AI-901 Microsoft Azure AI Fundamentals Study Cram

Key Concepts

• Generative AI (GenAI): Software that mimics human capabilities (prediction, creativity, reasoning) to generate new content (text, images, audio).
• Large Language Models (LLMs) vs. Small Language Models (SLMs): LLMs (billions/trillions of parameters) offer high reasoning; SLMs (millions of parameters) offer faster, cheaper, and more efficient inference.
• Agents: Autonomous systems that use models, tools, and knowledge to achieve specific user goals.
• Responsible AI: A framework ensuring AI is fair, reliable, safe, private, secure, inclusive, transparent, and accountable.
• Azure AI Foundry: The central platform for discovering, building, and operating AI applications and agents.
• Tokens & Embeddings: Tokens are numerical representations of text; embeddings are high-dimensional vectors representing semantic meaning.
• Multimodal vs. Multimodel: Multimodal refers to a single model handling multiple input/output types (e.g., text and image); Multimodel refers to an application using multiple distinct models to perform a task.
• Deterministic vs. Non-deterministic: Deterministic models (e.g., Azure Language) provide consistent outputs for the same input; Generative models are non-deterministic.

---

1. Azure AI Fundamentals & Infrastructure

• Identity Management: Microsoft Entra ID is the primary identity provider. Organizations use "tenants" to manage users and resources.
• Resource Hierarchy:
  • Subscription: The billing and access boundary.
  • Resource Group: A logical container for resources (cannot be nested).
  • Resources: Specific services (e.g., Azure AI Foundry, Storage Accounts).
• Authentication:
  • Entra ID (Recommended): Uses managed identities to eliminate the need for hardcoded secrets.
  • API Keys: Used for authentication but must be stored securely (e.g., Azure Key Vault) and never committed to code repositories.

2. Microsoft Azure AI Foundry

• Model Deployment: Users can deploy models globally, within a "Data Zone" (US/Europe for regulatory compliance), or regionally.
• Guardrails: Built-in safety mechanisms to prevent jailbreaking, content abuse, and exposure of protected materials.
• Agents:
  • Prompt-based: No-code; defined by instructions (system prompts).
  • Hosted: Pro-code; developers write code, package it as an image, and host it within the environment.
• Tools & Knowledge: Agents can be extended using Model Context Protocol (MCP) servers and Azure AI Search to access external data not present in the model's original training set.

3. AI Capabilities & Methodologies

• Natural Language Processing (NLP): Used for tokenization, entity extraction, and sentiment analysis. While GenAI can perform these, specialized Azure Language services are often preferred for deterministic, cost-effective results.
• Speech Services: Includes Speech-to-Text (transcription) and Text-to-Speech (synthesis).
• Computer Vision:
  • Image Classification: Assigns a single label to an image.
  • Object Detection: Identifies objects and their coordinates.
  • Semantic Segmentation: Identifies specific pixels belonging to an object.
• Image/Video Generation: Based on Diffusion models, which learn to reverse "noise" to materialize images from pure static.
• Information Extraction: Uses OCR and intelligent classifiers to turn unstructured data (receipts, invoices) into structured, usable formats.

4. Development Framework

• SDKs: Developers should use language-specific SDKs (e.g., Python openai library) to abstract REST API calls and handle authentication.
• System Prompts: Crucial for defining agent behavior (e.g., "answer like a pirate").
• Inference: The process where the model "thinks" and generates a response based on the input prompt.

5. Synthesis and Exam Strategy

• Preparation: Utilize the official AI-901 study guide, complete the free Microsoft online learning modules, and perform hands-on labs using an Azure trial subscription.
• Exam Focus: The exam tests the ability to identify which AI capability to use for a specific business problem rather than requiring complex architectural design or advanced coding.
• Key Takeaway: Always prioritize Responsible AI principles. When choosing between models, evaluate if the task requires the creative flexibility of a Generative model or the consistency and cost-efficiency of a specialized deterministic model.