Select Two (Multi-Select) Questions

Maximize model size for best accuracy

Why wrong: Larger models consume more resources and can amplify biases; size does not equate to responsibility.

Avoid fine-tuning the model to preserve original capabilities

Why wrong: Fine-tuning adjusts a model’s weights on domain-specific data to align outputs with organisational policies and safety guardrails; avoiding it leaves the model vulnerable to generating unconstrained or harmful content, which directly contradicts responsible deployment’s requirement for controlled output. This option is tempting because preserving a base model’s general capabilities is valuable for maintaining broad utility, and it would be correct in a scenario where the goal is to avoid catastrophic forgetting or maintain benchmark performance across diverse tasks.

Fine-tune a model on historical email data to ensure brand voice

Why wrong: Fine-tuning adds complexity and may not be necessary for drafting emails; prompt engineering may suffice.

Deploy a dedicated GPU cluster for inference

Why wrong: Deploying a dedicated GPU cluster for inference addresses the computational load of running large models locally, but the scenario explicitly states limited engineering resources. Managing GPU infrastructure—including orchestration, scaling, and maintenance—requires specialised DevOps skills the company lacks. This option is tempting because it offers full control over latency and data privacy, and would be correct if the company had an in-house ML engineering team and needed to avoid third-party API dependencies.

Build a custom web UI for the assistant from scratch

Why wrong: Custom UI development requires significant engineering resources.

Always provide multiple examples in the prompt

Why wrong: While few-shot learning can help, too many examples may lead to confusion or token limits.

Avoid any context in the prompt

Why wrong: Without context, the model lacks guidance and may produce irrelevant responses.

Set temperature to 1.0 for maximum creativity

Why wrong: High temperature increases randomness, which may reduce quality for customer service.

Deploy the model in a single region to reduce network costs

Why wrong: Single region may reduce network costs but is not a primary cost optimization strategy.

Store all model prediction logs indefinitely for auditing

Why wrong: Storing logs indefinitely increases storage costs; define retention policies.

Always use GPU instances for inference

Why wrong: GPU instances are more expensive; use CPU for latency-tolerant workloads.

Apply conditional formatting rules based on patterns

Why wrong: Conditional formatting is a standard manual feature, not Duet AI.

Create charts from images pasted into the sheet

Why wrong: Chart creation from images is not a Duet AI Sheets feature.

Write custom Apps Script code automatically

Why wrong: Duet AI does not generate Apps Script code; it assists with formulas.

Use only open-source models to ensure transparency

Why wrong: Open-source is not required by law.

Automatically delete all user logs after 30 days

Why wrong: GDPR does not specify a fixed retention period; it depends on purpose.

Use prompt engineering to tune the model without retraining

Why wrong: Prompt engineering is not part of RLHF.

Collect human-annotated demonstrations of ideal responses

Why wrong: That is supervised fine-tuning, not RLHF.

Deploy the model in A/B testing to gather implicit feedback

Why wrong: Implicit feedback is not part of the standard RLHF pipeline; RLHF uses explicit human rankings.

They allow the model to generate new training data automatically.

Why wrong: Embeddings are for retrieval, not data generation.

They provide deterministic outputs for the same query.

Why wrong: Embeddings themselves are deterministic, but the overall RAG output may vary due to generator randomness.

They are deterministic by design.

Why wrong: They use probabilistic generation.

They always produce the same output for the same input.

Why wrong: They are stochastic; output can vary.

They require no data for training.

Why wrong: Training requires large datasets.

Using only English prompts

Why wrong: Limiting to one language does not reduce bias; it may introduce language-related bias.

Increasing model size

Why wrong: Larger models may learn more biases from training data; size does not inherently reduce bias.

Increasing model size to learn more patterns

Why wrong: Larger models may amplify rather than reduce bias.

Relying solely on post-hoc filters

Why wrong: Post-hoc filters are insufficient and may not catch subtle biases.

Limiting the model to only factual prompts

Why wrong: This restricts input but does not address inherent bias in training.

Increase the model's temperature to make outputs more conservative.

Why wrong: Higher temperature increases randomness, making outputs less predictable and potentially more toxic.

Use a separate language model to rephrase the outputs before sending to users.

Why wrong: This adds latency and complexity; not the most immediate or direct measure.

Fine-tune the model on a curated dataset of polite conversations.

Why wrong: Fine-tuning is effective but takes time and resources; not an immediate fix.

Number of layers in the model

Why wrong: Number of layers is a technical detail not relevant to build vs. buy decision.

Brand reputation of the model provider

Why wrong: Brand reputation is secondary; technical fit and cost matter more.

Volume of expected inference requests

Why wrong: Volume affects cost but both custom and API can handle high volume; not a primary decision factor.

Disable logging to reduce cost

Why wrong: Logging is important for monitoring.

Manually scale instances based on expected load

Why wrong: Automatic scaling is preferred.

Serve the model directly without optimization

Why wrong: Optimization (e.g., quantization) reduces latency.

One-click deployment to a Vertex AI endpoint

Why wrong: Deployment is separate.

Fine-tuning models directly in the interface

Why wrong: Fine-tuning is done via other tools.

Building conversational agents with drag-and-drop

Why wrong: That's Agent Builder.

Sentiment analysis

Why wrong: Sentiment analysis is available but not a core conversational feature of Agent Builder.

Code execution

Why wrong: Code execution is not a built-in feature of Agent Builder.

Knowledge base integration

Why wrong: Knowledge base integration is a separate feature, not specifically for conversation flow.

Safety filters: Pro has stricter safety defaults

Why wrong: Safety settings are configurable for both.

Multimodal capability: Flash does not support image input

Why wrong: Both support multimodal input.

Lower cost compared to using generic APIs

Why wrong: Cost depends on usage, not inherently lower.

Ability to fine-tune models on custom data

Why wrong: Available but not unique to Model Garden.

Guaranteed data privacy and no data sharing

Why wrong: Not guaranteed; depends on model and settings.

Allow the AI to make autonomous decisions in time-sensitive emergencies

Why wrong: Autonomous decisions without human review are risky and not aligned with responsible AI in healthcare.

Publish all patient data used for training to ensure transparency

Why wrong: Publishing patient data violates privacy and is not allowed under HIPAA and GDPR.

Use the AI only for administrative tasks, not diagnosis

Why wrong: The scenario is about diagnosis; avoiding the core use case does not address responsible AI.

Cloud SQL

Why wrong: Relational database, not suited for vector search.

Bigtable

Why wrong: NoSQL database, not optimized for similarity search.

Cloud Functions

Why wrong: Can be used as middleware but not core component.