AWS Certified AI Practitioner AIF-C01 Practice Questions with Explanations
Questions
Which statement best describes how AI, ML, and deep learning relate to each other?
- Deep learning is a subset of ML, which is a subset of AI
- AI is a subset of ML, which is a subset of deep learning
- ML and deep learning are separate fields with no overlap
- Deep learning is a subset of AI but not of ML
What is the difference between training and inference?
- Training teaches a model from data; inference uses it to make predictions
- Training makes predictions on new data; inference adjusts the model's internal weights
- Training happens only in the cloud; inference happens only on mobile and edge devices
- Training uses unlabeled data only; inference uses labeled data only
What makes agentic AI different from a standard chatbot that only answers questions?
- It can plan steps, use tools, and take actions toward a goal with limited human input
- It is trained only on images instead of text, so it cannot generate written answers for users
- It always gives the same output for the same input because it has no randomness
- It runs only on local hardware and cannot connect to any external systems
A model learns to predict house prices from past sales where each record includes the sale price. Which type of learning is this?
- Supervised learning
- Unsupervised learning
- Reinforcement learning
- Self-supervised clustering
A retailer wants to group customers into segments based on purchase behavior, with no predefined labels. Which type of learning fits?
- Unsupervised learning
- Supervised learning
- Reinforcement learning
- Transfer learning
A robot learns to walk by receiving rewards for staying upright and penalties for falling. Which type of learning is this?
- Reinforcement learning
- Supervised learning
- Unsupervised learning
- Few-shot learning
Which type of inference fits scoring millions of records overnight when no one needs immediate answers?
- Batch inference
- Real-time inference
- Serverless inference
- Streaming inference
A fraud check must return a decision within milliseconds while a customer waits at checkout. Which inference type fits?
- Real-time inference
- Batch inference
- Asynchronous inference
- Offline inference
Requests have large payloads and can take several minutes to process, and the caller doesn't need to wait for the answer. Which SageMaker AI inference option fits?
- Asynchronous inference
- Real-time inference
- Serverless inference
- Batch transform
Traffic to a model is light and intermittent, and the team wants to pay nothing when it's idle and can tolerate occasional cold starts. Which option fits?
- Serverless inference
- Real-time inference on a dedicated instance
- Batch transform
- Asynchronous inference with a fixed instance
Which is an example of unstructured data?
- Customer support call recordings
- A table of daily sales by store
- A CSV of employee IDs and salaries
- Sensor readings in fixed columns
Hourly electricity usage recorded over three years is which type of data?
- Time-series data
- Image data
- Unlabeled text data
- Graph data
Which branch of AI focuses on understanding and generating written and spoken human communication?
- Natural language processing (NLP)
- Computer vision
- Reinforcement learning
- Robotic process automation (RPA)
A large language model (LLM) is a type of foundation model.
A model scores 99% accuracy on its training data but only 60% on new data. What is this called?
- Overfitting
- Underfitting
- Data drift
- Good generalization
A bank needs to predict a numeric value: how much a customer will spend next month. Which ML technique fits?
- Regression
- Classification
- Clustering
- Dimensionality reduction
An email system must label each message as spam or not spam. Which ML technique fits?
- Classification
- Regression
- Clustering
- Anomaly forecasting
In which situation is AI/ML a poor fit?
- When an exact, rule-based answer is required, such as statutory sales tax
- When the task involves finding patterns across millions of customer records
- When the workload is too large for staff to review every item by hand
- When decisions improve with predictions learned from historical data
A company needs to convert recorded customer calls into searchable text. Which AWS service fits?
- Amazon Transcribe
- Amazon Polly
- Amazon Translate
- Amazon Comprehend
Which AWS service finds sentiment, entities, and key phrases in text without building a model?
- Amazon Comprehend
- Amazon Lex
- Amazon Textract
- Amazon Polly
A company wants a voice and text chatbot for its call center menu. Which service is designed for this?
- Amazon Lex
- Amazon Polly
- Amazon Transcribe
- Amazon Rekognition
Which service converts text into natural-sounding speech for an e-learning app?
- Amazon Polly
- Amazon Transcribe
- Amazon Lex
- Amazon Translate
A bank must explain to regulators exactly why each loan was approved or denied, using structured application data. Which approach is usually more appropriate?
- A traditional ML model chosen for explainability
- A large foundation model prompted with each application
- A diffusion model that generates the decision
- A multimodal FM with high temperature settings
Which AWS service extracts printed text, handwriting, and table data from scanned forms?
- Amazon Textract
- Amazon Rekognition
- Amazon Comprehend
- Amazon Translate
An online store wants product recommendations based on each shopper's behavior, without ML expertise. Which service fits?
- Amazon Personalize
- Amazon Comprehend
- Amazon Lex
- Amazon Polly
What is the usual order of core steps in an ML pipeline?
- Collect data, prepare data, train, evaluate, deploy, monitor
- Deploy, train, collect data, monitor, evaluate, prepare data
- Train, deploy, collect data, prepare data, monitor, evaluate
- Evaluate, monitor, deploy, train, prepare data, collect data
A model's accuracy drops over months because customer behavior has changed since training. What is this called, and what is the usual fix?
- Model drift; monitor the model and retrain it on recent data
- Overfitting; delete the model and stop using ML for the task
- Hallucination; lower the temperature setting during inference
- Underfitting; deploy the same model to more instances
In a fraud model, catching as many fraud cases as possible matters most, even if some legitimate transactions get flagged. Which metric matters most?
- Recall
- Precision
- Accuracy
- Mean squared error
A spam filter must rarely put real emails in the spam folder. Which metric should be prioritized?
- Precision
- Recall
- Mean absolute error
- R-squared
Which metric combines precision and recall into one number?
- F1 score
- Accuracy
- Root mean squared error
- Perplexity
Which of these is a business metric, not a model performance metric?
- ROI per user
- F1 score on the test set
- Recall at a fixed threshold
- Precision
Which SageMaker AI capability lets a business analyst build ML predictions with a no-code visual interface?
- Amazon SageMaker Canvas
- Amazon SageMaker Model Monitor
- Amazon SageMaker Clarify
- Amazon SageMaker Ground Truth
Which practice is a core MLOps principle?
- Building repeatable, automated pipelines for training and deployment
- Training each model by hand on a laptop with no version control
- Deploying models once and never checking how they perform after release to production
- Keeping training data on personal drives so it stays private
A team wants to call a powerful model without managing any infrastructure. Which deployment method fits?
- A managed API service such as Amazon Bedrock
- Self-hosting the model on EC2 GPU instances
- Training a new model from scratch on SageMaker AI
- Running the model on employees' laptops
In an LLM, what is a token?
- A unit of text, such as a word or part of a word, that the model processes
- A security credential that grants an application access to the model's API endpoint
- A single complete document stored in a vector database
- A label that marks a training example as correct
What are embeddings?
- Numerical vectors that represent the meaning of text, images, or other data
- Encrypted copies of the training data that are stored in Amazon S3 for later audits
- Short prompts that tell a model what role to play
- Labels that humans add to images for training
Why are long documents split into chunks before they are stored for a RAG application?
- So relevant pieces can be retrieved and fit within the model's context window
- So the documents are encrypted before reaching the vector database
- So the model can be retrained on each chunk every night as the documents change over time
- So the documents use fewer tokens when they are stored in S3
Which architecture underlies most modern large language models?
- Transformer
- Decision tree
- K-means clustering
- Linear regression
Which type of model is commonly used to generate images by gradually removing noise?
- Diffusion model
- Linear regression model
- K-nearest neighbors model
- Recurrent decision tree
A model accepts an image and a text question and answers about the image. What kind of model is this?
- Multimodal model
- Unimodal text model
- Clustering model
- Regression model
Which is the correct general order of the foundation model lifecycle?
- Data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback
- Deployment, feedback, data selection, evaluation, fine-tuning, pre-training, model selection
- Fine-tuning, deployment, pre-training, evaluation, data selection, feedback, model selection
- Evaluation, pre-training, deployment, data selection, feedback, model selection, fine-tuning
With token-based pricing for an FM, which change most directly lowers inference cost?
- Shorter prompts and capped output length
- Raising the temperature setting
- Adding more example documents to every prompt
- Asking for longer, more detailed answers
What is context engineering in an FM application?
- Choosing what instructions, data, memory, and tool results the model sees
- Retraining the model's weights on new data every time a user sends a request
- Choosing which AWS Region hosts the model to reduce network latency for users
- Encrypting the prompt so other users of the model cannot read it
What is the Model Context Protocol (MCP)?
- An open standard for connecting AI agents to external tools and data
- An AWS billing model that charges for each token sent to an agent
- An encryption standard that protects prompts while they are in transit
- A benchmark dataset used to score how accurate an LLM's answers are
A complex task is split so one AI component plans, another researches, and another writes the final report. What pattern is this?
- A multi-agent system
- A single zero-shot prompt
- A batch inference job
- A supervised classification model
An agent needs to recall a customer's preferences from previous sessions. Which agentic concept addresses this?
- Memory management
- Temperature tuning
- Tokenization
- Model distillation
Which is a strong use case for generative AI?
- Summarizing long reports into short briefs
- Calculating payroll taxes by exact formula
- Sorting a list of numbers in ascending order
- Storing transaction records in a database
Agents can use tools, such as calling APIs or querying databases, to complete tasks.
An LLM confidently cites a court case that doesn't exist. What is this called?
- A hallucination
- Data drift
- Overfitting
- A prompt template
The same prompt gives slightly different answers each time. Which GenAI characteristic is this?
- Nondeterminism
- Interpretability
- Explainability
- Determinism