GCP Generative AI Leader Concept Map (Mind Map): Section 1

Every topic from Section 1 of the exam on one page, with the pairs the exam swaps to trick you. It's about 30% of your score.

1.1 Core gen AI concepts and use cases

The vocabulary the exam uses, and what gen AI is good for

From AI down to gen AI

Artificial intelligence (AI): Building machines that do tasks which normally need human intelligence: learning, solving problems, deciding.
Machine learning (ML): A part of AI where a machine learns patterns from data instead of being programmed with rules.
Deep learning: A part of machine learning that uses neural networks with many layers to find complex patterns.
Generative AI (gen AI): The use of machine learning that creates new content: text, images, code, audio, video.
Foundation model: A very large model trained on a huge amount of mostly unlabeled data, so it understands the world broadly and can be adapted to many tasks.
Large language model (LLM): A foundation model built to understand and generate human language.
Multimodal foundation model: A foundation model that takes in and produces more than one kind of data: text, images, audio, video, code.
Diffusion model: A model that creates an image by starting from noise and cleaning it step by step until it matches the prompt.
Natural language processing (NLP): The field of making computers work with human language: translation, sentiment, entity extraction, summaries.

How a model learns

Supervised learning: Train on labeled examples so the model predicts the right output for new inputs.
Unsupervised learning: Train on unlabeled data to find natural groups and patterns, such as customer segments.
Reinforcement learning: The model learns by acting, then getting rewards or penalties, and repeats what earns rewards.
Human in the loop (HITL): People review, correct or approve what the model does, and that feedback goes back into the workflow.

Prompting words that look alike

Prompt: The instruction or input you give a foundation model to get the output you want.
Prompt engineering: Writing and refining prompts so the model gives better, more useful answers. No change to the model itself.
Prompt tuning: A light form of tuning: the model's weights stay frozen and a small learned piece is added in front of the prompt to steer it for one task.
Fine-tuning: Train a foundation model further on your own examples so it gets better at your task or your domain language.

The ML lifecycle, stage by stage

Data ingestion: Collect raw data and land it somewhere usable: Cloud Storage for files, BigQuery for tables, Pub/Sub for streams.
Data preparation: Clean, join and transform the raw data into the format training needs, with BigQuery or Dataflow.
Model training: Build the model from the prepared data, or tune a foundation model, on Agent Platform (custom training, AutoML) or in BigQuery with SQL.
Model deployment: Make the trained model available for use, usually behind an endpoint an application can call.
Model management: Keep the model healthy over time: versions, metrics, drift, upgrades and retirement.

What gen AI does for a business

Create: Produce new content that did not exist: marketing copy, product descriptions, images, code.
Summarize: Condense long material into the short version a reader needs: call notes, contracts, reports.
Discover: Find the right information at the right moment inside large amounts of company content.
Automate: Let gen AI take over steps people did by hand: routing tickets, filling forms, drafting replies.
Text generation: Writing and rewriting language: emails, articles, answers, translations, tone changes.
Image generation: Create or edit pictures from a text description, for campaigns, mockups or product shots.
Code generation: Write, explain, test or convert code, which speeds up developers and modernizes old systems.
Video generation: Create video clips from a text description or a still image, for ads and training material.
Data analysis: Ask questions of your data in plain language and get charts, explanations or summaries back.
Personalized user experience: Adapt content, recommendations or support to each customer, at a scale people could not reach.

1.2 Data types and the business implications

What your data must look like before gen AI can help

The shape of the data

Structured data: Data organized in rows and columns with a fixed meaning per field, usually in a relational database or a warehouse: orders, payments, sensor readings.
Unstructured data: Data with no predefined structure: documents, emails, PDFs, images, audio, video. Most company data is this.
Semi-structured data: Data with some tags or keys but no strict table: JSON, XML, log lines. Not a guide term, but it explains why some sources sit between the two.
Data comes in many forms: Numbers, dates, text, images and sounds are all data, and gen AI can work with each when the model's modality matches.

Tagged or not

Labeled data: Data that carries a tag: a name, a class, a number. “This review is negative”, “this image is a defect”.
Unlabeled data: Raw data with no tags: a folder of photos, a stream of recordings. It has no meaning attached by itself.

Is the data good enough

Data quality: Data that is accurate, complete, consistent and relevant. Poor quality data limits any model, however good the model is.
Completeness: The data has the fields and the rows you need, with few gaps.
Consistency: The same thing is recorded the same way everywhere: one date format, one customer id, no contradictions between systems.
Relevance: The data actually relates to the question being asked. Extra unrelated data adds cost and noise.
Accuracy: The values are true: real prices, correct labels, no typos carried into training.

Can you actually use it

Data accessibility: The data is available to the team and the service that needs it, in a usable state, when it is needed.
Format: The data arrives in a form the tool accepts: supported file types, encodings, resolutions, sampling rates.
Cost of data: Collecting, storing, moving and labeling data all cost money, and that cost belongs in the business case.

1.3 The core layers of the gen AI landscape

Five layers, from the chips up to the app the user sees

The five layers, bottom to top

Infrastructure: The compute and storage that run AI: servers, GPUs, TPUs, data centers, and the software that drives them.
Models: The trained algorithms themselves, which learn patterns and generate content: Gemini, Gemma, and others.
Platforms: The layer with APIs, data management and deployment tools that connects models to agents and hides the infrastructure.
Agents: Software that pursues a goal: it observes, reasons, and uses tools to act, instead of only answering once.
Applications: The user-facing layer: the chat window, the portal, the feature inside an existing product.

Which layer is this?

Example: TPUs and GPUs: Custom accelerators and graphics processors sit on the infrastructure layer; you pick them to train or serve faster.
Example: Gemini and Gemma: A named model belongs to the model layer, whether you call it through an API or run it yourself.
Example: Agent Platform: Gemini Enterprise Agent Platform (formerly Vertex AI) is Google Cloud's platform layer: models, tuning, deployment, agents and governance in one place.
Example: a support agent: An assistant that reads a ticket, looks up the order, and issues a refund is on the agent layer: it uses tools and acts.
Example: the Gemini app: A chat product people open and type into sits on the application layer, even though a model powers it.
AI agent: An application that tries to reach a goal by observing its environment and acting on it with the tools it has.

1.4 Use cases and strengths of Google's foundation models

Which Google model for which job, and how to choose one

The model families

Gemini: Google's flagship multimodal family: understands and generates text, code, images, audio and video, with long context and strong reasoning.
Gemini Flash and Pro tiers: Within Gemini, Flash tiers are faster and cheaper for high volume, Pro tiers are stronger on hard reasoning.
Gemini Nano: Google's most compact Gemini model, built to run on a device rather than in the cloud.
Gemma: Google's family of lightweight open models, built with the same research as Gemini. You can download the weights, run them locally and customize them.
Gemini image models (formerly Imagen): Google's text-to-image generation and editing models. Imagen was the earlier name and was deprecated in 2026, with image generation moving to Gemini models. The exam guide still lists this as Imagen.
Veo: Google's video generation model: it creates video from a text description or from a still image.
Embedding models: Models that turn text or images into vectors of numbers so similar meanings end up close together. They power search and RAG rather than writing answers.

Open or managed

Open models: Models whose weights are published, so you can download, host, inspect and tune them. Gemma is Google's.
Managed model API: You call the model over an API and pay per use; Google runs, patches and upgrades it.

Choosing a model for the use case

Modality: Which kinds of data the model takes in and gives back: text, image, audio, video, code. Match it to the use case first.
Context window: How much text and media the model can consider in one request, prompt and answer together.
Cost: What the model costs to run at your volume: price per token or per image, plus the hardware if you host it.
Performance: Accuracy, speed and efficiency on your own task, measured on your own examples.
Availability and reliability: Whether the model is consistently reachable and steady under load: uptime, regions, redundancy, recovery.
Security and data control: Where the data goes, who can see it, and what the provider may do with it. Enterprise rules often narrow the choice before anything else.
Fine-tuning and customization: Whether the model can be adapted to your domain, and how much work that takes: tuning, grounding, tools.

Part of the GCP Generative AI Leader Study Kit: Concept Maps, Flashcards and practice questions. Also free: the Section 1 flashcards. New to the exam? Read the GCP Generative AI Leader study guide.