Domain 2: Fundamentals of Generative AI (24%)

← Domain 1 · Next Domain →

---

2.1: What is Generative AI?

Traditional ML: Analyzes and classifies (discriminative)

• Input: Email → Output: Spam/Not Spam

Generative AI: Creates new content (generative)

• Input: "Write a poem about clouds" → Output: Original poem

---

Key Concepts

GenAI Core Concepts

1 / 4

❓

What is a Token?

(Click to reveal)

💡

Basic unit of text for LLMs
Roughly 1 token ≈ 0.75 words
Example: "Hello world" = 2 tokens
Affects cost and processing.

Foundation Models (FMs)

• Large models trained on vast amounts of data
• Can perform multiple tasks without task-specific training
• Examples: GPT-4, Claude, Llama, Stable Diffusion

Large Language Models (LLMs)

• Foundation models specifically for text
• Trained on text from books, websites, articles
• Understand and generate human-like text

Transformers

• Architecture behind modern LLMs
• Key innovation: Attention mechanism
  • Allows model to focus on relevant parts of input
  • Enables understanding of context

---

2.2: Capabilities and Limitations

Capabilities of Generative AI

Text Generation:

• ✅ Write articles, stories, emails
• ✅ Summarize documents
• ✅ Translate languages
• ✅ Answer questions

Code Generation:

• ✅ Write code from descriptions
• ✅ Debug and explain code
• ✅ Generate unit tests

Reasoning:

• ✅ Multi-step problem solving
• ✅ Chain-of-thought reasoning
• ✅ Mathematical calculations (with limitations)

Few-Shot Learning:

• ✅ Learn from just a few examples in prompt
• ✅ Adapt to new tasks without retraining

Multimodal:

• ✅ Process text + images
• ✅ Generate images from text
• ✅ Understand and describe images

---

Limitations of Generative AI

GenAI Limitations

1 / 3

❓

What is Hallucination in GenAI?

(Click to reveal)

💡

Generating false or nonsensical information presented as fact
Why: Model predicts next token, not accessing knowledge database
Mitigation: Use RAG, implement guardrails, human review.

Critical Exam Concept: Hallucinations

Hallucination: When model generates false or nonsensical information presented as fact.

Why it happens:

• Model is predicting next token, not accessing knowledge database
• Trained to be helpful and complete responses
• No built-in fact-checking

Mitigation:

• Use RAG to provide factual context
• Implement guardrails
• Human review for critical applications

Other Limitations:

1. Training Data Cutoff

   • Models don't know events after training
   • Solution: Use RAG for current data

2. Context Window Limits

   • Can't process extremely long documents
   • Solution: Chunking, summarization

3. Computational Cost

   • Expensive to train and run
   • Larger models = higher cost

4. Bias in Training Data

   • Reflects biases in training data
   • Solution: Careful data curation, testing

5. No Real-Time Data

   • Can't access internet or databases
   • Solution: RAG, function calling

6. Lack of True Understanding

   • Pattern matching, not reasoning
   • Can make logical errors

---

2.3: AWS Generative AI Services

Amazon Bedrock

Fully managed service to access foundation models via API

Key Features:

• Multiple model providers (Anthropic, AI21, Cohere, Meta, Stability AI, Amazon)
• Pay-per-use pricing
• No infrastructure management
• Data privacy (your data doesn't train models)

Bedrock Models

1 / 3

❓

Which Bedrock model has the longest context window?

(Click to reveal)

💡

Anthropic Claude
200K token context window
Best for: Long document analysis, reasoning, code generation.

Models Available:

Provider	Model	Strengths
Anthropic	Claude	Long context (200K), reasoning, code
Amazon	Titan Text	Cost-effective, AWS-optimized
Amazon	Titan Embeddings	Text embeddings for search
Amazon	Titan Image Generator	Image generation
AI21 Labs	Jurassic	Multilingual text
Cohere	Command	Text generation, summarization
Meta	Llama 2	Open-source, customizable
Stability AI	Stable Diffusion	Image generation

Use Cases:

• Chatbots and virtual assistants
• Content generation
• Document summarization
• Code generation
• Image generation

---

Amazon Q

AI-powered assistant for business

Capabilities:

• Answer questions about your business data
• Generate content
• Summarize documents
• Connect to enterprise data sources

Use Cases:

• Internal knowledge base queries
• Customer support
• Employee assistance
• Business intelligence

---

Amazon CodeWhisperer

AI coding companion

Features:

• Real-time code suggestions
• Supports multiple languages (Python, Java, JavaScript, etc.)
• Security scanning
• Reference tracking (avoids IP issues)

Use Cases:

• Accelerate development
• Learn new APIs
• Generate boilerplate code
• Unit test generation

---

← Domain 1 · Next Domain →