AI fundamentals

0:00 But today, millions of people do exactly that every single day. But here's the surprising part. Most people know how to use AI. Very few know how it actually works. And that's exactly what you will learn in this complete generative AI course by Intellipath. Whether you are a student, a working professional, a software developer, a data scientist, or simply someone curious about AI. This course is designed to take you from the basics to advanced concepts step by step.

0:43 We will begin by understanding what generative AI actually is, and how it's different from traditional AI. Then we'll explore large language models or LLMs and understand how they generate text that sounds so natural. You will learn about transformers, the architecture behind model AI models, and understand concepts like tokens, embeddings, attention mechanism, and positional encoding in a way that's easy to follow. By the end of this course, you won't just know how to chat with AI, you'll understand how modern AI systems work behind the scenes.

1:27 And more importantly, you will know how to build applications powered by them. So if you're ready to go from simply using AI to actually understanding and creating it, let's get started. Alright. Now that we have seen AI systems interacting with each other, let's step back and understand the basics first. That's where our first question comes in. That is, what exactly is artificial intelligence? So let's study that. So in simple terms, artificial intelligence or AI is the ability of machines to perform the tasks that require human intelligence.

1:55 For example, these tasks include recognizing images, understanding language, making predictions, or even helping in decision making. Alright. However, there's a key difference between human intelligence and machine intelligence. As humans can learn through experiences, emotions, and understanding, machines don't think or feel like humans. Instead, they learn from data. Where AI systems analyze large amounts of information and find patterns within that data. For understanding this in a much more clear way, you can think of AI as a digital brain that learns from patterns and processing the data.

2:33 The more data it processes, the better it becomes at recognizing those patterns and producing more precise and useful results. Now the question comes is, why does AI matter so much today? Like you are hearing about it almost everywhere in almost every field. So the modern world today generates huge amounts of data through searches, photos, videos, and all the online activities. And humans alone cannot analyze this huge information by themselves. But AI can help us process this information and data quickly by turning it into useful insights and information.

3:18 Which is why it is becoming most important across almost every domain. So most of the AI systems today focus on three core capabilities. The first is prediction. Where AI analyzes past data to estimate what might happen next. And the second is recognition. Such as identifying faces, voices, or objects in images. And the third is decision making. Where AI evaluates different options and selects the best one based on the data. What's more interesting is that we are already using AI in almost everything.

3:56 For example, the smartphones uses AI for face unlock and camera improvements, while YouTube recommends videos using AI. Google Translate converts different languages instantly. And chatbots help answer questions online as well. However, AI has limits too. It does not truly understand the world like humans do, because it simply learns from the data. Like, for example, it has no real understanding where AI detects patterns but doesn't truly know what they mean, the way humans can understand it. The second is dependent on data.

4:45 It is completely AI dependent, which is why a poor or biased training data leads to inaccurate or unreliable outputs. The third is no emotions or creativity like humans does. AI cannot feel empathy, experience joy, or generate truly original thought or something that is based on the real human creativity. Lastly, AI can fail unexpectedly too. Edge cases and novel situations can trip up even the best AI models to fail, right? So before we continue to the next module, let me know which is your favorite AI tool that you use almost every day in your daily routine.

5:30 And that brings us to the module two, that is history and evolution of AI, which will help us understand how AI evolved over the time and how it came from basics to the intelligent AI that we live every day today. So the idea of artificial intelligence started in 1940s and 1950s when scientists began asking the simple question. Can machines really think? And one of the key pioneers in this field was Alan Turing. He proposed the famous Turing test, which suggested that if a machine could have conversation indistinguishable from a human, it could be considered as an intelligent machine.

5:00 So let's see the timeline now. So it all began in 1950s where the Turing test started to test early logic of the machines. After that it went to 1970s to 1980s, which was the era of expert systems, which was mostly rule-based. After that came 1990s, which led to machine learning era takeover. After that it came to 2000, where the deep learning revolution actually began. And lastly, into the 2020s, where generative AI and agentic AI started exploding.

5:00 Slowly, but steadily. So let's understand the timelines now. So in between 1970s to 1980s, AI research focused on the expert systems. So these systems tried to mimic the decision-making abilities of human experts that were using large set of rules. They were pretty useful in the fields like medicine and engineering but were difficult to scale at large levels. After that, later, the focus shifted to the machine learning era, where computers started learning through patterns and from data instead of following only fixed kind of rules that was based on expert systems.

5:20 So this approach soon become much more powerful than that. And in the recent years, major breakthroughs happened with deep learning, powering powerful GPUs and availability of big data. This advances allowed AI systems to process complex information and learn much faster than before. And this progress led to the generative AI era, where machines can create new content. For examples, the tools that we use today are effect of generative AI, which includes Chat GPT, which is conversational AI for everyone.

5:20 After that came Dali, which can generate images from the prompts. And after that, he also gave us tools like Stable Diffusion and Midjourney, copilot and even Sora that can generate videos just by a simple prompt. And today, we are moving towards the agentic AI era, where AI systems can now plan multistep tasks, coordinate workflows and act proactively. Where they don't just respond to prompts, but they work autonomously towards goals and takes actions as well. So now if we visualize this as a timeline, AI has moved from simple experiments to powerful intelligent machines and the future is still unfolding.

6:00 So far, we have covered two important things. First, we understood what AI is, what it can do, and why it matters in today's data-driven world. Then we explored how AI evolved over time. From early idea and rule-based system to modern AI powered by data, deep learning and massive computing power. All right, enough of theory for now. Let's make it a bit interesting. So here's the deal. There will be three question presented in this video. And yeah, this is a challenge.

6:00 If you can get all three right and drop your answer in the comment, we'll share you a free AI masterclass with you. All right. So let's just jump to the first question. Question number one: What makes modern AI significantly more powerful compared to earlier approaches? A. Only rule-based programming. B. Smaller datasets. C. Big data, GPUs and deep learning. Or D. Human emotions. Go ahead and drop your answer in the comment below. And now let's move to the next module.

6:30 Okay guys. Now jumping towards module three. That is, type of artificial intelligence. Now that we have seen how artificial intelligence evolve over time. Our next question is quite interesting. Are all artificial intelligence systems are same? The answer is no. Artificial intelligence is generally divided into three types, based on their capability. The first type is Narrow AI, also known as Weak AI. This is a type of artificial intelligence we use today. Narrow AI is designed to perform one specific task very well, but cannot go beyond that task.

6:30 For example, voice assistant, like Siri or Amazon Alexa can answer a question, they can set reminder or play music. Recommendation system on platforms like YouTube, this suggest video based on your watch history. These systems are powerful, but they are limited to the task that they are trained for. Okay. Now coming to the second type, that is Artificial General Intelligence. Also known as AGI. AGI refers to a system that would have human level intelligence. It could learn, reason, solve problem and perform many different tasks, just like a human.

6:30 However, AGI does not yet exist. Researchers are still working towards the goal. The third type is Super AI, which is more of a future concept. Super AI would be a system that become more intelligent than human in almost every area. While this idea is often discovered in research and science fiction, it also raise important of ethical and safety questions about how such powerful system should be controlled. So today, most of the artificial intelligence we are interacting with is Narrow AI.

6:30 But the future may bring much more advanced system like this. Let's look at another emerging concept that's gaining a lot of attention in 2026. Agentic AI. Now, generative AI can create content like text, images or code with a single. right let's take couple of examples there are um you know sort of broadly speaking there are four to five different personas here or four to five different types of stakeholders involved in this whole thing at the most bottom at the bottom most level you have companies like these are your um you know these are sort of your cloud providers so it can be companies like AWS it can be companies like Azure or Microsoft for that matter and so on and so forth so AWS Azure cloud providers these companies are providing compute right they are the ones that are providing the data centers right why are they important because they are the companies that are going to be providing the actual compute for these companies to host the models so it's it's these cloud providers where the models are actually hosted or trained at

7:20 okay but who is actually providing them these models so who's actually helping These consulting companies will will make money on your dead body as well. So these consulting companies are, you know, hopeless. These Bane's and McKinsey's of the world, they'll make money anyhow. Of course they're making a lot of money. Um, they just have a new thing to sell. So then they'll sell that thing, you know, to. It's not my perspective, the point is they, of course will, will make money any which ways, because that's their niche, right?

6:26 How much ever we think they're useless, they're equally useful, because you need them to sell something. They have a certain brand value, they, you know, they do sell a lot of things, you know, especially in an enterprise setup. As much as I hate to have them, they do a very good job of selling things within my company. So I do use them in my own company to sell things in my own company, with. Uh, let me just see accessing the Sora model.

5:40 I'm not sure if the Sora models are available for public consumption. Uh, but let's go here. Where are the models? Uh, API reference. There's audio. Um, yeah, the video models are not available to what I know. You you you need to access it through the interface. I'm not sure, but let me check. Now, whatever you're observing here, Q.K is what is getting computed here in this particular setup, in this particular in this particular step. Um, and that is being combined.

6:00 So whatever you get out of the key is multiplied by what is coming out of the query. Both of these are getting multiplied and then a very both that dot product is being computed here. So this is Z for you. This is the values in terms of Z for you. Whatever you're seeing and these Zs, you're applying a soft max over here. These Zs are essentially going through soft max over here. Right. That soft max. So if you see here, if you look at that computation over here, Q.K transpose or rather Q.K.

7:20 Your query and the key vectors are both getting multiplied and then there is a soft max being applied around it. It is divided by root of then an under root DK and there are some mathematical nuances over there that is for smoothing and and and and those kind of purposes. Um, don't worry about that for the moment. Again, minor detail. But as you can see, the output that you're getting out of here is a soft max output that ranges from minus one to zero.

8:20 Minus one to plus one, it'll typically be only, you know, zero to one in most cases. It just sums up to one. As you can see here, all of these values are simply going to sum up to one. That's it. So those are your attention weights. These are your W1, W2, W3, W4, W5 and so on and so forth. Is this part clear everyone? Do you at least understand how you can get how we got from this raw query and there is bias, there is the usual stuff out there for sure.

5:33 It is always a soft max. Okay. Number one. So how does if you're doing a binary class classification, you would apply a sigmoid. How does sigmoid work? You take sigmoid as an example, one by one plus E to the power of minus whatever is the input that comes from here. I'm going to put h2. Right? So the whatever input that you get from the previous layer, that's h2. Right? Before applying the sigmoid, whatever output you get, this h2, that's a logit.

6:23 Right? So the output before the sigmoid is referred to as a logit. It's the raw output. Then you pass it through some kind of a. Exactly. It's a weighted sum from the previous layer. Your activation from the last layer is referred to not the activation function, but the activated output from the previous layer is essentially referred to as a logit. You put it through a classifier, um, and you put it through some kind of an activation function and get the probability itself.

7:06 Just that part is referred to as a logit. Yeah. And and so just double clicking a little bit into this tokenization, right? is the context the question and the context there are two parameters that need to go in let me go back here again so so there has to be a question and then there has to be But anyway, actually, you can, we can actually see it here as well. That doesn't matter. Let's actually come back here.

5:40 So if you look at this, so almost 3.3 terabytes of data is what we are talking about. That's how big this dataset is. uh, for training and learning isn't, datasets and Kaggle enough? No, man. I mean, this is just to let you know, right? So the common crawl database is 3.3 terabytes. This is the whole of the internet. Publicly available internet. It's as big as that. Right? And of course, there is no way you and I can use any of that data on your on your machines, right?

6:30 3.3 terabytes. Right? And then there is bunch of others, GitHub code, 328 GB. Right, all of the publicly available GitHub. Wikipedia, 83 GB. Wikipedia is as small as that. Stack Exchange. Right, so your your archive, which is nothing but your publicly available papers, 92 GB. Books, 85 GB worth of book. on the pedal. Right? The person sitting next to you, who is your instructor, says, oh, great job. You suddenly see an obstacle. You turn left. Great job.

5:50 You suddenly see a cat. You don't press the press the foot on the. you you instead of breaking, you press the foot on the pedal and you actually accelerate. The cat dies and your and the guys on the the the instructor on the left side gives you one smack at the back and says, man, look, you you you've created an an error here. So now you've learned. Because you with the smack on the back, you've learned that oh, you know what, I'm not supposed to do this.

6:34 So the next time you see a cat, you again turn left. So now you get a thumbs up again. So your head, you're actually trying to course correct and make trying to always make the right choice given a certain certain circumstance. The circumstance here is the prompt and the response. And you're always trying to adjust the response in such a way that you always get a thumbs up. This is the safest route. Right? That you could use any of your large language models.

5:42 This is the only and the best way for you to access these. By the way, Lang chain plays here as well. So now you have the whole Lang chain setup here. So it doesn't matter. Lang chain accesses these models or these models. It doesn't matter. Lang chain also gives you a nice Lang chain or llama index for that matter. Both of them will give you a nice interface. as you can see, there's still some scope for us to improve the quality of the responses here.

5:40 But anyways, um, I hope you understand how the, uh, retrieval augmented generation part sort of works in this case. All right. We spoke about retrieval augmented generation. We looked at, um, how we could perform retrieval augmented generation using the OpenAI API, right? So we used the OpenAI API to actually walk ourselves through this, um, you know, through this example. The objective of this particular exercise was very simple. I think we, this exercise where we, when you, when a question is asked, you convert that particular question into an embedding.

6:20 Okay, the objective of this exercise, larger objective of this particular exercise was to be able to do question answering on a particular question. Is to be able to do question answering on a particular question that you have from a corpus of documents, not so to say, right? That's, that's where we kind of I've I've just stopped it right now. If you want to download, you can actually go ahead and download in the session. Right? You can take a couple of minutes to download.

5:40 It'll take a couple of minutes of course, to download and then maybe a few minutes for it to install as well. So if you want to do it with me, you can just hit it, you know, run the EXE. You just have to hit next, next. But how do I put it in my workflows? Right? So I was able to access the models through a chat and through this terminal. How do I put this in my lang chain sort of a workflow?

5:47 How do I access this particular model? Right? So we'll we're going to be spending the next few minutes doing exactly that. So I'll show you how. Yeah, you don't need I mean, okay. Do we understand why this is beneficial? I mean, is it beneficial? Yes, because it's cheap. Why is it cheap? It's cheap because you're not spending money on, you're actually running it on your compute. So you're not paying open AI, you're not paying anyone for cost.

6:26 You know, to access the model. Is it great? The thing is, I'm using a 2 billion model. A 2 billion parameter model. It's not the greatest of the models. It's okay. Maybe I can live with certain tasks with this. Will it perform on par with an open AI's GPT 4 or? Of course, no. Open AI's GPT 4 is in a different league. 40 is in a different league altogether. Right? Um, or even for that matter, the Gemini models that Google has.

7:11 That's in a different league altogether. This one is