02 - Google Cloud's generative AI offerings¶
The main memorization load. Learn each product by the problem it solves.
The adoption ladder¶
| Rung | What you use | Effort | Fits |
|---|---|---|---|
| Applied assistants | Gemini for Google Workspace, Gemini for Google Cloud | Lowest, no build | Productivity gains across an organization |
| Pre-built products | Vertex AI Search, Conversational Agents, NotebookLM | Low, configure not code | A common capability delivered quickly |
| Build on the platform | Vertex AI with Gemini, grounding, function calling | Medium | A differentiated product experience |
| Customize a model | Fine-tuning on Vertex AI | High | Consistent specialized behavior at scale |
| Train from scratch | Custom foundation model | Very high | Effectively never |
Start at the top. The exam consistently rewards the lowest rung that solves the stated problem.
Vertex AI¶
The unified platform for AI on Google Cloud.
- Model Garden - the catalog of available models: Google's own, open models, and third-party
- Vertex AI Studio - a console for prompt design, testing, and comparison without writing code
- Training and tuning - fine-tuning and custom training
- Prediction and endpoints - serving models
- Evaluation - measuring model and application quality
- Pipelines - orchestrating ML workflows
- Feature Store, Model Registry, Model Monitoring - the MLOps surfaces
- Vector Search - similarity search over embeddings at scale
- Agent Builder - building agents and search experiences
Models¶
| Model | Type | Used for |
|---|---|---|
| Gemini | Multimodal foundation model family | General generation, reasoning, and multimodal understanding across text, image, audio, video, and code |
| Gemma | Open models | Running yourself, on your own infrastructure or at the edge, where openness or control matters |
| Imagen | Image generation and editing | Creating and editing images from text |
| Veo | Video generation | Creating video from text or images |
| Chirp | Speech | Speech recognition and transcription |
| Embeddings models | Vector representations | Semantic search, RAG, clustering, recommendation |
Model families come in size tiers, trading capability against cost and latency. Matching the model to the task, rather than defaulting to the largest, is a recurring exam theme.
Pre-built products¶
Vertex AI Search - grounded enterprise search over your own content, with generative summaries and citations. The answer when the requirement is "let employees or customers ask questions about our documents".
Vertex AI Agent Builder / Conversational Agents - building conversational agents that can answer questions, take actions through tools, and hand off to humans. The answer for a customer-facing assistant or an internal helpdesk bot.
NotebookLM - grounded research over sources you supply, producing summaries, briefings, and question answering restricted to those sources. The answer for individual or small-team research over a document set.
BigQuery generative AI functions - calling models directly from SQL, so text generation, summarization, and embedding happen next to the data without building a pipeline. The answer when the data is already in BigQuery and the task is analytical.
Applied assistants¶
Gemini for Google Workspace brings generative AI into Docs, Gmail, Sheets, Slides, and Meet: drafting, summarizing, and meeting notes. It is the fastest path to organization-wide productivity gains, with no build at all.
Gemini for Google Cloud assists engineering work: code assistance, cloud operations, and troubleshooting inside the console and IDEs.
These matter on the exam because many scenarios describing a productivity problem are answered by an applied assistant rather than by a project.
Infrastructure¶
- AI Hypercomputer - Google's integrated architecture of hardware, software, and consumption models for large-scale AI
- TPUs - Google's purpose-built AI accelerators, designed for large-scale training and inference
- GPUs - the alternative accelerator, familiar and broadly compatible
- Cloud Storage and BigQuery - where training and grounding data usually lives
Most business-level scenarios do not reach infrastructure. What matters is recognizing that Google offers it and that infrastructure choice is a cost and scale decision, not a capability decision, for most adopters.
Enterprise controls¶
Frequently the deciding factor in a business scenario:
- Data governance: customer data submitted to Vertex AI is not used to train Google's foundation models
- Data residency and regional endpoints for regulatory requirements
- VPC Service Controls and Private Service Connect for private access
- CMEK for customer-managed encryption keys
- IAM for access control, and audit logging for accountability
- Safety filters configurable per application
When a scenario names a regulated industry or a data residency requirement, these are the differentiators to reach for.
Key terms¶
- Vertex AI - Google Cloud's unified platform for building, tuning, deploying, and evaluating AI
- Model Garden - the Vertex AI catalog of Google, open, and third-party models
- Vertex AI Studio - the console surface for designing and testing prompts without code
- Gemini - Google's multimodal foundation model family
- Gemma - Google's family of open models intended to be run by the user
- Imagen - Google's image generation and editing model
- Veo - Google's video generation model
- Chirp - Google's speech recognition model
- Vertex AI Search - grounded enterprise search with generative summaries and citations
- Vertex AI Agent Builder - the product for building conversational agents and search experiences
- NotebookLM - grounded research and question answering over sources the user supplies
- BigQuery generative AI functions - calling models directly from SQL inside the data warehouse
- Gemini for Google Workspace - generative AI applied inside Docs, Gmail, Sheets, Slides, and Meet
- Gemini for Google Cloud - generative AI assistance for engineering and cloud operations
- AI Hypercomputer - Google's integrated architecture for large-scale AI workloads
- TPU - Google's purpose-built accelerator for AI training and inference
- VPC Service Controls - the perimeter control restricting data movement out of a defined boundary