Google Cloud Generative AI Leader: примеры вопросов с ответами
20 примеров вопросов с правильными ответами и объяснениями. Вопросы взяты из банка и соответствуют формату экзамена.
Вопросы составлены на основе образцов, подготовленных специалистами в этой области, и соответствуют формату официального экзамена.
1. A gen AI system recommends which suppliers to drop. Who should be accountable for the final decisions?
- A) No one in particular, because the whole process is automated
- B) Named people who review the recommendations and own the outcome
- C) The model itself, because it produced the recommendation
- D) The cloud provider, because it hosts and operates the model
Показать ответ и объяснение
Правильный ответ: B
Ответственность несут конкретные люди, которые проверяют рекомендации и отвечают за итог.
2. A manufacturer wants a gen AI assistant that answers maintenance questions from 20 years of manuals, but the manuals are scanned PDFs in three versions with no owner. What should come first?
- A) Build the assistant now and let the model work out which manual version is correct
- B) Assign data ownership, select the current versions and make the documents searchable before building the assistant
- C) Fine-tune a model on all three versions so it learns every answer
- D) Skip the manuals and rely on the model's general knowledge of machines
Показать ответ и объяснение
Правильный ответ: B
Сначала назначить владельца данных, выбрать актуальные версии и сделать документы доступными для поиска.
3. A large company has dozens of teams building gen AI tools independently, with no common rules. What should leadership put in place first?
- A) A governance framework with policies, risk review, approved tools and owners
- B) A rule that internal tools can launch without any review at all
- C) A ban on all teams except IT building any gen AI tools from now on
- D) A single large model that every team must use for every use case
Показать ответ и объяснение
Правильный ответ: A
Нужна рамка управления ИИ: политики, оценка рисков, одобренные инструменты и ответственность.
4. What makes a gen AI pilot most useful for deciding whether to scale?
- A) Clear success criteria, a defined user group, a baseline and a time frame
- B) As many features as possible in the first release to impress users
- C) No measurement at all, so that users feel free to experiment openly
- D) The hardest problem in the company, to prove what the technology can do
Показать ответ и объяснение
Правильный ответ: A
Полезный пилот: критерии успеха, группа пользователей, базовая линия и срок.
5. A CV-screening model rates candidates from certain universities higher because past hires came mostly from them. What does this show?
- A) The model's temperature is set too low
- B) The model has a knowledge cutoff
- C) The model is hallucinating facts about the universities
- D) The model has learned historical bias from its training data
Показать ответ и объяснение
Правильный ответ: D
Модель усвоила историческое смещение из обучающих данных.
6. A team wants a model to answer questions about a 1,500-page technical manual in a single request. Which model characteristic matters most?
- A) Whether it uses supervised or unsupervised training
- B) A long context window measured in tokens
- C) A high temperature setting
- D) The number of image styles it supports
Показать ответ и объяснение
Правильный ответ: B
Важнее всего длинное контекстное окно, измеряемое в токенах.
7. What is a prompt in the context of a generative model?
- A) The hardware accelerators that run the model during inference
- B) The input instructions and context given to the model to guide its output
- C) The output text, image or code that the model returns to the user
- D) The large dataset used to pre-train the model before its release
Показать ответ и объяснение
Правильный ответ: B
Промпт — входные инструкции и контекст, направляющие ответ модели.
8. A regulated bank must keep all data and model weights inside its own environment and audit the model fully. Which option suits it best?
- A) An open model such as Gemma deployed in the bank's controlled environment
- B) A consumer chatbot app used through a browser
- C) Any managed API, since all APIs store data in the customer's data centre
- D) A model chosen only by benchmark ranking
Показать ответ и объяснение
Правильный ответ: A
Полный контроль весов и среды даёт открытая модель (Gemma) в собственной среде.
9. Which set of components does Google describe as the core of an AI agent?
- A) A database for storage, a dashboard for users and a tool for reporting results
- B) A training dataset, a team of human labellers and a dedicated GPU cluster
- C) A model for reasoning, tools to act on systems, and orchestration of the steps
- D) An image generator, a digital watermark and a public web search engine
Показать ответ и объяснение
Правильный ответ: C
Агент: модель для рассуждений, инструменты и оркестрация шагов.
10. What distinguishes an AI agent from a simple chatbot that answers questions?
- A) An agent always uses a larger and more expensive model than a chatbot does
- B) An agent cannot access external data, so it relies only on its training
- C) An agent only produces images and video, while a chatbot produces text
- D) An agent can plan steps toward a goal and use tools or APIs to take actions
Показать ответ и объяснение
Правильный ответ: D
Агент планирует шаги и использует инструменты и API для действий.
11. A retailer's existing system predicts next week's demand for each product, and a new tool writes product descriptions from a few keywords. How are the two best classified?
- A) The first is predictive machine learning; the second is generative AI
- B) Both are generative AI because both produce an output
- C) Both are traditional machine learning because both learn from data
- D) The first is rule-based AI; the second is reinforcement learning
Показать ответ и объяснение
Правильный ответ: A
Прогноз спроса — предиктивное ML, генерация описаний — генеративный ИИ.
12. An agent needs a standard way to connect to the company's databases and internal tools as data sources. Which open protocol is commonly used for this?
- A) Agent2Agent protocol (A2A)
- B) Model Context Protocol (MCP)
- C) Top-p (nucleus) sampling
- D) Secure AI Framework (SAIF)
Показать ответ и объяснение
Правильный ответ: B
MCP — стандарт подключения агента к инструментам и данным.
13. A hospital needs to extract text and fields from scanned referral forms quickly, without building a model. Which option fits best?
- A) Gemma fine-tuned on a small set of medical images
- B) A new extraction model trained from scratch on TPUs
- C) Document AI, a pre-trained document processing service
- D) Veo, used to turn the scanned forms into a video summary
Показать ответ и объяснение
Правильный ответ: C
Document AI — готовый сервис извлечения данных из документов.
14. An engineering team building its own RAG pipeline needs to store and search embeddings at large scale. Which Google Cloud capabilities support this?
- A) Google Slides, which stores embeddings in presentations
- B) SynthID, which tags each embedding with a watermark
- C) Cloud Billing reports, which index usage by project
- D) Vector Search, or vector search in BigQuery and AlloyDB
Показать ответ и объяснение
Правильный ответ: D
Эмбеддинги хранят и ищут в Vector Search или в BigQuery и AlloyDB.
15. A team wants the model to classify customer emails into its own five categories with a consistent format. The first attempt, with instructions only, is inconsistent. What should it try next?
- A) Removing the category names to let the model decide
- B) Raising the temperature for more varied categories
- C) Zero-shot prompting with shorter, fewer instructions
- D) Few-shot prompting: add a few labelled example emails
Показать ответ и объяснение
Правильный ответ: D
Few-shot: добавить в промпт несколько размеченных примеров.
16. A team has refined prompts for weeks but the model still lacks knowledge of the firm's internal procedures. What is the best next step?
- A) Raise the maximum output tokens for longer answers
- B) Keep rewriting the prompt with more detailed adjectives
- C) Ground the model in the procedure documents through RAG
- D) Switch to a larger model trained on more public data
Показать ответ и объяснение
Правильный ответ: C
Промпты не добавляют знаний; нужно заземление на внутренние документы через RAG.
17. In which case is fine-tuning more appropriate than prompt engineering or RAG?
- A) When the team has no example data at all and needs results within the same day
- B) When the model must reliably follow a specialised style or format that prompts cannot hold
- C) When answers must reflect information that changes every hour across many systems
- D) When the only goal is to shorten answers, which prompts and settings cannot do
Показать ответ и объяснение
Правильный ответ: B
Дообучение уместно для устойчивого специального стиля или формата, которого не дают промпты.
18. How does retrieval-augmented generation (RAG) improve answers?
- A) It compresses the model so it runs faster and can read more documents in each request
- B) It retrieves relevant passages from a knowledge source at query time and adds them to the prompt
- C) It raises the model's temperature so that answers draw on a wider range of knowledge
- D) It retrains the model's weights on all new company documents every night in a batch job
Показать ответ и объяснение
Правильный ответ: B
RAG находит релевантные фрагменты в момент запроса и добавляет их в промпт.
19. A finance analyst asks the model a multi-step question about loan repayments and often gets wrong totals. Which prompting technique is most likely to help?
- A) Role prompting, asking the model to act as a famous economist
- B) Asking for a shorter answer with only the final total shown
- C) Chain-of-thought prompting, asking for step-by-step reasoning
- D) Zero-shot prompting with no context and no instructions
Показать ответ и объяснение
Правильный ответ: C
Chain-of-thought: просить модель рассуждать по шагам.
20. A legal team needs consistent, factual clause extraction, while the marketing team wants varied slogan ideas. How should temperature be set?
- A) High for the legal team, low for the marketing team
- B) Low for both, because temperature only affects speed
- C) Low for the legal team, higher for the marketing team
- D) High for both, to maximise quality
Показать ответ и объяснение
Правильный ответ: C
Низкая температура — для стабильных фактов, высокая — для разнообразия идей.