Task 1.1: Describe core generative AI (gen AI) concepts and use cases.

• AGenerative AI is a field of machine learning, and machine learning is a subfield of AI

  Correct

  Google Cloud describes machine learning as one of the most popular subfields of AI, and generative AI as a field of machine learning that develops models for generating new content.

• BMachine learning is a field of generative AI, and generative AI is a subfield of AI

  Incorrect

  The nesting runs the other way: generative AI is built on machine learning, not the reverse.

• CGenerative AI is a branch of AI that does not use machine learning

  Incorrect

  Generative AI is described as a field of machine learning; its models learn patterns from large amounts of training data.

• DAI, machine learning and generative AI are three names for the same field

  Incorrect

  The terms are related but not interchangeable: each one is a narrower part of the one before it.

Think of nested circles: AI is the broadest field, machine learning is the data-driven approach inside it, and generative AI is the part of machine learning that creates new content such as text, images and code.

• AUnsupervised learning

  Incorrect

  Unsupervised learning looks for patterns in unlabeled data; here every invoice already carries the outcome the model must predict.

• BSupervised learning

  Correct

  The historical invoices carry the outcome the model must predict, so they are labeled examples. Supervised learning trains on labeled data to predict the label for new inputs.

• CReinforcement learning

  Incorrect

  Reinforcement learning learns by acting in an environment and receiving rewards or penalties; there is no environment to act in here, only past examples with known outcomes.

• DRetrieval-augmented generation

  Incorrect

  Retrieval-augmented generation grounds a generative model's answers in documents; it does not train a predictor on labeled history.

Labeled history plus a known target to predict is the signature of supervised learning; here it is a classification problem with two classes.

• AModel Registry

  Incorrect

  Model Registry is the central place to organize, version and deploy models. It tracks which model versions exist, not how production data changes.

• BModel Monitoring

  Correct

  Model Monitoring tracks deviations such as drift between the data a model sees in production and its baseline, and alerts when a metric crosses the threshold you set, so you can re-evaluate or retrain.

• CAgent Platform Pipelines

  Incorrect

  Pipelines are not a drift detector; they do not watch a deployed model's inputs or raise alerts on their own.

• DAgent Platform Feature Store

  Incorrect

  Feature Store can monitor the statistics of the feature data registered in it and detect feature drift, but it does not watch the requests that reach a deployed model's endpoint or compare them with the training data.

Model management in the ML lifecycle includes watching a deployed model: Model Monitoring detects drift and alerts; the other tools version, automate and serve data.

• AThe lowest price per output token

  Incorrect

  Cost matters, but a cheap model whose context window cannot hold the agreement cannot meet the requirement at all.

• BSupport for image and video output

  Incorrect

  Output modality is a selection factor when the use case needs it; this one needs long text input, not generated images or video.

• CA context window large enough to hold an entire agreement

  Correct

  The context window is the number of tokens a model can process in one prompt; the larger it is, the more of a long document the model can take into account at once.

• DThe option to fine-tune the model on the team's past contracts

  Incorrect

  Fine-tuning adapts behavior and style; it does not change how much text fits into one request.

Modality, context window, cost, performance and customization are all selection factors; the deciding one is the factor the use case cannot work without, here the size of the context window.

• ACreate: generating new content such as images or marketing copy

  Incorrect

  Creating new content is a different use case; the scientists need to find existing knowledge, not produce new material.

• BDiscover: finding the right information at the moment it is needed

  Correct

  Searching across many sources with one natural-language question and getting an answer from the right documents is the discovery use case, the one enterprise search and assistants such as Gemini Enterprise address.

• CAutomate: carrying out a previously manual task end to end

  Incorrect

  Automation replaces a manual process, such as handling an expense report; here the person still asks and decides, and the tool finds.

• DSummarize: condensing one known document into a short version

  Incorrect

  Summarizing works on content you already have in hand; the problem here is that the scientists do not know where the answer is.

Gen AI creates, summarizes, discovers and automates. When the starting point of a search is unknown and the value is in finding the answer across sources, the use case is discovery.

• AA text-only large language model

  Incorrect

  A text-only model can draft the summary, but it cannot take in the photos or the audio note.

• BA multimodal foundation model

  Correct

  Multimodal models process information from several modalities, such as text, images, audio and video, in the same request, which is what this use case needs.

• CAn AutoML image classification model

  Incorrect

  An AutoML image model sorts images into the categories it was trained on; it does not read forms, hear audio or write summaries.

• DThe Cloud Speech-to-Text API

  Incorrect

  Speech-to-Text transcribes the voice note into text, which is only one of the three inputs; it does not understand the photos or write the summary.

When one task combines text, images and audio, a single multimodal model such as Gemini replaces a chain of single-purpose models.

• ASupervised learning, training a model to predict each shopper's group

  Incorrect

  Supervised learning needs labeled examples of the groups to predict; the chain has no categories yet, so there is nothing to label.

• BReinforcement learning, rewarding the model when shoppers buy more

  Incorrect

  Reinforcement learning learns which actions earn rewards in an environment; the chain wants to find structure in past purchases, not choose actions.

• CUnsupervised learning, finding groups in the unlabeled purchases

  Correct

  Unsupervised learning works on unlabeled data and finds its natural structure, such as groups of buyers who purchase related products.

• DFew-shot prompting, showing a model examples of each shopper group

  Incorrect

  Few-shot prompting needs examples of the answers you want; the groups are what the chain is trying to discover, so it has no examples to give.

No labels and a goal of finding patterns or groups points to unsupervised learning, often clustering. Any approach that starts from examples of the groups assumes the answer the chain is looking for.

• ASupervised learning

  Incorrect

  Supervised learning would need routes labeled as best, which the operator does not have.

• BUnsupervised clustering

  Incorrect

  Clustering groups similar items; it does not learn which actions maximize a score.

• CFew-shot prompting of an LLM

  Incorrect

  Few-shot prompting shows an LLM examples in a prompt; it does not train a robot's behavior through rewards.

• DReinforcement learning

  Correct

  In reinforcement learning, an agent learns by trial and error in an environment, receiving rewards for good outcomes and refining its strategy to earn the most reward over time.

Actions, an environment and a reward signal are the signature of reinforcement learning.

• AThey are pre-trained on broad data and can be adapted to many tasks

  Correct

  Foundation models are pre-trained on massive, broad datasets, which lets them be adapted to a wide range of downstream tasks rather than built for one.

• BThey are fine-tuned on the team's own examples before each new task

  Incorrect

  Fine-tuning is optional and comes after pre-training; the team uses the model for all three tasks without retraining it.

• CThey are multimodal, so they accept text, images and audio as input

  Incorrect

  Multimodality is about the kinds of input a model can take; all three of these tasks are text tasks, so it does not explain the range.

• DThey have a long context window that holds a whole report at once

  Incorrect

  The context window limits how many tokens one prompt can hold; it helps with long inputs but does not explain why one model handles many tasks.

Breadth of pre-training is what makes a foundation model the base for many gen AI applications: one model, many tasks, with tuning only where a task needs it.

• AData ingestion

  Incorrect

  Ingestion is collecting the raw logs in the first place; the team has moved on to cleaning and transforming them.

• BModel training

  Incorrect

  Training uses the prepared data to create the model; no model is being trained yet.

• CModel deployment

  Incorrect

  Deployment makes a trained model available for use, which comes after training.

• DData preparation

  Correct

  Cleaning and transforming raw data into a usable format for training is data preparation; Google Cloud recommends tools such as Dataflow and BigQuery for this preprocessing.

The lifecycle runs from ingestion and preparation through training and deployment to management. Preparation is where raw data becomes training-ready.

• AThe frontier Gemini model, built for the hardest reasoning

  Incorrect

  A frontier model is built for deep, complex reasoning; on a simple task at this volume it adds cost and latency without a matching gain.

• BGemini Live, built for low-latency live interactions

  Incorrect

  Live models are optimized for real-time voice conversations; these reviews are short text tagged in bulk, with no one on the other end.

• CGemini 3.8 Flash-Lite TTS, built for high-throughput speech

  Incorrect

  Gemini 3.8 Flash-Lite TTS is a text-to-speech model; it produces spoken audio and cannot produce the text label each review needs.

• DGemini Flash-Lite, built for low-cost, high-volume traffic

  Correct

  Google positions its Flash-Lite models for high-volume, cost-sensitive, low-latency traffic, which fits a simple text task at massive scale.

Performance and cost trade off. Match the model's capability to the task's difficulty and modality, and let volume push you toward efficient models.