Prepare and test your skills
Prepare and test your skills
Core generative AI concepts explain how AI systems learn from data and how businesses apply them to create new content. Generative AI (opens in a new tab) is a subset of machine learning, which is a subset of artificial intelligence. Traditional AI mainly classifies or predicts from historical data, while generative AI learns patterns from training data to generate new text, images, audio, video, and code.
Artificial intelligence, machine learning, and generative AI form a nested hierarchy where each inner layer is more specialized. Artificial intelligence covers systems that learn from data to classify, predict, or generate. Machine learning provides models that learn patterns from data instead of following only hand-written rules, and generative AI sits inside machine learning as the part that creates new content.
Traditional machine learning focuses on prediction tasks such as classification that assigns a label like spam or not spam, regression that predicts a number like a house price, time series forecasting that forecasts sales and demand, and clustering that groups records like customer segments. Natural language processing (NLP) is the area of AI that works with human language, and large language models perform NLP tasks such as text generation, translation, summarization, and question answering. Generative AI expands beyond labeling to create summaries, uncover hidden correlations, and produce content that resembles its training data.
Exam tip: Traditional AI classifies or predicts from historical data, while generative AI creates new content that reflects patterns in its training data.
Large language models, multimodal models, and diffusion models are three foundation model capabilities that differ by input types and output types. A large language model (LLM) is a text-driven foundation model trained on vast amounts of data that uses embeddings to understand context and generate coherent text. A multimodal foundation model understands and generates content from multiple modalities including text, images, video, and audio, while a diffusion model learns visual patterns to create realistic images from text prompts.
| Capability | Input data types | What it generates |
|---|---|---|
| Large language model | Text | Generated text, translations, summaries, and answers |
| Multimodal foundation model | Text, images, video, and audio together | Responses that combine modalities, such as text with related images |
| Diffusion-based image model | Text descriptions, plus images for editing | New images, edited images, and short videos or animations |
Embeddings are arrays of floating point numbers called vectors that convert text, images, and video into a form models can process. Embeddings map data into a continuous vector space where closer proximity means similar meaning. Data flows from raw text and images into vectors, large language models then operate on language vectors to discern language patterns, and diffusion models operate on visual vectors to discern visual patterns.
Prompt engineering and model tuning are two ways to improve model output that differ by what changes and how much effort is needed. Prompt engineering changes only the request sent to the model, while model tuning changes the model itself by training it on a dataset. The practical order is fixed: first try prompt engineering, then move to tuning only if prompt design alone cannot handle the specialized task.
Prompt engineering is the iterative process of writing a prompt, reading the response, and adjusting the prompt until the output matches the goal. Common techniques are zero-shot prompting with no examples, one-shot prompting with one example, and few-shot prompting with multiple examples. Model tuning customizes default behavior through methods such as supervised tuning, reinforcement learning from human feedback (RLHF) tuning, or model distillation. A tuned model can then follow simpler prompts, which can reduce request cost and latency, but the tuned version must be deployed to an endpoint before use.
Machine learning approaches describe how models learn from labeled answers, unlabeled structure, and reward feedback. Supervised learning trains on labeled data where each example pairs an input with its correct answer, and the model adjusts to reduce the difference between its prediction and the label. Unsupervised learning finds structure in unlabeled data, where clustering groups similar examples such as customer segments and association finds rules such as shoppers who buy bread often buy butter. Reinforcement learning uses an agent that chooses actions in an environment and receives reward feedback from the environment, aiming to maximize total cumulative reward over many steps.
The machine learning lifecycle on Google Cloud moves in order from scattered raw data to clean tables to trained models to monitored endpoints. Data flows from source systems into Cloud Storage for staging raw files, through a Dataflow pipeline that cleans and reshapes records, into BigQuery tables for analysis and training, and then through Vertex AI training to an endpoint for predictions. The preparation flow runs in three ordered stages: first source files land in Cloud Storage buckets, then data flows through Dataflow for cleaning and enrichment, then curated output flows from Dataflow to BigQuery. Cloud Run (opens in a new tab) and Google Kubernetes Engine (GKE) can serve models when teams need more control, while Vertex AI Model Registry, Feature Store, logging, and monitoring keep models reusable and reliable.
A left-to-right process flow showing source data staged in Cloud Storage, cleaned by Dataflow, stored in BigQuery, trained and deployed in Vertex AI, served from an endpoint, with model management below.
Choosing a foundation model for a business use case (opens in a new tab) means filtering by technical fit, then cost and reliability, then security and customization. Modality and context window are pass-or-fail filters, where modality is the content type a model accepts and produces and context window is the maximum information in tokens a model can consider in one request. Only models whose modality matches and whose window can hold the largest expected input remain, and then teams compare latency, throughput, quality, availability, cost, and compliance.
Generative AI use cases group into synthesis, creation, and automation supported by embeddings and virtual agents. Information synthesis turns unstructured content into summaries and structured facts by generating embeddings, storing them, and using controlled generation with a schema to extract fields. Multimodal creation produces text, images, video, and code from prompts and tailors results to an audience, while virtual agents synthesize support cases and documentation into reasoned responses. Grounding with retrieval-augmented generation (RAG) (opens in a new tab) reduces hallucinations by converting the query into an embedding, comparing it to document embeddings, retrieving the most relevant facts, and adding those facts to the prompt.
Continue with this study guide: 1.1 Describe core generative AI (gen AI) concepts and use cases
How data types are used in generative AI explains why data quality, organization, and labeling decide accuracy, cost, and trust. High-quality, varied data improves model performance, and proprietary datasets a business uniquely owns can support features competitors cannot easily copy. If inputs are inaccurate or incomplete, the model can perform poorly or produce false outputs.
Data quality and accessibility decide whether generative AI is accurate, affordable, and trustworthy. Data quality means data is complete with all required records, consistent in format and location, and relevant to what the model is judged on. Data accessibility means AI services can reach that data in a usable format at an acceptable cost, and BigQuery (opens in a new tab) improves accessibility by bringing machine learning to the data instead of requiring data movement.
BigQuery ML lets teams train, evaluate, and run inference with SQL where data already lives, while BigQuery AI adds inference against large language models (LLMs) such as Gemini, embeddings and vector search, and agent tools. The repeatable evaluation workflow runs in a fixed order:
Cost and control shape which BigQuery AI function category is practical. General-purpose functions such as AI.GENERATE give full control over model, prompt, and parameters, managed functions such as AI.CLASSIFY, AI.SCORE, and AI.AGG let BigQuery choose settings to optimize cost and quality, and task-specific functions use Cloud AI APIs for translation, vision, and transcription. For large datasets, optimized mode trains a lightweight distilled model on a sample to avoid calling a large model for every row, and AI.COUNT_TOKENS estimates cost before running.
Structured and unstructured data serve different AI tasks because their organization is different. Structured data is organized in fixed rows, columns, and values such as sales figures, purchase history, and inventory lists, while unstructured data has no fixed schema and carries meaning through language, pixels, sound, or motion such as documents, images, video, and audio. Generative AI primarily learns from and creates unstructured data, while structured data grounds answers in facts and drives analysis with traditional AI.
| Data type | Organization | Typical AI use |
|---|---|---|
| Structured data | Tables with defined fields, such as tabular purchase and sales records | Classification, regression, forecasting, and segmentation with traditional AI |
| Unstructured data | Free-form text, images, video, and audio across multiple modalities | Training foundation models, summarization, content generation, chat, and transcription with generative AI |
Storage choice follows format and query needs. BigQuery manages large tabular datasets for analysis, while Cloud Storage stores diverse objects such as generated audio, video, and text for serving. Structured imports require JSON Lines or NDJSON with files under 2 GB, rows under 1 MB, and up to 1,000 files per request, while unstructured imports allow files under 200 MB and up to 100,000 files at a time. Grounding connects generation to verifiable sources through RAG, where the query becomes an embedding, similarity search finds relevant document embeddings, and retrieved documents enhance the response.
Exam tip: BigQuery is the home for structured datasets such as transaction logs and inventory lists, while Cloud Storage is the home for unstructured media such as social media sentiment and video archives.
Labeled data and unlabeled data differ in composition and in the learning method they support. Labeled data includes specific ground-truth annotations that state the correct answer and supports supervised learning, while unlabeled data consists of raw information without explicit descriptors and supports self-supervised learning. A confusion matrix compares predictions against ground-truth labels using true positives, true negatives, false positives, and false negatives, which helps reveal bias between data slices.
Business trade-offs balance human cost against scale and speed. Labeled data is slow and expensive because people must read material and assign correct answers consistently, and classes with too few examples cannot be predicted confidently. Unlabeled data scales easily because normal operations produce new content every day, such as conversation traffic that can be sampled without labeling first and scored quickly with LLM-based judges. Foundation models therefore use massive unlabeled datasets for pre-training breadth once with high compute, then use smaller high-quality labeled datasets for fine-tuning and RLHF precision with much less data and compute.
The core layers of the generative AI landscape move from infrastructure to models to platforms to agents to applications. Infrastructure supplies compute and storage, models generate content, platforms turn models into governed systems, agents carry out multi-step work, and applications face users directly. Each layer depends on the one below it, so weak infrastructure or poor model choice limits everything above.
Infrastructure is the bottom layer that supplies compute, storage, networking, and orchestration for training and serving models. Development starts in Colab Enterprise or Vertex AI Workbench notebooks, inputs are stored in managed datasets, repeatable work runs in Vertex AI Pipelines, and trained models are deployed to endpoints that serve inferences. Inference runs in two forms: batch inference processes a group of requests and writes results to one file, while online inference returns synchronous results for individual data points that need real-time answers.
Specialized compute uses graphics processing units (GPUs) and Tensor Processing Units (TPUs) for large matrix operations. Single-host TPU virtual machines (VMs) in a managed instance group (MIG) each handle separate workloads without depending on other VMs, so adding VMs adds independent capacity for inference. A multi-host TPU slice is a group of interconnected TPU VMs that work together as one unit on a single workload, so data and computation flow across connected VMs that depend on each other. The choice between generative AI and traditional AI (opens in a new tab) also affects infrastructure, since training and inference need different compute, storage, and networking resources.
| Platform | Management level | Best-fit workload |
|---|---|---|
| Google Kubernetes Engine (GKE) | Google-managed Kubernetes with deep control and portability | Large-scale training and inference and custom high-performance platforms |
| Vertex AI | Fully managed end-to-end MLOps platform | Rapid development and fast time-to-value without managing infrastructure |
| Cloud Run | Fully managed serverless platform for containers | Event-driven apps and smaller models with scale-to-zero serving |
GKE controls cost with Spot VMs and serves low-latency inference with model-aware routing but requires Kubernetes expertise. Cloud Run scales automatically and can scale to zero so teams pay only while serving, while Vertex AI offloads setup to shorten time-to-market.
Exam tip: Use single-host TPU VMs in a MIG for multiple independent inference workloads, and use a multi-host TPU slice when multiple interconnected TPU VMs must work together on one workload.
Models are the foundation-model layer that turns data into generated text, code, images, and multimodal outputs. A modality describes the kind of data a model accepts or generates, where text models use LLMs, code models turn prompts into functions, image models such as Imagen learn visual patterns through diffusion, and multimodal models such as Gemini handle text, images, video, and audio together. Shared mechanisms support all modalities: embeddings convert content to vectors, grounding connects outputs to verifiable sources to reduce hallucinations, and function calling lets the model name an external function and parameters, wait for the function to run, and use returned results in its final answer.
Vertex AI Model Garden is the central library where teams list deployable models, review verified machine type and accelerator settings, and deploy open models to endpoints. Deployment consumes quota in a designated quota project, so the project must be set before listing and deploying. Model groups trade management effort for control: first-party Google models such as Gemini and Imagen are called as managed APIs, partner models such as Anthropic Claude are called through a partner SDK, and open models such as Gemma and Llama are self-deployed with a chosen machine, accelerator, and license acceptance. In BigQuery, general-purpose functions give full control over model and prompt, managed functions optimize cost and quality over millions of rows, and task-specific functions provide fixed behavior for translation and vision.
Platforms are the unified layer that moves work from model selection to a deployed application with repeatable pipelines and governance. The enterprise deployment of generative AI models (opens in a new tab) follows an ordered lifecycle where data and decisions flow forward and early stages repeat until quality is good enough:
Vertex AI Pipelines automates, monitors, and governs ML systems, while Vertex AI Inference serves predictions from already trained models. Orchestration depends on control needs: Cloud Run runs workloads without managing Kubernetes clusters, while GKE orchestrates the complete lifecycle with Kubernetes control. Grounding pulls context at query time from vector stores such as Cloud SQL and AlloyDB for PostgreSQL, knowledge graphs, and real-world services such as Google Maps. Knowledge Catalog maps business meaning such as txn_qty to purchased quantity, and context flows from source systems into Knowledge Catalog, then to the agent, then into the LLM prompt before the model chooses a function.
Agents are the layer where AI shifts from answering questions to carrying out multi-step business processes. An AI agent is a system built on an LLM that interprets a goal, makes a plan, and executes steps while tracking context. The LLM acts as the reasoning core, planning breaks the goal into ordered sub-tasks, and memory carries context from earlier messages to later decisions. A typical workflow runs in fixed order: a trigger passes the goal to the agent, the agent gathers context from business systems, plans routine decisions within limits, executes actions through tools and APIs, then monitors results and escalates uncertain cases to a human.
Applications are the top layer that faces employees and customers in natural language through grounded support and domain-specific tools. Retrieval-augmented generation powers support by retrieving relevant resources from Cloud Storage first and generating a grounded answer with Gemini second. The grounded support flow runs in order from customer question to retriever to Gemini to fetched sources to generator to delivered answer, so generation depends on retrieval. Horizontal applications target general functions such as chat and summarization for broad reuse, while vertical applications target sectors such as marketing or manufacturing with industry data and compliance rules.
Google foundation models route work by output type: Gemini for language, code, and multimodal reasoning, Gemma for lightweight open deployment, Imagen for images, and Veo for video. Teams discover models in Model Garden and access them as managed APIs or self-deployed endpoints. Choosing correctly depends on matching modality, speed, cost, and control to the task.
Gemini is Google Cloud's family of large foundation models for language, code, and multimodal work. Business uses turn language understanding into content generation such as translation, summarization, creative writing, chatbots, and search help, plus sophisticated reasoning for complex questions and automated coding including through Gemini Code Assist. Human review remains important when output can affect rights, finances, employment, or access to services, because reasoning can be shallow on specialized topics and coding help can produce incorrect code from edge cases or weak prompts.
Long-context work must fit inside the maximum input and output token limit, often by splitting large jobs into smaller parts. Grounding improves accuracy by connecting output to real-world and business data, which reduces fluent but false answers. Safety controls include Vertex AI Studio filtering and API safety scoring, while customers remain responsible for Acceptable Use and Prohibited Use policies. Natively multimodal input lets one Gemini model accept text, images, video, audio, and PDFs together so it can reason across them, where each part carries a MIME type and a resolution of MEDIA_RESOLUTION_HIGH or MEDIA_RESOLUTION_LOW, and PDFs default to 560 resolution.
Advanced generation controls shape randomness and length. Top-K limits choice to the K most probable tokens, Top-P selects until probabilities sum to the chosen value, and Temperature controls randomness where lower values give less open-ended answers. Gemini model variants trade reasoning power against speed and cost:
| Variant | Reasoning strength | Latency and cost focus |
|---|---|---|
| Gemini Ultra | Highest for complex reasoning | Higher cost and latency trade-off |
| Gemini Pro | Strong and balanced | Moderate latency and cost |
| Gemini Flash | Sufficient for common tasks | Lowest latency and most cost-effective |
Gemini Embedding 2 accepts text, images, audio, and video as input only and produces embeddings up to 3,072 dimensions for search rather than generated text.
Exam tip: Inpainting edits content inside the image, such as inserting or removing an object, while outpainting expands content outside the original image boundaries.
Gemma is a family of small, lightweight open foundation models that teams deploy and serve inside Google Cloud. A self-deployable open model is deployed by the customer rather than consumed as a hosted service, so the model, its endpoint, and the network path stay inside the customer's project and virtual private cloud (VPC) network. The same model may be offered as a Google-hosted managed API reached by API call or as a self-deployable model reached through an endpoint the team deploys, and the model card tells which one is shown.
| Offering | Who runs the model | Where it runs |
|---|---|---|
| Managed API (MaaS) | Google-hosted service | |
| Self-deployable open model | Customer | Customer project and VPC network |
The Gemma family specializes by task from general text to code, safety, vision, and medical data:
Deployment follows a fixed order: first list deployable models and record the model ID, then view verified machine type, accelerator, and container image, then deploy to an endpoint and accept the license. After deployment the endpoint is identified by project ID, region, and endpoint ID, and inference sends a prompt to that endpoint path. Serving can use frameworks such as vLLM, Spot VMs, or Compute Engine reservations, while organization policy can allow only vetted models.
Imagen is Google Cloud's foundation model for high-quality image generation and editing. Text-to-image generation creates original images from a written prompt for creative production and marketing visuals. Inpainting edits inside existing image boundaries by inserting objects, removing objects, replacing the background, or editing details, while outpainting expands content beyond original edges. Image-to-image transformation is the broader pattern where an input image plus a text prompt produces a new output image.
Business capabilities map to content tasks: generation creates original assets, editing modifies images with prompts, captioning generates text descriptions, and Visual Question Answering answers questions about image content. Marketing teams produce campaign visuals quickly from prompts, while product teams use subject and style tuning to show the same product in new settings while keeping identity consistent. Enterprise deployment runs on fully managed Vertex AI infrastructure where the application sends a request to the Imagen API, Vertex AI runs inference, and the image or text returns. The shared responsibility model (opens in a new tab) divides duties so Google protects hardware, storage, and network while the customer protects data, identities, and configuration. Customization includes subject, style, controlled, instruct, and upscaling paths, plus production controls for aspect ratio, resolution, language, and deterministic generation.
Veo is Google's high-definition video generation model for creative storytelling and visual communication. A text prompt, still image, or existing clip flows into Veo as input, and Veo returns an MP4 video clip as output. Creative building blocks include text to video, first frame image to video, first and last frames to video, generation from up to three reference images, extending videos, and editing videos with text instructions.
Veo is built for short-form output assembled from shots rather than full films in one step. A single generation produces clips that are 4, 6, or 8 seconds long at 24 frames per second, with up to 4 variations per prompt. Supported formats are 16:9 landscape and 9:16 portrait at 720p or 1080p, with Veo 3.1 Preview adding a 4K option. Marketing uses include product teasers and social clips from brand imagery, internal communications uses include town-hall intros and training segments, and cinematic prototypes use first-and-last-frame control and reference images to pre-visualize scenes. A Veo request works as a long-running operation in a multi-tenant agentic AI system (opens in a new tab): first send the prompt and controls, then poll until completion, then receive video written to Cloud Storage or returned as bytes.
Traditional AI mainly classifies or predicts from historical data, while generative AI (gen AI) learns patterns from training data to generate new text, images, audio, video, and code. Generative AI is a subset of machine learning, which is a subset of artificial intelligence, with machine learning providing models that learn patterns from data instead of following only hand-written rules.
You should first try prompt engineering and move to model tuning only if prompt design alone cannot handle the specialized task. Prompt engineering changes only the request sent to the model through techniques such as zero-shot, one-shot, and few-shot prompting, while model tuning changes the model itself by training it on a dataset through methods such as supervised tuning, reinforcement learning from human feedback (RLHF) tuning, or model distillation.
Single-host Tensor Processing Unit (TPU) virtual machines (VMs) in a managed instance group (MIG) each handle separate workloads without depending on other VMs, so adding VMs adds independent capacity for inference. A multi-host TPU slice is a group of interconnected TPU VMs that work together as one unit on a single workload, so data and computation flow across connected VMs that depend on each other.
Veo returns a high-definition MP4 video clip from a text prompt, still image, or existing clip. A single generation produces clips that are 4, 6, or 8 seconds long at 24 frames per second, with up to 4 variations per prompt, in 16:9 landscape and 9:16 portrait at 720p or 1080p, with Veo 3.1 Preview adding a 4K option.
Last updated