Task 2.1: Describe Google Cloud's strengths in the field of gen AI.

• AThe training restriction in the Service Specific Terms

  Correct

  Under the training restriction, Google won't use customer data to train or fine-tune any AI/ML models without the customer's prior permission or instruction, for all managed models on Agent Platform.

• BPrompt logging for abuse monitoring by Google

  Incorrect

  Abuse monitoring is a reason Google may log prompts, to detect violations of its usage policies; it is not a promise about how data is used for training.

• CVPC Service Controls around the bank's project

  Incorrect

  VPC Service Controls are a network perimeter around the bank's resources; they do not govern what Google does with data inside its services.

• DCustomer-managed encryption keys (CMEK) for all of its stored data

  Incorrect

  CMEK is an encryption control; it does not address how the data may be used for model training.

Enterprise-ready AI starts with data commitments: the training restriction covers training use, while encryption and perimeters cover other risks.

• AGPUs on Compute Engine

  Incorrect

  GPUs fit a different workload profile; for this one, Google points to the hardware built around large matrix operations.

• BCPUs on Compute Engine

  Incorrect

  CPUs suit quick prototyping and small models that train quickly, not months-long training of a large model.

• CProvisioned Throughput on Agent Platform

  Incorrect

  Provisioned Throughput is not hardware for training your own model.

• DCloud TPUs on Compute Engine

  Correct

  TPUs are Google's custom-designed ASICs for large matrix operations, and Google recommends them for models dominated by matrix computations, with no custom operations in the training loop, that train for weeks or months.

Google's AI-optimized infrastructure includes custom TPUs as well as GPUs; which one fits depends on the workload's shape.

• AGemini models are available only through Google's own applications

  Incorrect

  Gemini is available to developers through the Gemini API and Agent Platform, not only inside Google's applications.

• BAgent Platform requires models to be fine-tuned before deployment

  Incorrect

  Many models in Model Garden are ready to use; tuning is optional.

• CGemini Enterprise replaces third-party models across the company

  Incorrect

  Gemini Enterprise is a work assistant and agent platform for employees; it is not a way to swap the models a retailer's own applications call.

• DModel Garden offers models from Google, partners and the open-source community

  Correct

  Model Garden is a single place to discover, test, customize and deploy more than 200 models from Google, Google partners and open source, so teams can pick the right model per task.

Openness means choice: first-party, partner and open models side by side, with one way to deploy and govern them.

• AStandard pay-as-you-go

  Incorrect

  Standard PayGo is flexible per-token pricing with no commitment, but it cannot guarantee a quality of service.

• BProvisioned Throughput

  Correct

  Provisioned Throughput provides guaranteed throughput for a commitment period and is meant for critical, steady-state, always-on workloads where an SLA is needed.

• CFlex pay-as-you-go

  Incorrect

  Flex trades slower responses and more throttling for a lower price, which is the opposite of what a storm-day assistant needs.

• DBatch inference

  Incorrect

  Batch inference is for large asynchronous jobs whose results can wait; travelers rebooking in real time cannot.

Reliability and scale are part of being enterprise-ready: the consumption option should match the workload's traffic pattern and how critical it is.

• AGoogle's custom TPU chips, designed for machine learning

  Incorrect

  TPUs are among the accelerators AI Hypercomputer uses, but the system also brings in open software, ML frameworks and consumption models; it is not one chip family.

• BA platform for building, deploying and governing AI agents

  Incorrect

  That describes Gemini Enterprise Agent Platform, which sits in the platform layer; AI Hypercomputer is the infrastructure underneath.

• CIntegrated hardware, open software and flexible consumption for AI

  Correct

  AI Hypercomputer is a supercomputing system that brings together performance-optimized hardware, open software, leading ML frameworks and flexible consumption models for AI training, tuning and serving.

• DA plan that reserves guaranteed model throughput for a term

  Incorrect

  That describes Provisioned Throughput, one consumption option for models on Agent Platform; AI Hypercomputer is a whole system, not a pricing plan.

Google's AI-optimized infrastructure is more than chips: AI Hypercomputer integrates hardware, software and consumption options into one system, and platforms and pricing plans build on top of it.

• AGemini in BigQuery

  Correct

  Gemini in BigQuery is part of Google Cloud with Gemini; it generates, completes and explains SQL and Python code and helps prepare and analyze data inside BigQuery.

• BGemini in Data Studio

  Incorrect

  Gemini in Data Studio helps analyze and export Data Studio data and reports, not SQL inside the analysts' BigQuery tools.

• CGemini in Looker

  Incorrect

  Gemini in Looker helps with visualizations, LookML and conversational analysis in Looker; the analysts work in BigQuery.

• DGemini Cloud Assist

  Incorrect

  Gemini Cloud Assist helps design, operate and troubleshoot cloud applications and resources; it is not the SQL helper in BigQuery.

Gen AI is built into many Google products, so help appears where people already work.

• AThe training restriction on customer data in the terms

  Incorrect

  The training restriction governs whether data may be used to train models; it gives the bank no control over encryption keys.

• BDefault encryption at rest with Google-managed keys

  Incorrect

  Data is encrypted at rest by default, but with keys Google manages; the regulator wants the bank to control them.

• CCustomer-managed encryption keys (CMEK) in Cloud KMS

  Correct

  With CMEK the customer manages the key in Cloud KMS that protects its data in supported services; RAG Engine on Agent Platform supports CMEK.

• DModel Armor templates on the application

  Incorrect

  Model Armor screens prompts and responses; it does not manage encryption keys.

Google Cloud's AI platform gives customers control over their data, including who holds the keys.

• AFine-tuning both agents on the same dataset

  Incorrect

  Tuning changes model behavior; it does not let two agents exchange tasks.

• BMoving both agents into one Gemini Notebook

  Incorrect

  Gemini Notebook is a research tool over curated sources; it does not connect agents.

• CReplacing both models with the same Gemini version

  Incorrect

  A shared model does not give agents a way to talk to each other.

• DSupport for the open Agent2Agent (A2A) protocol

  Correct

  A2A is an open standard for communication and collaboration between AI agents; Agent Platform supports inter-agent collaboration with A2A, and Agent Runtime can host A2A agents.

Openness covers agents as well as models: open protocols let agents built with different tools work together.

• ABy writing an agent with Agent Development Kit in Python

  Incorrect

  ADK is a code-first framework, which she cannot use.

• BBy using Workflow Builder in Gemini Enterprise

  Correct

  Workflow Builder is a no-code, low-code way to create chat agents and workflows from natural-language prompts or a visual editor, connect data sources and tools, and schedule recurring runs.

• CBy training an AutoML model on last year's weekly summaries

  Incorrect

  AutoML trains image and tabular models; it does not build a scheduled agent.

• DBy asking IT to deploy an open model on a set of GPUs

  Incorrect

  Deploying a model is infrastructure work that still leaves the agent unbuilt.

No-code tools are how Google Cloud puts agent building in the hands of business users, not only developers.