Describing Generative AI capabilities in Azure

Exam weight: 20–25%

Generative AI is the form of AI most people are familiar with today. It is what we use to chat, generate documents, ask questions, and perform everyday tasks.
Generative AI aims to get as close as possible to the way humans communicate.

Despite its almost magical ability to produce original content, it is grounded in mathematical techniques and advances in statistics, data science, and machine learning.
Understanding its core concepts helps us envision new possibilities for artificial intelligence.

LLMs (Large Language Models)

LLMs are the core of generative AI, encapsulating the linguistic and semantic relationships between words and sentences.

When you are typing on your phone and it suggests the next word — often correctly — you are already experiencing a form of language model in action.

The “magic” behind this lies in the model’s ability to learn linguistic patterns and predict likely tokens based on context.

Consider the following sentence:

I heard a dog barking loudly at a cat.

If we only heard “I heard a dog …”, you already know that the words “heard” and “dog” strongly indicate that the next word might be “barking”.

This happens because:

• We have a large vocabulary

• We have learned common linguistic structures

• We have semantic knowledge associated with words

Language models are trained to develop these same capabilities.

Tokenization

For a model to develop human-like abilities to understand relationships between words, large language models contain hundreds of thousands of tokens, representing a vast vocabulary of words, images, and sounds.

A simplified example of tokenization:

• I (1)

• heard (2)

• a (3)

• dog (4)

• barking (5)

• loudly (6)

• at (7)

• a (3) — already assigned

• cat (8)

To learn more about how models transform tokens using transformers, see the official documentation:
https://learn.microsoft.com/training/modules/fundamentals-generative-ai/3-language-models?pivots=text

Prompts

When we ask something of a generative AI model, we are providing a prompt as input to receive a response.
The more relevant information we provide, the more accurate and useful the response tends to be.

Through prompts, we can define roles that guide how the model should respond. There are two main types of prompts:

System prompt

Defines the personality or behavior the model should adopt when responding to users.
Example:

“You are a helpful assistant that responds in a cheerful and friendly manner.”

User prompt

Defines the specific request the user wants answered.
Example:

“Summarize the main points of document X into six bullet points.”

Conversation history

Today, many people share prompt examples on social media to extract the best possible results from generative AI models.
Using conversation history to enrich responses can lead to a much more interesting experience.

Have you ever asked the AI you use most often to generate a professional profile about you based on your entire conversation history? The result is usually quite interesting.

For example, suppose you send the prompt:

“What is the maximum amount I can claim for travel expenses on a business trip?”

Without additional context, the model will provide general guidance and likely suggest consulting your organization’s expense policy.
A better approach would be to include the relevant section of the organization’s expense policy along with the question.
With this added context, the model can deliver a much more accurate and relevant response.

Tips for better prompts

To get better results from generative AI models:

• Be clear and specific — explicit instructions or questions work better than vague language

• Add context — mention the topic, target audience, or desired format

• Use examples — if you want a specific style, provide an example

• Request structure — such as bullet points, tables, or numbered lists

Well-designed prompts can make a significant difference in the quality of the results produced by generative AI models.

Conclusion

This content summarizes the main concepts covered in the official Azure documentation for the AZ-900 / AI-900 exam.
The goal is to transform technical material into simple and practical explanations, helping build a solid understanding of the fundamentals before moving on to hands-on environments.

If you are studying for the certification, I recommend reviewing each topic carefully and experimenting in the Azure Portal whenever possible.