AWS Certified AI Practitioner practice questions

Question 1

An ML team is hitting a wall running a large, accurate foundation model in production because its per-token cost and latency are too high at scale. They want a smaller, faster model that approaches the large model's accuracy on their task, and they are willing to let Amazon Bedrock generate synthetic training data from the large model. Which customization technique fits this goal?

Accepted Answer

Model distillation, which transfers knowledge from a larger teacher model to a smaller, cheaper student model.

Answer

Fine-tuning the smaller model on a manually labeled dataset that the team curates by hand from production traffic.

Answer

Continued pre-training of the smaller model on the team's raw domain corpus to broaden its base knowledge further.

Answer

Retrieval Augmented Generation, so the smaller model can retrieve the larger model's answers from a vector store.

Question 2

A platform team is being asked which Amazon Bedrock endpoints permission they should attach to a role so that the role can call Bedrock APIs such as the Responses API, Chat Completions, and Messages API. According to the Bedrock User Guide on making inference requests, which managed policy can be attached as a shortcut for this access?

Accepted Answer

If your role has the AmazonBedrockMantleInferenceAccess AWS managed policy attached, you can skip the section on attaching individual permissions to invoke models through the bedrock-mantle endpoint.

Answer

All Bedrock inference calls require disabling IAM entirely so that no permissions are required. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Answer

Inference requests to Bedrock can only be made by the AWS root user and not by any IAM role. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Answer

Bedrock does not support managed policies and customers must write their own SCPs for every API call. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Question 3

A machine learning team is configuring Amazon SageMaker Feature Store and must support both low-latency feature reads for real-time predictions and a complete historical record of features for model training and batch scoring. According to AWS, which TWO statements correctly describe the online and offline stores? (Select TWO.)

Accepted Answer

The online store retains only the latest feature records and serves low-millisecond-latency reads that suit real-time predictions.

Accepted Answer

The offline store keeps all feature records as a historical database, supporting data exploration, model training, and batch predictions.

Answer

The online store keeps the full history of every feature record, making it the right place for data exploration and model training.

Answer

The offline store is optimized for high-throughput, low-latency reads and is the recommended source for live, real-time online model inference.

Question 4

A compliance officer is being asked which use cases are supported for Bedrock Guardrails contextual grounding checks against a reference source and a user query. According to the Bedrock User Guide on contextual grounding check, which set of use cases is documented as supported?

Accepted Answer

The supported use cases include summarization, paraphrasing, and question answering as defined in computer science discipline.

Answer

Contextual grounding checks support only image classification and not text question answering.

Answer

Contextual grounding checks are deprecated and no longer supported for any Bedrock workload.

Answer

Contextual grounding checks support only random text generation and exclude summarization.

Question 5

A platform team is hardening IAM for an Amazon Bedrock application and wants to follow AWS identity best practices for both human and workload access. Which TWO practices align with AWS guidance on least privilege and the use of IAM roles? (Choose two.)

Accepted Answer

Grant only the permissions required to perform a task and then reduce the permissions you grant over time as the use case matures, working steadily toward least privilege.

Accepted Answer

Use IAM roles, which provide temporary security credentials for a role session rather than long-term credentials such as a password or access keys associated with a user.

Answer

Attach a single broad administrator policy to every application principal so that no Amazon Bedrock API call is ever blocked at runtime by a permission that happens to be missing.

Answer

Embed long-term IAM user access keys directly inside the application source code so that the workload always holds durable static credentials that never expire and never rotate.

Question 6

A developer is iterating on a Bedrock prompt and a teammate asks how to compare different variants of the same prompt so the team can pick the best wording. According to the Bedrock User Guide on Prompt management, which capability supports comparing prompt variants and saving versions?

Accepted Answer

When you test your prompt, you have the option of comparing different variants of the prompt and choosing the variant that yields outputs that are best-suited for your use case, and while iterating on your prompt you can save versions of it.

Answer

Bedrock Prompt management does not support multiple prompt variants and only stores a single version. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Answer

Bedrock Prompt management forces every variant to be promoted to production immediately without comparison. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Bedrock Prompt management exclusively supports image generation and not text prompt comparison. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Question 7

A developer is being asked to explain to a non-technical stakeholder why the quality of prompts matters when invoking a generative AI model. According to the AWS overview of prompt engineering, which description best captures what prompt engineering is and why it matters?

Accepted Answer

Prompt engineering is the process where you guide generative AI solutions to generate desired outputs, choosing the most appropriate formats, phrases, words, and symbols that guide the AI to interact with your users more meaningfully.

Answer

Prompt engineering is a hardware tuning practice with no relationship to the wording or structure of user inputs. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Answer

Prompt engineering is identical to retraining the model from its base weights and requires running a new training pipeline for every prompt. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Prompt engineering is a billing concept that determines how AWS charges for inference rather than a way to shape model outputs. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Question 8

An AI product manager is evaluating which machine learning approach best fits a robotics use case where the software must learn the most effective sequence of actions to achieve a goal by interacting with an environment. According to the AWS overview of reinforcement learning, which description best matches this need?

Accepted Answer

Reinforcement learning is a machine learning technique that trains software to make decisions to achieve the most optimal results using a reward-and-punishment paradigm to learn from the feedback of each action.

Answer

Reinforcement learning requires a fully labeled dataset of correct decisions for every step the agent might take. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Answer

Reinforcement learning is a form of unsupervised clustering that does not involve any goal or feedback signal. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Answer

Reinforcement learning is exclusively used for image classification tasks and cannot be applied to robotics or decision sequences. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Question 9

A compliance officer is evaluating how Amazon Bedrock can help reduce hallucinations in a question-answering application that must stay grounded in a reference source. According to the Bedrock User Guide on contextual grounding check, which capability best fits this requirement?

Accepted Answer

Amazon Bedrock Guardrails supports contextual grounding checks to detect and filter hallucinations in model responses when a reference source and a user query is provided, with grounding checking if the model response is factually accurate based on the source and relevance checking if the model response is relevant to the user query.

Answer

Contextual grounding checks are unrelated to hallucination detection and only enforce token rate limits. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Answer

Contextual grounding checks require no reference source and operate purely on the foundation model's training data. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Contextual grounding checks are limited to image inputs and cannot be applied to text question-answering scenarios. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Question 10

A network engineer is being asked how the interface endpoint that provides private connectivity from a VPC to Bedrock is deployed inside the customer's subnets. According to the Bedrock User Guide on interface VPC endpoints, which deployment artifact does AWS create for each enabled subnet?

Accepted Answer

AWS creates an endpoint network interface in each subnet that you enable for the interface endpoint.

Answer

AWS creates a public IP address in each subnet for the interface endpoint.

Answer

AWS creates a Direct Connect circuit in each subnet for the interface endpoint.

Answer

AWS does not create any artifact in the customer's subnets and the connection is implicit.

Question 11

A team has decided to attach tags to a Bedrock application inference profile so finance can track on-demand model invocation costs by business unit. According to the Bedrock User Guide on inference profiles, which capability supports this cost tracking pattern?

Accepted Answer

You can use tags to monitor costs by attaching tags to an application inference profile to track costs when you submit on-demand model invocation requests.

Answer

Bedrock inference profiles do not support tagging and finance cannot track invocation costs by business unit.

Answer

Tagging an inference profile disables on-demand invocations entirely.

Answer

Inference profile tags only affect IAM permissions and have no impact on cost reporting.

Question 12

A developer evaluating Bedrock is being asked by leadership why teams now start with foundation models instead of training every model from scratch. According to the AWS overview of foundation models, which statement best summarizes how foundation models change the way data scientists approach machine learning?

Accepted Answer

Rather than develop artificial intelligence from scratch, data scientists use a foundation model as a starting point to develop ML models that power new applications more quickly and cost-effectively.

Answer

Foundation models eliminate the need for any data scientists by writing all business logic automatically with no human oversight. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Foundation models can only be used for image generation and cannot be used as a starting point for any other modality. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Answer

Foundation models must always be trained on a single customer's private data and cannot leverage broad pre-training. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Question 13

A team is evaluating a machine learning model that is giving accurate predictions for the training data but is performing poorly on new data the model has never seen before. According to the AWS explanation of overfitting, which statement best describes what is happening?

Accepted Answer

Overfitting is an undesirable machine learning behavior that occurs when the machine learning model gives accurate predictions for training data but not for new data, and an overfit model cannot perform well for all types of new data.

Answer

Overfitting is the desired behavior of any production model because it guarantees high training accuracy. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Overfitting only affects deep learning models and is impossible in classical machine learning algorithms. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

Answer

Overfitting is a hardware fault that occurs when GPUs run out of memory during training. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Question 14

An organization is standing up a model-governance catalog and must record each in-house ML model's intended use, risk rating, training details, and evaluation results in one place to support responsible-AI reporting and audits. Which AWS feature is built for documenting these model details?

Accepted Answer

Amazon SageMaker Model Cards, which document a model's intended use, risk rating, and evaluation results

Answer

AWS AI Service Cards, which publish intended use and limitations for AWS's own managed AI services and models

Answer

SageMaker Clarify reports, which quantify pre-training and post-training bias metrics for a given dataset

Answer

Amazon Bedrock Guardrails, which apply configurable safety policies to generative AI prompts and responses

Question 15

An architect is being asked to design a Bedrock workload that can keep running if a single Availability Zone has an outage. According to the Bedrock User Guide on resilience, which property of AWS global infrastructure should the architect rely on for this design?

Accepted Answer

AWS Regions provide multiple physically separated and isolated Availability Zones connected with low-latency, high-throughput, and highly redundant networking, and with Availability Zones you can design and operate applications and databases that automatically fail over between zones without interruption.

Answer

AWS Regions only contain a single Availability Zone, so cross-AZ failover for Bedrock is not possible. This statement misrepresents how AWS documents the service and does not align with the documented architecture or workflow.

Answer

Availability Zones are less highly available than single data center infrastructures and should not be used for resilient designs. This characterization conflicts with the documented service description and cannot be supported by the cited reference.

Answer

Bedrock is built without any underlying region or AZ concept and cannot tolerate any infrastructure failure. The cited AWS documentation does not describe this behavior and the option misstates the service's documented role.

AWS Certified AI Practitioner

Sample questions

Sources

Similar exams