AI for Beginners: Learn Tools, Prompts & Use Cases

1. Introduction for the Class "AI for Beginners": Hi. This is Rajamanickam Antonimuthu In this class "AI for beginners", you can learn the very basics of artificial intelligence. I created this class so that the beginners can also easily understand AI concepts. Apart from explaining AI jargon and listing various AI tools, I did my best to explain various AI concepts in an interesting way. Like using Google's teachable machines and TensorFlow's playground. You can learn about creating subtitle files locally using openAI's Whisper model. You will be learning about training a simple TensorFlow model using Google Colab, and you can learn the basics of training the model using scikit-learn. Overall, I tried to keep this class informative as well as simple for the easy understanding of beginners. 2. Introduction to AI: This video is for explaining about the very basics of artificial intelligence, AI, so that a beginner can also easily understand it. First of all, what is AI? AI stands for artificial intelligence. It is a way of making computers think and act like humans. But as of now, it is not up to the level of humans, but scientists are working towards that goal. Let us see some examples of AI usage. Apple Siri, Amazon Alexa and Google Assistant. These kinds of assistants are powered by AI. We can ask them to answer our questions. And if it is connected with Smart Home, we can ask them to switch on Light, that kind of activities also. Similar chatbots like ChaGPT and Google Gemini, they're helping us to give answers to our questions, and we can ask them to write articles, blog posts, even we can ask them to write coding . So all these things are powered by AI. And similarly, self driving cars, cars without drivers. It's enabled by AI. AI is now becoming part of our daily life. For example, whenever we do Google search, we are getting help from AI. AI is showing the best results. Okay. And similarly, Netflix and YouTube, that kind of streaming platforms are also using AI to suggest movies and videos that are more relevant to the Viewers. And chatbots, like customer support, chatbots are also powered by AI. So a lot of companies, banks, that kind of industries are using AI powered chatbots to give best customer support. And you might have used Google lens, right? So with the help of Google Lens, we can translate any sign boards into our desired language, as well as we can use Google Lens for recognizing any object. Suppose if you're seeing some new object, if you don't know about it, you can just take the picture of that, and with the help of Google Lens, you can just get details about that object, right? So all these things are powered by AI. Almost we are using them on daily basis. Now let us see about the types of AI. One is narrow AI. That is AI that does one task. well, for example, Google Translate, self driving cars. That means the AI used with Google Translate will not be useful for self driving cast. Each application or task will be using its own kind of AI. Okay? It's narrow AI. Mostly almost all the AI usage is narrow AI only as of now. And the next one is general AI, famously known as AGI, right? It means it will think like humans, but as of now, it is not available. It's still being developed. So if AGI is coming means, it can act similar to humans. And one more type also there, it is super intelligence ASI. Okay. In case of ASI, it's as of now, it's just theory. It means it will be I mean, AI will be acting better than humans, okay? So a lot of controversial discussions are going on, whether it will be good for humans or it'll be very bad. I will destroy the humans like that. A lot of discussions are going on. like only in theory. as of now, in our daily life, we are seeing the narrow AI, and scientists are working on to create general AI. Now let us see some basic AI concepts with examples, okay? You might have heard about the word machine learning many times, right? In case of machine learning, AI learns from data. That is the very fundamental of machine learning. So we have to provide data for teaching the AI, okay? For example, very simple example. Let us see some very simple example. You're getting emails, right? Some emails will be relevant to you, actual emails. Some of them are spam, okay? So once after seeing the spam, if you mark that particular email as spam, if you're doing it for a few emails, okay. But after that, that email client will be automatically marking any new spam emails, okay? So that is the behavior of AI, this kind of machine learning. So AI is learning from your input, okay? So, based on your input, it is learning about segregating spam email and actual emails, okay? This is the very basic of machine learning. Next is deep learning. In case of deep learning, AI is using neural networks. For example, face recognition. It's something similar to human brain, that concept. Next one is natural language processing NLP. It is like, understanding the natural language English, Tamil, like that, okay. So in the chatGPT, we can ask questions in any kind of sentence, right? There is no, like, standard commands. We need not follow any structure. We can just naturally, like, we'll be speaking with our friends, right? Like that we can talk with ChatGPT or any kind of AI tools powered by natural language processing NLP, okay? So machine learning is for learning, from data. Deep learning is using neural networks so that it can find the patterns automatically. And natural language processing is for understanding the natural language like English. So fun AI activities to try. As I explained previously, you can just try Google lens to recognize objects so that you can understand the power of AI. And similarly, ChatGPT or Gemini or DeepSeek to get answers for your questions and do some kind of brainstorming with that, and you can ask it to create images based on your own requirementss. And you can do audio chat, voice chat with it. Okay? And you can use it for creating code. So by doing these things, exploring these things, you can understand the power of AI, how it's using. What are the negatives also, okay? Because we won't get always the correct answer. So there are some certain restrictions out there. It's known as hallucination. AI won't always give correct answer. It hallucinate. It will give some random answers sometimes. Maybe even for simple questions, it will be giving irrelevant answers. We need to understand that thing, okay? So we should not blindly use the output provided by AI. Okay? We have to validate it manually. Okay. that is the important thing. as a beginner, you should be knowing, right? And use AI-generated images Dall-e, midjourney, that a lot of things are there, okay? So in case of image generation, you can, like, specify how the image should be, right? Say, for example, if you want to create a new animal, you can instruct, like, Eyes should be like this and legs should be like this, like that. If you describe the things, it will be automatically generating the image, okay? So not only for images, generativeAI, GenAI is playing important role these days, right? For text generation, everything. You can ask ChatGPT to write apart from writing articles, you can ask it to write a story like that also. Okay. Just start with that, okay? You'll be understanding about how it's working. And you can use it for actual work also for writing coding, for writing marketing content. Okay. Even you can create PowerPoint presentations. Actually, I created this PPT with the help of chatGPT. I just asked ChatGPT to come up with a PPT presentation to explain about AI basics. Then I got this PPT. Okay. So these are all some advanced things. I'm listing here just to know about these things are there, okay? But it may not be more relevant to the beginners, okay? there are different kinds of machine learning like supervised, unsupervised and reinforcement learning, okay? Deep learning. Okay. In case of supervised machine learning, it's spam email detection because we are labeling it, right? It is spam, it is not spam like that, we are labeling. So based on that supervised training, it will be behaving, okay. Similarly for fraud detection, house price prediction, that kind of stuff. UnSupervised, in this case, it'll be acting without our explicit labeling, okay? And another thing is in the bracket, you can see the logistic regression decision trees, linear regression, random forest, SVM, K-means clustering Isolation Forest, right? These are all the algorithms. These are all some advanced topics. Okay, for beginness, I'm just trying to give the overview. So AI is like we have to train the AI, okay? That is one thing. For that, we need training data. Okay. And we should have some algorithm. A lot of algorithms are there. You have to choose appropriate algorithm, okay? So that is what I'm trying to tell here. So everywhere, you'll be hearing about these two words machine learning and deep learning. as I explained previously, machine learning is for training AI with data. Okay. In case of machine learning, it will work with structured data, tables and small datasets. Okay. But in case of deep learning, it is powerful for unstructured data, images, text, speech like that. better you can view like this. ML is, like, learning with a teacher, okay? The teacher has to tell the students, like, these are the features like that, okay? To explain some concepts, okay? But in case of DL deep learning, it is like a self learning student who figures out the patterns automatically. Mostly because of the deep learning only, we are seeing a lot of growth in AI. Otherwise, we have to put a lot of human efforts to train the data, right? With the development of deep learning. So everything is done automatically. Deep learning will find the pattern, and they will be able to AI models will be able to use that patterns get trained, right? In that way, it will save a lot of time, right? So we will be able to train the models with a lot of data. So because of a lot of data, lot of data means better performance. So in that way, we are seeing fast growth in case of AI. Okay? So the future of AI, almost already AI is playing important role in medicine. started like in a very research level basis. But in the coming days, it will be taking place in the mainstream, okay? Like in case of medicine, it can find diseases just by, seeing the scans, right? as of now, many research papers are saying that it is behaving better than radiologists in finding the disease. And another thing is, it will be able to find the disease, in a very early stage. Though a lot of advantages are there, there are some ethical concerns. The key thing is AI replacing jobs. People are afraid that AI will be taking their jobs, okay? That is one thing and deep fakes. A lot of fraud images, fake images are there, not only images, videos are getting created to defame someone, right? So I can say that AI is powerful, but still it's evolving, okay? So it is better to understand that so that we can make our life easy. And it is like becoming a mandatory for our career also, okay? 3. Understand AI Basics with Easy Examples: This video is for explaining artificial intelligence AI with simple example, AI is simulation of human intelligence by the machines, it's an umbrella term okay, it's a overall ecosystem and machine learning is sub field of AI, it's like finding patterns in data, there are lot of various algorithms are there, machine learning algorithms, and deep learning is subset of machine learning, it's artificial neural networks with multiple hidden layers okay the purpose of this video is not to explain all the things, it is just to give overview okay so here I will be just giving the overview about things related to AI okay AI is an umbrella term and ML machine learning is finding patterns in data and deep learning is advanced thing like that we can assume okay and data science is interdisciplinary field using AI and other things to do the prediction okay LLM stands for large language models like chatgpt, deepseek,Qwen, it's a foundational model foundational model in the sense AI model trained with a lot of huge amount of data okay so that we can use it for various purposes by just fine-tuning it according to our need okay LLM is language model okay similarly other foundation models also there CLIP for images and lot of other things for video audio so now I believe the terms are clear right let me explain AI with a simple example okay just go through the data especially first start with this left side one simple linear relationship it is relationship between inches and cm 1 in is 2.54 cm and 2 in are 5.08 cm like that right so it is data right based on the data someone created this formula right 1 cm equal to inch multiplied by 2.54 so if we have solid formula we can just use the traditional programming any program C, python or whatever program okay we can just create code for implementing this formula so that we can find cm values for other inches also say For example if I give 6 as inch value it will be telling the equivalent centimeter value okay so it is a very simple concept if we have rule or formula it is formula right similarly we can have other rules also like if the age of the person is greater than or equal to 18 he can vote otherwise he cannot vote okay like that we can use simple if condition we can implement it okay and when we give the age value it can give the output whether eligible or not okay so it is the concept of traditional programming for that we have to have a solid formula or rule mostly that might have created based on the data right so here someone might have should have created it right just by seeing the data someone created the rules so that means all the data are fitting with that rule okay so we will be able to use the traditional programming so it is very simple example in case of like assume that there are two languages in the ancient times we won't be having dictionaries or any kind of translation right someone should have gone through the writings of both languages and he should have found a pattern right so based on the pattern he might have mapped the alphabets of both languages okay so now a rule has been created then we can use traditional programming for doing the translation okay So we have certain rules means we will be able to use the traditional programming for giving the response it can be prediction or translation anything okay so this is another example is Fahrenheit to Celsius conversion this one also rule based only right but the finding the formula is bit difficult okay in the first case we can easily find it but in the second case finding the formula is difficult but ultimately both are same okay that data is following some solid pattern so that we can create the formula and with the help of formula we will be able to process other unknown data also okay that is the fundamental of traditional programming okay but assume this case like data of grade and equavalent salary of various students okay this is a student and this is a grade out of 10 and it's a current salary in USD okay here it is just a sample data okay here we can see some pattern if we go through the graph we can easily understand but there is no solid rule okay we cannot create solid rule we understand that higher grade means the student will be getting higher salary but not all students okay so in that case we cannot create a rule or formula because it involves various factors right we know that right, apart from grade is one factor so other market job market and the students other skills communication skills other things okay and influence lot of things are there right so we cannot create a formula but we can find some pattern and based on that we can say the things right but traditional programming cannot work in this way okay AI is helping to do this okay that machine learning that machine learning algorithms can find the pattern from the data okay for that there are lot of various algorithms are there okay if the data is very huge amount of data and it's a very complex thing then we can use deep learning okay say for example if we are able to provide various other factors job market details and students other activities lot of data if we are able to give large amounts of data then deep learning can help to come up with some kind of model which will be able to do the prediction somewhat accurately okay now it's clear right in case of solid rule or formula we can use traditional programming if there is no solid formula then we have to find the patterns in the data with the help of machine learning models okay AI is the only way to handle this thing okay and generally if we provide more data then the patterns that accuracy of model will be better okay especially in case of deep learning we have to provide lot of data okay let me summarize the things we can say that someone with average knowledge can come up with the formula for converting inches into centimeter by going through little amount of data but it requires some good knowledge as well as good amount of data for coming up with the formula for converting the Fahrenheit into Celsius right now we can extend this further for converting the grade into salary it requires a lot of skill and huge amount of data so this is where AI is playing role so AI algorithms especially deep learning algorithms are providing the required skills for finding the patterns to come up with the formula and we have to provide a lot of data that's it okay but it is just an assumption it's not exact thing okay AI is not going to create the formula but somewhat equivalent of that okay next let me explain about data analytics and data science okay actually I used to see a lot of confusions related to these two terms actually both of them are entirely different okay data analytics is a very simple and straightforward thing it is just visualizing the existing data for example if you are having data sales data or any kind of data in Excel or CSV file or in database you can just visualize it using tools like Tableau, PowerBI like that okay and we can easily understand the data by going through the charts and graphs okay it's a very simple thing it's not related to AI it was there long back okay but similar to all other industries are using AI, data analytics also using the power of AI to improve it okay that is different thing but fundamentally it's not related to AI okay AI can improve the quality of data analytics okay but in case of data science it is completely different thing okay it's a interdisciplinary field it uses AI it uses the subject matter expertise and statistics like that so with the combination of all the domains we will be able to predict the future okay so it's a very interesting and complex thing so we should understand the terms and we have to use it properly okay but again another question will be coming okay the relation between data science and the AI okay as I said AI is one of the domains used for doing data science related predictions okay that is one thing so AI is used by data science okay again some confusions will be coming because it's a bilateral, data science is used for creating AI models okay so we need to understand this thing okay so data science plays both roles it is used by AI and it is also using AI okay so in case of data preparation that also data science okay we have to clean the data and we have to process data so that the machine learning algorithms can be trained effectively okay that is one thing and data science is also using AI for its functionality okay LLM it's part of generative AI right so it is making the work simple instead of doing all the things like training the data collecting the data doing everything by ourself We can just use API access to access the LLMs and we can immediately create the AI application but we have to pay for the API access so we need to learn both things entire things how to create AI models by ourself and we should know about how to use LLMs or any kind of API based things okay thanks 4. Famous and Useful AI Tools: This video is for giving an overview of a few famous AI tools. I'm going to list down the tools that are somewhat related to AI, not under any specific category. Okay. First one, open AI's GPT-4. It is a multi-modal model. Actually, chatGPT is using this model, okay So whenever you ask any question on chatGPT, actually, chatGPT is able to answer for your question using this model. Okay, So similarly, similar to chatGPT if you want to develop your own application, you can use this model through openAI's API. So you have to buy API access. So every request and response that will be charged based on the tokens. Token in the sense, it's number of characters, okay So you can easily start creating application, something similar to chatGPT using this model, okay It is for language generation, chatbots, content creation, and text analysis. So it's a very famous one as of now, almost most of the people are trying to create applications using GPT-4, okay? TensorFlow is a machine learning framework. It is provided by Google, okay? It'll be useful for building, training, and deploying machine learning models for all types of AI applications. So you might have heard about Keras, right? So with Keras aPI, you can access these things easily from many languages, including Python, okay? And another important thing about Tensorflow is mostly that tensorflow sample coding is written in notebooks. So you can just run that notebooks on Google Colab. So without doing any kind of setup, you can straightaway start using this tensorflow coding on a browser itself, web browser itself. So it's a very easy way of learning machine learning, right? As a beginner, you can start using TensorFlow, okay? And similar to tensorflow, PyTorch is also a framework, AI framework, it's a deep learning framework. Okay? It is provided by meta, okay? It is used in AI research, computer vision, NLP, and reinforcement learning. We can install it locally, or you can use it on a cloud platform also, okay? And hugging face, it's NLP library with transformer models. NLP stands for natural language processing. So mostly NLP related applications, it'll be more useful. And you can find a lot of types of models from their website, okay? Just if you want to develop AI application with NLP requirements. You can just get the model from there and you can use it, okay? It offers pre-trained models for text generation, sentiment analysis, and work, okay? Sentiment analysis, in the sense, if you give the Twitter content into for example, I'm telling into this model, it will say that whether it is positive or negative, that kind of stuff, maybe customer review. So that will be very useful in many applications, including ecommerce, okay? And next one is Google Cloud AI. It is cloud based AI service provider, okay? It will be useful for image analysis, text analysis, translation, and more. Okay? Similar to Google, Amazon also providing cloud based machine learning platform. It is called a AWS SageMaker, okay? It is fully managed service to build, train and deploy machine learning models, okay? And similarly, Microsoft also providing cloud based AI framework, okay? Um, it is called Azure AI. It provides tools for computer vision, NLP, speech recognition and AI integration. It will be more useful in case you are already using Microsoft ecosystem, like Office 365 something. Okay? And Dall-e is also from OpenAI. But it is specifically for image generation from text prompts. Okay? You can create unique images based on text descriptions. A lot of websites are using this Dall-e. Almost all of them are very inspiring, okay? Similarly, if you want to develop any kind of creative applications, you can image based applications. You can use Dall-e, okay? IBM Watson. It is for business solutions, AI Power business solutions, okay. It provides AI tools for data analysis, customer insights, and NLP. Almost all the companies are providing, uh this kind of AI services, right? Like Amazon, Google, okay? So similarly, IBM also providing. Microsoft also proding right? Runway, it is like it provides creative AI tools for video and image editing, okay? So designers, that kind of media persons can use this. Okay? It is useful for AI video editing, style transfer and creative design, okay? ChromaDB it is vector databe storing and researching embeddings, okay? Uh, vector database in the sense, it will be storing the content in embedding format. It's a mathematical. It's kind of matrix content, okay? Multi dimensional numbers. Okay. So it will be very useful in RAG like retrieval augmented generation. So for that kind of AI applications, we have to use, some kind of vector database. So Chromodb will be very useful because it's a very simple, open source. We can use it in our local so chromodb it's important from this perspective, okay? RAG perspective. And the langchain, it's a framework for building applications with LLMs, like LLM s , like GPT4, that kind of things are known as large language models, okay? So it is, as I said, previously, similar to Chromadb this framework also useful for RAG implementing RAG, okay? So not only Rag, for any kind of LLM or I can say that generative AI related applications, a langchain framework will be very easy. It's kind of like it will make your job easy by providing a lot of functionalities like your code will be very minimum in case if you go with this langchain framework. Okay? For example, previously, I was talking about ChromaDB vector database, right? So if you have created a chromoDB based rag system, in case if your ChromaDB is not capable of handling all your data, and if you want to go with some other vector database, some large scale vector database. In that case, if you have manually created the code, you have to completely rewrite the code. Instead using chroma DB, you have to , create the code for another database, vector database, okay? That will be taking significant time and effort, right? In case of frameworks like Langchain, it is very easy. Apart from Langchain, there are other frameworks also there, Ilamaindex like that, okay? So it's a framework for building LLM applications, okay? And pinecone. Similar to Chromadb pinecone also a vector database, but it is managed vectored database, hosted thing, okay? So it is more powerful than Chromadb, but we have to pay for that as you use. Suppose our usage is going beyond certain limitation. then we have to pay it because it is cloud based, right? But in case of chromadb it's open source thing. So we can just download it into our own computer also, okay? So it is also useful for implementing RAG application. Rag, in the sense, it's including semantic search, similarity search, okay? So Tableau with AI integration. So a lot of data visualization tools are there, right? So one such tool is tableau, it is provided with AI capabilities, okay? It integrates AI for predictive analytics and data insights. AutoML by Google Cloud. For users with minimal ML expertise, it'll be very useful to train custom models. Okay? It will make the job easy by automatating machine learning model development, okay? 5. NotebookLM: Google’s AI Tool for Notes, Docs & Podcasts – Made for Learners & Creators: In this video I will be explaining about Google's NotebookLM. It's a very useful AI application. It is provided by Google. We have free version also. And apart from that if you want to have more features you can upgrade to plus, paid version Okay you can access it just by typing notebooklm.google.com on the browser And I already created one notebook You can create many notebooks But as I said in free version there should be some limitation in creating notebooks and other limitations I will be talking about it later And if you want to create a notebook just click this create new First let me explain what is this basically Okay It's useful for research Okay And not only for academic research, for any kind of things. Say for example if you are planning to write an article and if you want to refer four, five PDF files and two, three audio files you can just upload everything here Okay And then you can ask questions and you can ask it to summarize that and even you can talk to that Okay And a lot of features are there So basically you can get complete details about that materials Okay So it will be useful for like if you are studying for something and if you have collected lot of notes and if you find it difficult to organize them or get information from them you can just upload everything here and you can create one single notebook and from there you can get a lot of information by chating or by creating summary other things Let me explain those things Okay so the first step is adding sources then using them Okay That is the concept of this Google notebookLM It's a very simple thing Okay Just visit notebooklm.google.com and you can just start using it Just create this new notebook and you can upload files PDF files and text file markdown audio Okay Or you can take it from Google Drive or you can provide link YouTube link or website links or even you can just simply paste the content Okay it's that much easy Just you have to add the sources and then you can start using it Okay to save the time I'll be showing already added thing Okay This is new one right so I'll be opening this already added one Okay Here I already added one PDF file we can add any number of sources any number of sources in the sense in the free version there will be some limitations Okay But it's okay for a regular usage It's okay Okay And here once after adding that you can chat with that okay you can ask questions okay you can ask it to write summary even it is suggesting some pre-made questions okay what are risk of generating AI so it's my book about using AI for business okay so there I mentioned those things so based on the content it automatically generated questions okay so here it's talking about risk right so it's listing the risks okay based on the content That is the important thing about notebookLM Okay we can suppose if you want to talk to something you can use chatgpt right they are trained with a lot of data huge amounts of data right so they can do effectively for that publicly available information But suppose if you want to do like this your own notes or private information if you want to talk to that information we cannot use chatGPT and Gemini directly we can go with RAG other things but that needs a lot of time and efforts coding like we have to create vector database embeddings we have to create and store that and we have to write coding right so that kind of things are there Okay but in this case NotebookLM it's very effective for end users anyone can easily start using it Just upload the content and talk with that the information output will be grounded Grounded in the sense that it won't be affected by hallucinations other things Okay Because you are providing information So based on that information only it's going to give the output Okay So you'll be getting citations like this Okay References Okay here is suppose saying something means it will be referring based on which chapter it's giving this answer So we can understand that it's reliable Okay this is a core concept Okay We can chat with this book content Not only this only book we can as I said we can add any number of sources Okay Suppose if you want to add another book you can just add that book also If it is PDF we can add here Okay "AI For beginners" I will add this book Okay it will be added Okay now we can ask chat with both books Okay chating is one feature Similarly it is having lot of features Okay I will show the already created one to save time Okay we can create study guide we can create FAQ we can create timeline we can create this one briefing doc So everything will be useful for research kind of thing for writing articles or academic research or business research Okay suppose if you're handling multiple documents and audio files or slides you can put everything here and you can combine get combined information in various formats through chat through these kinds of study guide FAQ briefing doc timeline Apart from that this audio it's a very very important one it will be very useful like podcast also I'll be showing the demo then only we can easily understand that okay so first let me show these things okay here we can ask questions the question itself is suggesting we can ask our own questions also the answers will be grounded we can get the reference citations so it's somewhat reliable okay that is the unique thing about this one okay and Here as I said we can create timeline like this Okay we can create in the sense we have to just click the button it will be automatically created So purely based on the book content Okay So based on the book content it is giving the timeline information according to this book it's just saying present future upcoming like that but if it's history or something it will be giving exact dates other things Okay So this one FAQ already created FAQ right okay Now I'm generating with both content Okay Based on two sources Previously I created one FAQ based on single source So based on the two ebooks it has created this FAQ What are the main types of artificial intelligence models and how do they differ okay How to say model and how is improved over time It's summarizing both book contents and creating question as well as answer what are the common application of ai in daily life and various industries So suppose if you are having 10, 12 books you can easily combine all the information and you can get the details Okay So if I similarly study guide also there study guide right So it has created a quiz also questions and it provided answers also answers P and essay questions glossary of key terms it has created automatically okay based on the book content so these are all really useful for learning right and the key thing is as I said this podcast kind of audio okay you have to just to click interactive mode okay you'll be getting two people talking about this content Okay So from that we can easily get lot of information Okay Welcome to the deep dive Okay Let's cut through the noise Today we're diving into something really practical AI for entrepreneurs It's based on Rajamanicam's guide Yeah The subtitle is a practical guide to using AI in your business And that's exactly our focus Right Our mission really is to pull out the key insights the stuff you as an entrepreneur can actually use Now no need for a computer science degree Exactly We're not just you know giving a book report We're trying to find those useful nuggets that could genuinely impact your business Antony's guide is pretty grounded Takes you from the basics right up to um generative AI which is everywhere now But yeah the whole AI landscape can feel a bit much right all the jargon the constant revolution totally overwhelming from times So we want to give you a clear path What really matters for making your business more efficient helping it grow Where can AI give you that actual edge let's find out So chapter one starts with the basics right understanding AI in simple terms Yeah And get straight to the point Forget the Hollywood robots We're talking about software machines doing tasks that usually need human intelligence It's about augmenting helping out not like total replacement right which is key for entrepreneurs I think it's about tools So what could it realistically do right now and maybe where are we getting ahead of ourselves good question Okay So the guide highlights automating routine stuff first Think scheduling answering those repetitive customer questions managing email Oh the email black hole Yeah exactly It's about getting back your time you know that's the most valuable thing Freeing you up from the mundane to focus on well growth Makes sense Then there's data Okay We can also join here Okay Finding patterns Oh hey I think our listeners got something to say Okay Okay can you give us some overview about this with us great to have you chiming in Where were we ah yes AI and data Right AI's ability to crunch numbers and spot trends can be a real gamecher for businesses Absolutely It can help you understand your customers better predict market shifts all that good stuff What stands out to you so far anything in particular sparking your interest well we were just about to move into how AI can impact customer experience instant support and more personalized marketing efforts What are your thoughts on that have you had any interesting experiences with AI oh yeah Uh-h Okay can you talk about hallucinations whether it is mentioned this right Absolutely That's a really important point to bring up Yes The guide by Antonimuthu does discuss hallucinations in AI Okay let's unpack this term a little What exactly are AI hallucinations essentially it's when an AI model generates information that sounds plausible but is actually false or nonsensical So it's not making things up maliciously but it's just wrong Exactly The AI is predicting responses based on patterns in its training data If it lacks accurate information it might guess or fabricate details That sounds a bit concerning for businesses relying on AI It can be The guide highlights the risk of misinformation poor customer experiences or even legal issues if you're not careful So what does Antonimuthu suggest to mitigate these hallucinations he emphasizes the importance of human oversight Always have experts review AI generated content especially for critical areas That makes sense You can't just set it and forget it Not at all Continuous monitoring is also key to flag any unusual or incorrect outputs for review And I imagine go on Okay Yeah Can we talk about how to start using AI in business definitely That's really the heart of what we're discussing Absolutely And the guide by AntoniMuthu has a whole chapter dedicated to this Chapter 4 Building your first AI system is all about getting started The very first step is to define a problem in your business that you think AI could solve Be really specific about the outcome you're hoping for Whether it's improving customer service or automating some marketing tasks Don't just say use AI Pinpoint an area where it can make a real difference Then the next step is to choose the right AI tool for that specific problem There are so many out there so it's about finding one that fits your needs You touched on some examples like chat bots for customer support Exactly Once you've picked a tool the next stage is training Okay I stopped it Okay So it's really useful right we can add lot of good sources and then we can use this podcast kind of things and even we can join and ask questions and guide the direction right So it's really useful Okay And if you want to use it just visit notebooklm.google.com basically it is internally using RAG kind of things okay maybe it might be using Gemini though they are not officially telling it, based on the response people are predicting that it should be using Gemini okay even we can set up the same thing with the RAG and API but this one already they provided in an effective way right so we can just make use of it as they providing it free Okay So we can just upload the files start asking questions or creating podcast or getting FAQ timeline and briefing documents and study guides Okay Thanks 6. Prompt engineering: In this video, I'm going to explain about prompt engineering. First of all, what is prompt engineering? Why we need to learn about prompt engineering? We can say that prompt engineering is the way of effectively framing the prompts, that is inputs to the LLMs, large language models like ChatGPT and Gemini, so that we can get the desired output required output. In other words, we can say, it is the ability to use the LLMs effectively. we can consider prompt engineering as both arts and science because we have to follow certain strategies and tactics and tips to use it effectively. And at the same time, we need to get a lot of experience. We have to gain a lot of experience by exploring various prompts to improve our prompting skills. These days, the need for having good prompting skills is becoming important. There are two reasons. One is chatGPT and Gemini. They are playing a key role in our life. Everyone is daily using it for creating articles, creating images, and even for writing code. So it is important to be familiar with prompt engineering. Another reason is for the AI developer for developing AI tools, they need to be familiar with prompt engineering so that they can effectively use the LLMs by preparing the prompts based on their application needs because the LLMs are more powerful compared to our individual models. The are trained with a lot of data and they are powered by a huge computing power. So it is better to access them through API for creating our own AI applications. In that case, mostly we have to deal with a lot of prompts. Okay, if you're still not clear about what is prompt, better I will explain that in simple way, okay? This is chatGPT. This AI tool is internally using the LLM GPT-4, okay? So whenever we ask questions here, here if you type something, it will be passed to the LLM GPT-4 for getting the response. So basically, this one is considered as prompt, right? What is 'one plus two?' This 'What is one plus two' is called as prompt. We can ask the same question in other way also, right? 'Add one and two', right? Or you can just say 'one plus two'. In all the three cases, we got the same result. But prompt is different, right? Here the prompt is what is one plus two, here the prompt is add one and two, here the prompt is, one plus two, right? Framing this kind of prompt is considered as prompt engineering. Say for example, I'm just saying as a simple example. It's instead of spending time for writing like this, if you are able to write like this, it will save a lot of time. Okay? So now it's clear, right? It is prompt. What is one plus two add one and two, one plus two are prompts. The way of preparing this skill or knowledge, thebility of creating like this is known as prompt engineering, basically the prompt engineering is useful in enhancing quality relevance and accuracy. It is useful in chatbots, content generation coding assistance. So as I explained, it controls AI behavior and output quality. So for getting quality output, we have to improve our skill in prompting. It reduces ambiguity and enhances clarity. So if our prompts are clear, then outputs will also be clear. It helps fine-tune AI for specific use cases, essential for automation and efficiency. Let me explain them with the examples later, okay? Components of a good prompt. Okay? So now it's clear, right prompts means that input text we send to the LLMs. It can be from user perspective, it can be a simple Question typed on chatGPT or from AI application developer perspective, it is the prompt we prepare for sending to the LLMs through API in our coding. So first one is clarity. Clarity, it is a very obvious thing. So whenever we ask any potion, it should be very specific and concise. Both things we need to follow. It should be very specific. It shouldn't be any random thing like um, it shouldn't be very generalized thing, okay? We need to ask specific questions, okay? Let me show a lot of examples. Even in the openAI, they provided a lot of examples. But it is very obvious one, right? It should be very specific. Instead of asking like, write a poem. You can ask, like write a poem about moon, right? Or you can ask a poem, write a poem highlighting happiness like that. And context. And you have to provide the background information also. The same thing, write a poem on moon. You can give the reason for coming up with this poem. I'm planning to use this poem in my blog post or something for whatever thing. So if you give some context, based on that context, it can fine tune the output, okay? Or you can say that it is for competition like that, okay? And the third one is constraints. We have to define the limits. In the prompts, we have to define the limits, like a number of word count and tone, format. For example, in case of article writing, you can say that write article about universal basic income in 20 lines or in 200 words like that. And in case of tone, you can say that write it in a formal tone or for professional way or in a funny way like that, right? And the format, whether you want it in Excel format, or CSV format or arranged in a table kind of stuff, okay? For example, if you're asking like differentiate the difference between machine learning and deep learning, you can specify whether you need the output in a comparison table like that, okay? And examples, it is a very key thing, okay? We have to provide sample outputs for guidance. Then only it will be effective. It is called as one-shot learning and few-shot learning, like that. Okay? Maybe I can explain it later in detail. Role assignment. We can ask the LLM to act as a particular role. For example, if you want to get poems, first you can say that act as a great poet. Maybe you can refer some people like that you can ask the chatGPT or any kind of LLM to act as a particular person, particular role, okay? If you can say that act as a teacher and prepare the guidance for students like that, okay? Types of prompts. Open-ended prompts, instruction-based prompts, role-based prompts, comparative promptss, step-by-step prompts. Open-ended prompts, it's like open questions. Discuss about climate change like that. It can write anything. And come up with any content. In case of instruction based prompts, you can give specific details like you have to write this article in this particular way with this word count, like that, right? So that kind of instruction based prompts, role based prompts, as I explained earlier, we can ask ChatGPT or any LLM to act as a particular role. And then we can ask questions, similarly, step by step prompts. Like it is called a chain of thoughts also. It's a very important thing. We have to ask chatGPT or any LLM to prepare the answer in a step by step way. It's maybe in any kind of thing, like if you want some answer for some question, just ask a question and then add, do it in a step by step way like it is very important thing. Otherwise, that mostly the Hallucination problem will occur. Even for simple questions, we may get some wrong answers. But if we ask specifically to do the processing step by step, we can get some solid correct output. It is very important. It is called as chain of thoughts. You might have heard that so before understanding prompts and prompt engineering, we need to understand the inherent issues with LLM. This hallucination is very, very important. So mostly for dealing with that, it will be useful step by step proms. Okay. So even though the LLMs are very powerful and they are able to give answers in a very effective way, that answers are not accurate. They are not reliable. They used to give some fake answers, okay? If you ask some references, it will give some random links or something, okay? So we need to be very careful in analyzing the output. At the same time to get somewhat better results, we have to follow the strategies. One important strategy is this step by step. chain of thought, okay? And giving examples and providing some reference things. Okay. All these things will guide the LLMs to act in a correct way. If we are not guiding, if you're just simply throwing the question and expecting answer, it won't be giving proper answer. Okay? So we need to guide it properly. Example prompts. Weak prompt is write about AI. It's a very generic. We need to be specific. So write a 200 word article on AI in healthcare. They provided the limitations, okay, the constraint, and um, they provided specificness, healthcare. Okay? So here, the best one is 'as a medical researcher, explain AI's impact on healthcare in 200 words' Here, the role based, we are asking it to act as a medical researcher, right? That part is there, and we are asking specific question like impact on healthcare in 200 words. That limitation is there. So it is the best prompt, okay? So just to learn the basic things about prompt engineering, then try to implement in your daily life prompts, okay? Instead of writing write about AI, we have to write as, as a medical researcher, explain AI's impact on healthcare and two hundred words. It's just example. Suppose if you're planning to write a blog post from technical perspective, maybe you can say that as a good blog writer. explain impact on health care in 200 words. Like that, you can ask. So it's based on our requirements. That's what I'm saying. Based on our requirements, if we're able to use the LLM properly, that ability, we can consider it as a prompt engineering, prompting skill. The ultimate aim is we have to effectively use the LLM to get desired output. A lot of prompt engineering techniques are there. One is zero-shot prompting, few-shot prompting, chain of thought prompting, multi-turn prompts, okay? Zero-shot means straight away asking question and getting an answer without providing any examples. Few-shot prompting is like we have to provide some examples. Suppose if you are planning to format the output in specific format, you can provide the few sample output. Then it is known as few-shot prompting. Okay. Then chain of thoughts prompting. As I said, we have to ask the LLM to specifically do the calculations step by step. And multi-turn prompts. So multi-turn prompts, it's nothing but interactively engaging with the LLM. We have to continuously talk with that instead of asking one question and then getting an answer. Instead of like that, we can continuously engaging with something like chatting, common challenges and solutions, ambiguous outputs. So in that case, obviously, we have to refine the prompts. We have to improve the prompts over there. Irrelevant responses. In that case, as I said, we have to use the few-shot learning, that kind of prompting, okay? We have to provide examples, so that it'll be useful for guiding the LLMs. Hallucination issues. Okay. So in that case, we have to check the output with other sources. So always don't trust the AI output. We have to manually verify the other sources also. So in case if it is providing lengthy or short responses, obviously, we have to provide the specific word limit. In that case, follow that. So a lot of applications are there for content creation, coding assistant, customer support education. So almost for any field, that prompt engineering skill is really important one. We can practice that with chatGPT, or you can use OpenAI's API for using this GPT-4 LLM, okay? But in case of API, you will be having additional option. So normally in the chatGPT, we are having two things, right? One is that chatGPT, that is giving answer. Another one is as a user, we are um, asking question. There are two roles, assistant that chatGP assistant and then user, ourself, user and assistant. But in case of API access, we have three options, three roles, system, user, and assistant, okay? So in that case, we will be having more flexibility that 'act as a role', other things we can specify in the system itself. Maybe I will explain the details with some sample. Obviously, Gemini the ultimate thing is we have to be very clear in writing prompt. It's a be specific. We have to experiment with different styles because it is dynamically changing. So whatever prompting skills you are gaining this month will not be applicable in the two, three months later at that time. Okay, I'm just telling you subtle things, but the core things always like if you go through Open AI's website, they listed six strategies. That six strategies, even the last year also they provided the same thing. So mostly the core concepts will be remaining same, but the subtle things will be changing. That we need to learn. That is what everyone is saying that prompt engineering is both arts and science. We have to understand the basic rules and strategies. Then we have to improve our ability because it's personally changing to person to person. So we need to stay updated with AI advancements. Okay. Let me show some example. In case of thumbnail generation, I can ask, say, for example, recently created one AI course for that I created one video in YouTube. Okay. Then in that case, I can say that 'create thumbnail . for AI course' It's not very specific, right? So it can create the Image in any way. Okay? But my purpose is I need to create the thumbnail for my youtube video. Here, that aspet ratio is square right. But in case of Youtube video, the aspect ratio ratio should be 16:9, right? In that case, the better way is, we have to specify the aspect ratio. or at least we have to provide the context for our video, YouTube video like that. But in that case, based on my personal experience, that is what I'm saying that always the rules, there are no solid rules, okay? As a best strategy, we can specify the context. But in some cases, it won't take that. from my personal experience, most of the time, it is not considering that even if I am specifically saying that it is for my YouTube video, it still used to generate the square image only. So I used to specify the aspect ratio, with the aspect ratio. So now it created with aspet ratio, 16:9 so that I can use it for my YouTube video. Okay. Apart from that, we can provide other details, what kind of color you should have. But it won't always follow the instructions properly. So we need to do a lot of trial and errors. Okay. In case of few-shot learning, maybe you can provide a few sample images and then ask chatGPT to create similar image like that, okay? Or in case of formatting, we can ask it to format in specific thing, like write about machine learning versus deep learning. We can say that format it table. So here, it's writing the differences in a table format, okay? Otherwise, if you want to get it in a blog post article kind of thing, you can ask it to write as an article in like 500 words like that, whatever. And as I said, we can provide specific samples also. Like for example, if you want to get the suppose if you're looking for some laptop or computer for buying, you can ask it to recommend recommend three laptops for my video creation. And you can ask it to specify the format. Like give details, we can give some examples. One is, you can just give laptop name like that you can give the label RAM, hard disk, display, CPU like that, you can ask, okay? It is format. In case of few-shot learning, we can provide some sample. It's giving the details, right? RAM, storage, display, CPU like that, right? Otherwise, as I said, we can give the samples. And you can use delimiters. That we can differentiate between the actual prompt and sample like that. We can use some markup or some star like that, right? So we can specify format like brand name, RAM, price. Okay. It recommended in specific format, okay? But I haven't specified the display. It included display also, because of my previous that effect will be there. That I will be talking about that, okay? It's context history, okay? It is remembering the contest history. Suppose if you want to clear that context history, you have to start new chart. suppose if your question is completely independent of your previous question. you have to start a new chat. And another thing also there that is memory. So among all your chat histories like say, for example, one week before, you might have asked some questions, okay? But still it'll be remembering that details, okay? Even if you start the new chat, if it is stored in memory, it won't store everything, it will be storing few things, okay? But context history is different. If it is in the same chat window, it will be remembering the old questions. fully, it will be remembering. Okay, that is context history. But in case of that memory, that will be selectively storing key important details in memory, okay? So if you want to have that option, you can keep it here settings, okay? So if you enable it, it will be remembering your key details. Okay? You can click Manage memories, and then if you want to selectively remove some details, you can remove, okay? That memory details will be available across various chat windows. Okay. And other than that, you can customize chatGPT. I'm just talking about chatGPT. The prompt engineering is common for any kind of LLM. Uh, even for this chatGPT, using GPT-4, right? So if you're using it from application through API, that is different case. But in case of chatGPT interface if you're using, you can specify these custom instructions like I know Tamil and English. In that case, you can straightaway ask translate this to something. It can understand that I know Tamil and English. Suppose, even if I'm not specifically saying Tamil or English, based on the context, it will be understanding that, okay? And here you can specify that it should be professional or funny like that. So based on that, it will give answers. If you are familiar with these kind of basic things, you can effectively use ChatGPT or any kind of LLM tools. And as an AI developer, you can effectively create your application. So in the Open AI, they provided a lot of strategies. They provided six strategies for getting better results, and they provided a lot of tactics. Okay? So almost whatever we explain here they provided in detail. In case of first let us go through all the six strategies, okay? Write clear instructions, it is obvious thing we already discussed, provide reference text, and split complex task into similar sub tasks, and give the model time to think, use external tools, test changes systematically. For each strategy, they provided list of tactics and go through that also. In case of write clear instructions, we can follow these tactics, include details in your query to get more relevant answers. They provided some examples. How do I add numbers in Excel. Instead of asking like this, we can ask specific things, right. How do add up a row of dollar amounts in Excel I want to do this automatically for a whole sheet of rows with all the totals ending up on the column called total. So we have to give the specific details. Instead of just asking who is president, we can ask who's the president of Mexico in 2021. Instead of blindly asking write code, we can ask write a typescript function. So it's like giving specific details, okay? And the next tactic is, ask the model to adopt a persona, its role as, right. So as I said earlier, the API is having this system option, okay? So as a developer, we can use that opportunity. So this I explained delimiters, right? We have to effectively use the delimiters to differentiate between the samples, and the command. In case if you're talking about two things, maybe you can find this format. So the delimiter. I need not be any standard thing. You can dynamically create based on your context. Next one is, speicy the steps required to complete a task. You have to provide intermediate steps like a step one, step two. It can be a part of chain of thoughts. Whenever we ask the LLM to follow certain path, then the output will be somewhat reliable. That is my observation. If we straightaway ask the end result, it won't be that much good. The next tactic is provide examples. Then, specify the desired length of the output. So we can limit that right, in about 50 words, 2 paragraphs, three bullet points. So for the next strategy, provide reference text. The tactic is, instruct the model to answer using reference text. So we have to provide some reference, and then we can ask the question. Okay? So this is one key thing, context window, limited context window, okay? We cannot, here, the reference text. Suppose if it is 100 pages article, we cannot put all the hundred pages here. So almost all the LLMs are having that context window limit. Okay? If you search for context window of the LLMs, you can find some details about the limitations. Every LLM is having its own limitation. Okay? So for that purpose, that embedding concept is there. So internally, we have to store the details as a embedding, and then we have to dynamically, get the required part from the embeddings. Okay? Embedding search we have to store them in the vectors database. It is called RAG retrieval augmented generation in case of generative AI, another tactic is, instruct the model to answer with citations from a reference text. Instead of staightaway giving answer abruptly, we can instruct the model to give the answer with citations, from which part it is taking that answer like this. Next strategy is, split complex task into simpler sub tasks. The tactic is, use intent classification to identify the most relevant instructions for a user query. If for example, in case of customer support, okay, there may be different sections like billing, technical support, account management, general query like that, right? So each primary category will be having its own thing, like in case of billing, unsubscribe, upgrade, add payment method like that. Technical support means trouble shooting device compatibility issues like software updates. So basically, even if the case of customer support, internally, it's having various sections, right? So based on that section, we have to classify it to identify the most relevant sections for user query. So if the user query is related to billing, then it can use the billing related instructions, right? Here, they provided some example. based on this example, if the user is asking, I need to get my Internet working again. That case, first of all, it can do the classification. If it is troubleshooting, it can again, ask I mean provide the instruction. only specific to that trouble shooting. It will be more effective, right? But in case of this approach, we won't be seeing it in chatGPT, that kind of thing because it needs system role, right? So mostly for API access, we can use it, effectively. In case of the dialog application, as I explained previously, each LLM is having its own context window, and we have to pass the context history whenever we write the prompt right. Even if we ask simple question, actually, the charGPT will be sending all our previous questions also to the GPT-4 model so that it can understand the context, okay. But because of the context window limit, we cannot send all the things, right? So after some time, it will be cutting the previous questions. In that case, we will be losing the context, right. To avoid that, we can summarize the previous context, okay? And then we can add it to our new prompt. Okay. That way we can handle the context Window. for keeping the context. In case of this strategy, you models time to think, here they provided one good example. Here just straight away asking whether the student's solution is correct or not. The user is providing the problem statement. Immediately, the assistant is saying that the student's solution is correct. But this solution is not actually correct. In that case, we can ask it to first work out your solution to the problem, then compare that solution with the student solution. Okay? So in that case, it will be saying that student solution is incorrect. As I explained, it's a very I can summarize it as we have to specifically guide the LLM. Then only we can get the proper reply. If it staightaway just put the question and then expect the answer, it won't happen. So they provide a lot of examples, okay? So mostly coding related, okay? I mean, API access. In this case, they provided some good prompt example also, okay. In case of grammar correction, so for converting grammatical statements into standard English, the prompt will be like you will be provided with statements and your task is to convert them to standard English. Okay? In that case, if the user says 'no went to the market', it is able to correct this 'she did not go to the market', okay. The API request is like this. It's using chat completions class. They're using GPT -4 model. As I explained previously, the system role is available with the API. In case of chatGPT, that interface, we won't be having that option, only assistant and user. Here we have the system role, okay? summarize content you are provided with for a second grade student so that it can explain the concept in a simple way. So instead of these technical things, it converted into simple, easy to understand response. Provided a lot of examples, maybe you can check that at openAI website, okay? 7. Beware of AI Hallucinations: Hello everyone, this video is for explaining about AI hallucinations, first of all, what are AI hallucinations, AI hallucinations occur when an AI generates incorrect, misleading or nonsensical information that appears plausible, AI tools like chatGPT, Google Gemini, Claude, Deepseek, Qwen, these kinds of LLMs, large language models, they're really useful okay they are very powerful, they will be useful for creating articles, creating images, and even they will be able to create coding programming, debugging the coding like that lot of applications are there okay, but the problem is they will hallucinate okay, that means they will provide the incorrect details, but it will be looking like a correct information, that is the dangerous part of a hallucination, if it is obviously incorrect then we can just ignore that part right, but that wrong information or misleading information will be looking like a genuine one okay, so we need to be very careful in identifying AI hallucinations and we should understand that AI hallucinations are inherent nature of a models as of now okay so we can do lot of things to manage them but there is no solid way of completely avoiding them so we need to be very careful with them okay, Causes of AI Hallucinatins, first one is, lack of training data accuracy, so we know that AI models work based on the training data so if there are any problem with training data then obviously it will reflect with the performance of AI models, AI hallucinations will occur if that training data is not accurate okay and similarly over generalization of patterns, if that data is not provided enough to do the generalization properly that time also it will hallucinate it can be like combination of data issues as well as model issues, biases in training data, again related to training data, if the training data is not having balanced information then hallucinations will occur, model limitations, it is important one as I explained previously it's a inherent nature of AI that models AI models are not cognitive, they are not like human brain or they are not having intelligence, they are not having actual intelligence they will work based on the prediction okay what should be coming next like that only they will behave so because of that reason that hallucinations will occur that that I mean lack of cognitive function okay, and prompt misinterpretation, prompt misinterpretation in the sense if we provide the prompt in a vauge or like ambiguous way then obviously that LLMs will not be able to give the correct answers but the problem is they will be giving answers looking like a correct one so again we need to be very careful okay, prompt engineering, we need to be familiar with prompt engineering like a lot of things are there we have to provide clear concise prompts with enough context and we have to use few-shot prompt that kind of approach and chain of thoughts okay step by step we if we are asking the models to do the process step by step then obviously it will help to improve the accuracy okay it will reduce the hallucinations okay so we need to provide a proper prompting okay if you are not providing the prompts correctly then AI hallucinations will occur so these are the causes of a hallucinations okay next, examples of AI hallucinations like AI hallucinations will occur in various formats like it will be like fabricated news articles it can create fabricated news articles looking like a genine one okay non-existent academic references even it can give some academic references which are not actually existing so these kinds of references will make us to believe that information okay so that also we have to cross check the existence of academic references, incorrect medical diagnosis so obviously it is better to avoid AI tools for medical and finance that kind of critical applications okay but we can use it but directly interpreting that response as a final output is not a good way of handling AI tools okay and other thing is false historical fact, incoherent chatbot responses okay so basically overall it will be like giving the wrong information in a believable way so we need to take additional care in identifying AI hallucinations okay this is one example I asked Google Gemini this question 'why did Google fire 250 employees?' okay so based on my Google search there is no such actual news okay actually I found a news saying that Google employees like 250 employees where protesting against Google's contract with Israel something okay so just found that number and framed this question okay but Gemini started giving the answer okay so someone who is not aware of these things if they see this question and this answer they will be believing that this is actually happened like that okay this is one example but a lot of solidly incorrect answers also it will give okay so we need to be very careful, impacts and risks, spread of misinformation okay so if you use chatgpt for example or any kind of AI application for creating your blog post your readers will be sharing that with other people also right and someone will be creating video based on that content so basically that misinformation will be spreading okay a lot of people will be start believing that the misinformation as a correct one okay so that is risks okay, legal and ethical issues, so if some person is getting affected because of that information they will go with lawsuits right so that kind of problems will occur even if they're not choosing to file Lawsuit then again anyway it's a ethical problem right, trust erosion in AI systems, if we are not managing AI systems properly without considering AI hallucinations and if you are taking the AI output as a final thing and you're using it in other systems then obviously people will be like losing their confidence or trust with AI systems okay, AI systems are really powerful there is no question about it okay but we have to use it in a proper way otherwise that overall Trust of the AI systems will be spoiled okay potential harm in critical domains, example Healthcare and finance okay so obviously in healthcare and finance if you are using incorrect or wrong information you will be seeing the impact immediately it will be heavy right, next one detecting and reducing AI hallucinations, okay as I mentioned earlier that hallucinations are inherent nature of AI models so there is no way of completely avoiding hallucinations okay but we can do something for detecting them and for reducing that impacts okay one is cross-check AI-generated content, this is obvious one so whenever you get answers from AI tools like ChatGPT, Gemni, Claude, Qwen, or any kind of other models also first step is we have to check the output by refering to other sources that is one thing and other thing also just we can get the output from one model and cross check with other model so that is another thing but ultimately it is better to cross check with some kind of original sources okay and improve training datasets, if we are using our own models we can provide the training data properly we have to add balanced unbiased training data with the enough data covering all the scenarios like that okay next one is Implement human in the loop verification, we have to use humans for verifying the AI output okay enhance AI interpretability and monitoring, best practices, AI hallucinations are a significant challenge so it's really true okay as I said AI tools are really powerful okay but because of this hallucinations their usages are limited right now in commercial setup right now we can't expect much usage of AI tools because of this hallucination issues and continuous Improvement and monitoring are needed ethical AI development is crucial for reliability awareness and responsible AI usages are key okay so right now the first thing is awareness okay that is the purpose of this video okay but from technical perspective we can do a lot of things okay especially RAG retrieval-augmented generation, we call us grounding the LLMs, okay by giving some external data we can drive the LLMs to avoid hallucinations by taking these external data okay that is one approach that we can use the RAG systems in case of AI applications or in case of simple prompting we can use a few-shot prompting by giving some solid examples so in that case obviously that LLMs will be trying to avoid the hallucinations okay so there are two things one is awareness we need to be very clear about the behavior of AI tools and another thing is using some kind of approaches for reducing the impact okay but as I said we cannot fully control it we need to keep it in mind always. 8. Google Teachable Machine The Easiest Way to Train AI: This video is for explaining about Google's teachable machine. Teachable Machine is a web based tool. We can access it just by typing teachablemachine.withgoogle.com on a browser. So it's a browser based tool. It will be very useful for explaining the machine learning model training. Okay? We can do the machine learning model training of images, audio, and pose. So this tool will be very much useful for beginners to understand the AI. Apart from explaining about AI, it can actually help us to create the trained model so that we can use it in our application. So there are two purposes. One is for explaining the concept. Another one, really, we can use this model in our application. Okay? So primarily, Google started this for explaining about AI concept, but still we can use it for our application. It's actually internally using tensorflow.js for the model. Similarly, while exporting it, we can export it as tensorflow.js file. Okay. You're seeing that explanation, right? So here, um, it is the detection process, right? Testing. So here, it's differentiating the metal and nonmetal, tree and wings, so it's pose detection. Okay? So it's audio. Snap and clap, it's differentiating. Okay? So basically, it'll work for images. I'll work for audio, it'll work for pose detection. Okay? So it's a very simple thing, easy to understand, and you can do everything on the browser itself. Okay? So let me explain it by showing one demo, ok? So let me start by clicking this Get started. So here we are seeing three options, right? Image project, audio project, and pose project. I'm just taking a very simple example of image project. Okay? So choosing standard image model. So here we are seeing class one, class two, training preview, that things, right. So basically, how it'll work is, like, we have to provide training data for each class. Okay? Once after uploading training data, we had to train the model with the data. Then we can start using the trained model to use it for detecting any other images. Okay, that is the overall concept. So here, class one and class two, I'm going to take very simple example cat and dog. Okay? Class one is cat, class two is dog. For giving the image samples. either we can use web camera, real time objects, we can just use that web camera feature or we can upload the already available images. Okay? So I'm just choosing the very simple thing. That is, I'm going to upload the data. I already downloaded a few cat images and dog images from Internet. I'm keeping it like three images for training this and one image for testing that. Okay? So I'm going to upload the training images now, okay? So choosing the cat so I'm loading three cats images, right. And similarly, for the dog also, I'm going to upload. So right now, three cat images labeled as cat, three dog images, labeled as dog. There are different kinds of images, right? So basically for training the model, we have to use various images, and it is better to have more number of images to get better accuracy. Okay? Now I'm going to train it. So the advanced, Epoc things are there. Maybe we need not worry about those things now. I'm going to click train model. Okay? So it's preparing training data. So now training the model, okay? Training got completed. So now we have to give the input image for testing it. Okay. So now I'm going to use a cat now. Testing cat, okay? So now it detected as cat, 100% with 100% confident, okay? So see here, we trained with these three cat images, but this cat is completely different from these three, but still it is able to identify this cat image as a cat, okay? Now we can test with another one. Okay, okay. So now it detected as a dog, okay? So that much simple. Okay? We can export the model. So if you click Export model, we can get the tensorflow.js file. Okay? You can download it. Okay? Even you can share the link also so that anyone who wants to use your model, they can use it. Okay? Actually, we selected training data and labeled as cat and dog and used those labeled images for training the model, once after completing the model. We tested it with input images by giving one cat image and dog image. When giving the cat image, it detected as a cat, 100% confidence, and when giving dog image, it detectd as a dog. Okay? So if you are able to understand this overview, you can try with various things. Actually, I have gone through various samples created with this, uh um, tool, okay? So a lot of very, very innovative things. Okay, because basically you will be able to train the model with a lot of different things, I'm just telling you about the entire thing. Apart from this image, it is allowing to train with audio and pose also, right? So I noticed one interesting example, like with fingers, detecting each finger position, um like they correlated with music notes. Okay? So by doing that, they will be able to generate music just by moving the fingers on the air. Okay? This is just an example. Similarly, you can do it for any number of work. Okay. So you can use it for creative purpose, as well as if there is any solid work needs to be you can just think about how to use this. Okay, I will help you to save a lot of time. Apart from learning purpose, even it can be used for real thing, okay? So I thought of telling this just to give overview about how the training will work. But a lot of other things also there, okay? Um, that are very complex things. Um, but mostly the basic process is this one. You have to choose the proper training data and label that and then train the model. With a trained model, we can detect the object, okay? 9. The Ultimate List of AI Tools to Explore in 2025: This video is for giving some overview about a few AI tools. Actually, I arranged it in a few sections, AI chatbots, AI for image generation, AI for video creation, AI for audio, AI for coding, AI for productivity, AI for research, and fun AI experiments. Okay. Let me show one by one. First one is chatGPT. It is provided by OpenAI, okay? I have already logged in with this. Otherwise, we have to first login, okay? We can use our Google account also for logging into this chatGPT. Okay? You can come to chatgpt.com and you can ask any questions. Okay. It will give answers for your questions. It is operated by LLM GPT-4 model, ok? And we can change the model also. I will show that. But generally, by default, it will be using GPT-4 model. It's a multi modal model. That means we can ask questions in text. We can use images. Okay. First of all, let me just start with a very simple thing. What is 4 plus 5. Okay. It will immediately give the answer. Okay? And we can ask it to write a poem, write a poem about moon. It will be writing the poem. If you feel that the poem is lengthy, you can ask it to rewrite by reducing the size. You can ask it make it short. Or even you can ask it to change the tone, other things. Okay? So you can interactively work with ChatGPT, okay? Not only text, you can use image also. For example, you can upload image, or you can get it from Google Drive. Okay. So you can upload any images, and you can ask questions related to that, okay? So I'm just uploading this 'introduction to AI' that thumbnail image, okay? So I can ask questions. Like, first of all, I can ask you to tell the text on that image, okay? Tell the text on this image. So it's kind of OCR. It will be recognized in the text, and it will give the sentence, okay, 'Introduction to AI', okay? Apart from that you can ask it to describe it what color, how it's looking like that. Okay? Not only this one, we can even ask ChatGPT to create images, okay? Maybe we can ask it to create a similar image, create an attractive image to show this text. Okay, it will try to create the image. Even you can give instructions, like the background should be like this. use this kind of font like that. Okay. Now, introduction to AI. So it's somewhat better than this, right? We can use it for promotion, other things. Okay. Suppose if you have created video like this, and if you want to promote it on Instagram and Twitter, you can use this image. So a lot of things are there. Not only this one, even in real time, also, we can use to get some details from images, text details. Okay? It will save a lot of time and time for typing the things. But one thing is sometimes it will be the answers may be wrong, so we need to be very careful because of hallucination, okay? And if you want, you can use this speaker icon to hear the text as Audio. And similarly, you can use this icon for Voice mode icon for speaking. Instead of typing, you can just ask questions by uh, clicking this mic. Okay? And moreover, you can use chatGPT for writing coding also. We can ask it to write a Python code to reverse string. Not only writing code, we can just copy our code here and we can ask it to debug it or if you want explanation, it will explain one by one each sentence. Okay. So it's really useful thing. You can use it for creating Ebooks also, okay? It is just to overview about ChatGPT. But a lot of things are there. We have to use prompt engineering. Next one is Gemini, Google Gemini. It is almost similar to ChatGPT. Here, also, you can ask any questions. You can use images and audio. Okay? So almost same as ChatGPT. Okay, create an image for my EBook cover. We can explore various LLM chatbots so that we can get better answer. We can just go through various things like ChatGPT, Gemini, Deepseek like that. Okay. We can use the best one, okay? And even you can cross check it. You can ask it to create an article from ChatGPT and paste that content in Gemini and ask Gemini to verify whether any error is there like that, okay. Almost all the things are similar, okay. And similarly, perplexityAI. It's viewed as search engine. We can get latest information also here. Okay. Here, also, we can ask questions. Almost all these things we have to log in, okay? I already logged in. Otherwise, it will ask to log in, okay? These are all AI chatbots. Like a lot of other things also there. Like, Deepseek is right now, it's growing very fast, okay. And next one is AI for image generation and Editing. Dall-e is image generation model provided by OpenAI. It is already integrated with chatGPT. That is why we are able to create images within chatGPT itself, okay? So we can give text descriptions, and it will automatically generate images based on our descriptions. Okay? A lot of applications are there using Dall-e. You can use them. Even you can use for creating your own application. Okay? Next is Canva AI, magic design, magic edit. So we can see various features. Make me an image, write my first draft, resize any design, remove backgrounds. Okay? All of them are using AI, okay? We can ask to create image, create an image with dog head and peacock legs. Okay? I'm just telling as an example, but you can use it for any kind of real things, okay, for your marketing purpose. Next one is remove.bg for removing the backgrounds from the images using AI, okay? So here, you can just upload the image, Okay, and then ask it to remove the background. Okay? It will remove it. Here. They provided some samples, okay? It will save a lot of time for doing image editing, right? Next one is AI for video creation. The famous one is runway ML. So here, we can start using it. They provided a lot of examples. Next one is, Synthesia This one also free AI video generator. You can just give text for creating videos, okay? They also provided samples, just go through that. And similarly, the Pika labs. So and AI is useful for audio generation as well as transcription This ElevanLabs is famous for creating voice. You can just give the text. I will convert into speech. Okay. You have to just give the text and then you can choose some options, then it will create the audio for you. And this one Whisper provided by Open AI. It is for speech to text transcription tool. So we have to upload the Voice, and it will convert into text. It will save a lot of time for doing transcription. Okay? This Voicify AI will clone your voice, okay? any voice to generate further text. So I'm just giving as an example, a lot of things can be done. Okay, so just a few examples, you can just go through those sites and try it by yourself. Okay. Apart from that, AI can be very useful for programming also. So for example, in case of Github copilot, it can help you to write the coding on your favorite IDE itself, okay? So if you provide the code and you can ask it to fix the error. Okay? It will be fixing the error and updating the code itself, okay? Similarly, other tools are out there Codeium, tabnine. And productivity for productivity, the notion is very famous. Almost we can do a lot of things within notion itself. Okay, so it will be saving a lot of time. So a lot of people are using this to improve the productivity. Okay? You can do the search, generate documents, analyze insights to get insights. We can chat with any kind of AI chat bots. Everything from within the notion itself, okay? You can see the samples here. And similarly, grammarly, I hope everyone will be using it for seeing grammar check, right? It'll be very useful for writing. And for research, a lot of research assistants are there for summarizing any academic papers. So it will save a lot of time. Okay. And a lot of fun experiments also there. So this site, this person does not exist. It will be randomly generating a human image. Actually, it's not image of any particular human. It is AI-generated. Okay. So it's looking like a real person. If you click refresh it, it will generate new images. Just for fun purpose so that we can understand the power of AI and this deep nostalgia you can give some old image. It will automatically generate video for that, for example, for this image, it created this video. A lot of things they provided. You can try by yourself. So all these things just for giving some overview about how AI can be used. But in real application, it's almost useful in almost all the industries, medical. So they're useful in identifying diseases by seeing the scans. Okay. Even it is useful for creating new kind of medicines. Okay. So just to understand this, just to get familiar, you can go through these things so that you can understand the AI power, okay? 10. AI Jargon for Beginners: In this video, I'm going to explain about a few artificial intelligence AI related terms. Let me first start with what is an AI model. An AI model is a computer program that learns patterns from data and makes predictions or generates responses. That means first we have to train the AI model using data, then the trained model will be able to make predictions. Some AI models are trained from scratch using raw data. And we already have a lot of pre-trained models also so that we can start doing the development quickly without spending time for training the models, okay? But we have to do some kind of fine-tuning to make it work for us. So next, pretrained model. So pretrained models, as I said, there are a lot of pre-trained models are there there like BERT for NLP tasks, natural language processing and whisper for speech recognition and Dall-e for image generation. Okay? So these kinds of pretrained models are trained with a lot of data, huge data with huge power, computing power. We need not spend that kind of effort for getting that model. So we can just fine tune the model for our own requirements, and then we can start using it. That is the next thing, fine-tuning. So fine-tuning is the process of taking a pre-trained AI model and training it further on a smaller specialized dataset, our own dataset to improve its performance for a specific task, our own task. For example, the general chatbot like GPT can be fine tuned on legal documents to provide more accurate answers to legal questions. Similarly, image recognition model can be fine tuned on medical scans to detect diseases more effectively. The next one is what is inference in AI. Inference is the process where an A model applies what it has learned during training to generate an output or make a decision. Basically, it's prediction or using the trained model. Whenever we say inference, that means we are going to use the trained model for our purpose. The examples are, when you ask chatGPT a question, it analyzes the input and generates a relevant response based on its training. Similarly, a self-driving car uses inference to recognize traffic signs and make driving decisions in real time. Next question is what is a dataset in AI? Dataset is a structured collection of text, images, numbers, or other types of data used to train and evaluate AI models. Whenever we do the training of AI model, we need to have a good quality data arranged in a proper way so that training process is effective. And once after completing the training, we have to evaluate test the AI models. For that also, we have to prepare data. All these data things together called as dataset. Next question is, what does bias in AI mean? Bias in AI is a very key topic because the AI model can be trained with data. But in case if the data is having some incorrect information, that functionality of AI model will be not correct. Okay. So if there are any issues with the data, it will reflect in that model's performance. It is called a bias. Then what is zero-shot learning? Zero-shot learning allows AI to handle tasks it hasn't seen before. For example, if a chatbot correctly answers a question on a new topic without specific training, it is using zero-shot learning. For example, if you are using a model, which is not trained with zebra images, but it is able to detect zebra images as zebra, then it is known as zero-shot learning. Basically, it will be using the text descriptions to identify. without image training, it'll be able to do that. This is known as zero-shot learning. Then what is neural network? Neural network is a type of AI model designed to mimic the way human brain processes information. It consists of layers of interconnected artificial neurons that learn patterns from data. Each neuron receives inputs, applies weights, and passes the result through an activation function to make decisions. Neural networks are widely used in deep learning for tasks like image recognition, speech processing, and language translation. Some examples are CNNs for image classification, conventional neural networks, CNN means and the RNNs for speech recognition. It's basically it's trying to copy the human brain activity. The weight in the sense, internally, based on the training data, it will apply the weights so that the loss will be less. So it's basically something similar to our biological neurons. Biological neurons that information is based on the strength of the connection between the neurons, right? The same concept it's applicable for this artificial neural network also. Then what is tokonization in AI? Tokonization is the process of breaking text into smaller tokens. So, AI can understand it. For example, 'AI is powerful'. it will be split into 'AI', 'is' , 'powerful'. AI processes these tokens instead of full sentences. Usually, in case of chatGPT, a token will be around four characters. Then what is generative AI? Generative AI is a type of artificial intelligence that creates new content such as text, images, music, or code rather than just analyzing existing data. It learns patterns from large datasets and generates output based on input prompts. For example, chatGPT creates human like text responses. Dall-e generates images from descriptions and musenet composes music. Generative AI is used in creative writing, design and automation to enhance productivity and innovation. So this generative AI is playing very important role these days, that large language models, LLMs like chatGPT, that GPT-4 model. They are trained with a lot of data. Instead of training the model by ourselves, we can just start using that kind of LLMs using API, and then we can create our AI applications. Okay? So genertive AI is playing important role these days. What is hallucination? A hallucination occurs when an AI model generates false, misleading or nonsensical information. that sounds convincing, but isn't true. This happens because AI models predict responses based on patterns in their training data and make guess when they lack accurate information. For example, a Chatbot might invent a fake citation and provide incorrect historical facts. Reducing the hallucinations requires high quality training data, better fact checking and techniques like grounding AI responses in verified sources. So basically, it is inherent nature of the AI. Okay? So as of now, there is no solid way of preventing it. If you are familiar with using ChatGPT and Gemini, you might be knowing this. Sometimes even for a small simple questions, chatGPT or Gemini, we'll be giving some wrong answers. The Hallucination is a very major problem in AI. But for handling this, we have different approaches, for example, prompting tricks like chain of thoughts and RAG Retrivel augmented generation. By giving some external data using RAG, we can just try to mitigate the hallucination issue. But still, it's a major issue. What does Overfitting mean in AI? Overfitting happens when an AI model learns patterns from training data too precisely, including noise or irrelevant details, making it perform poorly on new unseen data. It is like a student memorizing answers instead of understanding the subject. Overfitting can be reduced using techniques like regularization, dropout, data augmentation and cross validation. Overfitting is like say, for example, we will be having, um, training data or training the model, that model will be fully trained for best suitable for the training data. So if you use the training data for the testing purpose, it'll be working properly, 100% confidence like that it will be giving. But if you are going to use a similar image, but not exactly the same as training, in that case, it won't work. Okay? So it will be working perfectly for the training images, but it won't work for test images or any kind of actual usage images. For that we have to reduce this overfitting. Okay? So there are various approaches, data augmentation. Data augmentation in this sense, say, for example, if you're going to train the model with a few images, data augmentation in this sense, it will be like, flipping the image, creating mirror image, changing the colors like that. We can do some kind of processing, and then we can expand the data training data. With that expanded dataset, we can train the model to avoid overfititin, another reason for overfitting is if we use more than required hidden layers or number of neurons. The over usage of neurons also will cause overfitting. Okay? And then what is underfitting? Underfitting occurs when an AI model fails to learn meaningful patterns from the training data, leading to poor performance on both the training set and new data. This happens when the model is too simple or lacks enough training time to capture the underlying relationships. It is like a student studying too little and struggling with both practice and real test. Underfitting can be addressed by using a more complex model, increasing training time or providing more relevant features in the data. Underfitting in this sense, the model cannot work properly even for the training data itself. That means the data is not that much good and the model is not having the ability to process it properly, and the training time is not enough. In that case, we got to use some good model, complex model and we have to increase the training. And the training data, you got to add more features. In the training data in the sense that dataset will be having features and labels. I'm just telling simple example. In that case, the number of features or quality of the features should be increased to avoid underfitting. Then what is transfer learning? Transfer learning is when an AI model trained on one task is re used for another similar task. For example, a model trained on general images can be fine tuned to identify medical scans. It can help us to save a lot of time. We need not do the training fully again. Then what is embeddings in AI? Embeddings are numerical representations of words, images, or other data that help AI models understand relationships and similarities between them. They convert complex information into dense vectors in a way that preserves meaning. For example, in a word embedding model, the word king is mapped to a vector close to queen, capturing their semantic relationship. Embeddings are widely used in natural language processing recommendation systems and the image recognition to improve AI peformance, normally, we will save store the embeddings in vector databases. It is very much useful in retrieval augmented generation RAG. It's semantic meaning. It's for keeping the semantic meaning. It will be useful for semantic search also. What is vector database? A vector database is a specialized database designed to store and search embeddings. It enables AI models to quickly find similar images, documents or text by comparing vectors based on their mathematical proximity, something like cosine similarity like that. This makes it ideal for tasks like image recognition, recommendation systems and semantic search. Popular vector databases include FAISS provided by Facebook, pineCone, managed vector database, and ChromaDB, it's open source. Even we can use it in our local machine. What is natural language processing, NLP. NLP is a branch of AI that enables computers to understand, interpret and generate human language. It powers applications like chatbots, voice assistance, Siri, Alexa, language translation, like Google Translate and sentiment analysis. NLP combines linguistics and machine learning to help AI process text and speech in a way that feels natural to humans. Whenever we use A application like chatGPT, we can ask questions in natural way using languages like English. what is supervised learning? Supervised learning is a type of machine learning where an AI model is trained using labeled data, meaning each input has a known correct output. The model learns by mapping inputs to the correct labels and improving its predictions over time. For example, training an AI to recognize cats and dogs involves showing images labeled as cat or dog, so it can learn to classify new images accurately. Supervised learning is widely used in image recognition, speech recognition, and spam dection. Almost all the traditional machine learnings are using this approach. Okay. So we have to just label the data. And then we have to use that labeled data for training the model. What is unsupervised learning? Unsupervised learning is a type of machine learning where an AI model discovers patterns and structures and data without labeled answers. Instead of learning from pre defined categories, the model identifies hidden relationships such as grouping similar items together For example in customer segmentation, you can cluster customers based on purchasing behavior without being told which group they belong to. Unsupervised learning is commonly used in anomaly detection, recommendation systems and data compression. In case of unsupervised learning, we need not specify the labels. It will find the patterns automatically. What is reinforcement learning? Reinforcement learning, RL is a type of machine learning where an AI learns by trial and error receiving rewards for good actins and penalties for mistakes. Over time, it optimizes its decisions to maximize long term rewards. RL is widely used in robotics, game playing AI, example, Alpha go, self-driving cars and personalized recommendations. Basically, it's like, giving feedback kind of stuff, right? If it is behaving correctly, then it will be rewarded. Otherwise, penalties for mistakes. What is LLM? LLM, it's a very key thing. Okay. It's a large language model. What is explainability in AI? Explainabilty refers to how we can understand and interpret an AI models decisions. Some AI models are like black boxes making decisions without clear explanations, while others provide insights into their reasoning. This explainability is becoming important factor from the perspective of ethical perspective, regulations perspective. It is important to know how the AI is internally taking decision about the output. Why it's giving that information. But there will be some trade off. If you're focusing on that thing, the efficiency of effectiveness of actual working may be reduced. What is deep learning? Deep learning is a subset of machine learning that uses artificial neural networks with multiple layers deep neural networks to process data and learn complex patterns. It excels at tasks like image recognition, speech processing and natural language understanding. Deep learning powers advanced AI applications such as facial recognition, self driving cars, chatbots, example, chatGPT, and medical image analysis. Deep learning it's neural network. What is a large context window in AI? Large context window means an AI model can remember and process more words in a conversation or document at once. for example, a chatbot with a small window might forget earlier parts of a discussion, while one with a large window can maintain context better. So in case of ChatGPT, if you are discussing with ChatGPT it will be remembering your previous questions, also, right? So if you're asking a question about your computer, then you can ask about it further without specifying the computer, right? it isable to recognize your question properly because it is able to maintain the context window. So if it is a large context window means, it will be keeping many number of previous questions or discussions, right? So whenever LLM is specified, they used to this context window also so that we can understand their performance. What is few-shot learning? Few-shot learning allows AI to learn a new task with very few examples. For instance, if you show an AI, just a couple of labeled images of a new object, it can recognize similar ones without extensive training. In case of few shot-learning, we need not do a lot of training. Just showing few examples will help. Even in case of prompt engineering, for using ChatGPT, something in the prompt itself, we can give some examples, and then we can ask questions. Mostly chatGPT we will be able to answer our questions. What is multimodal AI? Multimodal AI can process and understand different types of data like text, images, and audio together. For example, ChatGPT can analyze both text and images in a conversation. GPT-4 is a multimodal model. We'll be able to handle both texts, we can ask questions, we can ask questions by speech, we can ask to generate image,s we can upload the image and we can ask questions. about the image. What is computer vision? Computer vision is a field of AI that helps machines interpret and analyze images and videos. It is used in applications like facial recognition, self-driving cars, and medical images. It is also important field, especially in self driving cars. It plays important role and medical imaging also, it's a growing field. Computer vision techniques are performing better than the radiologists in identifying the diseases from the scan images. Then what is pre training versus fine tuning? Pre-Training is when an AI learns general knowledge from a massive dataset before being adapted for a specific task. Fine tuning is the process of further training the AI on a smaller specialized dataset to improve performance on a particular task. Pre training in that sense, if image recognition model is there, we can fine tune it for specifically identifying the medical image. It's a good way of balancing both things. The pre training that huge dataset will help in some way, and narrowing down our specific field with a fine tuning dataset will help in some perspective. It's combining power of both things. What is data augmentation in AI? Data augmentation is a technique used to expand the dataset by modifying existing data. For example, in image recognition, flipping, rotating or changing brightness in images helps AI models plan better without needing new data. If you observe the Yolo model training, you can find this thing. Even if you use some ten images for the training purpose, if you go through the files created by Yolo model, you can find a lot of images. Some of them are flipped, some of them are mirror images, rotated images, okay. So just by having very little data, the data augmentation process will increase the dataset. What is the Loss function in AI? Loss function measures how far the AI predictions are from the correct answers. AI models adjust the learning based on the loss function to improve accuracy. It's a key thing. Okay? So when we're measuring the performance of accuracy of the model, the loss function gives the details about the error, basically, the error Based on that value, that model will be changing their weights. What is gradient descent? Gradient descent is an optimization algorithm that helps AI models adjust their internal settings, that is weights to minimize errors and improve learning over time. It's something similar to backprobagation algorithm, right? What is model drift? Model drift happens when an AI model's accuracy decreases over time because the real world data it encounters has changed from the data it was trained on. This often happens in fraud detection and the recommendation system. It's a deviation of model output, okay? It is because of the usage nature. So in case of recommendation systems, if the viewers are viewing in a different way, then obviously the model output, previously trained model will not be a suitable in this new scenario, right? So it is known as model drift. What is catastrophic forgetting? catastrophic forgetting happens when an AI model trained on new data forgets what is learned from previous data. This is common in Deep learning models that don't store past knowledge effectively. If you train the data for cat, then after some time, if you train it for dog, then it will forget the identification of cat, for that we can follow different approaches. Whenever we train new data, maybe we can train with the previous thing also. or lot of different approaches there. Okay? What is explainable AI? It is about giving details about how the model is working and why it is giving this out like that. It will help to improve that trust and accountability. What is reinforcement learning with human feedback RLHF? It is a technique where AI is trained using human feedback to improve its responses. This is how models like chatGPT are fine-tuned to be more helpful and accurate. Whenever you use chatGPT, you'll be having that thumbs up and thumbs down buttons. It's by clicking that buttons, you're giving your feedback. It's reinforcement learning with human feedback. What is federated learning? Federate learning is privacy focused training method. Data remains on users devices instead of being sent to central server. AI models learn from distributed data without compromising user privacy. It's keeping training data on the user's device itself. Instead of centralizing it, you are keeping the distributed system. What is a transformer model, a transformer is a deep learning model architecture designed to handle large scale language processing tasks. Models like GPT, BERT and T5 use transformers to understand and generate text. Mostly the transformer models revolunaized LLMs Because they are able to process all the text simultaneously. What is a knowledge graph. A knowledge graph is a structured information that connects concepts and relationships. Google search and chatbots use knowledge graphs to provide relevant answers by linking related topics. So it's connecting the information, node, that kind of stuff, what is retrieval augmented generation RAG. RAG is an AI approach that improves the responses by retrieving relevant information from an external database before generating and it helps chatbots provide more accurate and up to date responses. In case of RAG apart from sending our prompts to the LLMs, we will be able to send some external data also along with the prompt relevant external data. So for that, that retriever model is working. For that, we have to use the vector databases like ChromaDB, or PineCone with this setup, the RAG will be able to improve the accuracy. What is an AI accelerator? AI accelerators like GPUs and TPUs are specialized hardware designed to speed up AI computations, making training and inference faster. It's a hardware related stuff, so that AI training and inference can work faster. What is Edge AI? Edge AI runs AI models directly on local devices like smartphones and cameras instead of relying on Cloud servers. This allows for faster processing and the privacy. In case of Edge AI, the models will run on the mobile phones itself. What is Quantization AI? Quantization reduces the size of AI models by using low-precision numbers, making them run faster and more efficiently, especially on edge devices. Quantitation is helping to achieve this Edge AI, what are parameters in AI models? Parameters are the internal variables of an AI model that are learned during training. They define how the model processes input data to make predictions. For example, in a Neural Network, parameters include weights and biases that connect neurons. The more parameters a model has, the more complex it can be. When you hear about models with billions of parameters, for example, GPT-3 with 175 billion parameters, it means the model has a vast number of internal variables, allowing it to capture intricate patterns in data. So in case of more parameters, then it can work better. What is precision in machine learning? Precision is a metric used to evaluate the performance of a classification model. It measures the accuracy of the past predictions made by the model. So these precision and recall that kind of measurements, metrics are very much important. These metrics are very useful in estimating evaluating the quality of the model. Precision is it's kind of accuracy indicating accuracy. If your model predicts 100 emails as spam and the 90 of them are actually spam, the precision is 90 percentage. The formula is precision equal to true positives that is 90 of them actually spam 90 divided by true positives plus false positives. 90 by 100, 90 percentage precision. High precision means the model is good at avoiding false positives. The recall, it is also called as sensitivity. It measures how well your model identifies all relevant positive cases. For example, if there are 100 spam emails in a dataset and the model correctly identifies 80 of them, recall is 80 percentage. High recall means the model is good at finding most of the positive cases. Precision focuses on accuracy of the positive predictions. That is avoiding false positives. Recall focuses on capturing as many postive cases as possible, avoiding false negatives. so, basically sensitive In a medical test, high precision means most positive test results are correct. High recall means most actual cases of the disease are detected. Often, there is a trade off between precision and recall and the right balance depends on the use case. What is the F1 score? The F1 score is a single metric that combines precision and a recall. It is the harmonic mean of the two and is useful when you want to balance both metrics. The F1 score ranges 0-1 where one is the best possible score. It is commonly used in classification tasks, especially when the dataset is imbalanced. What is the confusion matrix? confusion matrix is a table used to evaluate the performance of a classification model. It is basically helping to calculate metrics like precision, recall and accuracy. Okay? Accuracy measures the percentage of correct predictions made by the model. So basically accuracy true positives plus true negatives divided by total predictions. While accuracy is a useful metric, it can be misleading in imbalanced dataset. For example, when 90 percentage of the data belongs to one class. In such cases, precision, recall, and F1 score are more informative. What is a loss function? A loss function measures how well an AI model performs by comparing its predictions to the actual values. So basically, it's error. 11. Python for AI Developers: Hello everyone, this video is for telling about using Python for doing AI development. First of all, we need to understand why we have to choose Python for doing AI development, right? The first reason is simplicity. Python syntax reads like English, so we can focus more on doing AI development instead of worrying about syntax or grammar of that language, okay? So that is one of the reasons for choosing Python. The main reason is that library power. Python is having a lot of libraries, especially libraries related to AI. We can find a lot of libraries because many people are using Python, a lot of open source networks are there for supporting Python. So we have a lot of related libraries. Library in the sense we can just import that library and we can just start using their methods or functions in our application. So it will help us to save a lot of time. Let us see some important libraries, NumPy for numerical computing and Pandas for data analysis. Pandas is a very, very important library. It is widely used for data science related projects because data science is important for AI for training the model. We have to provide a very quality data. Then only the model performance will be good. For creating quality data, we have to do a lot of data processing. So for that, Pandas will be very helpful. It will be helping us to save a lot of time in preparing data in a required way, okay? And for plotting graphs, we can use matplotlib for simple graphs, for advanced features We can go with seaborn library, okay? And scikit-learn, and scikit-learn is useful for training and using traditional machine learning models. That means other than deep learning models, okay? For deep learning models, we can use with tensorflow and PyTorch that kind of libraries we can use for training deep learning models, for training, for fine tuning, for using DL models. We can use tensorflow or PyTorch, okay? And hugging face is providing transformers library that will be useful for NLP, natural language processing. Especially the pipeline method of transformers libraries is really a very simple way of doing NLP, okay? Even we need not specify the model name. We can just specify our requirement like classification sentiment analysis. And within like by writing few lines of code, we can create a useful application, okay? Next reason is massive community support. Obviously, we will be getting a lot of support because a lot of people are using Python, right? So if we have any questions related to Python, we can just post our questions in any forums and we can get a reply, relevant reply, okay? Or already someone might have asked the same question, we can just search it and get the answer, okay? And one more reason is Cross-Platform compatibility, okay? So Python will be used on various operating systems like Windows, Mac, Android Linux. That is one reason for using Python. And industry approved, top companies like Google Meta, that companies are supporting Python. That means whenever they release any kind of AI related applications, they provide sample coding in Python also. So we can easily copy that code and start exploring their applications. So these are all the reasons for using Python for a development, okay? Let us go through in detail, but this video is not for teaching about Python, just to give some more to be about using Python for AI development, okay? I'm not going to teach anything specifically here, okay? So first of all, for installing Python, we can go to Python.org and then we can download it and we can use it on any operating system, right? And IDE integrated development environment that code editor, a lot of things are there. I am using VS code, okay? It's lightweight, a lot of extensions are there, so beginner friendly, okay? So here we can get a lot of extensions, Python related extensions are there, okay? And other things like PyCharm, Jupyter notebooks, it's important one, it's very useful. Okay? So for installing libraries, we can use pip command, pip install library name, okay? For installing Jupyter notebook, we can install like this pip install Jupyter notebook. So as I said, Jupyter notebook is very important concept, it's IDE, but it will be very useful for learning Python, especially for beginners, it will be very useful. And it will be useful for not only for learning, for showcasing our work to other people, non-technical people, it's very easy because it's mix of code, notes and output, okay? I will show that, okay? And before going into Jupyter notebooks, first we can check this Google collab. It's very, very beginner friendly. Jupyter notebook, we have to install, we have to set up in our local, right? The same thing, it's available with cloud, Google collab. So we can just, we see that website, colab.research.google.com. And we can just start using that notebook, okay? In this case, there are two advantages. One is, no need to do installation setup other things. And another thing is, instead of using it in our local, we are going to use it in Google's environment, right? So that GPU, TPUs, RAM, that things will be better than our local, okay? So that will be an additional advantage of using Google colab, okay? So we can just start using it by visiting colab.google.com. It will be redirected to colab.research.google.com, okay? I will show that first. So this is that, okay? It's a new notebook, I created it. And here, as I said, it's combination of code and text and output, okay? We can create, for example, text, text saying that it's a demo, okay? And here we can create code, okay? And we can execute it. Since it's a first time, it is connecting to Google that hardware, right? That is GPU. It's taking little time, okay? Output is shown now, okay? So now, second time, it will be very fast, okay? Here we can choose various options, okay? So this is the basic concept. We can enter text markup here and add the code. And we can see the output, okay? Everything is in one place. And similarly, we can have more number of such cells. We can add any number of code cells, any number of text cells, okay? And we can execute them individually also, okay? So that will be very useful for studying purpose, learning purpose, as well as, as I said, for showing our work to some non-technical people also, okay? So it's a very useful one. That file extension will be like IPYNB, okay? We can download it. It's all about Google Colab. And the same thing can be used locally using Jupyter Notebook. So we can just start Jupyter Notebook. So we have to just start like this, Jupyter Notebook. It will be opening the Notebook. I already started creating here, okay? So to save time, I already prepared everything here. And we can just execute one by one, okay? So in case if you're a first time coming in, it will be like this new notebook. Here we can use it like colab, right? Here also we can add text. That's it, okay? Similarly, we can add any number of cells, okay? So a lot of features are there. You can just explore that. It's very useful for beginners, as I said. If you're a very beginner, start with a colab, and once you're familiar with setting up like a pip install other things, just install Jupyter Notebook, and then you can start using it. And the same Jupyter Notebook, you can use it from here like VS Code also. While creating file, you can create Jupyter Notebook. But for that, you have to install Jupyter Notebook that extension, okay? Okay, since it's the first time, it's connected to Python kernel, right? So it's taking some time. Next time I'll use it, it will be very fast, okay? Here we can add more down as well as code, okay? Or many number of cells we can have, okay? So this is all about Google Colab, and Jupyter Notebook. It will be very useful for beginners for learning Python. So explore that, and you can get a lot of sample IPYNB files from various places, okay? just get that files, and you can just open here, and then start using that. Or you can step by step, you can learn, okay? So these are the Python basics, okay? So as I said, I'm not going to focus here, just to give some overview, okay? So Python supports various data types, like string, integer, float, Boolean, and various operators, okay? So just as we are assigning the values right for A and B, and just printing A plus B, and a division, A divided by B, we can see the result by running it. And similarly, it supports a lot of different kinds of data structures, like list, tuples, sets, dictionaries, okay? Each data structure is having its own characters, okay? So mostly used one is list, and the dictionary is also getting used, okay? So just go through that, how to use it And their key features, like list is mutable, tuple is immutable, okay? Set will be having unique values, dictionaries, key value pair, okay? It's a very basic data structure with Python, okay? And the list means we have to use this bracket, for tuples this one. And for the set it is brace, okay? Key-value pair means Key and colon, in that way okay? And key, colon and value, okay? So it is very basic data structure with Python, just to familarize yourself. Especially list and dictionary, okay? And functions in Python is like this, def, and function name, and argument, and return, that return value, okay? It's a basic function in Python, and we can call it like this function name, and then sending the parameters or arguments, okay? So the function and printing that return value. So it's passing this rajamanickam.com visitor, okay? And adding that hello using this F string, okay? So put this hello rajamanickam.com visitor, okay? So learning Python is very interesting only. You have to keep on exploring various things by really seeing the value, and then changing the things that will be very useful to learn quickly. Say for example, you can suppose if you want to add two parameters, how to do that, like that, if you keep on changing the same code, means you can get better in understanding those things, okay? You try various things like this, okay? So as Python is a very general purpose programming language, it is having all the fundamental things, like if condition, for loop, while loop, okay? Just I provided a few samples here. It is for addition and the power. This is conditional statements, okay? Assigning value to the age, and using if condition, we are checking whether age is greater than or equal to 18. If it is greater than or equal to 18, printing 'you are eligible to vote' otherwise, 'sorry not eligible', okay? And we need to understand the intent in the Python. To make it simple, there is no need to use any kind of brackets or things, but we are achieving that by providing this intent. So we have to do it properly. So it is always better to use the supporting editors, okay? In case of VS code everything, it will be done automatically. Okay, we can run it. So since the value is 18, it's saying that 'you are eligible to vote' if you make it 17, saying that, 'sorry not eligible', okay? In case we are dividing by 0 means it's giving error, right? So in that case, we have to do the proper error handling, okay? Here the exceptions are there, okay? We have to use the Try block and Except, okay? So here in this case, it will be giving our custom error message, "can't divide by 0", okay? Similarly, we can use for loop, especially this range that function will be very useful. If we are using range 5, we can use 5 iterations, okay? So 5 iterations are there. Similarly, similar to any kind of language, we can use while loop. So it's giving the output. So mostly the syntax will be very simple only, but we need to take care of the intent, okay? And it is object-oriented language. Class is blueprint, right? Object instance of class. Here no need to use any kind of keywords, okay? Just so we can define in this way. Just so we have to define the class like this, and we can just create the instance of that class as object without specifying any keywords here, okay? Here, class definition is here, and it's having this method bark, saying that the name says Woof, okay? The name, it is passing through here, constructor, right? So while creating the object, that name is passed. And when we use this bark with that object, it is saying that, "Buddy says Woof" Right? "Buddy says Woof" Just to pass this name, I mean this string, and here, using F string, okay? So as I said, a lot of libraries, AI related libraries are there. For example, this numpy library, we can see one sample, or we can give alias name import numpy as np, okay? Then we can just use the alias name only, np.array, it will be easy, right? Instead of numpy, we can just use as np, and here just we can create arrays like this, and we can just print it. And we can find the mean value just by using this mean, okay, arrays there, and we are showing in the matrix form, and then mean of that. And the next one is pandas. As I said, it is very useful for data analysis, and especially that data frames that are very useful for processing data. We can easily import a CSV file data here, and then do a lot of processing using very simple coding, okay? We are importing pandas as pd, okay? And the data, we are providing this dictionary, right? key and value, and we are loading to data frame, just printing the data frame, and then we can find the average score just using mean. As similar to numpy, but in case of pandas, the data frame is providing additional things. We can get data from any source, CSV, other things, okay? It's with the average. And as I said, this matplotlib and the seaborn are useful for data visualization, let us see one example using matplotlib, a pyplot, we are aliasing it as plt, alias name plt, and then we are just plotting it, giving the title and the label names, and then showing it, okay? So it's creating this graph, okay? If you want to change anything, you can get change it. And this example is for using seaborn chart, okay? We have to import seaborn, okay? And we are using boxplot , it's creating this graph, okay? So for simple things, we can go with matplotlib, if it is involving any complex things, or advanced features, we can go with seaborn, okay? scikit-learn, it's very important library for machine learning models, okay? It's also called as sklearn It is useful for classification, regression, clustering, okay? As I said, I'm not going to give importance for explaining these things, just to give some overview, okay? I already uploaded various videos for individual topics. it's one sample thing, okay? So here, so for that, we have to first install a scikit-learn, okay? I already installed it. So this sample is loading iris data, it's data about flowers, okay? The features or properties are that dimensions of flowers, and the classification will be done based on the dimensions, okay? So here, first we are like splitting the loading the data, okay? Then splitting the data using train_test_split, and then training the model, okay? This is DecisionTreeClassifier model, okay? We're training with the fit, okay? And then evaluating the model to find the accuracy of things, and then using it for actual thing, we are just passing sample dimensions of one flower, and that model is predicting that, finding that flower name, okay? So all these things we are just doing from here, running it at evaluation 0.93, okay? pridicted as setosa, okay? If we change these numbers, it will predict as something else, okay? It's predicted something, okay? So this is the fundamental of AI, right? So everything we are able to do it in a simple way, because of these Python libraries, okay? So similarly, we can do it for deep learning models also, okay? It's for traditional machine learning model. For deep learning model, we can use PyTorch tensorflow, I think, okay? And this one is working with data, okay? So as I said, processing data is important for effectively training the models. So for that, we can use Pandas effectively, so let me show an example. So here we are loading this sample data CSV file using read_csv method of pandas, okay? And we are just showing the first five rows. Then we are doing some kind of data processing, just for example, I'm showing removing extra spaces, okay? And then replacing the dollar sign and then showing it again, okay? So the sample file is this one, okay? Here, this price is having space, right? This is a column, right? And one is having a dollar, another row is not having that. So we are just trying to make it clean, okay? So that the training will be effective, okay? So I'm just running it. First time it's saying it is second time, okay? So first time it's showing that space, right? Okay? And here now it's removed that space, unnecessary space, and the dollar symbol also removed. So now it's clear, right? So we can use it for training a model, okay? So just trying saying as examples, similarly we can do a lot of data cleaning, other processes, okay? A lot of steps are there. So for that we can use Python libraries effectively, okay? A lot of things, renaming columns, removing unwanted characters, converting data types, filtering rows, columns, okay? A lot of things we can do, and handling missing data, outliers okay. So if you are not doing that, then that output may not be effective, okay? And similarly we can create this BMI column with just one line of code, right? So this kinds of feature engineering can be done very easily, okay? Okay, as I said, there for deep learning also, we can use Python libraries, like TensorFlow and the PyTorch. And for NLP natural language processing also, Python is having various libraries. Especially the tokenization is very important, okay? That is just breaking text into smaller pieces like words and sentences. For that we can use the NLTK library, okay? So I'm just running this using NLTK. And the text is "AI is amazing, NLP helps machines understand language". And using this Word_tokenize, we are getting words, reading as words, and here we are splitting as sentences, and here showing it, okay? These are the words, sentences, okay? So it's a very simple task. Just for example, but a lot of things are involving like text processing, okay? This is just for example, but we have a lot of libraries for doing that kind of stuff, Python libraries. That is the reason why we have to choose Python for AI developments. And the word embedding is an important concept. Like embedding, it's like vectors, okay? We can turn any words into vectors, okay? That is like it will be keeping some meanings and relations, it's also semantic, search other thinks it will be useful, okay? And for that lot of techniques are there, TFID is common one, okay? Let us see one simple TFID this Term frequency that one, okay? So for that also, the sklearn is providing support. We can just use this TFID vectorizer. vectorization is like it's like changing the words into some kind of numbers, numerical representation, okay? Based on how many times they are appearing, but after doing some kind of normalization, okay? So that will be giving some overview about that document, okay? So that we can do using few lines of code, I will just show that. Here it's just as numbers. It will be useful when we go with the RAG, the Retrival Augmented Generation, right? So we can keep knowledge based in some vector database and we can use these kinds of approaches for doing cosine similarity search, that kind of stuff. So when we deal with numbers, it will be very fast and because it's having semantic meaning also, that the retrieval process will be better, okay? And as I said, the transformers library of Huggingface is providing this pipeline that is very powerful. We can just do simple sentiment analysis with few lines of code, okay? Just to be able to specify the purpose as sentiment analysis. Here we are not going to give the model name anything, but internally it will be picking the relevant model. Huggingface is having more than million models, okay? And we are just passing the input text and it will be classiffied, okay? Since no model was supplied, it's choosing one model, okay, this model. But it's not recommended in production, that's what you're saying. But for learning purpose, it's okay, okay? So it's labeling this sentence as positive, okay? With 100% confidence, okay? We change it as something, okay? So it's negative, again, confidence is 100%. Something may, something we are saying is I think, like that means the confidence level may be reduced, right? If we add I think probably, in that case, it's reduced in the confidence level, right? Okay, these are very simple examples, okay? And practice more for understanding it. We have a lot of libraries, try to learn the libraries and their purpose, and then use them appropriately, okay? That will help you to save a lot of time in developing code, as well as for testing, because that libraries are tested rigorously, right? So there won't be any errors with that library. Suppose if you're going to develop by our own with that, that will take time, as well as that will be having errors, right? That is what I wanted to highlight here. So, we explore about AI related libraries, learn about that, and how to use that, and for what purpose, like that, okay? And then you can effectively develop a lot of AI related things. These are all basic things. Similarly, for RAG development, other things, Langchain thre, ecently, Agentic AI is becoming more popular, right? For that, LangGraph, a lot of new libraries are coming up, okay? Just try to understand the power of Python libraries for doing AI development. We can just easily use that, for example, here I'm just showing one example, right? My_module, I just placed this function here, and I'll be calling it from this another file, test.py. Just importing that thing, and just using that function, This greet is available with this file. My_module.PY, I'm using it from test.py file, okay? Let me run it. So it's saying that hello, Rajamanickam. Rajamanickam is passed from here, and hello is added here, right? So this is the fundamental of using a library. So we can create a function in one file, and then import that into another file, and then we can start using that function, okay? So just try to understand the available AI libraries, and make use of them effectively, okay? Thanks. 12. Creating Subtitle files locally using openAI's Whisper model: This video is for explaining the usage of AI with a very simple example. It is for creating caption files, subtitle files. Normally, whenever we watch videos on YouTube or any kind of video platforms, we can read the text while watching the video, right? Transcribed text. It'll be very useful in case the speaker's accent is not understandable, right? But for providing this feature, the video creator has to spend significant time for creating the caption files, okay? But these days, platforms like YouTube are automatically generating captions, I mean, subtitles. But in some cases, it is not possible to use platforms like YouTube. As I explained in this blog post, I already created one video and uploaded it in YouTube, and then I modified something for uploading in another platform. But in this case, if I upload the same video into YouTube for the purpose of getting captions, YouTube will be considering it as a duplicate video, and it will be a problem, okay? So in that case, I explored, like, other options. A lot of other things are there, okay. But they are charging money. So they are providing free option for a very few minutes of videos. But for lengthier video, longer video, they are charging money. Okay? Again, I need to upload the video there that things are there, right? So I was exploring AI models, whether any AI models are there. And I found that this Whisper model, it is provided by OpenAI. It is a free and open source model, okay? It's very simple model, also. I trained with a lot of data. And really, it's very useful. Okay. I was surprised by seeing the quality of its transcription. Okay? We can do it in our local machine itself. Once after creating the video, we have to just load the model, and then we have to call the transcribe by passing our file. Okay? So in case, if you are having MP4 that kind of video file, first of all, we have to convert it into Wave file. Okay? For that purpose, we can use the tools like FFMPEG, that kind of tools. We can convert the video file into a audio file, okay? So after that, we can just create one Python code. this transcribe.py I created this, okay? Um, so we have to input the whisper model. Suppose if you are using it first time, you need to install it using PIP command, Pip install openAI-whisper. Okay? So once after installing this whisper, you can start using it. Just import that whisper and just load the base model. Okay? Once after creating this model, you can just call the transcribe function for transcribing this audio file. We already created this audio file using FFMPEG right? So with that audio file, we have to just call this transcribe, okay? So it will be transcribing that audio, okay? I mean, it will be converting that speech content into text content, okay? But in case of captions, we have to do that time coding, also, right? Start time and end time. Then only it will be effective when we read the text while watching the video, right? So for that purpose, a lot of different models are there. One is famous one is .SRT format. Another one is .VTT format. Each of them are having some little differences, okay? Some platforms accept .SRT files mostly, but some platforms accept only .VTT format, okay? So this one is for creating .SRT. So just we are getting that text and opening .SRT file and creating this start time and end time and putting it in writing it in this Okay. It's a very simple thing. So now let us see that by experimenting it. I have audio file, this one, and these are the files for transcribing and for converting into other formats. Python code, okay? It's a very simple. Just we are creating the model, okay, from the base model. And then just calling this transcribe, okay? Just so we can call it from here. Okay. I'm just running this transcribe.py Python code for this audio.wav , okay? So it will be taking significant time, because it's a very lengthy file. My computer is not having GPU. That is why it's throwing these warnings, okay? If you are having GPU, maybe it will be fast. So now we can understand the real power up AI, right? We're not doing anything. We're just inputting this whisper. Before that, we are just installing PIP install. It's not a huge file. It can be loaded locally into any small computers, also, low configuration computer also. Okay. Comparatively, it's better than any cloud platforms also, okay? So you can use it for any number of files without any restriction. Okay, in case if you are going with Cloud platform, they are having different plans, right, so we need to purchase that. Okay. But if you know that how to use it locally, it's a very simple thing. Even this code, you can just ask ChatGPT to write a code for using whisper for your purpose. Okay. Suppose if you want that output in some other format, you can just copy this code into chatGPT and ask it to update the code to get the desired format, okay? Initially, I tried to create I mean, convert the .srt file into .vtt file using chatGPT itself. But that format is not accepted by a few platforms. Better I just provided the desired output format and asked chatGPT to create this code. So it created this Python code for converting the format. Okay? So after using this code, I was able to use that .VTT Capption file. But the key thing is the quality is really good. Okay? Normally, I'm non native English speaker, okay. even YouTube will be finding difficulty in understanding my accent for doing transcription. Whenever it automatically generates that transcription or subtitle content for my video, I used to spend a lot of time in correcting the text, okay? And moreover, I need to add punctuations manually. Comma, like that. Okay, then only it will be easy to read the content, right? YouTube won't do that kind of stuff. But I was really surprised to see the output of this whisper thing. It is able to understand the concept, and accordingly, it is able to add the punctuation. Okay, it's really good thing and it is removing unnecessary filler words, like, Okay, so like that. But YouTube used to include them all, I need to manually spend time for removing those stuff. Okay? So really good tool, okay? And moreover, it is free and open source. We can use it in our local, a lot of advantages are there, since we can do it programmatically, we can automate the things. Okay. Suppose if we have 100 files, we can just create one code for converting all of them one by one without manual intervention. Okay? So this is the power of understanding or learning about AI. Suppose if you are able to know how to use this whisper model, then you get a lot of advantages, right? That is this is one example. Similarly, a lot of models, open source models, pre trained models are there. Okay. So the only thing is we have to take the model and find the appropriate function for our purpose or appropriate models. If there are parameters, we need to understand that one. That is the part we have to learn. As a AI developer, we need not spend more in understanding how it's working. It is good to know that, but just I'm telling, as a beginner, you can just take the model, okay, and understand how to use that. Then just to start using it. Whenever you face any difficulties or if you want to improve further, then you can go further for understanding how it's working okay and how to improve that further. Okay. So now it's completed. Let me check the folder. Okay, I created this subtitles.srt File. So it created, right? But this timestamping that time coding is that format is different from .VTT, okay? So for that purpose, now, I'll be running it convert. So if you go through that subtitle text, it's really, like quality is really good, right? It added punctuation everything. Okay. Now let us run this convert.py It's very fast within, like, fraction of the second, it completed, right? It converted the time formats, and it added other required things, okay? That much easy. Now we can use it in platform supported by this format. Okay? I'm just telling it as an example, okay? It's a very simple one, but it could save a lot of time. So, for example, if I prepare your course with 1 hour, mostly, I have to spend more than 2 hours for creating this kind of time coding of transcript. Even if I'm going to type the transcript, then it will be like it will be taking many hours, right? So in that case, it is helping to save time by automatically converting the speech into text and automatically doing the time coding, ok? This is how AI models are helping us to save a lot of time, okay? It is one example, a lot of things. A lot of models are there. In case of computer vision, we can go with the Yolo model. Okay. And a lot of pretrained models also the sentiment analysis for that. If you go with HuggingFace, even in Kaggle, even Roboflow also having some models, like datasets models. Everything is there. So if you're familiar with those things, and if you just check them regularly, you can find a lot of information. Without having, like, a deep knowledge about AI, you can create a lot of AI applications, useful application. In the long run, you can improve your knowledge related to how it's working. But for start using AI in your coding, you need not be that much expert in understanding about the AI models, okay? First of all, we should be able to know how to use the AI model, already available AE models. Okay? It will be very simple, okay? You have to install that package and load the model and then start using it that much easy, okay? 13. Training a Simple TensorFlow AI Model with Google Colab: In this video, I'll be explaining about training an AI model using dataset and then explain about how to use the trained model. I'm trying to keep it very simple. I don't want to give more details here because I want the beginner to learn about the overview about training an AI model. I don't want to give any minute details here, okay? So this website is tensorflow.org. TensorFlow is a famous machine learning framework provided by Google. We can use tensor flow for developing various AI applications. As a beginner, you can find a lot of tutorials to learn AI, okay? In the home page itself, they provided a very simple Python code for training the AI model, okay? We can run it by clicking this link. If we click this link, it will open this code as a notebook on Google Colab. Okay. So there we can just run the code, okay? If you are not familiar with Colab, I will be just giving some basic details about it now, okay? We have to visit colab.google.com on a web browser. And we can open new notebook. It is a notebook, .IPYNB file. Colab is having options to add code cells as well as text cells. Okay? We can have various cells. In the code cell, we can include Python code. I'm just adding a very simple code, okay? And then we can execute it. Okay, now we got the result, right. Actually, that RAM and GPU are provided by Google. So we need not worry about the infrastructure. We can just focus on writing coding, okay. And similarly, we can add the text also for explaining the code or giving any instructions, okay? And we can add another code cell and we can add another piece of coding, okay? So we can run each section separately. It will be very useful for debgging purpose and for understanding the coding. Okay. As a beginner, it is very simple and easy way and even I can say effective way of learning, okay. But after that, you can start using any kind of your favorite IDE, like VS code, something, okay? Okay. Now I will open this link. Okay, I opens that notebook on Google Colab, okay? Here, the flow is like it is for loading a prebuilt dataset and then building a neural network machine learning model that classifies images and then train this neural network, then evaluating the accuracy of the model. That is the purpose of this example, okay? For setup tensorflow, they're importing the tensor flow, okay? And for loading the dataset, they're using MNIST dataset. Okay? MNIST is handwritten numbers dataset. If we check the Wikipedia, you can see some samples. It's a handwritten numbers zero to nine, okay? Even Kaggle also providing details about it, okay? And actually, it is having 60,000 training examples and 10,000 test examples, okay? Um, for training any kind of AI model, it is required to provide both training dataset as well as test dataset, okay? So we have to load the dataset, okay? And then build a machine learning model, okay? For that, we can use Keras. Keras API is a very easy way of doing the things. As of now, it is becoming part of tensorflow, okay? It is easy for using tensor flow using Keras API. So all these steps are for setting up the model. Okay. And then we can train the model using this fit. Okay? And then evaluate the model. Okay. This is the basic of training the model, okay? But they provided a lot of other information also, right? So as a beginner, it will be confusing to go through all the details. So what I'm doing is, I just created a simple uh, notebook, just by copying the very basic coding, okay? We can just execute this one so that we can just to focus on the overview. Okay? The first cell, we are importing the required libraries, Keras from TensorFlow and Matplotlib for showing the image. Okay. And if we click this, it will be loading the libraries. Okay. Then loading dataset, the MNIST dataset, right? So we have to load it using Keras. Okay? Once after loading the dataset, we have to normalize data. Normally, the color data pixels will be like up to 255 scale, right? But for neural networks, it is easy for them if the scaling is 0-1 like that. Okay? So we got to scale it. So for that, first we are scaling it. We are loading the MNIST dataset and then we are scaling the data. Okay, this is just for learning purpose. I am just showing sample image, okay? If we run it, it'll show the sample image, okay? It's showing one sample image. That means we understand that it got loaded. Okay? It is for defining the model. For any kind of neural network, there will be one input layer and one output layer in between many hidden layers. Okay. So while defining the model, we are setting those parameters hidden layer with 128 neurons like that. Okay. And then we have to compile that to specify the optimizer and specify about the loss. And then we can train the model using this fit function, okay? So while training the model, it's training for five epoch. Epoch in the sense, it's kind of iteration, okay? So it'll be training for five times. So the purpose is, it has to improve the accuracy each time. Okay? So first time the accuracy is 0.87. Loss is 0.43. Loss in the sense, error, okay? In the next iteration, the accuracy got improved. Up to 0.96. Okay? And the error got reduced loss got reduced, right? So, finally, accuracy reached up to 0.98, okay? Loss reduced up to 0.04. Okay. So this is the basic of training the model with a specific number of epochs. Okay. So now the training got completed, okay? Now we can test it, okay? Because that MNIST dataset is having test data also. I tested with one data. So for that accurate is 0.97, loss is 0.09, okay? It's okay, right? So now the final step is we have to use it. For this input image, it predicted as seven correctly, right? It's a very simple thing, right? So as a beginner, you can easily understand the overall view. Just we are loading the data set, and then we are setting up the primilinary things. I mean, defining the model. Then we are training the model using the loaded dataset with a specific number of EPOCH then we are testing the accuracy. And finally, we are doing the prediction. Okay. So for prediction, they're using in this example, we are using the data taken from the same dataset. Suppose if you want to predict with your own image, we have to do some pre processing. Then only it will be predicting properly. Okay? So as of now, it is able to predict the, predict in the sense here, identifying the I mean, reading the number from the image, okay? So it's doing properly. Here it's saying seven and the image is looking like seven, right? So that's it. It's a very simple thing, right? So similarly, we can take any model and we can just train with our dataset, and then we can use it, right? 14. Scikit-Learn (sklearn) Example: In this video, I will be explaining about the famous machine learning framework, scikit-learn. It is more suitable for traditional machine learning models like classification, regression, clustering, dimensionality reduction, model selection, preprocessing. So in this website, they provided algorithm supported by them, right? And in case of deep learning models, it is better to go with Google's tensorflow or metas PyTorch, okay? This scikit-learn is more suitable for traditional machine learning models, okay? But it's a very simple thing. It's Python library. The Python library is known as sklearn, okay? Um, it is built on top of Python's, NumPy, SciPy, and matplotlib libraries, okay. It is more suitable for small scale projects, okay. And it'll be very simple and easy to use. So better as a beginner, you can try many things with scikit-learn, okay. In case of classification, it is just identifying, I mean, in case of spam detection, identifying whether it's a spam email or not. It's like, using various algorithms. And we can use them, uh in scikit-learn, okay? So we can choose our required model and then we can train it and we can use it. Okay. Similarly, regression. Regression is predicting the continuous values, in case of stock prices, like predicting the upcoming stock price value. Okay? So it's a continuous value prediction. And Clustering is just grouping similar objects into sets, okay? For every kind of thing, there is related algorithm, okay? And based on our requirements, we can choose appropriate model, and then we can use scikit-learn for training it and for using that trained model. Okay? Similarly, it is applicable for dimensionality reduction, model selection and preprocessing. But this scikit-learn is not that much suitable for deep learning models. In case of deep learning models, we have to go with tensor flow or Pytorch, okay? Let me just to show you a very simple example to explain about scikit-learn. let me just take a classification thing. Okay? I already put the code here in the Google Colab, okay. Data dataset also sklearn is providing datasets also. We are just importing the datasets. In case of this example, we are going to use the Iris dataset. It's like details about some Iris flowers, okay? It will be keeping the dimensions, sepal and petal dimensions and labels as flower type, okay? So based on this training data, our model will be trained. And then we can use some sample data for getting the prediction so that the model can identify the flower, right? It's a very simple example just for expanding this sklearn. I'm just taking this example. Okay? So first, we are importing the data sets, okay? And mostly for training, any kind of model, we have to use some, training data as well as test data, right? For training any model, first of all, we have to choose that algorithm, okay? Um and then we have to provide training data as well as testing data. Then we have to train the model using the training data. Then we have to evaluate the model using test data. Then only we will be able to start using it for prediction. I mean, actual usage, okay? This is the very, very basic of any kind of a models, okay? So in that perspective, sklearn is providing this train_test_split so that it can split the data, like train data and test data automatically. So it will help us to save time. And this algorithm, this DecisionTreeClassifier, it's a classifier algorithm, right? So Sklearn.tree from that package, we are importing this method and metrics for measuring the accuracy, okay? For any kind of model, we have to measure the accuracy. It's some kind of testing. Then only we will be able to confidently start using that model, right? So we have to get that data and then we have to train it, and then we have to test it. Then only we'll be able to use it, right? This is for any kind of model. I'm highlighting here again, okay? In case of Sklearn what I'm trying to say is it's providing all the supporting libraries so that we can do everything very easily with this Sklearn itself. Okay? We are just loading the IRIS data. So normally as a standard way, X is always features, right? And Y is labels. So for any kind of classification, we have to say in case of dog, we can say that it has four legs, okay, two ears like that, right? So features features of the object then the label the CAT means, what are the features of CAT and then label like CAT. So it is a very fundamental way of doing this kind of data training. So for that we normally use X for the features and Y for the labels. So similarly, in case of this IRIS flowers data, features are like length and width of sepal and petal, and Y is obviously label flower type. Okay. And next step is, we have to split the training data and test data. For that, we are using this train_test_split, okay? So we're giving the X and Y as a arguments, and it will be just splitting into X_train and X_test, y_train and y_test. Okay. So obviously, it's simply it's splitting the dataset into two things train and test. Then we have to choose the model. Here we are choosing the decisiontree, okay? So DecisionTreeClassifier model, okay? And next step is using fit, we had to train the model. So while training the model, we have to provide the training data, right, these features and then label, right? So training data. Okay, X_train, y_train. Okay. As I said earlier, we have to messure accuracy. Okay? It's kind of evaluation before putting it for actual usage, okay? So here, we are using this accuracy score. So now we have completed loading the data and then splitting it into training and testing and then creating the model and then uh, training the model using training data, and then, doing the prediction, okay, with that trained model. And then, based on the data, we are calculating the accuracy, okay? Once we are satisfied with the accuracy, we can move forward with actual usage predicting any kind of new samples. So the new sample, we are just using this sample example, okay? Um, it is the length of sepal and the width of the sepal and length of the petal and width of the petal like that. Okay. So if you give these numbers as a features of this sample flower, then we can just call this predict for predicting the new sample, using this trained model. It's very simple, right? But it covers all the basic things of training, any kind of machine learning model. So this is a very fundamental. any kind of even for computer vision thing, the basic is same, okay? So with the help of sklearn, you can easily learn it, okay? So we can just run it. Okay, so the accuracy is 100%, okay? So we can move forward with actual usage. Actual usage for this sample, it is classifying it as setosa, okay? Suppose if we change that parameters, it will be detecting as some other thing. Maybe I have some other sample. It is some other thing, okay? Maybe we can just randomly change Okay. So it's able to predict the flower, okay, based on the dimensions of the flower, okay? That's it. It is about sklearn, start using it for practicing. But in actual usage, as I explained earlier, for any kind of deep learning, it is better to go with tensorflow because that are providing production implementation that deliver everything they make easy. But in case of sklearn, we can learn it easily, okay? 15. Hugging Face Tutorial: In this video, I will be explaining about Hugging Face. Hugging Face is a popular platform for artificial intelligence AI, especially for machine learning models. It is claiming that it is democratizing AI because it's a huge open source AI community. Here, we can find a lot of open source models and datasets available. Okay. And we can access various applications here. So basically, it's like, three things are very important here. One is models, another one is datasets and spaces. Here in the models, we can find a lot of models, okay, hundreds of thousands of models. Huge number of models are available, and it is categorized based on, like, multimodal computer vision, natural language processing, NLP, audio, tabular, reinforcement learning, others like so a lot of models are there in case of multi model here, audio, text to text, image, text to text, visual question answering, okay? And the computer vision, here, depth estimation, image classification, object detection, image segmentation, text to image, image to text, image to image, like image to video. So almost it's covering all the things related to computer vision and especially it is very, famous for for natural language processing. Apart from having models here, it is providing library also, in case of transformer library, by using transformer library, we can easily access many models using this pipeline, we can just mention what we want instead of giving the model name itself. It will be automatically finding relevant models. Okay? That much easy, within a few lines of coding, we can do a lot of things. Okay? That is the key advantage of hugging face. And we can start using it freely, but based on the usage, suppose if you are using it using GPU heavily, then they will be charging that based on the usage. Okay? so it's mostly research people like research in this sense. Even if you want to develop something and if you're exploring something, this will be very useful. But for production, I don't think whether it is suitable or not. Okay. Anyway, for whatever you are doing something, you can freely host with the spaces also, okay? Let me explain one by one. Here, models, we are having a lot of models, okay? AI models and categorized with various topics so that we can easily find the relevant models. And we can get a lot of things. If you check one particular model, you can find the sample code, or the details about the model. so that we can start using the model easily. Even they provided sample applications. So just straighaway, you can just click that and how other people are using this model. Okay? That much convenience, this hugging face is providing Okay? So if you're new to AI development, I would recommend exploring hugging face heavily. Okay? Spend more time with hugging face. Suppose if you want to use this model, just click use this model. And with the transformer libraries, we can use that they're providing the code. Okay. So only thing is mostly we have to install this transformers library provided by Hugging Face. And then just you can use the sample code provided by them, okay? That much easy. So maybe we can test some models. We can do it from our local also. So some models may be requiring some like significant GPU usage. So I will be explaining the things through Google Colab, okay? To avoid that computing power issues. But mostly almost all, I'm having very low end machine only here. But when I tried with various models, it is working smoothly. Okay. But some models may be requiring some kind of resources. So better for Demo purpose, I will be using Google Colab, but you can mostly use in your local machine also. Some models will be recurring API access, okay? Mostly other models will be automatically downloaded to your local. Okay? Some will be using API. And API, you can just get a token from here, okay? Just log into hugging face, get the token, and then just to export the API key in your environment. Then you can just use the sample codes, for exploring various models just to be familiar with AI development, hugging face is easy way of doing that. Okay? A lot of models are there, okay? You can explore and especially they categorize like this so that you can easily find your model or you can just search the model, okay? Here they sorted with trending so that we can find the popular ones, okay. And similarly, next one is datasets. It is providing a lot of datasets, also. Suppose if you are developing your own model, and if you want to train with some datasets, you can just use the datasets. Here also something similar to transformer, you can just use the dataset library. Okay? Here, suppose if you want to use any kind of dataset, you can just use that library. Okay? So from datasets, you can just import load datasets like this. Okay? So that easy. Okay. So for beginner, it will be really useful, okay? So you can explore a lot of models, like do by yourself easily. Within a few minutes, you can use any kind of models. Okay. The spaces, suppose you local is not having enough GPU or other things, you can explore the spaces. Here, already many people used some models and created some applications. Okay. So you can straight away, explore this one. And when you are developing any kind of application using this, models, you can also host this hugging face space freely. So that others can also go through your thing. A lot of applications are there. It's because of image, it's taking some time. You can explore that things. Lot of applications are there. Similarly, we can just create one application and host it in the space. We can create new space, and then we can just put our.. Okay. We can use streamlit or gradio for developing our UI. Okay? So this is about spaces. So similarly we can use datasets and models. Okay? So we can say there's three pillars of this hugging face, okay? So models, datasets and spaces. And importantly, it is having its own library like transformer, dataset, diffusion like that so that we can handle various coding very easy way. Okay, an easy way. Okay. Let us first start with simple example, okay? So here we can filter based on the task and even libraries used, based on datasets also and languages, natural languages and based on licenses like that. Okay. So we can easily find out our required model, okay? Let me start with some simple model first. Maybe text classification we can take right. And here it's listing a lot of models for that purpose. Okay. I'm just randomly choosing this one. But we can just go through the details to understand which model will be more suitable for our own requirements. Say, for example, some models will be trained with general common text, some will be finance related things, some will be e commerce related things, better, you can choose appropriate one. I'm just for a demo purpose, I'm just taking this one. Okay? So here they provided all the details. So if you want to use this transformers through transformers. As I said previously, transformers is very simple way of using it, especially this pipeline feature will be very useful. I just trying to use this. Otherwise, if you want to load the model directly, you can go like this also. Okay. But pipeline will be a very simple way of use Google Colab. You can just run it. It's a positive. The 99 percentage confidant It's a negative. Okay? So it is very easy. You can give any text. Okay? It will be classifying this. Okay? Maybe we can copy something from, things positive, okay? So we can simply use this pipeline of this transformer library, just give the model and what task we have to do, Text classification. That's it. Okay. Then we can start using that model. Okay? That much easy. Okay. Similarly, we can do it for any models for any purpose. I already listed something. Maybe we can go with this one, okay? Sentiment analysis. So here we provided as text classification and specified the model. In this example, we are just asking as sentimental analysis, sentimental analysis, okay. And then we are passing this text, and it's doing that Okay. So basically, it's using this text classification model, okay? doing sentimental analysis. And using pipeline, we can do the zero-shot classification also. That means without labeling, we can categorize the input, okay? Using pipeline, we are just calling this zero-shot classification. Okay? So it will find appropriate model, okay? And we're giving the inputs. Like text is 'I recently bought a new laptop and it is super fast and lightweight'. The labels, technology, fashions forces. Okay? We are not doing any labeling, okay? We just provided the input and the list of labels. And this zero-shot classification will be finding their proper label, okay? That is what zero-shot. We are not giving any examples or any label things, okay. So based on the text itself, based on the meaning meaning of the text and the meaning of the labels, it will be able to classify So here it provided technology, for technology. the score is 97 percentage, for sports, one percentage, and fashion, it is zero percentage. Basically, it is technology, right? It's a correct . 'I recently bought a new latop and it is super fast and lightweight'. It's technology, right? So far that it is using this Facebook model and revision number is this one. Okay. Just to save the time, I'm taking from this sample code. Otherwise, you can from the model itself we can take, instead of using pipeline, we can specify the Specific model. But to save time, I'm using this way. Here we are giving this prompt 'artificial intelligence will change the future by' and we're asking to generate text with the maximum length of 50. In this case, we have to specify text generation. It will choose appropriate model. It is choosing this open AI community GPT2 model, 'the artificial intelligence will change the future by' That is what we provided. Here it's saying that 'showing you the possibilities of being a more complex person'. The second text is, 'will change the future by taking care of all the humans because artificial intelligence will change everything. We will have a human consciousness, which is different type of technology..' Something. So basically, it's able to generate text, okay. So what I'm trying to say is with the help of this hugging face, we can save a lot of time. We have to just use transformer and pipeline and just to specify what we want. It will be automatically finding the appropriate model. Even if you're going to specify the model specifically, we can use this transformer so that we can do the coding in an easy way. We need not understand how each and every model is working, right? So in that way, it can save a lot of time, and we can just simply change the model in an easy way. Okay, that code maintenance will be easy. this one, question answering, okay? We get this word in context and asking question. Just in the pipeline, we mentioned that question answering, okay, so that we can get answer for this question. It is using that model. It's giving answer as 'New York and Paris' Here we mentioned 'New York and Paris', right? So it correctly taken that answer. So we can do it for various things. In case of named entity recognition, it'll recognize whether it's person or place, that kind of stuff, right? Okay. So just go through various models, okay? And you can test it in your local machine or Google Colab, or you can just see some samples from these spaces. And if you're creating some good applications, you can just post it here to showcase your skills, okay? 16. Deploying an AI Application (e.g AI Chatbot) on Hugging Face Spaces : Hello everyone, This video is for telling about how to deploy a simple AI application Okay I created this AI application It's a very simple application It's chatbot for talking or discussing about wellness and fitness it uses Gemini model LLM and it uses Streamlit for user interface Okay actually I tested it in my local Okay And it will be available from my local only that Streamlit will be running from my local only but for making it available to everyone we have to deploy it Okay that is the purpose of deploying the AI application and for that there are various ways In this video I'll be telling about a simple and free way of doing the deployment Okay So for that we have to go to hugging face okay you can type hf.co okay it will be redirecting to huggingface.co co there you can find models, datasets, spaces right so click spaces if you're doing first time you have to login okay for creating spaces okay there you can find lot of other spaces you can create your own space for that you have to click new space okay there you have to give the space name I created wellness bot right so I create gave the name I already created it I'm just repeating it Okay I just give the name as getwellness Okay and you can give the description Okay and you can choose some license Okay and I created with the Streamlit right So I choose Streamlit Okay and you can choose other things also Gradio, Docker Okay select hardware for free means you can choose the CPU based Okay Otherwise you can go with GPU but that will be paid one Okay And we can make it as public or we can keep it as private Okay And then click create space Okay Once after clicking create space it will be asking to upload files I already uploaded so I'm just showing here Okay we can upload the that app file app.py and other than that we have to upload the requirements.txt also there we have to provide the modules okay here I provided Streamlit and Google-generativeai okay because I'm using those two libraries I mentioned those things okay and these things will be automatically created okay so we have to upload the app file and the requirements.txt file okay and if there are in your application if you have any other things you can upload those things also okay So once after uploading the things if you want to make some changes we can directly make changes here also by clicking this okay edit icon and that's it it will be run and you can see it here apps actually the Hugging Face will be loading all the required moduels and it will be showing here and one more thing is since this thing is using this API key secret Right this this one we have to take it from environment So we have to provide that thing in the settings Okay So there you can find the option for providing the secret and you can choose that secret and create one secret variable Okay And then you can put the API key there Okay Here you can find that option Okay So secret option is there Right So you can, I already created that Okay So that's it We have to upload the file we have to set the secret if any secret API key or any other thing is there and then you can come to this app tab to see the app Okay It will be built automatically Okay So we can chat, start chatting Okay "Give some wellness tips" Okay It's giving wellness tips as well as it's recommending the products Okay That's it Right now we can access it from this URL but since it's a chatbot we will be requiring it in our website right so for that we can just use this like iFrame within iframe we can set up it okay I set up in this website when clicking this button it will be opening this iframe okay and we can chat here that's it okay very simple right so we have to upload the files into Hugging Face by creating new space Okay And then we can get that URL and using iframe we can set up in our website Okay You can put the iframe directly in the website itself in the page itself But that loading issues will happen in case of viewers viewing from mobile devices Okay So better we can put as a button and then on clicking the button we can load it Okay that will be giving good user experience Okay Actually it's seamlessly including my book name also right Apart from giving obvious recommendation it's including here also Okay So that is the reason we have to create these kinds of chatbots right Otherwise we can directly just put the Gemini here right So if you want to have our custom logic we can develop our own chat bots and then put there I'm just telling as a example 17. TensorFlow Playground Explained: This video, I'm going to tell about TensorFlow Playground. Tensflow is a machine learning framework provided by Google. It's a very useful framework for developing AI applications. And the playground is just for learning purpose, okay? It is a web based tool. We can use it from browser just by typing playground.tensorflow.org. Okay? Once after typing that in the address bar, you will be seeing this page. It's a single page. Here you can find the way to learn the deep learning by doing experiments. First of all, I will explain what are the components available here. So here it is for start and I mean, running and pausing the training process, okay? And this area is this left side area is for data. Here, we can choose the data from various kinds of datasets like circular and exclusive or, Gaussian, spiral. There are various kinds of data that's available here. We can choose one of them. And here we can choose the ratio between training data and test data. We can split the data into training and testing. And if you want, we can introduce noise into the data. So if we introduce noise, the data will be corrupted with some noise. Okay. And the batch size also we can change. If we click regenerate, it will stop the training process. Okay, and it will load different kinds of data. Uh, here, it's for, like, resetting the network, okay? Epoch, here we can't enter the Epoch. Epoch is nothing but like a number of steps or iteration, okay? Training iterations. I mean, training steps here. Here, it's only for display purpose. We won't have any control to stop the training by setting the Epoch, okay? But in actual coding, we can do that. Either we can do early stop by stopping the training process if the loss is not changing for certain consecutive steps epochs, okay? But this learning purpose, it will keep on running. We have to manually stop the training process by seeing the boundary and the loss values. Okay. Learning rate, it is the speed speed of learning. So if we keep it slow, then I mean, low value, that learning process will be slow, but the stability will be good. But if we increase the learning rate, it will learn faster. But that model will become unstable. Okay? And the activation function ReLU lot of different kinds of activation functions are there. Activation functions will control how the model handles the weights. Weights are like internal parameters weights will be changed based on the training that loss parameter, other things, okay? So that is the very core concept of neural network, right? So here we can see that training values. Okay. Even we can edit it. There will be weight. This is regularization, L one, L two, like that. None also we can choose. And we have to choose that value. Okay. So the regularization is for avoiding the overfitting. Overfitting is, like, if we choose like very powerful things like large number of neurons or many numbers of hidden layers, if we keep uh, set that neural network in that way, that neural network will be learning the training data more than enough. That means, it will be very focused on the training data. It won't be generalized. Okay? If that model is trained in a generalized way, then only it will be effective at dealing with unknown data. If it is specific to the training data, it will be detecting, I mean, it will be predicting or handling the training data effectively, but it won't be able to deal the new data. even if the data is very similar to the training data, it won't be handling it properly. So in that case, we have to avoid the overfitting. So for that, some kind of regularization are there, just for controlling the increase of weight in specific places, okay? So for that this regularization will be very useful. And this is problem type, whether it's a classification or regression, okay? We can explore with various parameters and we can find how the model is working and how the output is looking like, how much time it is taking, and how the weights are calculated. So this one is features, like properties. Which properties do you want to feed in input properties. Like any kind of neural network that features are important from the training data, right? Here, we can set what kind of features we are going to give to the training And here, hidden layers. It is input layer, it is output layer. Between that, we can have hidden layers. Without hidden layer also, it can work, but adding hidden layers will improve the performance effectiveness, okay? And each hidden layer can have its own number of neurons. We can increase the neurons or decrease them, okay? In each layer. Here we can keep four neurons. Here we can keep two neurons. And then we can find the pattern how it's working. All these things will be able to teach you how it's working. Based on this learning, we can implement it in our actual coding so that we can save a lot of time. Okay. Otherwise, if we are practically doing everything means, it will take lot of time, right? So here it's kind of simulation. Here we can know how the model will be behaving. based on that, we can plan our AI development. Okay? In that perspective, this playground tensor flow playground will be more useful. Okay? More useful in this sense, we have to do a lot of practices. We have to change the learning rate. Learning rate, as I said, if we keep on increasing the learning rate, it will affect the stability, okay? Sometimes even the convergence won't happen. Okay? Regularization, as I said, if we increase the hidden layers, we got to apply the regulation. Okay. So always, there will be a trade off. So we need to find the appropriate best point using in our coding, okay? So for that purpose, we have to explore various things. We have to select I mean, change the dataset and then do the same thing with different datasets. So here, the color coding also is having certain meaning. Orange means negative, okay? Blue means positive, okay, generally. But other than that, it is used for differentiating, various classes, also, okay? So here, as I said, here we can see the weight as well as we can change the weight. So that output it is doing test loss as well as training loss. So Test loss is like loss is error. Okay. Our goal is reducing the loss up to zero. That means there is no error. But in ideal case, getting zero is bit difficult. Okay? In case if you're planning to write a coding. In that case, for stopping the train obviously, we'll be thinking that once we reach the loss zero, then we can stop the training like that we'll be thinking. But in some cases, most of the cases, reaching zero is bit difficult. Either we can as I said early stopping we can use. That means we have to continuously monitor the loss values for each epoch. And if it is not getting changed for few consecutive epochs, then we can stop the training. Otherwise, no meaning of running the training further. Okay. And otherwise, if you believe that it will reach zero, maybe you can use that condition also. But instead of putting exact zero, maybe we can put some minimum value, okay? We can run it and we can see how that loss is changing, okay? Here you can see the epoch. It's running continuously. As I said earlier, it won't stop. Okay? It will keep on running. Even if it reaches the zero, it won't stop, okay? We have to see the boundary, how it's behaving because if we introduce noise, what it happens. Noise in the sense here, the blue is here also the red. The oranges are here, it's like a tedious task. If we want to regenerate, we can stop the entire training and can load new dataset. We can use simple this Gaussian. The noise is high, right? will reduce the noise. In that case, I think mostly it will be reaching near zero loss very easily as expected, loss becoming zero very quickly, it's able to differentiate clearly. It's reached zero, okay? So there is no requirement for further running. From the experiments, we can understand how many epoch it will be taking. So based on that, we can write the code. Okay. That perspective, it will be very useful. And another perspective as a beginner, just to know about what is learning rate, what is activation, what is regularaization. For that purpose also, we can use this, okay. And if you are advanced in AI, maybe you can study how the weights are getting changed. If you manually change the weight, how it's behaving, so you can learn, okay? So it's for all kinds of people. So based on our skill set, current level of skill, we can make use of this. So it's really useful tool. I would recommend practicing it, spending more time. But you have to keep on, like, noting the values, what kind of values currently selected, how it's behaving like that. If you're just randomly changing something means, it won't be useful. 18. RAG (Retrieval-Augmented Generation) Tutorial: In this video, I'll be explaining about retrieval augmented generation RAG. It is the technique of combining retrieval and generation for better AI. In this video, we'll be covering these topics. What is RAG? Why is RAG important? How does RAG work? Components of RAG, applications of RAG, advantages and limitations. And then a simple demo. So first of all, what is RAG? RAG is a hybrid AI model that combines retrieval based methods, fetching relevant information from your knowledge source with generative methods. That is creating new text. So basically, there are two things. One is traditional retrieval things like searching a text from your database or from PDF file like that, right? So we'll be having some knowledge base. There from there, we can do the search to get required information. Okay? That is fetching relevant information from knowledge source. That is retrieval. Another is generation. Generation, it's a very important thing these days. Like, you might have heard about the word generative AI, right? So these chatbots like chatGPT and Gemini, they have the ability to generate new content. They will be able to create new articles, new blog posts, and even they can create new images. In that case, it will be called a multi-modal. Okay, internally, it is using large language model like GPT four. In case of GPT, it is using GPT four model. These LLMs are powerful in generation things, but they have some limitations. In case of RAG, it is combining power of both retrieval and generation. That is the key idea of combining the power of both retrieval and generation. The key idea is it enhances language models by grounding their responses in external knowledge. Okay? So it's something like guiding LLMs with a help up, external knowledge, okay? So why is RAG important? The traditional models, AI models are having few shortcomings. One is limited knowledge cutoff. Another one is Halish nations. Limited knowledge cutoff in that sense. Actually, as I said, the chatGPT is internally using GPT four model, right? That GPT four model is a pre-trained model. That means it was trained with a lot of data. So in case of these LLMs, they are trained with huge amounts of data. With the help of a lot of computing power, they got trained. In that case, there should be some cut off dates. Say, for example, they trained it two months back, then it will be having data up to that time period only. Okay so it won't be having the latest information. Okay? So that is one key problem with LLMs. And another one is hallucinations. Hallucinations are like inherent nature of LLMs. They used to generate incorrect or nonsensical information. If you are frequently using chatGPT, you should be knowing about that. Sometimes it used to give very irrelevant answer for simple questions also. It is because of hallucinations. Okay. And other than that, a lot of things are there because of wrong information with the training data, it can hallucinate. Okay? So there are two problems. One is limited knowledge cut off. That means there is no latest information and then hallucinations. Okay. So both things are important because for using any kind of LLM, it's important to have latest information, right? Similarly, if that information is not reliable, then the entire system will be facing problem, right? The problems are very important. So we have to find better solution, right? So that is where RAG is playing an important role. RAG helps to access up to date information, and it is trying to avoid hallucinations. Okay? How does RAG work? The step by step process is, input query. So user first asks a question, right? And then retrieval. The model retrieves relevant documents from a knowledge source. And then generation. The model generates response based on the retrieval information. So for the user input query. First, the retrieval system will retrieve relevant documents from a knowledge source or external knowledge. Then once after getting the context information, it will be sending that context information along with the input query to the LLMs for getting the output. So here, it is the overall workflow of RAG. But there are two things. One is one time task, another one is regular usage. One time task is like we had to load the documents, right? So suppose assume that you are running an organization within your department, you may be having assume that you have five departments. Okay? Each department is having its own policies and some details about the departments, right? Suppose someone is asking some questions about your company. He cannot straightaway ask the LLM. LLM won't be having that data, right? You'll be having that data. But we cannot send all the data to the LLMs. Okay? Suppose in this example, if each department is having one PDF file, we cannot send five PDF files along with user query to the LLM, okay? Because of that context window limitation. Each LLM is having certain context limit. We cannot send a huge amount of data. Okay? There is a limit in that case, we have to send only the relevant document along with the user query. So for that purpose, first of all, we have to keep all the documents in a vector database. Okay? There are different approaches. I'm just talking about the common approach, okay? First of all, we have to store the documents in a vector database. Vector database, it is the search will be happening very faster because the reason is it will be keeping the data in a numerical format, okay? So it will convert the text content into our image content into equivalent number. Okay. It'll be storing only the numbers, numbers and along with corresponding associated text. But the search will be happening based on the numbers. So basically, it'll be very quick, okay? And it will be effective. Effective in that sense, I will explain. Okay. So first of all, it is very fast, okay? So for that we had to convert the documents into embedding embeddings, as I said, it's a numerical representation of that content. For that, we had to use some LLMs for embedding. That LLM will be different. They will be providing embedding services. We have to use the embedding services and then convert the documents into equivalent embeddings and store it in a vector database. There are various vector databases like ChromaDB, Facebook's FAISS and the pinecone managed vector database. And in case of ChromaDB we can locally use. It's open source. There are various vector databases are there. We have to store the data in a vector database. It is one time process. One time before start using our system, preliminarily as a prelminary step, we have to set up this. As I said, that vector database, that search will be faster, and another advantage is it's a semantic search. Okay. So not only keywords search, it will support semantic search also. That the effectiveness of that semantic search will be depending on what kind of LLM we are using for creating the embeddings. That is a completely different topic. Like in the RAG in the case of RAG, the overall structure is very simple. LLM is having some shortcomings. We are going to solve that problems with the help of external data. So for that we are using this RAG. So we are just augmenting the generative process with the retrieval step. That is the very key. That is a RAG. Okay. That is very simple. But in case of implementations and other things, like, based on the requirements and usage, there are various things up there, okay? But the scope of this video is for giving the basics. I'm just addressing the beginners in this video. So I'm just keeping it very simple, okay? So the documents are converted into embeddings and stored in vector database. That is one time step, okay? And then come to the actual daily process. So user ask a question, that query also converted into embedding, that is called query embedding. Okay, Query vector, so that query vector will be searched, will be used to search relevant entry in the vector database, okay. Say for example, in our example, we search five departments. So if you are asking if the user is asking a question related to one particular department, say example, he's asking about second department and that query will be converted into that embeddings, that query embedding will be searched against these vector database doc embeddings, so it will be like finding the data related to the second department. Okay? So that is the nature of this vector search. Okay. Here, it will be fast as well as it will be doing semantic search. That is the key here. So once after getting the relevant information, now that is called a context chunks. That will be like with the augmentation, it will be added along with the query and then sent to the LLMs for getting the response. Now it will be like something related response. Okay. The overall process, one time processes, storing the documents external data as embedding using LLM embedding. Okay? And the actual process is converting the user query into embedding, then do a vector search. There are various things cosine similarity that kind of approaches are there. Okay. So based on that, we have to find the relevant chunks, and we had to augment the prompt with that context chunks, okay? The prompt will be like we got to provide the query and the chunks and then provide additional information like use the context information to give the answer for this query. Like that we provide some system prompt also, okay. So finally, we will be getting the relevant response. That is the core concept of RAG. So if you are not using this RAG thing, then we may be getting irrelevant response. This process is grounding the LLM. That is one thing. And as I said, previously, it will be useful for getting the latest information. The LLM may not be having the latest information, but the context chunk will be having that latest information. It will be using that information. So the components of RAG are retriever generator knowledge source. So the retriever searches large knowledge base for relevant information. There are various retrievers there. Generator. It is like GPT model. It's a normal LLM, and the knowledge source. So for a retriever to get the data, we have to have the data right. So for that data, in case of example vector database, we have to store the details right. So for that we have to use some knowledge source. It can be as I said previously. It can be any document simple cases, or it can be any crawler crawling latest information from various web sources, or any kind of database also applications of RAG. So a lot of applications are there, okay? One is obvious one is question answering for getting accurate answers to the user queries and chatbots for more informed and the context of our conversations because we will be able to send the context information so that it can work properly based on the context information and content creation, creating articles, summaries, or reports. We have to provide external data so that LLM will be able to create the content with more information. Okay? So customer support, obviously, whenever the user asks any questions, the customer asks the questions, that system will be able to give proper reply with the help of this external knowledge. For example, if a hospital is running a customer support system, the user will be able to ask any questions to the system in natural language, and that RAG system will be able to search the content, get the relevant content from the knowledge source, and then reply to the answer. So basically, I can say in two ways. One is like the retrieval system previously, in this case of this hospital customer support system. Previously, if someone is asking any question, the representative in the hospital has to go through some PDF files or database. He has to do the keyword search and then give the answer, right? So that is traditional way, okay? In that case, he will be able to get the answer only when the keywords are matching. That is the limitation. He cannot use the natural language. He cannot use every day simple languages. He has to use the keywords properly. In case of LLM, we can use natural language. But the problem is, it won't be having the data. To get benefit of both things, that is the name RAG. We are augmenting the generative process with the help of the retrieval of that content. Augmenting the generation using retrieval augmented generation. Advantages and limitations. Advantages are, obviously we can get the latest information, real time or updated information, reduces hallucinations, okay? It combines the best of retrieval and generation. So it's obvious advantages of RAG. Limitations. Again, it depends on the quality of the knowledge source. Even if you are able to use the external knowledge source, if the quality is not good, then again, the problems will continue. So it depends on the quality of the knowledge source and computationally expensive. Converting the documents into a vector database that will take significant computing power. And if you are using API, then the API usage will be somewhat costly, not up to the level of generative that cost, but still we have to pay for that. Okay? So computationally expensive. And compared to other approaches, few-shot learning other things that requirement for data set for training is comparatively high here. We have to have significant amount of data in the externals. So let me show you a simple demo. Okay. So normally, there are two approaches, okay. One is, like, using the open source free models. Okay? We have a lot of things. Okay. You can search in the HuggingFace or Kaggle. Like, we can get a lot of embedding models. Okay. But based on my experience, if we use the embedding models provided by large organization like openAI, then it will be the performance will be good and basically, it'll be better, okay. But we have to pay for them, okay? That is one thing. This example, I'll be showing simple example using HuggingFace. But in actual usage, mainly I used openAI API. Okay. So here, the coding, first of all, we are importing the required libraries. So this code, we use langchain. Langchain is a framework or library for AI developments. It will be easy to handle various models and various kinds of vector databases. the coding will be simple, okay? But the maintenance, based on my experience, the maintenance is continuously, like, the continuously changing their libraries, things, okay? But the obvious reason is the AI is developing very fast, right? So they too, like, catch that, they're continuously updating their libraries, also, okay? and moreover, if you use the coding individually, then, understanding the concept will be easy. But in case of langchain, you suppose if you're developing if you want to save a lot of time, then you can go with Langchain. But for learning purpose, I would recommend doing everything separately, without using any kind of framework or libraries. Okay. But as a beginner, for giving some overview, I'm using Langchain here, okay, so that you can understand the full flow easily without getting deviated by each and every individual coding, okay. Here, the overall, we are like importing all the required libraries. Okay? So the steps are first, we have to, as I explained in the diagram, Uh, there are two things. One is one time step, and another thing is the actual flow like retrieval and augmenting the prompts and then generation, right? So that three steps. So totally one step is first step is one time step, then again, three steps for every query, right? So the first thing is, we are preparing the documents. So here, it's a very simple. So just for explanation. Otherwise, assume that each document is like, two, three pages lengthy. Okay? So here, vitamin C helps boost immunity, exercise, improves mental and physical health, drinking enough water keeps you hydrated and improve focus. Okay? It's a document and chunking is important concept in a RAG, okay? So we have to Chunk means splitting the document into different sections so that it can fit into the context window of the LLL, okay? So for that, we have to provide overlap also. So I don't want to give all the details here so that we can, like, fully focus on the overview, basically, we prepare the documents and chunking that, okay? And then embedding, creating the embeddings and storing it in FAISS. Okay, vector database. Okay. So for creating embedding, we have to use the embedding model, huggingface embedding we are using. So for using huggingface, we have to use the huggingface API. Okay? So I already loaded I mean exported the huggingface API key in the Environment. So suppose if you are going to use this code in your environmnt make sure that you loaded huggingface API key, okay? So you have to get the details from access tokens from here, huggingface. And then you have to load it in the environment. Then you can use this code. Okay? So first we are preparing the documents, chunking it, I mean, splitting it, and then converting the chunk into embedding and then storing the embeddings in the vector database. Okay? This is one step one time setup. And then actual chatting, for that, we are using this LLM, okay? And like we are augmenting the query by giving this prompt, preparing this prompt. we're using this prompt template, you're using basically we are asking the LLM to use this context for answering this question. And it has to give answer in this format. This is the context. This is the question. You have to give the answer like this, like that we are asking, It is very important because otherwise, it'll be giving answers in different format. For preparing the example. I didn't give this format first time, so it behaved differently. So for giving the example u , I just I arrived with this template. Okay? So now we are creating the rag chain. And here, this is a query user query. How does exercise affect health, okay? So once after running this chain, it should give the answer. We are printing the answer. Okay? So here, three things are there, right? So let us see what is happening. So I'm running this script. After running this, it gave the answer, Exercise improves both mental and physical skill. For the question, it provided the proper answer. Apart from answer, it provided explanation also. 19. Using AI for digital marketing: In this video, I am going to explain about digital marketing using artificial intelligence. Here, I am going to explain it with ChatGPT, Gemini, DeepSeek, and Qwen, these kinds of famous large language models (LLMs). Mostly, we can use certain features freely here. Apart from these things, we have a lot of specific tools also there, like for writing articles, we can go with Jasper.ai and Copy.ai, which have a lot of features, but I am not going to go in that way. So, anyone can easily start using digital marketing with the help of these freely available and common AI tools. First, we can start by asking ChatGPT about doing digital marketing using AI. It gives a lot of information. In this way, we can just do brainstorming with ChatGPT to get a lot of information. Here, it listed various things. We can go through this way. One important thing is instead of typing the text here, maybe you can go with the voice mode so that we can talk continuously, and ChatGPT will be answering through voice. You can just download the mobile app and you can simply talk with ChatGPT. At the end of the talk, we can go through the transcription, the text also. So, that will be very useful. Here, I am just for demo purposes and just typing the text here and these things. In ChatGPT, it is the first starting point. You can just talk with ChatGPT, and always remember that these kinds of LLMs—ChatGPT, Gemini, DeepSeek, Qwen—all these LLMs are not reliable. So, you need to be very careful about that. You can just use it for doing some kind of work, but don't rely fully on them, and don't use that output as a final thing. Manually review those things. Let me explain how to use it effectively. It is true that all these LLMs are really powerful. They can write articles, they can create images, they can create videos, and they can do any kind of format changes, and they can do a lot of calculations, and they can create coding. A lot of things that are useful for doing digital marketing, but they are not reliable. That we have to always remember because they are subject to hallucination and biases also there. A lot of things are there. So, before starting to use these kinds of LLMs for real business, we need to understand that kind of limitation first. So, we need to know the ways to handle them also. To some extent, we can ground these LLMs to give some kind of, up to some level, we can make it reliable. So, just let us go through how to do that kind of stuff. First, start with brainstorming. It will give you a lot of things, and we have to apply the prompt engineering tips. We have to provide clear questions, concise, and it should be clear. We should provide context, and we have to use some kind of few-shot prompting, like giving some examples, and we have to use the chain-of-thoughts thing, like if we have to ask ChatGPT to do the processing step by step. That kind of stuff. That is the basic thing. Once you are familiar with prompt engineering and know the limitations of ChatGPT or any kind of LLMs, then it is better to start using this for doing digital marketing. It is really very useful. First, we can start with creating articles, blog posts. That is the first point in the digital marketing funnel. Right now, I am selling an artificial intelligence course for beginners. I am planning to promote it. So, in that case, I am just asking ChatGPT to suggest some blog posts for promoting my course. I am just giving a simple example, but you have to use the prompting skills properly. I have to give all the context and clear information and everything. But here, I am already talking with ChatGPT. It stored a lot of things in the memory and in the context also, it is having some information. So, it will be able to give a proper reply. It is suggesting some good article titles. They provided the title and some suggestions also on how to use that. So, all are really relevant ones for my course. Then, we can ask it to create real blog posts also. Maybe you can ask it to create a blog post for number one, number two, like that. You can ask, or you can just if you want to create this one. Let me ask it to write a blog post for that title. We have to provide the number of characters or any kind of guidelines targeting beginners, non-technical persons like that, we have to give. Here, you can give comments like change this section, change that section, and replace these words with like that. You can give comments here. So, it's like a collaboration kind of. This feature will be really useful for writing articles and eBooks. For doing digital marketing, publishing an eBook is also a way of doing effective marketing. So, for that also, we can use ChatGPT or Gemini or DeepSeek, Qwen. Here, it's saying that 'let me know if you'd like any tweaks or additions'. So, we can say that add this section. So, if you want to have some additional section or if you feel that it is lengthy, you can ask it to remove this 'network...' this one. Or we can say like combine 4 and 5, just for the example, I'm saying okay, combine sections 4 and 5. So, it will rewrite it. Okay, so this feature, as I said, it's a really good one. But if you don't want to use this feature, you can open a new chat and then there you can ask to do the corrections. So, you have to just copy the existing content and then you can ask ChatGPT to make the modifications, but instead of doing that, this will be very useful. Here, it combined those things. Okay, like if you want to expand, then ask to expand this section with more details. You can ask to expand here like that. Okay, so here you can give more details here. Okay, so this kind of collaborative drafting feature is useful for writing a blog post. Okay. Since I already put the links, it's using that link. Okay, that we need to be very careful about. Suppose if you want to do it for some completely different client, maybe you have to reset the memory. Okay, that we need to remember. That's that thing. Okay, and this one, it's clear, right? So, we can use ChatGPT for writing blog posts and improving the blog posts and things. This is one step, but the best practice is you have to just copy this and put it in Gemini and ask Gemini to improve it. And if you can ask Gemini if there is any mistake here, because as I said, the LLMs are prone to mistakes. Okay, we can ask Gemini whether any incorrect information is there. So, by combining multiple LLMs, we can effectively do the blog post preparation easily. Okay, so it's giving some kind of suggestions. Okay, so either you can manually do those things or you can just ask Gemini to apply these things. Okay, so it will rewrite the article by applying those suggestions. Okay, but anyway, finally, you have to fully review the article; otherwise, still, there may be some issues with that. Okay, and obviously, you can ask Gemini or ChatGPT to follow like for a professional tone or like a funny way of writing, so that you can always specify. Okay, so it's up to you, based on your requirements. So, it's kind of part of prompt engineering, right? So, you need to, as I said, you need to be very familiar with prompt engineering first, and then we can start using it for various purposes, writing blog posts, things. Okay, so you can ask Gemini to write like even for if you need like image representation, image for that, you can ask it to create an image for this post. Okay, so it will create an image, and again, you can specify the aspect ratio. Okay, if you want for image are 16 by 9 aspect ratio, just to provide that thing also. Okay, and if you already have the image, you can just upload the image. Okay. It provided an image, but you need to be very careful. Almost all the spellings will be completely wrong. That is the current status of image creation. Okay, so you can ask Gemini to create images without any kind of words. Okay, that will be one approach, or you can manually change the things, and as I said, you can upload any kind of images and then you can ask Gemini to make the corrections or take the inspiration from that image like that. Okay, a lot of things can be done. Okay, so you have to first know the basics, what are the features available with this. We can upload files, we can upload images, we can do the voice chat, that kind of thing. We have to be familiar, and then it's up to our imagination. Okay, we can effectively use by combining various things. Okay, we can create quality blog posts and relevant images, and if you are creating a video, YouTube video, you can create thumbnail images from here. Okay, you have to provide the aspect ratio in some places, but in Qwen, it is having that option like in case of image generation, it is like asking the aspect ratio. So, in case of YouTube thumbnail, you can provide as 16:9. Okay, thumbnail for video promoting my course. Okay, so here we selected that image generation. The aspect ratio need not worry about the aspect ratio, but in case of Gemini, we have to manually mention the aspect ratio. Okay, 'Unlock AI potential - Enroll now'. Okay, similarly, we can, while running PPC campaigns, for example, if you are running Google ads, you can ask like I'm running Google ads, suggest me the headlines like that. Say, for example, I'm asking it in Qwen. Similarly, we can ask it in ChatGPT also. For that, better we can start with a new chat, and we need to understand that these LLMs are having certain knowledge cutoff. Okay, they won't be having the latest information, but they have the ability to search the web, but it won't be as natural as pre-trained information. So, whenever you are dealing with the latest information, just take additional care. Okay, so here it provided a headline, description, it provided all the details. Okay, you can provide a list of high-conversion keywords also. Okay, so here it is able to create ad copies, and it's suggesting keywords. Okay, but as I said, it's not that much reliable. Anyway, always you have to check with the keyword tools for search volume and other things. Okay, just for like it's a guidance only. Okay, and similarly, here also in DeepSeek, Gemini, here also since we were talking about the image, it created an image copy. Okay, so you can also say this thing. Just you have to manually go through it, or you can copy everything and put it in one LLM and ask it to combine everything and select the best one like that also. Just explore various things, whichever is working for you, continue using it. Okay, so here it is suggesting some keywords. Maybe we can give the list of keywords. So, we are getting some keywords from here and some keywords from there. Right, suppose if you want to find a unique thing, we can ask either, ask, if you are familiar with some kind of Python knowledge, you can ask Python or any kind of development, you can ask ChatGPT or Gemini to create a simple code for doing it. Otherwise, you can ask the LLM itself to do the things. Okay. Merge these things like, include these keywords also and give me the final list, give me the merged list without any duplicates okay, it will combine both things and it will give you the unique list of keywords okay so it is one way of using LLMs for keywords creation and they are providing the broad match right, suppose you want to convert them as phrase matches either you can ask it to do that, or you can create a simple python code okay here it provided, it added quotes, right and similarly you can ask it to exact match, it converted as exact match okay, so these are all very simple things but once you start using it you can understand or you can explore various ways of using it effectively based on your requirements okay so it will be able to create blog post suggest blog posts titles and create keywords, ad copies okay, in case of social media promotion you can ask it to create tweets or Instagram posts, anything, just whatever you want to do you can just ask it okay but we have to give the prompts properly and verify the results properly that's the things we have to do okay for some activities chatgpt may be better for something Geini will be better, in case of extracting details from YouTube videos Gemini will be more suitable like suppose if you want to get some detailed summary of any YouTube video summary you can just give the, here I provided one video for bundle offer right, so I just will ask Gemini write article on this video okay it will help me to save a lot of time once I create the video I can automatically create blog posts or any kind of Articles okay so press release or anything I can just send this okay it's able to extract that books details also okay so this feature will be useful uh from Gman okay and similarly the Qwen it is able to like generate image on specific aspect ratio here we can that okay so just we have to explore, from my personal uh experience each LLM is having its own strength okay we need to explore okay okay right 20. Google AI Studio: Hello everyone, this video is for explaining about Google AI Studio. Google AI Studio is a very useful application. It's a browser-based web application. We can easily access it by typing aistudio.google.com on web browser. It is useful for developing AI applications. Especially it is useful for exploring models provided by Google. Here we can explore various things and then we can get the code of whatever settings we provided here. We can easily implement it in our AI application and coding. That is the main purpose of this AI studio. For example, here we can go through all the models provided by Google, Gemini models, as well as open source, Gemma models. Here we can get the overview about each and every model. If you mouse over on any particular model, here it will be giving the details about the price details input output pricing and the purpose of the model. Best for, best useful for what purpose like that and use cases and a knowledge cut off date, rate limit. We can get all the essential details for all the models so that we can choose appropriate model based on our own requirements. That is one key thing here. We can choose the model. Here we can do the other settings like temperature or other things. Here if we set the temperature as 0.9 and we can just use the get code. Here in the get code, it will be set as 0.9. We can just copy the code. If we are going to use Python code, we can copy this one or if you are planning to use in JavaScript, Go, Android. We have to choose the relevant language or you can straight away use the cUrl also. Normally Python commonly used one. This is the overview about Google AI Studio. We can do the exploration here and we can just copy the code and use it in our application. For that we just need this API key. You can get the API key for using Google Gemini. Just create the API key and then we can start using it. For exploring prompts, here various options are there. The starting point is for if you are a beginner, you can go with this prompt gallery. Here a lot of examples are provided. You can just go through them and just click that and just run the sample prompts and see how it's working for various models. We have to just choose one prompt 'which is bigger'. We can here now I selected Gemini 1.5. If we click it, it will work with this model. This model is saying that this explanation and let us try some advanced models. We can use this thing in the experimental. It's giving some detailed explanation. By doing exploration of various models, you can find out which model is suitable for you. For that you can find it's performance and you can see the details about pricing other details. Finally you can decide. Once you find those things, as I said, we can easily use this 'get code' option and copy the code and use it in our application. This is the overview about this Google AI Studio. Here we can find a lot of sample prompts and you can also explore various things. This one is an experimental image generation. It's a really useful one because as of now it is not available with regular Google Gemini application. Here we can use it for our regular prompting also. For example, if you want to create a birthday card, you can use it or if you want to create a comic it will generate story with images and image editing is a really powerful one. Say for example, if you want to do some kind of image editing, you can use that for example here I am uploading a file, I am uploading my photo and then I will ask to change the color of change the shirt color to red. If you enter it will go to next line. If you want to submit it, you have to use the control enter or click this one. This model is Gemini 2.0 Flash Image Generation. It's immediately a changed the shirt color. Similarly, we can do a lot of things. Say for example, we can ask it to change the background color or remove the background. It will be useful for various purposes. Image editing will be made very simple here. For that purpose also, we can use it. This feature is right now not available with regular Google Gemini. Here we can use it that. But primarily it is for development purpose. But still we can use it for these kind of things also. Similarly, for any kind of further future introductions will be also happening here first. It's a simple web browser application. We can just go through it. We can just start using it easily. Other features also there, we can just share the screen and we can ask any questions using a microphone we can have a detailed discussion. It's like exploring all the models in an easy way. That is the purpose here. In case of sharing in your screen, I think it's available with this model only. For example, I have published a blog post about here. We can ask it what is the screen is showing like that. I'm just sharing the screen. What is about this blog post? It highlights rapid prototyping and experimentation, access to cutting edge Gemini models, versatile prompting interfaces, seamless SDK export, and built in safety controls. Would you like me to go into any of these points in more detail? Can you tell about how to use it effectively? You'd like to know how to use Google AI Studio effectively. Is that right? Yeah, correct. Based on the blog post, it seems that effective use involves leveraging the rapid prototyping and experimentation capabilities. This means testing ideas quickly, iterating on prompts and models, and making use of the quick feedback loops. Additionally, using the versatile prompting interfaces to communicate effectively with the models is important. Finally, you might need to leverage the seamless SDK export feature for integration into larger applications. Is there anything in particular that you'd like to know more about? Okay, that's from the model by visual story also. It's really useful, for this prompt generated story about here. It's creating the image as well as it's really useful, internally using this model. Okay. You create visual story for AI It will be useful for creating videos also, right? Creating some useful images also. We can download the image. Okay, anyway, it's a just one example, but the overall purpose of Google AI Studio is exploring the new models or any kind of models. It's really for new models and we can test with various settings, by changing that parameters that response will be varying. So we can visually test various things. Suppose if temperature in the sense it's creativity. If we reduce the temperature, that creativity will be reduced. So if you want some factual information, then you can set low temperature. That is the onset, but we can visually test it. And then finally, we can choose appropriate things. Okay, here you can take these things also. So basically we can say that it is for exploring the Gemini models. Okay, so just to various models and we can explore them. Okay, and then we can use this get code and use the code in our AI applications using this API key, Gemini API key. 21. Google Gemini Gems: Hello everyone, this video is for telling about Google Gemini's Gems Feature, Okay so it is gemini.google.com, here we can find this Gems manager and you can see the gems also okay if you click the gem manager you can find the pre-made Gems or you can create your new Gem by using this button okay so first of all what is Gem, Gem is something similar to chatgpt's custom GPT okay so we can write the prompts in the gems and we can use it further okay so it will help to save a lot of time here we can see the premade gems so it is coding partner right I selected this coding partner so it's customized towards coding okay so if you want to customize further for python coding or Java coding here what you can do is make a copy option right so you can just make a copy of that Gem okay so coding partner maybe if it is for python you can choose python like that okay here in the instruction you can say that write the code using python like that you can add so instead of typing this entire instructions every time with the gems we can pre-build this prompts okay python code so here we can customize it to python right code to reverse string, I haven't specified as a python but still it's able to create the python code right so similarly you can use it for various purposes just you can explore various use cases of these gems for example if you're planning to write reply for your emails you can just write email replier like email responder so here you can give the instructions right and email reply with the professional tone like that we can say and add signatures it is simple example okay but you can write a lot of things even if you are not comfortable in writing by yourself you can use this assistant okay it will rewrite your simple explanation okay here suppose if I write it so even for writing instruction for this gems we can use this a feature okay here we can provide signature so after setting this up we can use it or here even we need not specify that send a reply prepare reply for email like that just we can copy some sample email so if it is a email then it will just send a reply so it added the signature also okay so this will save a lot of time suppose if you're using Gemini for your business purpose and you may be using it for your personal reason also so for that you can have separate gems right suppose if you have two three customers maybe you can create gems for each customers okay so it's really useful feature it will save a lot of time even if you are are not comfortable in writing the instructions you can use this feature and then you can correct the mistakes and another feature is here you can upload files also knowledge files so it will refer that data also that also one important thing suppose if you are having list of policy details something you can just include that so it will consider that information also when sending reply okay in this case sending reply generally for creating the response it will be using that knowledge also so just you can upload the files okay image and files you can take it from drive also okay so we can just explore various gems lot of gems are there right and writing editor learning coach and if you want you can just make a copy of it and then you can customize it according to your own needs okay so it's a very simple thing only just instead of rewriting the same instructions multiple times in the prompts we can keep it in the centralized place as a gem okay but it is in practical life it will help us to save a lot of time we can give the instructions and as well as we can give files also explore it thanks 22. Getting Started with Google Gemini API: Hello everyone, this video is for explaining about how to use Google Gemini API to create a simple AI application. I am going to keep it very simple so that anyone can easily understand it, just to highlight how to use the API. That is the key behind this video. For using API for creating a application, first of all we should have API key. The basic is like you might be knowing about Gemini.google.com. It is the web interface for using the models, Gemini models available with Google. Here you can see the Gemini models. That are LLM's large language models. We are using that model using this Gemini interface. Similarly, we also can use those models, this 2.0 flash, that 2.0 flash thinking, that kind of models. We can also use that model using that API key. We can also create an application, something similar to this gemini.google.com. That is the thing. We can use it based on our own requirements. Even we can do something better than this one also. But with the API key, we can do those things. That is what I wanted to highlight here. For that first of all, we should get the API key. Google is providing that API key for freely. But with a lot of limitations, token limitations, rating limitations, other things are there. But anyway, for a learner, we can easily start using it freely. For that, just visit this aistudio.google.com website. Here we can do a lot of things. But right now, I am just going to highlight about how to create a simple thing. Just click this, get API key. Here you can create a API key. I have already created a API key. I will just copy it and I will be using it. Suppose if you are first time coming here, you have to just click this create API key. Here we can do a lot of things. We can just explore various models. Here a lot of models are there. So, initially you can choose various things. Based on your requirements, here you can see the model details. The model is best for what kind of things. Say for example, if you are planning to use it for coding, you can use this model like that. Just explore it. But if you are going with some better model means, you will be having limitations. Just explore that things. That is the purpose of this Google AI Studio. But once you decide the models everything, you can do all the settings here. You can take the sample code from here. You can use it. That is the basic. We can get API key and then we can get the code and we can use it. Let me first start with a very simple example. I already created that one. This is that example. It is a very very simple one. We are using the API key and then asking questions to that model and displaying that response. But for that also we have to follow something. The first one is we have to import this Google Generative AI. For that importing that library, first of all we have to install that package. We have to use pip install for installing that one. We have to pip install Google hyphen generative AI. I already installed it. It will be listing that one. Otherwise it will be installing it. It needs a lot of dependency. It may take some time if you are doing it first time. Similarly, you have to export the API key to your environment. I already did that. Then only we will be able to use that. Don't show your API key to anyone. Because if they are using, it will affect your usage. Token limitations will be there. Now I exported that API key. We installed the package and we exported the API key. Now we can use it. It is a very simple one. Similarly, here that model, if that model is not supported by your API access, it is better to list all the models. Then you can use the command for listing that models. Then you can pick that model name from that. By doing that, only I picked this model name. We are just sending this question "Tell the capital of India?" to the Gemini model. We got the answer. If we ask Tell about India, it is a very simple thing. Just I wanted to explain about how to use that API. But we can do a lot of things. It gave the answer. Here, we can do a lot of settings. We can explore from Google AI Studio. Once you change the temperature and other things. Once you finish those things, you can copy the code and do the configuration of that model. This is a very simple thing. It is importing that GenaAI and then setting the API key and configuring GenAI and then model, defining that model, using GenerativeModel. Then using this Generate_content, we are getting the response from that model. That much simple. With just a few lines of code, we will be able to interact with that model. But according to our own requirements, we can improve it. Even we can use Google Gemini or chatGPT, Claude for creating coding. Or you can refer to documents. But it is better to go with the document. If you rely on LLMs, they will be giving coding based on the old things. This one also, old one. Right now, they changed it again. If you see the sample code from this documentation, that will be something different. Maybe in the future, we have to follow that one here. That sample code. They are using this format "from Google import GenAI". They are using GenaAI.client. Normally, it will be keeping on changing. We need to update ourselves. But if you are randomly using some kind of LLMs to create the code, then it's a bit difficult. You need to understand how to use it. That is the key thing. Right now, we just use the Gemini API for getting answers for our question. The next example is, it is from Commandline. Next example, I created a simple example for using a streamlit. Streamlit is a powerful Python library for creating a simple UI for Python AI applications. Here also, we have to first load the API key. But here, the approach is different. We have to use this streamlit folder and secrets.toml file. There, we have to specify that. We need to configure using the API key, then creating model. Same approach. Using streamlit title, we are giving the title. It's a simple chatbot. For that, we are using this steamlit chat_message. This is a default message for getting input. We have to maintain the session state for showing the old discussions also. For that, we are using this. We have to specify the role whether it is user or system or assistant. We have to follow the same format here. In the session state, we are just appending the messages from both assistants as well as from the user. We can see the chat format like user question and assistant answer, then user question and assistant answer. For that, we are doing like this. It's a overview about this simple thing. It will be simply creating simple chatbot using Gemini API. Let me just run it. It is running the streamlit app. Let us wait for opening the browser. It will open another browser. The st.title is showing here. The st.chat_input. Here, we can ask questions like. Here it is showing right. The user question is getting displayed here. It is waiting for assistant answer. It is giving that answer. Here, I set as srteam as false. It is waiting for getting all the reply. If it is true, then it will be updating the character by character. Let me rerun it. I will ask,"Tell about India". Since now, we set the steam as true. It is giving the reply dynamically. I will set it back to false. I am running it again. What is capital of India? I am asking what is capital of India? It is just in the chat format. We are updating st.session_state.messages. We are using the user question as well as assistant answers. It is getting displayed. Actually, it is every time it is displaying. We are appending that into the session state. It is reading that content every time. It is updating in the display. If we ask another question like. What is the capital of USA? Yes, it is giving answer. It is in the chat format. If we ask the question like tell more about it. Here, it is not giving the context. Actually, it is storing the context. But only for the purpose of showing the display in the screen. Whenever we ask the question, it is freshly taking that question and just sending that question into the Gemini model. It is not able to understand the context. It needs context. We have to provide the context. In that case, we have to update the code. I will update the code to keep the context. Keep the context in the sentence, sending the previous questions along with the new question when sending the text into the Gemini model through API. Now, I updated the code with this context full_prompt. Just joining all the messages and then putting it in the prompt. Tell more about it. Now, it is giving details about New Delhi. We have somewhat a complete chatbot. It is able to display the previous content. It is able to include the context also whenever giving answers. That is much easy. Within a few minutes, we will be able to create a complete chatbot. In case if you are planning to improve it further, we can do it. For example, we can provide options to use different models. Similarly, we can improve a lot by improving the prompts. The system prompts we can add more details about our application. We just created a API key for using Google Gemini models. Then we just created the chatbot. Thanks. 23. Chatgpt's GPT-4o modles creates Marketing Images: hello everyone, recently OpenAI has provided its latest GPT4o model uh within chatgpt.com itself that means whenever you use chatgpt.com uh it will be using its latest model GPT4o actually GPT-4o can generate images also, previously chatgpt was using separate Dall-e model uh but this GPT4o model is a visual-language model multimodal model it can generate images also so it seems it's doing it in an effective way, previously whenever chatgpt creates images using dall-e there will be a lot of spelling mistakes okay but now we are not seeing that kind of things so it's very effective and we can use it for creating any kind of promo images I mean advertisements and even infographics also we can create and profile pictures banners so it it is almost doing the graphic designer's work in an effective way we can give the proper prompts and we can get the required output so just try that uh even the free option uh we have but limited daily limits are there but it's okay for regular usage share your thoughts about this new feature whether it's really useful for you I believe for digital marketers affiliate marketers it will be really useful okay thanks bye 24. The Basics of Machine Learning: A Non-Technical Introduction: Machine Learning (ML) is one of the most exciting and rapidly growing fields within Artificial Intelligence (AI), but it often seems complex and intimidating. Whether you’re a business owner, marketer, or simply curious about AI, understanding the basics of machine learning doesn’t have to be overwhelming. In this video, we’ll break down what machine learning is, how it works, and how it can be applied in simple, non-technical terms. What is Machine Learning? In its simplest form, Machine Learning is a subset of AI that enables computers to learn from data and make predictions or decisions without being explicitly programmed. Instead of following specific instructions written by a human, machine learning systems improve over time by identifying patterns in data and using these patterns to make informed decisions. Imagine teaching a child to recognize pictures of cats. At first, you show them various images labeled as “cat” and “not cat.” Over time, the child starts to understand the common features of a cat, like its shape, fur, and ears, and can recognize new pictures of cats without needing to be told each time. This is similar to how machine learning works: the system “learns” from the data and applies that learning to make predictions or decisions. Types of Machine Learning Machine learning can be divided into three main types, each with its own approach to learning from data: 1. Supervised Learning: Supervised learning is the most common type of machine learning. In this approach, the machine is trained using a labeled dataset — meaning the data includes both the inputs (features) and the correct outputs (labels). The goal is to learn a mapping from inputs to outputs, so that the model can predict the correct output for new, unseen data. Example: A common use case is email spam detection. The system is trained on emails labeled as “spam” or “not spam” and learns to classify future emails into the correct category based on features like subject lines, content, and sender information. 2. Unsupervised Learning: In unsupervised learning, the machine is given data without explicit labels. The model must find patterns and relationships in the data on its own. The aim is to group similar data points together or reduce the complexity of the data. Example: Customer segmentation in marketing is a good example of unsupervised learning. A business might use unsupervised learning to identify different groups of customers based on their purchasing behavior, even without knowing in advance what these groups will be. 3. Reinforcement Learning Reinforcement learning is a type of machine learning where an agent learns by interacting with its environment and receiving feedback in the form of rewards or penalties. The goal is to learn strategies that maximize the total reward over time. Example: Reinforcement learning is often used in training AI for games. For instance, a machine learning model can play a video game like chess, learn from each move it makes, and gradually improve its strategy based on the feedback it gets after each game. How Does Machine Learning Work? At the core of machine learning is data. The process of training a machine learning model involves the following basic steps: Data Collection: The first step is gathering data. This can come from a variety of sources, such as customer databases, websites, sensors, or even social media. Data Preparation: Once the data is collected, it must be cleaned and formatted properly. This step often involves removing duplicates, filling in missing values, and converting data into a usable format for the machine learning algorithm. Choosing the Algorithm: There are various algorithms (or models) used in machine learning. Each algorithm has its strengths and weaknesses depending on the problem you’re trying to solve. Some common algorithms include decision trees, linear regression, and neural networks. Training the Model: During training, the machine learning algorithm uses the labeled data (in supervised learning) or the input data (in unsupervised learning) to “learn” by finding patterns and relationships. Evaluation: Once the model is trained, it’s tested on new, unseen data to evaluate how well it performs. This is where the machine’s ability to generalize from its training data is assessed. Making Predictions: After the model is trained and evaluated, it’s ready to be used for making predictions or decisions based on new input data. Real-World Applications of Machine Learning Now that you understand the basics of machine learning, let’s take a look at how it’s applied in the real world: 1. Recommendation Systems: Services like Netflix, YouTube, and Amazon use machine learning to recommend shows, videos, and products based on your past behavior and preferences. 2. Image and Speech Recognition: Machine learning powers image recognition systems in apps like Google Photos, where the system can identify faces or objects in your photos. Similarly, voice assistants like Siri and Alexa use machine learning for speech recognition. 3. Fraud Detection: Banks and credit card companies use machine learning to detect fraudulent transactions. By analyzing patterns in spending behavior, these systems can flag unusual activity and prevent fraud. 4. Autonomous Vehicles: Self-driving cars rely on machine learning to understand their surroundings, make decisions, and navigate safely without human intervention. 5. Healthcare Diagnostics: Machine learning is increasingly used in healthcare to assist doctors in diagnosing diseases. Algorithms can analyze medical images and predict conditions like cancer based on historical patient data. Benefits of Machine Learning Automation: Machine learning automates repetitive tasks, saving time and reducing the need for human intervention. Improved Decision Making: By analyzing large amounts of data, machine learning helps businesses make more informed, data-driven decisions. Personalization: Machine learning enables businesses to personalize services and products for individual users, improving customer experience. Why You Should Care About Machine Learning Machine learning is transforming industries by enabling smarter decision-making, automating processes, and improving efficiency. Whether you’re a business owner looking to leverage AI or simply someone interested in the future of technology, understanding machine learning basics will help you better navigate the rapidly evolving tech landscape. With AI becoming a fundamental part of various industries, getting familiar with machine learning concepts can give you a competitive edge. Don’t worry if you don’t have a technical background — the core principles are straightforward and can be applied in a wide range of fields. Thanks. 25. Data driven insights: In today’s fast-paced world, making informed decisions is crucial for success — whether in business, marketing, or even personal life. But with so much data available, it can be overwhelming to process all the information and make the best choices. That’s where Artificial Intelligence (AI) comes in. AI can be a game-changer when it comes to making data-driven decisions. By analyzing large amounts of data quickly and accurately, AI tools can provide insights that humans might miss, 00:00:35.200 --> 00:00:40.400 offering a clearer picture and guiding smarter, more informed decisions. What is Data-Driven Decision Making? Data-driven decision making (DDDM) is the process of making decisions based on data analysis and interpretation rather than intuition or gut feelings. It involves collecting relevant data, analyzing it, and using the insights to guide decisions that improve outcomes. AI enhances this process by automating data collection, running complex analyses, and generating actionable insights at a speed and scale that is impossible for humans to match manually. How AI Helps in Making Better Decisions 00:01:17.680 --> 00:01:23.680 1. Faster Data Processing AI can process vast amounts of data in a fraction of the time it would take a human to analyze manually. This means businesses can quickly adapt to new information and trends without delay. Whether it's customer behavior, sales data, or industry trends, AI ensures that decision-makers have up-to-date insights to guide their choices. Example: AI-powered tools like Google Analytics or Tableau can analyze website traffic patterns in real-time, helping marketers adjust their campaigns immediately based on what’s working and what’s not. AI excels at identifying hidden patterns in data that might otherwise go unnoticed. By leveraging machine learning algorithms, AI can recognize correlations and trends that humans may miss. This is especially useful in industries like finance, retail, and healthcare, where spotting 00:02:16.560 --> 00:02:22.800 early warning signs or opportunities can mean the difference between success and failure. Example: AI can identify purchasing trends and customer preferences, helping businesses predict what products will be in demand, optimize inventory, and personalize recommendations. 3. Predictive Analytics Predictive analytics is a powerful feature of AI that uses historical data to predict future outcomes. By analyzing past performance, AI can forecast trends, behaviors, and even potential risks. This helps decision-makers anticipate problems and opportunities before they arise, allowing them to make proactive decisions. Example: Salesforce Einstein, an AI-powered tool, uses predictive analytics to help sales 00:03:03.200 --> 00:03:08.240 teams identify which leads are most likely to convert, allowing them to focus their efforts where they matter most. One of the challenges in decision-making is bias — our personal preferences or experiences often cloud our judgment. AI, on the other hand, can analyze data objectively and make recommendations based solely on facts and patterns, rather than emotions or biases. Example: In hiring, AI-powered tools like HireVue use data-driven insights to assess candidates based on their qualifications and experience, reducing bias in the recruitment process and ensuring a more diverse and qualified workforce. Anyway, if you train your AI model with biased data, then AI will also show bias. So, it is important to prepare training data properly. 5. Improving Operational Efficiency AI can analyze business operations and identify inefficiencies or areas for improvement. By automating tasks, optimizing workflows, and providing real-time data, AI helps businesses streamline operations and make decisions that maximize productivity and reduce costs. Example: AI tools like Zapier and Trello can automate repetitive tasks and improve project management, giving decision-makers more time to focus on high-impact activities. 6. Providing Real-Time Insights AI doesn’t just analyze past data — it can also process data in real-time. This allows businesses 00:04:34.240 --> 00:04:39.440 to make decisions based on live data, which is especially important in fast-changing environments like stock trading or social media marketing. Example: AI in financial trading uses real-time market data to make split-second decisions, executing trades faster than a human ever could, which can be the key to success in the stock market. How to Leverage AI for Better Decision Making 1. Adopt AI-Powered Tools: Invest in AI-powered platforms that cater to your business needs. From CRM software to data analysis platforms, many tools are available that can help you make smarter decisions. 2. Train Your Team: Equip your team with the knowledge to interpret AI-generated insights effectively. Understanding how to apply data-driven insights is just as important as gathering the data itself. 3. Start Small: If you’re new to AI, start with one department or use case (such as marketing or sales) and gradually expand as you see results. 4. Monitor and Adjust: AI tools can provide valuable insights, but it’s important to continuously monitor the results and adjust your approach as needed. AI is an evolving field, so staying on top of new developments is crucial. Conclusion: Making Smarter Decisions with AI AI can significantly enhance your decision-making process by providing accurate, real-time, and data-driven insights. By leveraging AI tools to process data faster, identify patterns, predict future trends, and minimize bias, you can make more informed and strategic decisions that drive business success. Whether you’re in marketing, sales, finance, or any other industry, incorporating AI into your decision-making process can help you stay ahead of the competition and make choices that lead to long-term growth.