Hard Difficulty Questions

Number of model versions

Why wrong: Model versioning does not directly affect latency.

Number of model instances

Why wrong: While important, autoscaling handles instance count dynamically.

Increase the beam search width to 10.

Why wrong: Wider beam search improves fluency but not factual accuracy.

Fine-tune the model on the conversation dataset.

Why wrong: Fine-tuning may help but doesn't guarantee factual grounding for new queries.

Set the temperature to 0.1.

Why wrong: Low temperature reduces creativity but doesn't add factual grounding.

Free with advertising.

Why wrong: Ad-supported models are rare for specialized AI tools and may not generate sufficient revenue.

One-time license fee for the model.

Why wrong: One-time fees are uncommon for cloud-hosted AI; ongoing costs for inference and maintenance require recurring revenue.

Selling user data collected from interactions.

Why wrong: Selling user data raises privacy concerns and is illegal in many jurisdictions.

Store all patient data in Cloud Storage with object versioning.

Why wrong: Data storage is separate from model deployment; versioning does not address access restrictions.

Enable encryption at rest for all resources.

Why wrong: Encryption at rest is necessary but not sufficient; network access control is also required.

Use a publicly accessible endpoint for faster response times.

Why wrong: Public endpoints violate HIPAA by exposing PHI to the internet.

The user is not authorized to use Vertex AI at all

Why wrong: The aiplatform.user role does authorize use of Vertex AI, just not deployment.

The model artifact is not in the same region as the endpoint

Why wrong: While region mismatch can cause issues, the most common IAM-related deployment failure is missing permissions.

The user needs the roles/aiplatform.admin role

Why wrong: Admin role is more permissive but not specifically required for deployment.

Use a larger batch size

Why wrong: Larger batch size can provide more stable gradients but doesn't prevent overfitting.

Increase the number of training epochs

Why wrong: More epochs typically increase overfitting.

Reduce the learning rate

Why wrong: Lower learning rate slows learning but does not directly address overfitting.

Memorization of sensitive training data

Why wrong: While a risk, it is less commonly cited than the others; only three correct answers are required.

Increased model accuracy

Why wrong: Increased accuracy is a benefit, not a risk.

Use a larger model with more parameters

Why wrong: Larger models are slower.

Disable safety filters

Why wrong: Disabling safety filters does not reduce latency significantly and risks harm.

Increase the number of tokens

Why wrong: More tokens increase processing time.

Persistent Disk

Why wrong: Persistent Disk is block storage but not as performant for massive training data access patterns.

Cloud Filestore

Why wrong: Filestore is a file storage service but may not match Cloud Storage's throughput for training data.

Cloud Spanner

Why wrong: Cloud Spanner is a relational database, not designed for training data storage.

Use a zero-shot model with context in prompt

Why wrong: Zero-shot with context is limited by context length and may not scale.

Fine-tune the model on the documents

Why wrong: Fine-tuning embeds data into the model, risking exposure.

Use prompt engineering to instruct the model

Why wrong: Prompt engineering alone cannot handle large volumes of proprietary data effectively.

Alice can view models; Bob can delete models

Why wrong: aiplatform.user includes more than view; delete is not granted by customCodeModelAdmin.

Both have full access to all Vertex AI resources

Why wrong: Neither role grants full access; aiplatform.user is limited, customCodeModelAdmin is specific.

Alice can train models; Bob can deploy models

Why wrong: Training is not typically part of aiplatform.user; customCodeModelAdmin includes deployment of custom models, but not necessarily all models.

The endpoint will automatically scale based on GPU utilization

Why wrong: GPU utilization is not a scaling metric in dedicated resources.

The endpoint will scale from 1 to 3 replicas based on load using automatic scaling

Why wrong: It uses dedicated resources, not automatic scaling.

The endpoint can scale to zero when not in use

Why wrong: Dedicated resources do not support scaling to zero.

Upgrade to GPU-accelerated machines for all replicas.

Why wrong: A is wrong because GPUs increase cost and may not fix cold start latency.

Implement client-side request batching to reduce the number of inference calls.

Why wrong: C is wrong because batching increases response time and is not suitable for real-time apps.

Use prewarmed containers by setting an idle timeout to keep instances alive.

Why wrong: D is wrong because Vertex AI serverless endpoints do not have an idle timeout setting; they scale to zero.

Reduction in model error rate on a test set

Why wrong: Model accuracy improvement alone does not guarantee business ROI if the task is already well-performed.

Increase in user satisfaction scores

Why wrong: User satisfaction is important but hard to monetize and may not correlate with cost savings.

Cost per inference compared to historical average

Why wrong: Cost per inference is a component but not a comprehensive ROI metric; it ignores benefits.

Rapidly deploy an MVP and iterate based on user feedback

Why wrong: Rapid deployment without compliance checks could lead to regulatory penalties.

Open-source the model to gain community trust

Why wrong: Open-sourcing does not address regulatory compliance and may introduce security risks.

Partner with a cloud provider that offers indemnification for model outputs

Why wrong: Indemnification protects against lawsuits but does not ensure regulatory compliance with financial advice rules.

Reduce the temperature to 0 and set top_p to 1.

Why wrong: Deterministic sampling reduces randomness but does not enhance reasoning.

Increase the number of output tokens and enable beam search with width 4.

Why wrong: Beam search can improve likelihood but not necessarily reasoning quality.

Apply reinforcement learning from human feedback (RLHF) using a reward model trained on the existing evaluation dataset.

Why wrong: RLHF requires new human preference data, not just the evaluation dataset.

Create a new regional bucket in australia-southeast1 and store a copy of the index there, then redeploy the vector search index endpoint to use the local bucket.

Why wrong: While this might help, it requires manual synchronization and is not the most scalable or reliable solution. The recommended approach is to use a multi-region bucket.

Deploy a separate vector search index endpoint for each region with its own index copy stored in a regional bucket in that region.

Why wrong: This would increase complexity and cost, and still requires managing consistency across regions. A multi-region bucket is simpler and more effective.

Increase the number of replicas for the vector search index in all regions to improve throughput and reduce latency.

Why wrong: Increasing replicas does not address the underlying issue of index update latency caused by a single regional bucket. It may also increase costs unnecessarily.

Use Cloud Audit Logs to capture all API calls to Vertex AI, but do not log the actual prompts and responses

Why wrong: Audit Logs do not capture model input/output, failing the audit requirement.

Log all prompts and responses to Cloud Storage and use a Cloud DLP job to scan and redact sensitive data periodically

Why wrong: Periodic redaction means sensitive data remains in logs temporarily, violating security requirements.

Implement a custom proxy that logs all requests after stripping sensitive data, then forward to the model

Why wrong: A proxy adds latency and complexity; not ideal for performance.

Deploy the model on a dedicated Vertex AI endpoint with autoscaling.

Why wrong: Dedicated endpoints are for custom models, not the API; autoscaling does not reduce request volume.

Switch to a lower-tier model like Gemini 1.0 Pro to reduce quota consumption.

Why wrong: Lower-tier models may have reduced accuracy or capabilities, impacting moderation quality.

Request a quota increase from Google Cloud support.

Why wrong: Quota increases can take time and may not be granted immediately; also increases potential cost.

Vertex AI Feature Store

Why wrong: Feature Store manages features, not model evaluation.

Vertex AI Prediction

Why wrong: Prediction serves models but does not evaluate performance over time.

Vertex AI Pipelines

Why wrong: Pipelines orchestrate workflows but do not provide continuous evaluation by themselves.