Free AIF-C01 Practice Questions

• (Natural Language Processing [NLP])
  NLP is a branch of AI focused on enabling machines to understand and process human language. It's used in chatbots, sentiment analysis, and speech recognition.
• (Deep Learning)
  Deep Learning is a subset of Machine Learning that uses neural networks with many layers. It powers applications like image recognition, self-driving cars, and voice assistants.

Incorrect Options

• (Overfitting)
  Overfitting is a model issue during training where it memorizes the training data too well, failing to generalize. It’s a concept in ML, not a technique.
• (Bias)
  Bias refers to unfair tendencies in data or models, often leading to unethical outcomes. It’s a risk in AI, not a technique.
• (Large Language Model [LLM])
  LLMs like GPT are examples of models, not techniques. They are built using deep learning techniques, but aren't techniques themselves.

• (Personalized, predictive, adaptive → Best AI/ML fit)
  Machine learning analyzes individual patterns (like past attendance, appointment times, demographics) to dynamically predict the no-show likelihood per patient, enabling smarter overbooking strategies. This creates personalized and scalable optimizations, rather than using averages or manual intervention.

Incorrect Options

• Automated text reminder system
  Helpful, but not predictive. It sends reminders but doesn't assess risk or optimize scheduling dynamically.
• Static overbooking policy
  Using overall averages is rigid. It doesn’t adapt to individual patient behavior or changing patterns over time.
• Hiring staff to call patients
  Manual labor-intensive process. Not scalable or adaptive to trends in no-show behavior without ML analysis.

• (Recommendation systems → AI/ML)
  A recommendation system uses machine learning to suggest relevant items (in this case, jobs) to users based on historical behavior, preferences, and patterns — exactly the role of an AI-powered recommendation engine.

Incorrect Options

• Music streaming app categorizing songs by genre
  This involves audio feature analysis and unsupervised learning (like clustering), but it is not personalized recommendations for individual users.
• E-commerce chatbot about return policies
  A rule-based or scripted chatbot answering policy questions is a form of automation, not a recommendation engine.
• Weather app predicting future temperatures
  Weather forecasting uses time-series prediction models. It is predictive analytics, not recommending items or actions based on user behavior.

Incorrect Options:

• Model Training
  Model training involves using the features provided to teach the model, but it assumes that the features have already been prepared. It doesn’t involve creating or refining input features.
• Hyperparameter Tuning
  Hyperparameter tuning adjusts model settings (like learning rate, tree depth) to optimize performance. It doesn’t modify the input data features.
• Model Deployment
  Deployment refers to moving the trained model into a production environment to serve predictions. It is unrelated to transforming input data.

Incorrect Options:

• Tokens
  Tokens are pieces of input text (such as words, subwords, or characters) broken down for model processing, but they aren't numeric representations of meaning.
• Diffusion models
  Diffusion models are generative models used primarily for image generation tasks, not for representing text semantics.
• Prompt engineering
  Prompt engineering focuses on crafting input instructions to guide a model’s output, not on numerical representations of text.

Incorrect Options:

• Detecting fraudulent transactions in real time
  This is a classification problem using traditional supervised machine learning, not generative AI.
• Sorting customer support tickets by topic
  This is typically a clustering or classification task, not a content generation task.
• Monitoring server performance metrics for anomalies
  This involves anomaly detection, a standard machine learning task, not generative content creation.

Incorrect Options:

• High memory consumption
  This is a disadvantage of many large models, not an advantage related to adaptability or flexibility.
• Complexity
  While generative AI models can be complex, complexity itself is not an advantage; simplicity and ease of interaction are typically preferred.
• Static rule enforcement
  Static rules refer to rigid systems that don’t adapt dynamically, which is the opposite of what makes generative AI advantageous.

Incorrect Options:

• AWS Shield
  AWS Shield is a security service focused on DDoS (Distributed Denial of Service) protection, unrelated to building generative AI applications.
• AWS Direct Connect
  Direct Connect provides private network connections between on-premises environments and AWS - it has nothing to do with model development.
• Amazon Macie
  Macie is used for discovering and protecting sensitive data like PII (Personally Identifiable Information) - not for developing generative AI applications.

Incorrect Options:

• Perplexity: Perplexity measures how surprised a language model is by the next word it predicts. It’s useful for evaluating how well a model fits a dataset - not for checking whether summaries keep the same meaning.
• BLEU: Bilingual Evaluation Understudy (BLEU) checks how similar the wording is to a reference, which works best for direct translations. But this question is about keeping meaning, even if the words change.
• F1 Score: The F1 Score is used in classification tasks to measure how well a model balances precision (how many selected items are relevant) and recall (how many relevant items are selected). It’s not suitable for evaluating generated text like summaries, because it doesn't measure how well the meaning or quality of text is preserved.

Incorrect Options:

• Support for multi-lingual capabilities
  Multi-language support is valuable but not the startup’s main priority if they are focused on speed and device efficiency.
• High model complexity for deeper analysis
  Complex models are typically slower and larger, making them poorly suited for lightweight mobile applications needing quick outputs.
• Maximum input/output length for large documents
  Transcription of short spoken sentences doesn't require handling extremely large inputs or outputs, so maximizing document size is unnecessary here.

Incorrect Options:

• It guides the model to produce longer, more detailed responses
  This is influenced by prompt wording or parameters like maximum token limits, not negative prompts.
• It increases randomness in the model's output by adjusting temperature settings
  Temperature affects creativity/randomness, but it’s unrelated to the concept of negative prompting.
• It fine-tunes the model's internal embeddings to represent external knowledge
  Fine-tuning involves training processes, not the crafting of prompts during inference.

Incorrect Options:

• Training loss evaluation after model optimization
  Training loss measures how well the model fits the training data — it doesn't tell you how users perceive the outputs.
• BLEU score comparison against benchmark datasets
  BLEU is useful for measuring linguistic similarity, especially in translation tasks, but it can miss deeper quality issues that humans would spot.
• Automated model performance using F1 and accuracy scores
  These scores help with objective classification metrics but are limited when evaluating subjective tasks like creativity, style, or nuance.

Incorrect Options:

• High bias resulting in underfitting to the evaluation data
  High bias leads to both poor training and evaluation performance, not just evaluation performance. Underfitting typically shows low accuracy across the board.
• Poor data diversity leading to biased results
  Lack of data diversity can cause bias issues, but the key symptom described here (training vs evaluation performance gap) is more closely linked to variance and overfitting.
• Low variance creating instability in the model
  Low variance would make the model more stable, not less. Instability comes from high variance, not low.

Incorrect Options:

• Amazon SageMaker Clarify
  Clarify helps detect bias and explain model behavior during training and evaluation phases, but it doesn't actively monitor a deployed model in production like Model Monitor does.
• Amazon Augmented AI (Amazon A2I)
  A2I supports human review workflows for AI predictions but isn’t designed to automatically detect bias or drift without human involvement.
• Amazon Bedrock Guardrails
  Guardrails are designed for controlling generative AI content, not for monitoring traditional ML models for bias or drift.

Incorrect Options:

• Deep neural network
  Deep neural networks are complex "black box" models. While highly accurate for some tasks, it's extremely difficult to explain their internal reasoning for individual predictions.
• Transformer model
  Transformers are used heavily in NLP and generative AI, but they operate through deeply layered attention mechanisms that make explanations difficult without specialized techniques.
• Convolutional neural network (CNN)
  CNNs are excellent for processing images but have internal feature extraction layers that are not naturally interpretable without extra analysis like saliency maps.

Incorrect Options:

• Amazon SageMaker Clarify
  While Clarify detects bias and explains model predictions, it focuses on bias mitigation and fairness analysis rather than documenting the full lifecycle or transparency like Model Cards do.
• Amazon SageMaker Ground Truth
  Ground Truth is used to create labeled datasets for model training, not to assess or document transparency or explainability.
• Amazon Kendra
  Amazon Kendra is an intelligent search service for retrieving information, not a tool for evaluating machine learning model transparency or documentation.

Incorrect Options:

• Amazon Macie
  Amazon Macie is used for identifying and protecting sensitive data at rest (like in S3), not for securing traffic or endpoints.
• IAM Roles Anywhere
  IAM Roles Anywhere extends AWS identity access to workloads outside of AWS but doesn't secure private network traffic to AI endpoints.
• AWS Key Management Service (AWS KMS)
  KMS provides encryption key management, but it doesn't secure endpoint access directly over private networking like PrivateLink does.

Incorrect Options:

• SageMaker Model Cards
  Model Cards document model details such as intended uses, evaluation results, and ethical considerations, but they do not track dataset transformations and lineage at the dataset level.
• Amazon Macie
  Amazon Macie identifies and protects sensitive data (like PII) but does not provide a full lineage or transformation history of datasets.
• AWS Config
  AWS Config monitors configuration changes across AWS resources for compliance auditing, but it is not intended for tracking dataset origins or transformations.

Incorrect Options:

• System and Organization Controls (SOC)
  SOC reports demonstrate internal control effectiveness but are mainly for audit transparency, not standardized international security certifications.
• Payment Card Industry Data Security Standard (PCI DSS)
  PCI DSS applies narrowly to payment card data, not general AI system compliance or broad operational security.
• General Data Protection Regulation (GDPR)
  GDPR governs personal data privacy and protection within the EU. It's essential for compliance but not a security operational standard like ISO 27001.

Incorrect Options:

• AWS encrypts data at rest by default, but customers must manually enable encryption in transit
  Data at rest is usually encrypted by customer configuration per service, and encryption in transit is enabled by default, not manual in most AWS-managed services.
• AWS provides encryption options, but it is the customer’s responsibility to encrypt data in transit
  While customers can add extra encryption, AWS itself already encrypts traffic between services by default.
• AWS automatically applies AES-256 encryption for all data moving inside the AWS network without customer involvement
  TLS (for transit) and AES-256 (for rest) serve different roles — AES-256 isn't used for all in-transit data.