Select Two (Multi-Select) Questions

Maximize model throughput to handle high volumes.

Why wrong: Throughput is secondary to safety.

Use the cheapest model to reduce costs.

Why wrong: Cost should not compromise quality or safety.

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.

Adjusting the temperature

Why wrong: Temperature controls randomness, not style.

Modifying the top_k value

Why wrong: Top_k affects token selection diversity.

Changing the top_p value

Why wrong: Top_p controls nucleus sampling.

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.

Generative models require less training data than discriminative models.

Why wrong: Generative models often require more data to capture the full distribution.

Generative models cannot be used for supervised learning tasks like classification.

Why wrong: Generative models can be used for classification by computing P(Y|X) via Bayes rule.

Discriminative models always outperform generative models on tasks like image classification.

Why wrong: Performance varies; generative models can sometimes excel.

Decrease the model's temperature parameter to make outputs more deterministic.

Why wrong: Temperature does not address bias; it affects randomness.

Conduct a legal review of all generated outputs before release.

Why wrong: This is a process after deployment, not a technique to train the model.

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.

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.

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.

Allow clinicians to share AI-generated summaries with anyone in the organization.

Why wrong: Access must be restricted to authorized personnel only.

Store raw patient data in model training logs for auditing.

Why wrong: Storing raw data violates HIPAA privacy rules.

Higher accuracy on all tasks

Why wrong: RAG may not outperform fine-tuning on specialized tasks.

Reduced model size

Why wrong: RAG uses the same model; it doesn't reduce model size.

Lower latency

Why wrong: RAG may have higher latency due to retrieval step, not lower.

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.

Training dataset size

Why wrong: Training dataset size is not directly evaluated at inference.

Cost per token

Why wrong: Cost is important but often secondary to quality and latency.

Model size in parameters

Why wrong: Model size does not directly correlate with summarization performance.

Increase the model's max output tokens to 2048.

Why wrong: Increasing output tokens increases cost per request.

Implement retry logic with exponential backoff.

Why wrong: Retries increase the number of requests, raising costs.

Lower the temperature parameter to 0.

Why wrong: Temperature does not affect cost; it only changes output randomness.

Reduce top-k to 10

Why wrong: Lower top-k restricts token pool, reducing diversity.

Reduce frequency penalty to 0.0

Why wrong: Lower frequency penalty reduces the penalty on repeated phrases, increasing repetition.

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 can improve safety and relevance; avoiding it may lead to generic and less appropriate outputs.

Use of only open-source models

Why wrong: Open-source models are not inherently more responsible; proprietary models can also be responsibly deployed.

Maximum model size

Why wrong: Model size does not affect responsibility; it is a technical parameter.

Set the threshold for 'SEXUALLY_EXPLICIT' category to BLOCK_NONE.

Why wrong: No blocking allows potentially inappropriate content.

Set the threshold for 'SEXUALLY_EXPLICIT' category to BLOCK_LOW_AND_ABOVE.

Why wrong: Blocks even low probability sexually explicit content, likely too restrictive.

Set the threshold for 'HARASSMENT' category to BLOCK_LOW_AND_ABOVE.

Why wrong: Blocks even low probability harassment, which may overly restrict creative content.

Deploy on TPU instead of GPU

Why wrong: TPUs may not always be faster for generative models.

Increase the max output tokens

Why wrong: Longer outputs increase latency.

Use larger batch sizes

Why wrong: Increases per-request latency.