Section 1: Fundamentals of generative AI約30% · 6 topics

Q1. Which best describes natural language processing (NLP)?

1. The AI field that finds objects in images and video and judges their position and type
2. The AI field that lets computers understand and generate human language
3. The security field that detects unauthorized network intrusion and blocks it automatically
4. The statistics field that predicts future figures such as sales from tabular numeric data

Q2. Which best states the difference between prompt engineering and prompt tuning?

1. Prompt engineering retrains all model parameters; prompt tuning only rewrites the instruction text
2. Prompt engineering is only for image generation; prompt tuning is only for text generation
3. The former crafts how instructions are written; the latter trains a small number of extra parameters
4. Both retrain model weights from scratch, differing only in the amount of training data

Q1. A retailer wants to find why customer satisfaction is falling. Which use of gen AI best combines structured and unstructured data?

1. Analyze the purchase history table together with inquiry email text, and summarize trends in complaints
2. Aggregate the purchase table in BigQuery and chart spending trends of unhappy customers
3. Summarize inquiry emails with gen AI and report a list of the most frequent topics
4. Use the survey score table to predict low-scoring stores with regression

Q2. Which is the best example of reinforcement learning used in generative AI?

1. Using human ratings as rewards to tune LLM answers to be helpful and safe
2. Grouping unlabeled text by similar topics
3. Filling missing values in tabular data with the average
4. Labeling images “cat” or “dog” to train a classification model

Q1. At a board meeting you explain gen AI's impact on the business. Using the exam guide's create/summarize/discover/automate view, which explanation is most appropriate?

1. Impact is mostly limited to IT development efficiency; other departments can keep working as before
2. Use of unstructured data such as documents and images grows, affecting productivity and customer experience in every department
3. The main effect is faster aggregation of tabular data; impact on sales or HR document work is small
4. The main value is headcount reduction, replacing human judgment starting with routine work

Q2. What is the most appropriate advantage of a model with a large context window?

1. Answers are always shorter, and usage fees automatically drop
2. Hallucinations stop entirely, so checking is unnecessary
3. It always knows information newer than its training data without outside search
4. It can take long documents or many materials at once and answer with the whole in view

Q1. You have many scanned PDFs of paper contracts and want to search clauses with gen AI. From the data format viewpoint, what should you consider first?

1. Contracts are structured data, so they must be re-entered into tables before use
2. Volume matters more than format, so first collect every contract and check quality later
3. Multimodal models can read scans, so skip format considerations
4. Images are hard to search, so estimate the effort and cost of extracting or converting text

Q2. Which is the most appropriate business challenge in creating labeled data?

1. Assigning answers takes time and money, and inconsistent criteria hurt quality
2. Labeling always cuts the amount of data in half
3. Labeling work doesn't affect data quality
4. Labeled data can't be used in generative AI at all

Q1. You are designing a system that takes an inquiry, checks inventory and places an order if needed. Which gen AI layer is mainly required?

1. Agents (use tools and data to carry out multiple steps autonomously)
2. Applications (provide the screens and features users operate)
3. Models (pretrained on large data to generate text and more)
4. Platforms (provide the base for using, tuning and managing models)

Q2. What is the most appropriate advantage of choosing an application-layer generative AI product?

1. You can start using it at work right away with no development
2. You can freely choose the model's training data
3. You can design the computing hardware yourself
4. You can freely rewrite the model's weights

Q1. Which is the most appropriate reason to choose Gemma?

1. It specializes in image generation, and using it always requires a Google Workspace contract
2. It specializes in video generation and can't be used for text
3. It's the largest of all Google's models and gives top accuracy on any task
4. Its weights are published, making it easy to run and customize in your own environment

Q2. Which best describes the TPU?

1. A processor Google designed itself for AI
2. An admin screen for humans to evaluate generative AI output
3. A dedicated API for translating Gemini's answers
4. Storage for saving Google Workspace documents