Página traducida y verificada automáticamente. Prevalece la versión en inglés.Algunas partes siguen en inglés mientras se verifica su traducción.Ver en inglés · Informar un error de traducción
How does machine learning differ from traditional programming?
In machine learning, an algorithm learns the mapping from inputs to outputs from data, while in traditional programming developers code the rules directly. Hand-written if-then rules describe traditional programming, not ML. ML outputs are usually probabilistic predictions rather than exact answers, and unsupervised learning works without labels, so labeled data is not needed for every problem.
Question 2 of 10 · Fundamentals of Generative AI
What is an embedding in generative AI?
An embedding is a vector of numbers that captures the meaning of text or other data, so similar content has similar vectors; embeddings power semantic search and RAG retrieval. A unit of text is a token, and a known output attached to an example is a label. Response length is limited by the maximum tokens setting.
Question 3 of 10 · Applications of Foundation Models
What should you consider about latency when choosing a model?
Model selection weighs latency against accuracy and cost: a real-time chatbot needs fast responses, while a batch summarization job can tolerate slower ones. Benchmark scores, context window size and training cutoff are separate selection criteria. None of them tells you whether responses arrive fast enough for the use case.
Question 4 of 10 · Guidelines for Responsible AI
Why involve humans in high-stakes AI decisions?
Human oversight places an accountable person where AI errors could seriously affect people, such as in lending, hiring or health care, so mistakes can be caught and corrected; Amazon Augmented AI supports this kind of review workflow. Oversight does not replace testing before deployment. Labeling training data is a separate task, and review does not change how repeatable the model's outputs are.
Question 5 of 10 · Security, Compliance, and Governance for AI Solutions
Which principle limits what an AI service account can access?
Least privilege grants an identity, including an AI service role, only the permissions needed for its task, which limits the damage if it is misused or compromised. Separation of duties splits sensitive tasks among different people. Defense in depth layers several controls, and single sign-on simplifies authentication without limiting what an identity can access.
Question 6 of 10 · Fundamentals of AI and ML
What is machine learning?
Machine learning learns patterns from data rather than following explicitly coded rules.
Question 7 of 10 · Fundamentals of Generative AI
What does zero-shot prompting mean?
Zero-shot prompting gives the model only an instruction and relies on what it learned in pre-training, with no examples. Including worked examples is few-shot prompting. Fine-tuning changes the model's weights rather than the prompt, and asking for explicit reasoning steps is chain-of-thought prompting.
Question 8 of 10 · Applications of Foundation Models
Which factor should guide choosing a foundation model for a use case?
Model selection weighs whether the model can do the task well, what it costs, how fast it responds and whether its license terms permit the intended use. Leaderboard rank, parameter count and release date can inform the choice, but none of them alone shows fit for a specific use case. A smaller or older model is often the better choice when it meets the need at lower cost.
Question 9 of 10 · Guidelines for Responsible AI
What does explainability provide for an AI decision?
Explainability describes why a model produced a given output, for example which features most influenced a prediction; Amazon SageMaker Clarify provides feature attribution for this. An accuracy score shows how often the model is right, not why. An explanation does not prove a decision is unbiased, and a list of training records is data lineage.
Question 10 of 10 · Security, Compliance, and Governance for AI Solutions
What is a key privacy concern when using generative AI?
Users may paste personal or confidential data into prompts, so organizations must know whether a service stores prompts or uses them for training, and must control that data. Token charges are a cost matter. Prompts do not change a model's weights, and output variation comes from sampling settings rather than a privacy weakness.
0 of 10
Fundamentals of AI and ML
Fundamentals of Generative AI
Applications of Foundation Models
Guidelines for Responsible AI
Security, Compliance, and Governance for AI Solutions
Which domains cost you the points? Members see a breakdown by domain and a study plan built from it.