AWS Certified AI Practitioner AIF-C01 Practice Test 20
Question 190
A team is selecting how to serve a production FM under predictable sustained load. A pilot uses this approach instead: Always buy dedicated throughput for a one-off experiment. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Reserved capacity can waste money under sporadic load
- Training per inference is not a viable serving strategy
- On-demand services can still have limits and performance constraints
- Capacity choice should reflect sustained traffic, latency and pricing
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Always buy dedicated throughput for a one-off experiment. Reserved capacity can waste money under sporadic load
Incorrect Answers
B:This limitation concerns Train a new model on every request to guarantee throughput; it does not diagnose the proposed Always buy dedicated throughput for a one-off experiment approach.
C:This limitation concerns Assume on-demand never has rate or quota limits; it does not diagnose the proposed Always buy dedicated throughput for a one-off experiment approach.
D:Capacity choice should reflect sustained traffic, latency and pricing This explains the recommended design, but it is not the requested diagnosis of Always buy dedicated throughput for a one-off experiment.
Question 191
An agent must access approved external APIs through a common tool interface. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A reviewed design record approves this competing approach: Assume mentioning a tool in a prompt makes it callable
- The agent discovers approved tools and is denied unauthorized calls
- A component test invokes this alternate approach successfully: Use vector relevance as permission to transfer money
- A system inventory lists this competing approach as active: Use tokenization as the API authorization mechanism
Explanation
Correct Answer
B:The agent discovers approved tools and is denied unauthorized calls. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Prompt text cannot establish secure tool registration Design approval is not evidence that the alternative meets the actual requirement.
C:Similarity is not authorization A component invocation does not validate the necessary end-to-end outcome.
D:Breaking text into tokens does not authenticate API calls An active setting cannot independently prove the requested business outcome.
Question 192
An engineer describes how a modern LLM relates distant words in a sentence. A pilot uses this approach instead: Describe an LLM as only a static rule table. Which limitation MOST DIRECTLY explains why the requirement is not met?
- Transformer attention represents dependencies across token positions
- Retrieval indexes are external to in-model attention
- ASR is an application task, not the model architecture
- Rules do not capture learned contextual representations
Explanation
Correct Answer
D:This is the actual failure mode of the proposed approach: Describe an LLM as only a static rule table. Rules do not capture learned contextual representations
Incorrect Answers
A:Transformer attention represents dependencies across token positions This explains the recommended design, but it is not the requested diagnosis of Describe an LLM as only a static rule table.
B:This limitation concerns Describe vector indexing as the whole network; it does not diagnose the proposed Describe an LLM as only a static rule table approach.
C:This limitation concerns Treat speech transcription as the transformer architecture; it does not diagnose the proposed Describe an LLM as only a static rule table approach.
Question 193
Several specialized agents must research claims, calculate a payment and request final human approval. Which option BEST meets the business requirement?
- Use isolated chatbots with no handoff state
- Grant every agent unrestricted payout permissions
- Orchestrate agent handoffs with explicit tool and approval controls
- Use embeddings search as a final claim settlement
Explanation
Correct Answer
C:Coordinated agents can combine outputs while preserving authority checks The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:Disconnected assistants cannot reliably compose the process This prevents the approach from meeting the stated requirement.
B:Unrestricted write access bypasses human approval This prevents the approach from meeting the stated requirement.
D:Retrieval alone does not authorize or execute payment This prevents the approach from meeting the stated requirement.
Question 194
An assistant gives convincing but invented policy clauses. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A reviewed design record approves this competing approach: Treat output length as a correctness certificate
- A component test invokes this alternate approach successfully: Increase sampling randomness to ensure correct facts
- A system inventory lists this competing approach as active: Assume fluent wording guarantees factual accuracy
- Fact checks identify ungrounded clauses against approved sources
Explanation
Correct Answer
D:Fact checks identify ungrounded clauses against approved sources. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Long responses may contain unsupported content Design approval is not evidence that the alternative meets the actual requirement.
B:Higher randomness cannot create missing evidence A component invocation does not validate the necessary end-to-end outcome.
C:Style is not factual evidence An active setting cannot independently prove the requested business outcome.
Question 195
A support agent must recall prior steps for the same customer without exposing other customers. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- Cross-user access tests confirm isolated state and same-user continuity
- A reviewed design record approves this competing approach: Discard all prior context despite multi-turn requirements
- A system inventory lists this competing approach as active: Put every user's conversation in one unrestricted context
- A component test invokes this alternate approach successfully: Use model temperature to isolate sessions
Explanation
Correct Answer
A:Cross-user access tests confirm isolated state and same-user continuity. This checks the business result rather than just a configured feature.
Incorrect Answers
B:No session state breaks follow-up references Design approval is not evidence that the alternative meets the actual requirement.
C:Shared raw history violates isolation An active setting cannot independently prove the requested business outcome.
D:Sampling randomness cannot enforce access control A component invocation does not validate the necessary end-to-end outcome.
Question 196
A team is selecting how to serve a production FM under predictable sustained load. After selecting the recommended approach, which evidence MOST DIRECTLY demonstrates that the actual required outcome has been achieved?
- A reviewed design record approves this competing approach: Train a new model on every request to guarantee throughput
- Peak load tests and observed unit economics justify the chosen capacity mode
- A system inventory lists this competing approach as active: Always buy dedicated throughput for a one-off experiment
- A component test invokes this alternate approach successfully: Assume on-demand never has rate or quota limits
Explanation
Correct Answer
B:Peak load tests and observed unit economics justify the chosen capacity mode. This checks the business result rather than just a configured feature.
Incorrect Answers
A:Training per inference is not a viable serving strategy Design approval is not evidence that the alternative meets the actual requirement.
C:Reserved capacity can waste money under sporadic load An active setting cannot independently prove the requested business outcome.
D:On-demand services can still have limits and performance constraints A component invocation does not validate the necessary end-to-end outcome.
Question 197
An engineer describes how a modern LLM relates distant words in a sentence. Before approving a rollout, an architect suggests Describe vector indexing as the whole network. Which mismatch should be documented for this precise design?
- Transformer attention represents dependencies across token positions
- ASR is an application task, not the model architecture
- Retrieval indexes are external to in-model attention
- Rules do not capture learned contextual representations
Explanation
Correct Answer
C:This is the actual failure mode of the proposed approach: Describe vector indexing as the whole network. Retrieval indexes are external to in-model attention
Incorrect Answers
A:Transformer attention represents dependencies across token positions This explains the recommended design, but it is not the requested diagnosis of Describe vector indexing as the whole network.
B:This limitation concerns Treat speech transcription as the transformer architecture; it does not diagnose the proposed Describe vector indexing as the whole network approach.
D:This limitation concerns Describe an LLM as only a static rule table; it does not diagnose the proposed Describe vector indexing as the whole network approach.
Question 198
A legal assistant must draft a new paragraph from a policy topic instead of assigning a label. A pilot uses this approach instead: Use binary classification as the entire draft. Which limitation MOST DIRECTLY explains why the requirement is not met?
- A category label does not supply prose
- Grouping documents cannot compose the new paragraph
- A number does not generate requested policy wording
- A generative model can compose new natural-language text
Explanation
Correct Answer
A:This is the actual failure mode of the proposed approach: Use binary classification as the entire draft. A category label does not supply prose
Incorrect Answers
B:This limitation concerns Use clustering without a text generator; it does not diagnose the proposed Use binary classification as the entire draft approach.
C:This limitation concerns Use regression to output a monetary score as the paragraph; it does not diagnose the proposed Use binary classification as the entire draft approach.
D:A generative model can compose new natural-language text This explains the recommended design, but it is not the requested diagnosis of Use binary classification as the entire draft.
Question 199
A team wants managed access to several pretrained FMs through an AWS API. During a design review, someone proposes Use AWS CloudTrail as the inference endpoint. Which technical objection is specifically relevant to this proposal?
- Textract focuses on document extraction
- CloudTrail logs account activity rather than generating text
- Storage does not by itself run FM inference
- Bedrock provides managed access to supported FMs
Explanation
Correct Answer
B:This is the actual failure mode of the proposed approach: Use AWS CloudTrail as the inference endpoint. CloudTrail logs account activity rather than generating text
Incorrect Answers
A:This limitation concerns Use Amazon Textract as the FM catalog; it does not diagnose the proposed Use AWS CloudTrail as the inference endpoint approach.
C:This limitation concerns Treat an S3 object as an automatic model runtime; it does not diagnose the proposed Use AWS CloudTrail as the inference endpoint approach.
D:Bedrock provides managed access to supported FMs This explains the recommended design, but it is not the requested diagnosis of Use AWS CloudTrail as the inference endpoint.
Question 200
A product owner is choosing between text-only and image-capable models for photo-assisted replies. Which option BEST meets the business requirement?
- Use text-only just because it is a foundation model
- Select a model with the required image input modality
- Assume language support automatically includes photos
- Select by model popularity rather than interface capabilities
Explanation
Correct Answer
B:Models differ in supported input/output types The specified scenario makes this the most appropriate approach.
Incorrect Answers
A:FM status does not imply image input This prevents the approach from meeting the stated requirement.
C:Multilingual and multimodal support are independent This prevents the approach from meeting the stated requirement.
D:Popularity cannot guarantee image understanding This prevents the approach from meeting the stated requirement.
