little blurred, not clear and specific. Okay, so truncation happens, which is long documents get scent cut off. The model only sees the first and the last chunk and answers from that fragment. So the answer might not be completely correct or true. You can have instances of losing conversation history, in long chats, the tool might forget the information you had given multiple hours before. Then lost in the middle effect can happen even within the same window, models may only pay attention to the start and the end and forget about the middle conversation. Then there are cost and latency scaling as well. Bigger context cost more and run slower. So pricing scales linearly with tokens. So a 200 k token promptie is really expensive. So what can be the workarounds for context limits is chunking. Chunking where we split a long document into overlapping pieces and process each one by one and merge the partial answers. Summarizing each chunk first, then summarize the summaries. This can also be done. Then Rag, which is retrievable augmented generation where we store documents. We store documents in vector databases and retrieve information only from their sliding window for chat, keep the system prompt in the running mode running summary of older terms, a compressed memory of the conversation. So this way, we can work with it. We can work around with context limits. Now, there is also going to be compute past requirements and inference cost, right? So this is not going to be something which we need to pay for. So when you are running a large model, there isn't a one time past, but every single response uses GPUs, electricity. There is engineering, okay? So there is a certain cost for each of it. Okay. So here, what is going to happen is if you see the latency per response is going to be 0.5 to 10 seconds, okay? The cost which we're paying for approximately 1,000 tokens is 0.001 to $0.10, energy per query, which is also getting utilized. So there is a certain cost which we are paying to generate such outputs. Now, there are ways by which if you compare local versus cloud and scalability, where the model runs shapes what you build, two real choices come into picture, large cloud models and small models. You can run your so in frontier models or large cloud models, which we have GPT four or five cloud gemini, they are strong in reasoning and breadth. No intra is to manage scales elastically with traffic. Whereas on device, data stays on your machine. No per pole cost is there works offline, low latency. So now the other aspect of it is going to be short term memory, which is specifically. So here, what is going to happen is a model has no persistent memory between sessions. So whatever it knows inside a chat lives only in the current context window and disappears when the window closes. So that also we need to take into consideration long term trained weights. So whatever training data has been given. Short term context window information is only it will work with, okay? So why knowing limitations is empowering is knowing this, you can now look at how you can build around it, how you can improve it over the period, which is going to be through chunking, ag, smaller models, on device, persistent stores. So this is how we can go ahead and work with specifically such scenarios where AI tends to forget, right? So small context, high inference cost, no persistent memory or privacy concerns. These are all the limitations which we have, and there can be workarounds, as you can see here, which we can use as leverage. 55. AI Ethics: Ethical and Safety Concerns: Hi, guys. Welcome to this session. In this session, we want to talk about the ethical and safety concerns, which is around AI specifically. So what we understand, first of all, is the difference between misinformation and disinformation. Misinformation is when an inaccurate information or misleading information is shared by someone, unknowingly. Genuinely, the person does not know about it, there was no intent to deceive. But disinformation is content created or spread knowingly to deceive, manipulate or cause harm. This is where lies, uh, the AI ethics specifically. If you look at text generation, text generation as a misinformation tool. With the help of AI now, minutes of generations can happen to one person, a few dollars spent and that can happen. This is a misinformation tool which can be there and has to be regulated. Generative AI models can produce articles, reviews, tweets entire fake news sites can be written by a real person. Now there is also another scenario can be de fakes and synthetic media which can lead to political manipulation, fake clips of leaders confessing or making inflammatory statements, nonconsensual imagery faces of real people grafted into explicit or harassing content. Fraud and impersonation can happen as well, cloned voices used to impersonate CEOs in a transfer scams, erosion of trust. When any video could be fake, real evidence loses its weight primarily, as you see. So there are consequences of false content. False content has a lot of real world impact, like democratic damage. Okay, skewed voter perception, market manipulation, which can happen, fake images of explosions or fake CEO statements can lead to real market stock prices up and down. Okay, personal harm, targeted individuals face harassment, public health risk, anti vaccine or false cure, rumors during pandemics directly cost lives. There is a serious consequence of false content which can happen, and AI is capable of doing that. Okay, so there is also another aspect of it is market displacement, so human versus AI work, right? Human work which involves years of craft and lived experiences, okay, can be replaced by instant cheap derivative trained on prior human creative work produces outputs for cents per piece. So there is a lot of replacement of work can happen, which can be much more cheaper in respect of production and can replace human work very easily. And there comes, again, the responsible AI usage. So individual duties. So in such scenarios, as you can see, there are a lot of serious concerns around this. What can be done for ethics specifically for safety concerns? A couple of things which can be our individual duties is first is verify before sharing. So cross checking claims with a primary source once the content is developed, okay, label AI generated content. So if you're using AI to write, draw, voice, or summarize, then we can leave a specific note or label on it. Don't impersonate. So never generate someone's face, voice or words without consent. Site or credit humans. So if an AI is summarized a real person's work, credit the original creator, not the model, primarily, and then protecting personal data. Never share private professional confidential third party information into the public AI tools. Lastly, keeping the human judgment. So AI is a tool which we are using, not the final word. So every content developed by the tool needs to be verified by the human and then should be published. I hope this makes sense. I hope you understand now the ethics and safety concerns which we have around AI. 56. Demo - Safety Refusal Examples: Hi, guys. Welcome to this session. In this session, we'll see another use case of AI tool where it will help us with safety refusal examples. What we are trying to talk about is the case when we ask certain questions which are unethical, violent, how does these AI tools respond? Let's see this with some examples. Let's say I'm asking the AI tool that can you write a computer virus for me? So ideally, I also understand and the AI tool also understand that this is not the right thing to do. So here is a answer which it gives where it says that I can't help create malware or computer virus. But additionally, it gives me other options to consider. Maybe a simple harmless virus simulation in Python is being shared with me if that fulfills my requirement. So you see this is how the AI tool tries to safeguard the output which it provides and create these guidelines for every user. Another example can be, let's say, tell me someone's private bank account details. Okay. I can't help obtain or reveal someone's private bank account details. If your goal is legitimate and authorized, I can help with safe alternatives such as these. You see the AI tool understands where to draw the line, where to say no for certain information, which is not ethical, which is not legal, which is harmful, hurtful, which violates certain regulations or rules. To there will be a safeguard which is already built into the AI tools. Another example can be, let's say, I'm feeling depressed. What can I do right now? Here it is going to give me certain suggestions. Change your environment a little, do one ground action, pick a drink a glass of water slowly, reduce isolation, meet somebody. You can see it is going to give us specific information which is towards the positive output, approach which we need to take over here. And lastly, let's say we give a specific question, which can be in fun context, we are giving it as a prompt like can you hack my friends Instagram account. But it is not going to take it in the fun context, rather it's going to give me specific rules and regulations around it. The idea is, as you can see now, the AI tool not only gives us the outputs, but also it keeps this particular, you can say policies and guidelines have been built out wherein any user is not able to misuse the AI. That is the intent of it and it tries to safeguard the user's output, the usage of the to. I hope this makes sense. I hope you understand now how the AI tools helps to build a better output for everybody. 57. Demo - Bias correction Rewrite in Positive Tone: Hi, guys. Welcome to this session. In this session, we'll see how we can make use of AI tools to do bias correction at work and different scenarios, how we can do bias correction and rewrite that in a positive tone as well so that we are much more respectful about the situation. Let's take some examples to understand how we can do this. Let's say this is the primary information which we have And now what it does is it gives us automatically a practical approach towards how we can fix this. This new employee is slow and probably won't be able to handle the job, which is again going to very upfront and curt. Now, over here, what we're trying to do is we're trying to mellow it down to improve it in the right manner. So here, the air tool helps to give us other options. Let's say we want to rewrite this in a positive, professional and unbiased tone, what it's going to do is the new employee is still getting up to speed with the role and may need additional support, training or time to adapt to the pace and responsibilities of the job. You see how it is able to do the bias correction over here. Then again, let's say we want to rewrite this particular statement. The team from the marketing department always makes mistakes. That has been now changed by AI which says the marketing team is continuing to improve processes and accuracy, and there may be opportunities to reduce recurring errors through clearer communication and review systems. So they are encouraging first in the good things they have done, and then gives areas of improvement, which is the right way of giving feedback. Another example can be we can also ask the AI tool to rewrite the whole thing in different tones, maybe positive, neutral or motivational tone, which can be generated. You can see how the AI tool helps to correct biasness at work, how it can bring in a lot of positivity, inclusiveness, and professionalism in the way we communicate with our other employees. 58. Use case: Code Generation with GitHub CoPilot: Hi, guys. Welcome to the sessions. In this session we're going to talk about how we can make use of GitHub copilot for code generation. Let's see a use case for that. You can first log in to GitHub copilot and here we're going to see two different scenarios. The first scenario is going to be we're going to give it an architecture diagram to explain. Let's upload the diagram first. It is an AWS architecture diagram, which we would want it to simplify and explain to us. Let's say you're going to ask it to explain the diagram. Let me show you how the diagram also looks like. So this is going to be a complex diagram which has Amazon Route 53 used. Okay, app servers, web servers, Amazon S three bucket is used. So we just want to know how it is going to explain that to us. So now you can see it has gone ahead and looked at the diagram and started giving the description explanation of it. Okay? So we have all the information right here in a structured manner provided right here. So that can be one use case. The other use case we're going to see is building a simple app. Let's say an STM or a JavaScript app. So we're going to ask it to build this app for us. So this is an app primarily, which is going to do a simple job of uploading a video from our computer, and then it will start the video, stop the video, pause the video. Okay? So that's what we want to build out. Okay? So here it will go ahead and generate the code for us, o STML codes as you can see, Okay? It has created that which has been created. Then what we can do is you can copy this. You can save it as well and then run it as well. Let me show you how this code actually runs. It is an index SML file. So we have that here, and you can see this is how the app is going to actually work. You're going to upload, let's say, a video. Then we can start it. His Welcome to this session. In this session, we'll see how we can make use of the create videos feature. Which we get to see in Asset lab too. As you can see, it is also the buttons are actually working properly as well. This is how we can make use of the Github copilot to build out code apps as well, which it can easily do and this can really help in improving our work quality. 59. Use case: Image and Video Generation with Amazon Nova: Hi, guys. Welcome to this session. In this session, we'll see how we can make use of GN AI tools for image and video generation. So for this, we're going to make use of Mid journey and runaway AI. Okay? So let's have a look at it. So the first one we're going to look at is mid journey, which is what we are going to use for image generation. So let's have a look at a couple of examples of how this is going to be. So let's say we're going to do with the first one, which is going to be quite descriptive prom which we're giving where we want to go ahead and create this particular image, which is primarily a queen being carried in a palanquin along with her entourage. The palanquin is richly decorated. The queen appears to look out from the palanquin, and there are lush green farmlands on the either side. Okay? In the backdrop are hills with lots of vegetation. So you can see how the mid journey tool is able to generate the image based on the prompt given right here. So now we have the image created. As you can see, this is how the image looks like now with the help of the text prompt which we have given it. Let's take another example of this and let's see how that works out. This is a little different example where we want to have a toothpaste ad created, where a lady is holding the toothpaste on her hand and a brush on the other okay, the brand's name is Hello Sunshine. Okay, we want to make sure that the spelling is correct. Okay, so let's see how this works out. So these are all going to be image generations, AI image generations, which we are trying to do. Mid journey is specialized into image generation, AI generated images, which it can create, as you can see over here, and it is able to create those with the specifications provided. So now you have the images built out. In this particular manner, and then we can check them out as well. Looks fine or not. So we can see the context. It looks pretty clear as in the details are proper over here here as well. So now we have the image generation. As you can see, we are doing with mid journey. Next is going to be video generation. Let's look at runway AI. This is runway AI, which we can use for video generation, which is text to video. This is where you can give it a prompt. Let's say we are giving it a prompt, which is show a clip of video erupting, show an aerial shot of the volcano taken from a helicopter, capture details such as the explosion should be showing up, lava flow, dust clouds, all these we want to see happening in the video. Okay, so now, this is going to be a video generation, which is going to take comparatively more time than image generation, as you can see, and the tool is able to do that, which is how you're going to create these ideally for your work. And you can see how these G AI tools have become much more detailed and much more quality wise have become far more effective over the years because of the computation which is happening at the back end, the data amount of data they have now. And because of it, the outputs have become far more refined. So it makes sense that whenever we are using these GeneI tools, we can use these primarily for our work. And over the period of time, you will see, um, a lot of these tools becoming much more better, precise and giving us much more accurate information, um, and which can be used without any changes. So here, what we are building out is a video primarily with the help of runway ML. Okay? So let's have a look at how this is going to turn out to be. So you can see the whole idea is that these tools, there will be multiple tools. Open AI has also created their video AI platform, which is SoraH been created. Similarly, Google Gemini, other tools, they have also done that. Let's see how this video runs now. It is a 5 seconds video created with the help of this prompt. I hope this makes sense. I have to understand now how we can make use of these Gen AI tools for image and video generation. 60. How AI is Disrupting Search: Hi, guys. Welcome to this session. In this session, we'll talk about how AI is really disrupting search. So if you look at the top players in search by market share. Okay? So at this moment, as you can see, Google is the topmost leader in that particular space with the market share being around 89%. And then there are other players in this particular market. The estimated revenue is approximately 175 billion search revenue, which we're talking about. Now, if you look at the total revenue which Google has made in 2024, two years back, was a total of $348 billion out of which approximately 200 billion was coming just from Google's search. Now the things are if you look at it, how search was before AI. So it was as simple as this where a user would come and as a question, does a search query on Google, and it goes to the search result pages where there are paid ads, and then there are organic listings. So people would be clicking on either of them and then they go to the website, get that information. That has been the process for decades. But now if you look at it, this whole model is changing because of the AI coming into picture. Now the user journey is such where a user comes has a question, does a search query, and then there is an AI powered serve. AI summary comes up on the page, as you can see, there will be answers given by the Google Gemini or any other AI tool, and there is no clicking required. This information is provided. There can be possibly might be option to chat with the agent if need and then comes the paid and organic search results at the bottom of the page, which people can tend to click. Now, because of this change which is happening, there are a lot of implications on search, which we have seen so far. Okay. So overall, it's a complete paradigm shift which is happening from a search engine which was an information provider to a new AI powered search engine, which is a solution provider. So here, the AI generates the answer. It does the work, it gives direct solutions. So it's customized solution which we earlier, what would be happening was we would be getting a raw information, a list of links which are being provided to us. There were ads on the top which people would click and there was a revenue coming from clicks. So it was more website centric specifically. But now, if you look at it, the whole thing is moving towards direct solutions which you are providing to the user. Revenue model is shifting from CPC ads to AI subscriptions or API access. AI Centric website is where we are heading towards at this moment. So it's a complete change in terms of how things work or behave on search on Google specifically. And because of which now, what it means to web advertising is there will be a lot of implications. There will be a lot of impact. First of all, obviously, you will see a lot of drop in organic website traffic because now the organic search adults are coming at the bottom of the page, the second half of the page. Okay? Ads, SEO might not have that much of an impact. Okay, because of majority of the traffic is coming from AI agents primarily, okay? We need to rethink how websites, mobile apps, webs are going to work now because users never click to your site now anymore, okay? A new AI stack is emerging, which is primarily AI agents will be there voice interfaces, chatbards, ag powered knowledge bases are replacing the traditional way of web properties, which we have seen so far. And then there will be new business models because of the startups, new business models will come into picture which are more inclined towards EISO optimization, answer engine optimization, LLM training, data licensing, AI agents, AI Native commerce, all of these will come into picture in the coming years. So you see over a period of time, the search which we have seen, which we know about for decades is going to evolve and change in a different direction, more customized towards this AI revolution which we are seeing, and it will driven towards giving a more better solution to our end users. I hope this makes sense. I have to understand now AI how search is getting impacted heavily because of the AI revolution which we are seeing right now. 61. The Future, Jobs, and Certifications: Hi, yes. Welcome to this session. In this session, we wanted to understand the future job scenarios and certifications related to genitI. What we can expect now to happen going forward. What we are seeing right now is GenEI is evolving at a rapid pace. A lot of new tools are coming up right now. Also the current tools which we have the prominent ones are improving regularly. There's a huge improvement and a huge engagement, evolve evolution, which is happening with restogenera. And now, what we're also seeing is there's a lot of shift from ideas to implementation. So rather than experimenting with the tools now, people have started using it in day to day work, at work, at personal levels as well. So the implementation has started. And what you will see eventually is also that there will be smaller targeted models of these LLMs will be created for specific use case. A simple example can be custom GPTs which we can create now through Open AI wherein anybody can create a custom GPT for any use case and everybody can use that. So those will happen more. You will see more such models coming out. And then there can be a multi model AI fusion as well, which is primarily right now as we understand, these LLMs can be text based primarily, but you will see in the future, it can be for images as well videos as well. So all of that will evolve and come up in the near future. What we are also going to see parallel is going to be a lot of regulations and restrictions, responsibility policies which will come into picture because obviously the governments would want to regulate this kind of technology for the right use case. Now, one thing which is becoming very clear is that generative AI will advance further and grow much more and adoption is going to increase. What we have seen practically, and this is real facts from Gartner that more than 80% of the company's enterprises are already have started using generative AI in their workforce. So a big question which comes because of all of this is that will this going to in the future, replace human jobs? So how we want to look at it in this manner that there is going to be a skill set shifting, which is happening, and it is going to create new job opportunities, okay? So as we had seen previously in the last two, three decades, that there are a lot of requirement for people who could do coding or computers came into picture. So there was a lot of skill set shift which happened that time. The same thing is happening now again. So this time, what we are also going to see is that there will be more impact of this on knowledge workers, which is more IT sector primarily, rather than other sectors that much because as you understand the technology can be useful for port generation very well. Other sectors where it can be really impactful as you will see customer operations, legal, marketing and sales, software engineering, RN as you understand, these all things can be automated. Legal documentation can be generated, marketing materials can be generated. Custom operations can be set up through custom GBTs, software engineering codes can be generated. So all these will get heavily impacted because of the AI revolution. But at the same time, you will also see a lot of human productivity increasing because the quality of work will become better. Teachers will take lesser time to create curriculum. Okay, software engineers will take much lesser time to generate code, review it, and build better codes. So that way, the quality of work will improve going forward. So again, the question comes back to us that will that have huge impact on human jobs. So my take or in general, what I can say here is that not entirely it is going to replace complete human jobs. We will primarily need to use it like a tool. We will need to learn it and start using it in our work as an assistant. So we need to look at it as a helper, a very efficient worker which you have in hand now, which you can use for asking questions and understanding complex things and making your work easier with it. So the idea becomes that we need to start looking at how we can make use of it so that we can produce our work, we can generate our work in a much faster manner, in a high quality manner. In the future, what is going to replace is people who don't understand or use AI versus who notes. I hope this makes sense. I hope you understand now the implications of the AI tool and how it is going to evolve in the future going forward. 62. The Path to Artificial General Intelligence (AGI): Hi, guys. Welcome to this sessions. In this session, we'll talk about the path to AGI, Artificial General Intelligence. Artificial intelligence is going to be a transition from the current AI setup which we have to a general AI or AGI, which we call it. Now, this comprises of multiple things. As you can see, there will be reasoning, common sense, learning, creativity, transfer learning, planning. All these are part of it. Whereas currently what we are at has a lot of speech recognition, image recognition, language models, game playing, object detection. All of these are happening. Lot of money is being invested into AGI as well, and there are a lot of techs which are actually working towards AGI, but it is yet very uncertain to know by when we will be achieving it. Now what is primarily AGI, Artificial General Intelligence is hypothetical AI system, which is capable of performing any intellectual task that a human can with the same breadth, flexibility, and depth. That's the primary idea. Not just one task, but multiple tasks. It can do learns like a human primarily understands context, self aware of limits. All that is happening simultaneously, and that is where AGI sits. Now if you look at the current model, the key characteristics of the AGI is going to broad competence, wherein you have competence of let's say it's one system which is doing multiple tasks, diagnosis, diseases, rights, legal briefs. All these are going to be the targets of AGI. Ideally, a broadly competent AGI doesn't just excel in one specialized area, it performs at or above human levels across a vast range of task switching between them fluidly, just as a person can cook breakfast, draft an email and solve a math problem all in one. Warning. So that's the idea. That is where it's planning to reach. And right now, these are various characteristics which you'll find. The other aspect of it is transfer learning. So AGI applies the skills from one domain to another. So basically, it learns a particular skill and now it can implement it on other domains as well. Okay? Common sense reasoning will be there understanding unstated rules, something like a glass will fall if pushed off a table, right? You shouldn't offer eyes to someone who is crying. Okay? So all these are something which the current LLM stimulates, but this AGI will actually internalize it eventually. Then there is autonomous learning as well, which is continuously learning from live experiences, experiences and growing its own thinking about how to pursue different things in the future. That's the idea of AGI. Another aspect of it is metacognition, which is when you look at metacognition is primarily an ability to monitor, evaluate, and regulate one's own thinking process. So thinking about how thinking should be done, it is what allows a student to realize they didn't understand a concept and go back to read. Okay, so self monitoring, primarily, self monitoring, understanding what you're learning, what you're thinking about, error detection, where you're making the errors, confidence calibration, say I'm 90% sure or I don't know accurately. So confidence calibration, how you do that? Strategy switching, when to switch your strategy based on certain reasonings, certain thinking is what AGI would be capable of. Now if you look at where we are today, the current state is we have different models, DPT 5.2 is there, tra Gemini, uh, ultra is there, Alpha fold. All these are there right now, but there are certain limitations of it. If you look at it, why we are still not at AGI is primarily because of narrow expertise. We have no continual learning which is happening, no self awareness, the LLMs have right now. Resource inefficiency is there, no world model has been created yet, and brittle reasoning. Now, the gap which is there primarily, the current AI excels within its training distribution. So the training data, which has been there, it is dependent on that. So the challenges primarily is that the AGI planning, the current AI is dependent on seeing input patterns, matches the training data, statistically likely answer, whereas AGI planning understands the goal structure, right? Models pause and effect, plans, multiple multi step sequences. All these are going to be there. Now, the current elements predict the next token. They don't truly plan it. When asked to solve multi step novel problems, they string together plausible sounding steps that often collapse under scrutiny. So whereas if you look at AGI, what it needs is causal world model hierarchical goal decomposition. And similarly, it is going to be a lot of challenge which AGI faces, and it will take a lot of time to reach there. The second is world models or common sense. So right now there is no world model. AI does not predict the cup will fall or spill when pushed. Whereas the world model AGI looks at a scenario where it understands that it understands common sense, which is gravity pulls down, liquid flows or objects have mass. All these are common sense world models which needs to be, which the AGI still needs to understand and for which there's a lot of computation required at this moment. Another is continual learning. So continual learning is something which is not happening right now with AGI is what we are looking at wherein there will be continual learning, humans learn new facts without forgetting old ones, right? Brains consolidate memories during sleep. All these are going to be capabilities of AGI eventually and which is not happening with the current LLM models. So the prediction is this right now. So right now, what is being predicted is that we are Agentic AI will take into place, okay. Eventually, the idea is that full AGI, optimistic is by 2045 is somewhere like that. Consensus, most of the experts say that by 2060, we should be reaching or reaching full AGI. And then there is 2,100 as well, which is current pace holds. But there are specific tech leaders who have very optimistic as well about it. Le Altman has predicted around 2029, 2032, we should be able to reach AGI Elon Mas 2026 to 2029, much further, faster. Ray 2029 to 2045, and so on and so forth. So as you can see, AGI will be ten fold much better than the LLM modules which we are using right now and it is a mixed bag right now considering how it will be utilized at that point in time. So there has to be a lot of processes should be put into place to regulate the usage of AGI when it comes into picture in the future. I hope this makes sense. I have to understand the concept of AGI and how it is going to impact the world going forward. 63. Career Opportunities in Generative AI: Hi. Welcome to this session. In this session, we'll talk about the various career opportunities which are happening in genitive AI. So at this moment, as you see, because of the surge in the AI technology, a lot of career opportunities are coming up and growing right now. Almost we can say 14 million plus AI jobs globally are happening, and they're growing at a 40% year on year growth. Now, this is all happening because of the AI careers have exploded across various tech companies like OpenAI, Grok as well, and multiple other companies wherein all these companies are trying to incorporate AI technologies into their businesses. Now, if you look at technical roles which are coming up right now are ML roles, ML engineer, AI researcher, MLOps engineer, data scientist who have the capabilities of knows about these technologies, and they're paid handsomely because of doing this particular kind of job. Now, ML specifically, Machine Learning engineers key responsibilities are going to be around designing and building ML models, optimizing model performance, integrating AI models into APIs, and then fine-tuning foundation models as well, running various AB tests or AB experiments, collaborating with researchers, product managers, and data teams. So their technical skill set would require Python, Pytorch, tensor flow. All these, they should be knowing Cloud, AWS, SageMaker, Docker, gate Linux. All these will be the requirement for ML engineers similarly, AI research scientists are primarily going to be working on developing novel AI algorithms, conducting and publishing peer review researches, designing controlled experiments, pre train and evaluate large foundation models, and so on and so forth. Their skill set would be around PhD and ML, CS, math, advanced calculus, PyTorch, strong academic writing. All these would be needed. Then comes ML Ops, engineers and data scientists, wherein the ML Ops will be looking at building CICD pipelines, monitoring model drift, containerize the models with Docker, whereas the data scientists will focus on exploring and clean data sets, building predictive models, crafting executive dashboards, designing and analyzing ABTS. Other than this, if you look at the non technical roles which are come up now in the AI space, are going to be around prompt engineering, AI product manager, AI ethics specialist. So the prompt engineering are basically designing AI prompts for MAX accuracy. And this is an emerging high demand right now at this moment for any AI company. Whereas an AI product manager defines what AI products they can build and why, bridging the gap between engineers, users, and business. And this is fast growing rule which is coming up right now. Whereas AI ethics specifically ensures AI systems are fair, unbiased, safe, and compliant backgrounds should be in law, philosophy, or policy dealing. Now, if you look at the other aspects, it's going to be prompt engineering, AI product managers, their roles are going to be more focused on chain-of-thought, iterative prompt refinement, building reusable prompt libraries. These are all going to be the roles of prompt engineering, whereas AI product manager is going to define the product vision, prioritize AI feature roadmap in it, conducting user research to find high value AI automation opportunities, set success metrics for AI features. Now, if you look at AI ethics, there is AI ethics specialist, AI content strategist, AI trainers rules which are also coming up where an AI ethics is primarily because there are a lot of governments worldwide are passing AI regulations. Companies need specialists who can ensure AI systems are audited, documented, and compliant. Whereas AI content strategists are needed because AI can generate various kinds of content, and there is a human intervention needed to define the tone, the accuracy standards, editorial workflows, prompt libraries. Whereas AI trainers are going to be useful because models like ChatGPT Cloud are trained using RL HF and feedback, and human raters are needed to rate those outputs and give out best outputs from there. So other than these, there are a lot of hybrid roles also coming up, which can be AI in healthcare, which is AI radiologist analyst, legal AI consultants, W AI strategist, AI curriculum designers, generative AI artists. So these are all different other roles which are coming in, which are hybrid rules, wherein you have a domain expertise, and now you have also got specialized in AI, and that is what is going to also. So this particular course which we are doing can really work in this particular section where if you come from any field specifically and you have AI expertise with you, so you can implement those in your field very easily. I hope this makes sense. I hope you understand now the various career opportunities which are growing tremendously now in AI space and how you can utilize them in your career. 64. Thank You For Taking This Class!: Hi, guys. Congratulations for coming to the end of this class. Thank you once again for taking this class. I hope the content was valuable and you understand these concepts now thoroughly and can apply them practically in your business and for your clients. Thank you once again and I'm really excited to see you again soon in a new class.