Table of Contents
Generative AI and Large Language Models
A Cross-Vendor Training Guide
Certification Alignment: Azure AI-102, AWS ML Specialty, Google ML Engineer, Salesforce AI Specialist, NVIDIA DLI GenAI
Introduction
Generative AI represents a paradigm shift in artificial intelligence. Rather than simply classifying or predicting, these models create new content—text, images, code, audio, and video. Large Language Models (LLMs) like GPT-4, Claude, Gemini, and Llama have demonstrated remarkable capabilities in understanding and generating human-like text.
This guide provides comprehensive coverage of generative AI concepts, architectures, and practical implementation across all major cloud platforms.
What Is Generative AI?
Generative AI refers to AI systems that can create new content rather than just analyzing existing data. These models learn patterns from training data and use that knowledge to generate novel outputs.
Generative vs. Discriminative Models
|
Aspect |
Discriminative |
Generative |
|
Goal |
Classify or predict labels |
Generate new data samples |
|
Learns |
P(y|x) – probability of label given input |
P(x) or P(x|y) – data distribution |
|
Examples |
Logistic Regression, SVM, CNN classifiers |
GPT, DALL-E, Stable Diffusion, VAEs |
|
Output |
Class labels, predictions |
New text, images, audio, code |
Types of Generative AI
|
Type |
Description |
Examples |
|
Text Generation |
Create human-like text, answer questions, write code |
GPT-4, Claude, Gemini, Llama |
|
Image Generation |
Create images from text descriptions |
DALL-E 3, Midjourney, Stable Diffusion |
|
Code Generation |
Write and explain programming code |
GitHub Copilot, CodeLlama, StarCoder |
|
Audio Generation |
Create speech, music, sound effects |
Eleven Labs, Suno, AudioCraft |
|
Video Generation |
Create video content from prompts |
Sora, Runway, Pika |
|
Multimodal |
Process and generate multiple modalities |
GPT-4V, Gemini, Claude 3 |
Foundation Models
Foundation models are large AI models trained on broad data that can be adapted to many downstream tasks. They represent a shift from task-specific models to general-purpose AI.
Characteristics of Foundation Models
- Scale: Billions to trillions of parameters
- Pre-training: Trained on massive, diverse datasets
- Adaptability: Fine-tuned or prompted for specific tasks
- Emergence: Capabilities that emerge only at scale
- Transfer Learning: Knowledge transfers across domains
Major Foundation Model Providers
|
Provider |
Models |
Access |
|
OpenAI |
GPT-4, GPT-4o, DALL-E 3 |
API, Azure OpenAI Service |
|
Anthropic |
Claude 3.5 Sonnet, Claude 3 Opus |
API, AWS Bedrock, Google Vertex |
|
|
Gemini Pro, Gemini Ultra, PaLM 2 |
Vertex AI, Google AI Studio |
|
Meta |
Llama 3, Llama 2, Code Llama |
Open source, cloud platforms |
|
Mistral |
Mistral Large, Mixtral 8x7B |
API, AWS Bedrock, Azure |
|
Cohere |
Command R+, Embed |
API, AWS Bedrock |
The Transformer Architecture
The Transformer architecture, introduced in the 2017 paper “Attention Is All You Need,” is the foundation of modern LLMs. It revolutionized NLP by enabling parallel processing of sequences through self-attention.
Key Components
1. Self-Attention Mechanism
Self-attention allows each position in a sequence to attend to all other positions, capturing long-range dependencies.
How it works:
- Each token is transformed into Query (Q), Key (K), and Value (V) vectors
- Attention scores computed: Attention(Q,K,V) = softmax(QK^T / √d_k) × V
- Scores determine how much each token attends to every other token
2. Multi-Head Attention
Instead of single attention, use multiple attention “heads” in parallel, each learning different relationship patterns.
Benefit: Captures different types of relationships (syntactic, semantic, positional)
3. Positional Encoding
Since attention has no inherent notion of order, positional information is added to embeddings.
Methods: Sinusoidal encoding (original), learned positional embeddings, Rotary Position Embedding (RoPE)
4. Feed-Forward Networks
After attention, each position passes through identical feed-forward networks (typically 2-layer MLPs).
5. Layer Normalization and Residual Connections
Stabilize training and enable very deep networks through skip connections and normalization.
Transformer Variants
|
Type |
Architecture |
Examples |
|
Encoder-Only |
Bidirectional attention, good for understanding |
BERT, RoBERTa |
|
Decoder-Only |
Causal attention, good for generation |
GPT, Llama, Claude |
|
Encoder-Decoder |
Both components, good for translation |
T5, BART |
Vendor References:
- NVIDIA: nvidia.com/blog/understanding-transformer-model-architectures/
- Google: org/text/tutorials/transformer
Tokenization
Tokenization converts text into numerical tokens that models can process. The tokenization strategy significantly impacts model performance and efficiency.
Tokenization Methods
|
Method |
Description |
Used By |
|
Word-level |
Each word is a token. Large vocabulary, OOV issues. |
Older models |
|
Character-level |
Each character is a token. Small vocab, long sequences. |
Specialized tasks |
|
BPE |
Byte-Pair Encoding. Merges frequent character pairs. |
GPT, RoBERTa |
|
WordPiece |
Similar to BPE with likelihood-based merging. |
BERT |
|
SentencePiece |
Language-agnostic, treats text as raw bytes. |
T5, Llama |
Context Windows
The context window defines how much text a model can process at once. Longer contexts enable better understanding but increase computational cost quadratically.
|
Model |
Context Window |
Approx. Words |
|
GPT-3.5 |
4K – 16K tokens |
3K – 12K words |
|
GPT-4 Turbo |
128K tokens |
~96K words |
|
Claude 3 |
200K tokens |
~150K words |
|
Gemini 1.5 Pro |
1M+ tokens |
~750K words |
Pre-training and Fine-tuning
Pre-training
Foundation models are pre-trained on massive datasets using self-supervised objectives.
Common Pre-training Objectives:
- Causal Language Modeling (CLM): Predict next token given previous tokens (GPT-style)
- Masked Language Modeling (MLM): Predict masked tokens (BERT-style)
- Span Corruption: Predict corrupted spans (T5-style)
Fine-tuning Methods
Adapt pre-trained models to specific tasks or domains.
Full Fine-tuning
Update all model parameters. Highest quality but most expensive.
Use when: You have sufficient data and compute, need maximum customization
Parameter-Efficient Fine-tuning (PEFT)
Update only a small subset of parameters, keeping most frozen.
LoRA (Low-Rank Adaptation):
- Inject trainable low-rank matrices into attention layers
- Typically <1% of original parameters trained
- Can be merged back for inference efficiency
QLoRA:
- Combines LoRA with 4-bit quantization
- Enables fine-tuning large models on consumer GPUs
Adapter Layers:
- Insert small trainable modules between frozen layers
Instruction Fine-tuning
Train model to follow instructions using instruction-response pairs.
Result: Models that are better at following user requests
RLHF (Reinforcement Learning from Human Feedback)
Align model outputs with human preferences using reinforcement learning.
Process:
- Collect human preferences on model outputs
- Train a reward model on these preferences
- Use PPO or similar to optimize LLM against reward model
Vendor Fine-tuning Services
|
Vendor |
Service |
Documentation |
|
Microsoft |
Azure OpenAI Fine-tuning |
learn.microsoft.com/azure/ai-services/openai/how-to/fine-tuning |
|
AWS |
Bedrock Custom Models |
docs.aws.amazon.com/bedrock/latest/userguide/custom-models.html |
|
|
Vertex AI Model Tuning |
cloud.google.com/vertex-ai/docs/generative-ai/models/tune-models |
|
NVIDIA |
NeMo Framework |
docs.nvidia.com/nemo-framework/user-guide/latest/ |
Prompt Engineering
Prompt engineering is the practice of designing inputs that elicit desired outputs from LLMs. It’s a critical skill for working effectively with generative AI.
Prompting Techniques
1. Zero-Shot Prompting
Ask the model to perform a task without examples.
Example: “Classify this review as positive or negative: ‘Great product!'”
2. Few-Shot Prompting
Provide examples to guide the model’s behavior.
Example: “Review: ‘Loved it!’ → Positive. Review: ‘Terrible.’ → Negative. Review: ‘Pretty good’ → “
3. Chain-of-Thought (CoT)
Ask the model to show its reasoning step-by-step.
Trigger: Add “Let’s think step by step” or provide reasoning examples
Benefit: Dramatically improves performance on complex reasoning tasks
4. ReAct (Reasoning + Acting)
Combine reasoning traces with actions (like tool use).
Pattern: Thought → Action → Observation → Thought → …
5. Self-Consistency
Generate multiple reasoning paths and take majority vote.
Benefit: Reduces errors by leveraging diverse reasoning
Prompt Structure Best Practices
- Role/Persona: “You are an expert data scientist…”
- Context: Provide relevant background information
- Task: Clear, specific instruction
- Format: Specify desired output structure
- Examples: Provide few-shot examples when helpful
- Constraints: Define limitations and requirements
Vendor Prompt Engineering Resources
|
Vendor |
Documentation |
|
Microsoft |
learn.microsoft.com/azure/ai-services/openai/concepts/prompt-engineering |
|
|
cloud.google.com/vertex-ai/docs/generative-ai/learn/prompts/introduction-prompt-design |
|
AWS |
docs.aws.amazon.com/bedrock/latest/userguide/prompt-engineering-guidelines.html |
|
Anthropic |
docs.anthropic.com/en/docs/build-with-claude/prompt-engineering |
Retrieval-Augmented Generation (RAG)
RAG combines LLMs with external knowledge retrieval to provide more accurate, up-to-date, and verifiable responses.
Why RAG?
LLM Limitations:
- Knowledge cutoff – can’t access recent information
- Hallucinations – generate plausible but false information
- No access to private/proprietary data
RAG Solutions:
- Ground responses in retrieved documents
- Access current and private information
- Provide citations for verification
RAG Architecture
- Indexing: Documents → Chunking → Embedding → Vector Store
- Retrieval: Query → Embed → Similarity Search → Relevant Chunks
- Generation: Prompt + Context → LLM → Response
Key RAG Components
Chunking Strategies
|
Strategy |
Description |
Best For |
|
Fixed-size |
Split at fixed token/character count |
Simple documents |
|
Semantic |
Split at natural boundaries (paragraphs) |
Structured content |
|
Recursive |
Hierarchical splitting with overlap |
General purpose |
|
Document-aware |
Respect document structure (headers) |
Technical docs |
Vector Embeddings
Convert text to dense numerical vectors that capture semantic meaning.
Popular Embedding Models:
- OpenAI text-embedding-3-large
- Cohere Embed v3
- Google Vertex Embeddings
- Open source: BGE, E5, GTE
Vector Databases
|
Database |
Type |
Best For |
|
Pinecone |
Managed cloud service |
Production, scale |
|
Weaviate |
Open source, managed |
Hybrid search |
|
Chroma |
Open source, embedded |
Development, prototyping |
|
pgvector |
PostgreSQL extension |
Existing Postgres infra |
|
Qdrant |
Open source, managed |
Performance, filtering |
Vendor RAG Services
|
Vendor |
Service |
Documentation |
|
AWS |
Bedrock Knowledge Bases |
docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html |
|
Microsoft |
Azure AI Search + OpenAI |
learn.microsoft.com/azure/search/retrieval-augmented-generation-overview |
|
|
Vertex AI RAG Engine |
cloud.google.com/vertex-ai/docs/generative-ai/rag-overview |
|
NVIDIA |
NeMo Retriever |
developer.nvidia.com/blog/rag-101-retrieval-augmented-generation-questions-answered/ |
AI Agents and Tool Use
AI agents are LLM-powered systems that can take actions, use tools, and work autonomously toward goals.
Agent Capabilities
- Tool Use: Call APIs, search web, execute code, query databases
- Planning: Break complex tasks into steps
- Memory: Maintain context across interactions
- Reasoning: Decide which actions to take
- Self-correction: Learn from errors and adjust
Function Calling / Tool Use
Modern LLMs can be trained to output structured function calls that applications can execute.
Pattern:
- Define available tools with schemas
- LLM decides when/which tool to use
- Application executes the function
- Results returned to LLM for continued reasoning
Agent Frameworks
|
Framework |
Description |
|
LangChain |
Popular framework for building LLM applications with agents, chains, and tools |
|
LlamaIndex |
Framework focused on data ingestion and RAG with agent capabilities |
|
AutoGen |
Microsoft’s multi-agent conversation framework |
|
CrewAI |
Framework for orchestrating role-playing AI agents |
Vendor Agent Services
|
Vendor |
Service |
Documentation |
|
Microsoft |
Azure AI Agent Service |
learn.microsoft.com/azure/ai-services/agents/ |
|
AWS |
Bedrock Agents |
docs.aws.amazon.com/bedrock/latest/userguide/agents.html |
|
|
Vertex AI Agent Builder |
cloud.google.com/vertex-ai/docs/generative-ai/agent-builder/overview |
|
Salesforce |
Agentforce |
salesforce.com/agentforce/ |
Generative AI Safety
Ensuring safe and responsible use of generative AI requires addressing unique risks.
Key Risks
- Hallucinations: Model generates false but plausible information
- Harmful Content: Generation of toxic, illegal, or dangerous content
- Prompt Injection: Malicious inputs that hijack model behavior
- Data Leakage: Model reveals training data or private information
- Bias Amplification: Model amplifies biases present in training data
Mitigation Strategies
- Input Validation: Filter and sanitize user inputs
- Output Filtering: Detect and block harmful outputs
- Grounding: Use RAG to ground responses in verified sources
- Guardrails: Implement content policies and system prompts
- Human Review: Human-in-the-loop for high-stakes outputs
- Red Teaming: Proactively test for vulnerabilities
Key Takeaways
- Generative AI creates new content – text, images, code, audio, video
- Foundation models are general-purpose – adapt through fine-tuning or prompting
- Transformers power modern LLMs – self-attention enables parallel processing
- Prompt engineering is essential – technique choice significantly impacts results
- RAG grounds responses in facts – reduces hallucinations, enables private data
- Fine-tuning customizes models – PEFT methods enable efficient adaptation
- Agents extend LLM capabilities – tool use, planning, autonomous action
- Safety requires active measures – guardrails, filtering, human oversight
Additional Learning Resources
Official Vendor Documentation
- Azure OpenAI: microsoft.com/azure/ai-services/openai/
- AWS Bedrock: aws.amazon.com/bedrock/
- Google Vertex AI: google.com/vertex-ai/docs/generative-ai/learn/overview
- Salesforce Einstein GPT: salesforce.com/s/articleView?id=sf.generative_ai_trust_layer.htm
- NVIDIA NeMo: nvidia.com/nemo-framework/user-guide/latest/
Certification Preparation
- Azure AI-102: microsoft.com/certifications/exams/ai-102
- AWS ML Specialty: amazon.com/certification/certified-machine-learning-specialty/
- Google ML Engineer: google.com/learn/certification/machine-learning-engineer
- Salesforce AI Specialist: salesforce.com/credentials/aispecialist
- NVIDIA DLI GenAI: nvidia.com
Article 6 | AI/ML Training Series – Generative AI Track
PowerKram Career Preparation Resources
Preparing for a certification exam aligned with this content? PowerKram offers objective-based practice exams built by industry experts, with detailed explanations for every question and scoring by vendor domain. Start with a free 24-hour trial:
- Salesforce Agentforce Specialist Practice Tests — Prompt engineering and generative AI objectives for the Agentforce Specialist (AI-201) exam
- Databricks Generative AI Engineer Associate Practice Tests — LLM architecture and GenAI pipeline objectives for Databricks certification
- Azure AI-102 Practice Tests — Generative AI service objectives for the Azure AI Engineer Associate exam
- AWS ML Specialty Practice Tests — Foundation model and GenAI objectives for the AWS ML Specialty
Level: Intermediate to Advanced | Estimated Reading Time: 35 minutes | Last Updated: February 2025
Part of the Complete AI & Machine Learning Guide
This article is part of The Complete Guide to AI and Machine Learning, a comprehensive pillar guide covering every essential AI/ML discipline from foundations to production deployment. The pillar guide maps how this topic connects to the broader AI/ML ecosystem and provides business context, common misconceptions, and underutilized capabilities for each area.
Continue Your Learning
Explore these related articles in the AI/ML training series to deepen your expertise across the full stack:
- Deep Learning and Neural Networks — For the neural network and transformer fundamentals that underpin LLMs
- RAG Architecture Deep Dive — For a comprehensive guide to grounding LLM responses in your organization’s data
- Advanced Prompt Engineering — For production-grade prompting techniques including CoT, self-consistency, and function calling
- AI Agents and Orchestration — To explore how LLMs power autonomous agent systems with tools and memory
- Implementation specialist
← Return to the Complete AI & Machine Learning Guide for the full topic map and all supporting articles.
Question #1
A data science team at a consumer lending company is building an AI model to approve or deny personal loan applications. The compliance officer insists the model must achieve Demographic Parity, Equalized Odds, AND Predictive Parity simultaneously to satisfy all stakeholders. The lead ML engineer pushes back, citing a fundamental limitation.
Why is the compliance officer’s requirement problematic?
A) These three metrics can only be satisfied simultaneously if the model uses protected attributes as direct input features.
B) Achieving all three metrics requires an interpretable model architecture such as logistic regression, which would sacrifice accuracy.
C) These metrics are designed for classification tasks only and cannot be applied to the continuous probability scores used in lending decisions.
D) It is mathematically proven that — except in trivial cases — Demographic Parity, Equalized Odds, and Predictive Parity cannot all be satisfied simultaneously, so the organization must choose which definition of fairness is most appropriate for their context.
Solution
Correct Answer: D
Explanation: This reflects the Impossibility Theorem described in the Fairness Metrics section. These three fairness definitions are mathematically incompatible in all but trivial cases (e.g., when base rates are identical across groups). Organizations must make a deliberate, documented choice about which fairness metric best fits their use case, regulatory requirements, and stakeholder values. The other options introduce incorrect preconditions — using protected attributes, requiring specific architectures, or limiting metric applicability — none of which are the actual constraint.
Question #2
A consortium of five hospitals wants to collaboratively train a diagnostic AI model for a rare disease. Data privacy regulations such as HIPAA prohibit sharing patient records across institutions, and no single hospital has enough data to train an accurate model independently. The consortium needs a technique that enables collaborative model training while keeping all patient data within each hospital’s infrastructure.
Which privacy-preserving technique is BEST suited to this scenario?
A) Homomorphic encryption, which allows the hospitals to upload encrypted patient records to a shared cloud server where the model is trained on ciphertext without ever decrypting the data.
B) Federated learning, where a global model is sent to each hospital, trained locally on that hospital’s patient data, and only aggregated model updates — not raw data — are shared with a central server.
C) Differential privacy, which adds calibrated noise to each hospital’s patient records before they are combined into a single centralized training dataset.
D) Synthetic data generation, where each hospital creates artificial patient records that mimic statistical patterns and then shares the synthetic datasets for centralized model training.
Solution
Correct Answer: B
Explanation: Federated learning is specifically designed for this scenario — it enables collaborative model training across decentralized data sources without centralizing the raw data. The model travels to the data, not the other way around. Each hospital trains locally, and only model gradients (updates) are aggregated centrally. While homomorphic encryption is a valid privacy technique, it is computationally expensive and does not directly address the distributed training challenge. Differential privacy with centralized data still requires sharing records. Synthetic data loses fidelity for rare diseases where subtle clinical patterns matter most.
Question #3
A corporate legal department has deployed an AI system to review vendor contracts and flag potentially risky clauses. After initial deployment as a fully automated system (human-out-of-the-loop), the tool missed several unusual liability clauses that fell outside its training patterns, exposing the company to significant financial risk. Leadership wants to redesign the system to balance efficiency with risk mitigation.
Which approach BEST addresses this situation while maintaining operational efficiency?
A) Retrain the model on a larger dataset of contracts that includes the unusual liability clauses it missed, then redeploy as a fully automated system with quarterly accuracy audits.
B) Replace the AI system entirely with a team of paralegals who manually review all contracts, since AI has proven unreliable for legal document analysis.
C) Implement a human-on-the-loop model with confidence-based routing, where high-confidence contract reviews are auto-approved with sampling, and low-confidence or high-value contracts are escalated to attorneys for review.
D) Switch to an interpretable rule-based system that uses keyword matching to flag risky clauses, since black-box AI models cannot be trusted for legal decisions.
Solution
Correct Answer: C
Explanation: The human-on-the-loop model with confidence-based routing directly addresses the core problem: fully automated systems miss edge cases, while fully manual review is inefficient. By routing decisions based on the model’s confidence level, the organization captures the efficiency benefits of automation for routine contracts while ensuring human expertise is applied to uncertain or high-value cases. This matches the document’s guidance that the appropriate level of human oversight should be calibrated to the risk, impact, and reversibility of decisions. Simply retraining doesn’t prevent future novel patterns from being missed. Abandoning AI entirely sacrifices the efficiency gains. Rule-based keyword matching is too rigid for complex legal language.
Question #4
A fintech company uses a gradient-boosted ensemble model to evaluate personal loan applications. A financial regulator has issued an inquiry requiring the company to provide individual-level explanations for each applicant who was denied credit — specifically, they must cite the top contributing factors for every adverse decision and show applicants what changes would improve their outcome.
Which combination of explainability techniques BEST satisfies both regulatory requirements?
A) SHAP values to identify the top features contributing to each denial, combined with counterfactual explanations to show applicants the smallest changes that would produce a different outcome.
B) Global feature importance rankings to show which factors the model weighs most heavily across all decisions, combined with partial dependence plots to illustrate how each feature affects predictions on average.
C) A global surrogate model (decision tree) trained to approximate the ensemble’s behavior, which can then be presented to regulators as the actual decision logic.
D) Attention visualization to show which parts of the application the model focuses on, combined with LIME to fit a local linear model around each prediction.
Solution
Correct Answer: A
Explanation: The regulator requires two things: (1) individual-level factor attribution for each denial, and (2) actionable guidance for applicants. SHAP values provide mathematically rigorous, game-theoretic feature contributions for individual predictions — making them the gold standard for per-decision explanations. Counterfactual explanations identify the smallest input changes needed to flip the outcome, directly addressing the ‘what would need to change’ requirement. Global feature importance and PDP are aggregate techniques that do not explain individual decisions. A surrogate model is an approximation and misrepresents the actual decision process. Attention visualization applies to neural networks and transformers, not gradient-boosted ensembles.
Question #5
A global consumer brand is deploying a generative AI system to create personalized marketing emails at scale across diverse international markets. During pilot testing, the system occasionally produces culturally insensitive content when targeting specific demographic segments, including stereotypical references and tone-deaf messaging that could damage the brand’s reputation.
Which set of safeguards is MOST comprehensive for responsible deployment of this generative AI system?
A) Translate all marketing content into English first, run it through a single toxicity filter, and then translate it back into the target language before sending.
B) Restrict the generative AI to producing content only in English for all markets, and hire local translators to manually adapt every email for cultural relevance.
C) Add a disclaimer to each email stating that the content was generated by AI, which satisfies transparency requirements and shifts responsibility away from the brand.
D) Implement a multi-layer pipeline: prompt engineering with cultural sensitivity guidelines, automated toxicity and bias detection on outputs, human review sampling with higher rates for diverse segments, and a recipient feedback mechanism to flag inappropriate content.
Solution
Correct Answer: D
Explanation: The multi-layer pipeline approach addresses the problem at every stage — from input (prompt engineering with cultural guidelines), through processing (automated toxicity and bias detection), to output (human review sampling and recipient feedback). This aligns with the document’s guidance on responsible generative AI deployment, which emphasizes content filtering, human review for high-stakes content, transparent disclosure, and red-team testing. Translating to English and back introduces translation artifacts and misses cultural nuance. Restricting to English ignores the reality of global marketing. A disclaimer alone does not prevent the harm — it merely attempts to deflect accountability, which contradicts the core principle of accountability in responsible AI.
Choose Your AI Certification Path
Whether you’re exploring AI on Google Cloud, Azure, Salesforce, AWS, or Databricks, PowerKram gives you vendor‑aligned practice exams built from real exam objectives — not dumps.
Start with a free 24‑hour trial for the vendor that matches your goals.
- All
- AWS
- Microsoft
- DataBricks
- Salesforce




