Notes for my Cloud certifications.
These are my personal notes for the Google Cloud Generative AI Leader Certification, taken during following the Cloud Skills Boost Generative AI Leader path.
Overview:
Helpful resources:
Data is the foundation of any AI system. Data quality and accessibility are essential for effective AI development.
Data can be structured or unstructured, each requiring different analysis techniques.
Key dimensions of data quality:
Accuracy
Completeness
Consistency
Relevance
Availability
Cost
Format
Understanding the types and quality of your data is crucial for successful AI initiatives.
Machine learning models can be trained using:
Supervised learning
Unsupervised learning
Reinforcement learning
The choice of approach depends on the specific task and the nature of the data available.
The ML lifecycle encompasses several key stages:
Data ingestion and preparation
Model training
Model deployment
Model management
Google Cloud provides a comprehensive suite of tools to support each stage of this lifecycle.
Vertex AI helps with model training and deployment, while various data tools support ingestion, preparation, and management.
By understanding and effectively managing this lifecycle, organizations can maximize the value of their initiatives and ensure long-term success.
The process of creating your ML model using data is called model training.
Vertex AI provides:
A managed environment for training ML models
Prebuilt containers for popular frameworks
Custom training jobs
Tools for model evaluation
Powerful computing resources to speed up training
Model deployment is the process of making a trained model available for use.
Vertex AI simplifies this with:
Tools to deploy models for generating predictions
Options to scale deployments by adjusting resources based on demand
Managing and maintaining your models over time is critical.
Google Cloud offers:
Versioning: Track different model versions
Performance Tracking: Monitor model metrics
Drift Monitoring: Watch for accuracy changes over time
Data Management: Use Vertex AI Feature Store to manage data features
Storage: Vertex AI Model Garden to organize models
Automation: Vertex AI Pipelines to automate ML tasks
Deep learning provides the core technology.
Foundation models are powerful architectures built on deep learning.
Generative AI is the application of these models to create new, original content.
Vertex AI streamlines integration of advanced AI capabilities into business applications:
Seamless discovery, deployment, and customization
Access to many models without extensive in-house development
These models empower businesses to enhance customer experiences, increase productivity, foster innovation, and improve decision-making.
Gemini: Multimodal; processes text, images, audio, and video.
Gemma: Lightweight, open models for local deployments and specialized AI applications.
Imagen: Text-to-image generation.
Veo: Video generation.
Gemini is designed to handle multiple data types, while Gemma is optimized for lighter, specialized deployments.
Modality
Context window
Security
Availability
Cost
Performance
Fine-tuning
Ease of integration
Google Cloud offers a suite of foundation models with unique strengths and capabilities.
Data Dependency
Performance depends on large, high-quality datasets. Biases or incompleteness in the data will seep into outputs.
Example: It’s like asking a student to write an essay on a book they haven’t read.
Knowledge Cutoff
AI models are only aware of information up to their training date.
Example: A model trained in 2022 won’t know about events after 2022.
Bias
LLMs can amplify biases present in their training data.
Even subtle biases can be magnified in outputs.
Fairness
Defining fairness is complex.
Fairness assessments can miss some forms of bias.
Hallucinations
Models may produce plausible-sounding but incorrect or nonsensical answers.
This is a major concern in accuracy-critical applications.
Edge Cases
Rare or unusual scenarios can reveal model weaknesses and lead to errors.
Connect the AI’s output to verifiable sources—like giving AI a reality check.
Benefits:
Reduces hallucinations
Anchors responses in real data
Builds trust with citations and confidence scores
Retrieval: Search engine finds relevant information using semantic understanding.
Augmentation: Retrieved data is added to the prompt.
Generation: The model uses this context to produce informed, accurate responses.
RAG grounds outputs in real, verifiable sources, improving accuracy and relevance.
The most rapid, straightforward approach to guide models.
Involves crafting precise prompts
Limited by the model’s existing knowledge
When prompting isn’t enough, fine-tuning adapts a model to specific needs.
Further trains a pre-trained model on task-specific data
Adjusts parameters for specialized performance
Use Cases:
Generating content in a specific style
Code generation in specific languages
Domain-specific translation
Vertex AI provides tooling to facilitate tuning.
Even the best models benefit from human oversight.
Key use cases:
Content Moderation: Ensures accurate, appropriate filtering of user-generated content.
Sensitive Applications: Provides oversight in healthcare, finance, etc.
High-Risk Decisions: Adds accountability for decisions with serious consequences.
Pre-Generation Review: Validates outputs before deployment.
Post-Generation Review: Continuous human feedback to improve models over time.
Preventing intentional harm to AI applications.
Protect AI systems from malicious attacks and misuse.
Ensure security throughout the entire lifecycle, from development through deployment.
Key risks:
Data poisoning
Model theft
Prompt injection
Google Cloud’s SAIF framework provides tools to help build and maintain secure AI systems.
Ensuring AI avoids both intentional and unintentional harm.
Users need to know how their information is used and how AI systems work.
Protecting privacy often involves anonymization or pseudonymization.
High-quality data is essential for ethical AI.
Poor data quality can lead to biased, unfair outcomes.
AI systems can amplify societal biases.
Example: A resume-screening tool favoring certain demographics due to biased training data.
Fairness requires accountability.
Know who is responsible for AI outputs.
Make AI decision-making transparent and understandable.
Vertex Explainable AI helps:
Debug errors
Uncover hidden biases
Build user trust
AI development is governed by evolving legal frameworks.
Key considerations:
Data privacy
Non-discrimination
Intellectual property
Product liability
Legal compliance is essential for building trustworthy AI systems.
Gen AI agents process information, reason over complex concepts, and take action.
Applications include:
Customer service
Employee productivity
Creative tasks
An application that observes the world and acts on it using its tools to achieve goals.
Capabilities:
Understanding and responding to natural language
Automating complex tasks
Personalization
Conversational Agents
Input: User types or speaks
Understand: AI interprets meaning and intent
Call Tool: Searches web, accesses databases, triggers actions
Generate Response: Produces a relevant answer
Deliver: Provides the output
Workflow Agents
Input: User triggers a task (form submission, upload, event)
Understand: Defines steps needed
Call Tool: Executes integrations, transformations
Generate Result: Compiles output
Deliver: Sends via email, dashboard, database
Rule-based calculations
Thought chains
Machine learning algorithms
Probabilistic reasoning
Examples include ReAct and Chain-of-Thought (CoT).
Manage the ML lifecycle with built-in tools.
Feature Store: Share and serve ML features consistently.
Model Registry: Track changes, manage versions.
Model Evaluation: Compare model performance.
Workflow Orchestration: Automate processes with Vertex AI Pipelines.
Model Monitoring: Detect performance degradation and drift.
Two main options:
Fully Custom: Train at scale with any framework (PyTorch, TensorFlow, scikit-learn, XGBoost).
AutoML: Minimal effort, guided training.
Google’s most efficient, compact AI model for edge deployment.
Designed for smartphones, embedded systems.
Runs locally for real-time responsiveness and data control.
Tools: Lite Runtime (LiteRT), Gemini Nano
Access Gemini’s generative AI features within Gmail, Docs, Sheets, Meet, and Slides.
Features vary by Workspace plan.
Zero-shot: No prior examples.
One-shot: Single example.
Few-shot: Multiple examples to improve understanding.
Guide the model by assigning a persona.
Examples:
Business analyst
Shakespearean actor
Customer service agent
Create complex interactions where each prompt builds on the last.
Ensures outputs are based on verifiable, specific sources.
Accesses external knowledge sources.
Produces more accurate, relevant, transparent outputs.
Cites sources used for generation.
An AI-first notebook grounded in your own documents.
Capabilities:
Summarize findings
Identify connections and contradictions
Generate outlines and drafts
Answer questions about content
Plus: Adds capacity, customization, usage analytics.
Enterprise: Extra privacy, compliance, IAM controls.
Token Count: Controls conversation length.
Temperature: Controls randomness and creativity.
Top-p: Limits probability spread to most likely tokens.
Safety Settings: Filters harmful or inappropriate content.
Output Length: Defines maximum generated text length.
| Feature | Google AI Studio | Vertex AI Studio |
|---|---|---|
| Audience | Experimenters, early-stage users | Developers building production systems |
| Features | Easy Gemini API access | Advanced tools for the ML lifecycle |
Combines reasoning and action.
Steps:
Think: Generate thoughts about the problem.
Act: Take actions (e.g., search the web).
Observe: Receive feedback.
Respond: Formulate an answer.
Benefits:
Dynamic problem-solving
Reduced hallucination
Increased trustworthiness
Guides the model through step-by-step reasoning.
Benefits:
Improved problem-solving
Better accuracy
Enhanced explainability
Techniques:
Self-consistency
Active prompting
Multimodal CoT
ReAct Cycle:
Reasoning (Tool Selection)
Acting (Tool Execution)
Observation
Iteration
Retrieval:
Data stores
Vector databases
Search engines
Knowledge graphs
Augmentation:
Generation:
Define step-by-step behaviors using linked external tools and data stores.
Enables dynamic, adaptable prompt creation and interpretation.
Centralized platform to manage AI agents using company data.
Integrates with internal websites and dashboards.
Acts as personal research assistants for employees.
Agentspace vs. NotebookLM
| Feature | NotebookLM | Agentspace |
|---|---|---|
| Purpose | Deep dive into specific documents | Enterprise AI assistant across systems |
| Scope | Only user-provided sources | All connected business systems |
| Integration | Can connect with NotebookLM Enterprise | Unified search and automation |
Additional Helpful Resources: