[00:07] Simply Learn. There was a time when using artificial intelligence meant building a machine learning model, collecting a data set, training that model, and programming it for a very specific task. Today, you can simply
[00:21] describe what you need. You can even ask an AI system to summarize a 100-page document, analyze data, generate code, create an image, explain a difficult concept, draft a marketing campaign, or even help build an application. But,
[00:37] here's the important part. Getting an answer from AI and getting the right answer from AI are two completely different things. You may have already experienced this, right? You ask ChatGPT or another AI tool a question and
[00:51] receive a generic answer. Then, you modify just a few words in your prompt, add context, specify the format you want, and suddenly the response becomes dramatically better. So, have you ever wondered, why does that happen? To
[01:05] understand it, we need to understand what is happening behind the interface. Actually, how does generative AI generate the content? What exactly is a LLM, that is large language model? What are tokens? Why are transformers so
[01:20] important? And most importantly, how can we communicate with these systems more effectively through prompt engineering? That is exactly what we are going to understand in this session. We will start from the foundation and gradually
[01:34] move towards practical prompt engineering techniques that you can use while working with modern generative AI tools. By the end of this video, you won't just be using it, you will be mastering AI. Also, just a quick
[01:47] information, if you are serious about moving beyond simply using AI tools and want to understand how AI systems are actually built and deployed, then you Certificate Program in Artificial Intelligence and Machine Learning. The
[02:01] program takes you through the complete AI learning journey from machine learning and deep learning to generative AI, large language models, Rag, MLOps, and Agentic AI. You will also get exposure to modern AI technologies and
[02:15] framework while working on hands-on real-world projects. The program also includes masterclasses from IIT BHU faculty along with a program completion faculty along with a program completion certificate from the IIIT BHU Varanasi
[02:28] and Microsoft Learn along with Simply Learn. So, instead of learning AI as a collection of isolated tools, you get a structured pathway that helps you understand how the entire modern AI ecosystem fits together. You can find
[02:42] through the link given in the description box below. Go check it out, enroll now, and start your generative AI journey today with Simply Learn. So, quiz question for you. And the question is, which prompting technique provides
[02:57] the AI model with multiple examples before asking it to perform a similar task? And your options are zero-shot prompting, option B few-shot prompting, option C tokenization, or the option D fine-tuning. Think about your answer and
[03:11] let us know your answers in the comment section below. And in case you don't know the answer, then do not worry. Stay till the very end because you'll be finding the answer to this question in this video itself. So, let's begin with
[03:24] generative AI itself. What exactly is generative AI? So, artificial intelligence is the larger field concerned with creating systems capable of performing tasks that normally require some form of human intelligence.
[03:38] Within AI, we have machine learning. Machine learning allows systems to recognize patterns from data instead of requiring every decision to be explicitly performed. Within machine learning, deep learning introduced
[03:52] capable of working with much more complex information such as images, speech, and text. And generative AI extends these capabilities further. Instead of only asking which category does this belong to or what value should
[04:09] we predict, we can now ask create something new. A generative AI system can produce text, code, images, audio, designs, summaries, explanations, and many other forms of content. For example, traditional machine learning
[04:24] example, traditional machine learning might classify an email as spam or not as spam, but generative AI can actually write that email. A conversational computer vision model might identify the objects inside an image, but a
[04:37] multimodal generative AI system can examine the image and explain what's happening inside it. The shift from mainly predicting and classifying towards generating and interacting is one of the reasons generative AI has
[04:51] become so influential. But generative AI didn't appear overnight. The journey began with early ideas around whether machines could demonstrate intelligent behavior. Then machine learning allowed computers
[05:05] to learn patterns from the data. Neural networks and deep learning made it possible to process increasingly complex information. Architectures such complex information. Architectures such as recurrent neural networks or RNNs
[05:19] improved the processing of sequential information such as text. Then long short-term memory networks, better known as LSTMs, helped these systems retain as LSTMs, helped these systems retain information across longer sequences. But
[05:34] language still presented a major challenge. Consider the sequence. The developer updated the application because it had security issue. To understand that sentence correctly, a model needs to recognize relationships
[05:48] between word even when those words aren't next to each other. This is where the attention mechanism became extremely important. Attention allows a model to determine which parts of an input deserves greater focus when processing
[06:02] other parts and eventually this idea became center of transformer architecture. Transformers fundamentally changed natural language processing and became the architecture behind many modern large language models. That's how
[06:17] we reach system capable of generating natural contextual responses at the enormous scale. Now, let's understand the term large language model that is LLM. We can break it into three words where first word is large. So, large
[06:33] refers to the enormous scale of these systems. Their training data, computational requirements, and model parameters. Now, the second word is language. It refers to their ability to process and
[06:47] generate human language. And model refers to mathematical system that has training data. But, an LLM doesn't process language exactly the way humans see it. Suppose you enter artificial
[07:02] intelligence is changing business. Before the model processes this sentence, the text is divided into smaller units called as tokens. A token might represent an entire word, part of a word, punctuation, or another textual
[07:17] unit depending upon the depending on the tokenizer being used. Those tokens are then represented numerically so that the neural networks can process them. When an LLM generates an answer, it essentially estimates
[07:32] which token is most appropriate to generate next based on the context it has already received. It generates one token then uses that token along with the existing context to predict the next one. Then the next and then the next.
[07:47] complete. That means an LLM is fundamentally a probabilistic system. It isn't retrieving one permanently stored answer
[07:59] to every question. Instead, it is generating an output based on patterns information available in the current context. This also explains why changing your prompt can produce a different response. Now, let's return to
[08:14] transformers. Why did the transformers models become so important? One major reason is attention. Consider the sentence. Ria gave Priya the laptop because she needed it for the presentation. There
[08:28] are several relationships the model must understand. Who does she refer to here? What does it refer to here? Which word matters the most when interpretating the sentence? Attention mechanism helps model evaluate
[08:43] relationship between different elements of the sequence. Instead of treating every word as equally important, the model can assign different levels of importance depending on the context. Think about the way you study in a
[08:56] textbook. You don't treat every sentence as equally important, right? You underline definitions, you highlight important arguments, you connect one concept with another concept several paragraphs earlier.
[09:10] >> So, attention gives neural networks a mechanism that helps them make similar contextual connection mathematically. Modern transformer based models use this mechanism extensively, which is one of the reasons they are capable of handling
[09:25] large languages. Once we have understood how LLM processes information, prompt engineering begins to make some more sense. Now, let's look at some of the basic fundamentals of prompt engineering. A prompt isn't simply a
[09:38] question. It is information and instruction you provide to the model so it can understand what you want. See, compare these two prompts. First prompt says, "Write about AI." Now, compare that with explain generative AI to a
[09:55] beginner in approximately 200 words. "Use simple English, provide one business example, and end with three key takeaways." Both prompts discuss the same subject, but the second prompt give you the model far more useful guidance.
[10:10] So, the first one was a weak prompt and the second one is a strong prompt. Which one would you choose? Let me know in the comment section below. See, a strong prompt usually communicates four important things. First, context. What
[10:25] situation are we dealing with? Then the task. What should the model do? Third one is constraints. What rules should it follow? And fourth one is output format. How should the final response be presented? You can also specify things
[10:40] like audience, tone, length, examples, terminologies, and structure. For example, you're explaining machine learning to the first year college student. "Explain supervised and unsupervised learning using simple
[10:54] language. Give an example of each and present the final comparison in a table." Now, the model has much more information about the expected result. That's the fundamental principle of prompt engineering. The clearer the
[11:08] instruction, the easier it becomes for the model to generate a useful response. Now, let's move into different prompting approaches, and here we will be discussing zero-shot, one-shot, and few-shot prompting. So, let's start with
[11:23] zero-shot prompting. Zero-shot means that you give the model an instruction without providing an example. For instance, "Write a professional LinkedIn post announcing the launch of a new analytical product. There is instruction
[11:37] analytical product. There is instruction but no example. Next comes one-shot Here you provide one example showing the model what kind of response you are expecting from it. For example, you can write, "Here is an example of the style
[11:51] I want and so and so on." Then you provide a simple post before asking the model to generate another one. So, this was one-shot prompting. And finally, we have few-shot prompting. Here, instead of providing one example, you provide
[12:07] several examples. This helps establish a clearer pattern. It is important to understand that these examples are not retraining the underlying model. Instead, you are giving the model additional information within the
[12:20] current context. This is commonly referred as in-context learning. If consistency of tone, structure, classification, or formatting is important, examples can significantly clarify what you expect. So, these were
[12:35] the three types of prompting, and I would always recommend using few-shot prompting because the better the prompt, the better would be the result. See, some problems cannot be handled effectively with one broad instruction.
[12:49] For complicated task, it can be helpful to break the work into stages. Suppose you're creating a product strategy. Instead of asking, "Create a complete strategy for launching a product," you could break the workflow down. First,
[13:04] could break the workflow down. First, identify the target users. Then, based upon those users, identify their top problems. Then, based upon those problems, suggest three positioning approaches. And then,
[13:18] compare those approaches and recommend the best one. The output from one stage provide context for the next. Breaking a complex task into smaller stages makes the process easier to inspect and refine. So, if something goes wrong, you
[13:34] can identify exactly which stage needs improvement rather than rewriting the entire request. So, this was the benefit of chain-based reasoning. Now, the of chain-based reasoning. Now, the another useful idea is self-consistency.
[13:48] For reasoning-heavy tasks, instead of relying immediately on one generated solution, multiple candidate approaches can be explored and compared before settling on the final result. The underlying principle is simple. If a
[14:01] problem can be approached in several ways, comparing those approaches may produce a more dependable conclusion. Then, we have tree-based exploration. Rather than following only one path, the problem can be explored through several
[14:16] possible branches. For example, suppose you're trying to reduce customer churn. One path might investigate product problems, another path might investigate pricing, another could examine customer support, and another could examine
[14:30] onboarding. Each branch represents a potential direction. The most useful directions can be evaluated further. So, be valuable for planning, problem-solving strategy, and other
[14:44] complex tasks where there may not be immediately obvious path. Now, let's look at how we can improve our prompts. So, what should you do when an AI
[14:56] response isn't good enough? See, don't immediately assume the model can perform the task. First, inspect the prompt. Ask yourself, "Did I clearly prompt. Ask yourself, "Did I clearly explain the objective?" Second, "Did I
[15:10] provide enough context?" Third, "Did I specify the intended audience?" And the fourth point, "Did I explain the output format easily?" Next, "Would an example help?" And is the task too large and should be broken
[15:25] into smaller stages? For example, instead of writing improve my resume, try writing review the following resume for a data analyst role, identify the five weakest bullet points, rewrite each using action verbs, measurable impact
[15:40] where supported, and concise professional language, do not invent achievements. That gives the model far clearer boundaries. So, this is how you should write your prompt. Prompt engineering is therefore less
[15:54] about finding some secret collection of magical words and more about communicating requirements clearly and systematically. And that's ultimately the skill that makes generative AI more useful. And that brings us to the end of
[16:09] this session. We started with generative AI and followed the technology all the way to large language models, tokens, transformers, and attention. Then we moved from understanding how these models work to understanding how we can
[16:23] communicate with them more effectively through structured prompts, examples, and advanced prompting approaches. The biggest takeaway is simple. AI capability matters, but how you communicate your requirements matters,
[16:35] too. So, the next time an AI tool gives you an average response, don't just regenerate the answer. Look at your prompt, improve the context, define the task more clearly, and experiment with the technique you are using. So, if you
[16:50] found this session helpful, then try one of the prompting approaches we covered today. And tell us in the comment section below how the results change. If you have any doubt or any queries, let us know in the comment section below,
[17:02] and our team of experts will be happy to help you as soon as possible. Thanks for watching, and I'll see you in the next session.