Current product names first; names the exam guide or quizzes use are in italics.

§1 · 30%Fundamentals of gen AI

Nesting

AI (machines doing tasks that need human intelligence) ⊃ ML (learn from data) ⊃ Deep learning (multi-layer neural networks) → Foundation models (huge, diverse data; adaptable) → LLMs (language) · Diffusion models (refine noise into images, audio, video).

Gen AI = creates new content. NLP = understand and generate human language. Multimodal = several data types combined.

Learning types

Supervised	labeled data → predict (spam filter, house prices)
Unsupervised	unlabeled → find groupings (segments, topics)
Reinforcement	trial and error, rewards/penalties
Semi-supervised	small labeled + large unlabeled

Data

Quality (5): accuracy · completeness · representativeness · consistency · relevance.

Accessibility (3): availability · cost · format.

Structured = rows/columns (orders, CSV). Unstructured = text, images, audio, video.

ML lifecycle → tools

Gather (ingest)	Pub/Sub (streaming) · Cloud Storage (unstructured) · Cloud SQL, Spanner (structured)
Prepare	BigQuery · Knowledge Catalog (course: Data Catalog)
Train	Agent Platform (custom training, AutoML)
Deploy & predict	Agent Platform (endpoints, scaling)
Manage	Model Registry (versions) · Model Monitoring (drift) · Feature Store · Pipelines

Choosing a model (8)

Modality · context window · security · availability & reliability · cost · performance · fine-tuning & customisation · ease of integration. Bigger isn't always better.

Google models

Gemini	multimodal reasoning
Gemma	lightweight, open; local/custom
Imagen	text-to-image diffusion (exam's image model)
Veo	video from text or images
Nano Banana	Gemini image generation/editing (course)
Chirp	speech-to-text
Gemini Nano	on-device (edge)

5 layers

Infrastructure → Models → Platform → Agents → Gen AI powered applications.

4 uses

Create · Summarize · Discover · Automate.

§2 · 35%Google Cloud's gen AI offerings

Why Google Cloud

AI-first · enterprise-ready (responsible, secure, private, reliable, scalable) · integrated ecosystem · open approach (TensorFlow, PyTorch, open models, open standards).

Infrastructure: hypercomputer · TPUs (Google's AI chips) · GPUs · data centres · cloud computing. Democratises AI: low-code/no-code · pre-trained models · APIs.

Prebuilt: work productivity

Gemini app / Gemini Advanced	chat assistant · Gems (task-specific assistants) · Saved info (applies to all chats)
Google Workspace with Gemini (Gemini for Google Workspace)	Gmail, Docs, Sheets (formulas), Slides (images), Meet (transcribe/translate), Drive side panel, Vids, AppSheet
Gemini Notebook (NotebookLM)	answers only from your sources, citations, Audio Overviews · Plus · Enterprise (IAM, compliance)
Gemini Enterprise app (Agentspace)	agents + unified search across all company systems · Gemini Notebook API · multimodal search · custom agents

Gemini for Google Cloud

Cloud Assist (design/manage apps) · Code Assist (AI pair programmer) · in BigQuery (no SQL needed) · in Looker · in Colab Enterprise · in Databases · in Security. Prompts and responses are not used to train models.

Customer experience

Agent Search (Vertex AI Search) + Google Search: external search.

Gemini Enterprise for CX (Customer Engagement Suite)
CX agents / CX Agent Studio	Conversational Agents: chatbots; deterministic, generative or hybrid
Agent Assist	helps live human agents
CX Insights	Conversational Insights: sentiment, topics, scoring
CCaaS	Contact Center as a Service: omnichannel, routing

Build: Agent Platform (Vertex AI)

Model Garden = discover/deploy models · AutoML = train with little ML skill · Model Registry = versions · Model Evaluation · Pipelines = automate · Feature Store · Model Monitoring = drift/skew.

RAG offerings: pre-built RAG with Agent Search (fast, managed) · RAG APIs (custom pipeline).

Agent tooling

4 tool types: extensions (APIs) · functions (actions) · data stores (information) · plugins (new skills).

Services: Cloud Storage · databases · Cloud Run functions (Cloud Functions) · Cloud Run · Agent Platform · APIs: Speech-to-Text, Text-to-Speech, Translation, Document Translation, Document AI, Vision, Video Intelligence, Natural Language · API Library.

Google AI Studio = quick prototyping, Google account, usage limits. Agent Studio = professional, Google Cloud, enterprise security.

§3 · 20%Techniques to improve output

Limitations → fixes

Limitations: data dependency · knowledge cutoff · bias · fairness · hallucinations · edge cases.

Fixes: grounding · RAG · prompt engineering · fine-tuning · humans in the loop.

Which fix?

Quick improvement	prompt engineering
Fresh or private facts, citations	RAG / grounding
Consistent style, format, behaviour	fine-tuning (prompt tuning = lightweight variant)
High stakes	human in the loop

Prompting

Zero-shot (no examples) · one-shot · few-shot · role prompting (persona) · prompt chaining (steps build on each other) · chain-of-thought (step-by-step reasoning) · ReAct (reason + act with tools: think → act → observe → respond) · metaprompting (AI writes prompts) · system instructions.

Grounding

Sources: first-party enterprise data · third-party data · world data (Google Search).

RAG = retrieve → augment prompt → generate. Offerings: pre-built RAG with Agent Search · RAG APIs · Grounding with Google Search.

Sampling parameters

Temperature	low = focused, high = creative
Top-p	smallest set of likely tokens reaching p
Top-k	pick from k most likely tokens
Token count	~4 characters per token
Safety settings · output length

Monitoring

Automatic model upgrades · KPIs · security patches · versioning · performance tracking · drift monitoring · Feature Store.

§4 · 15%Business strategy, secure and responsible AI

Strategy

Top-down (leaders: vision, funding, ethics) + bottom-up (managers and ICs: use cases, experiments). Start low-risk; prioritise feasible, actionable, affordable, high-ROI.

6 steps: clear vision · high-impact use cases · invest in capabilities · drive change · measure value · champion responsible AI. Impact: define metrics → collect data → iterate.

People and needs

Augment strategic work, automate repetitive work; humans stay in the loop.

Needs: scale · customisation · user interaction · privacy · latency · connectivity · accuracy · explainability. Factors: business requirements + technical constraints.

Cost: training · deploying · using. Pricing by usage (tokens, characters, requests, compute time), subscription, licence, free tier.

Secure AI (intentional harm)

Data poisoning · model theft · adversarial attacks / prompt injection. Protect every lifecycle stage.

SAIF (Secure AI Framework) · secure-by-design infrastructure · IAM · Security Command Center · workload monitoring · Mandiant.

Responsible AI (intentional + unintentional)

Built on security. Transparency · privacy (anonymization = can't link back; pseudonymization = re-linkable with a protected key) · data quality, bias, fairness · accountability and explainability · legal (privacy, non-discrimination, IP, liability).