Free AIF-C01 practice questions

AI, ML, and deep learning nest from broadest to narrowest, and AI also includes systems that do not learn.

AI covers any technique that lets a machine mimic human intelligence, including hand-written rule engines. Machine learning is the subset that learns patterns from data, and deep learning is the subset of ML that uses multi-layer neural networks. A tool that only applies fixed rules is therefore AI but not ML. Calling it neither AI nor ML is the tempting mistake: it treats AI as if it meant only learning systems.

1. Ask whether the system learns from data or only follows rules it was given.
2. Recall that AI is the widest circle and includes rule-based systems.
3. Place ML inside AI and deep learning inside ML, so a non-learning system can be AI without being ML.

Exam tip: AI contains ML, ML contains deep learning, and rule-based systems are AI without being ML.

A token is the small unit of text, often a word, part of a word or a punctuation mark, that a language model reads and generates.

Models do not read whole documents or single letters; they read and produce sequences of tokens, and limits such as the context window are counted in tokens. A list of numbers representing meaning is an embedding, a signed credential is an API access token (a security term that only shares the name), and a section split from a long document is a chunk.

1. Notice that the request limit is expressed in tokens, so a token is a unit of text the model processes.
2. Recall that a token is roughly a word or word piece, about four characters of English.
3. Separate it from embeddings (meaning as numbers), chunks (document sections) and API credentials.

Exam tip: A token is the small unit of text a model reads and generates; context windows and bills are counted in tokens.

Retrieval Augmented Generation (RAG) retrieves relevant information from an external source at query time and adds it to the prompt.

RAG works in front of the model: the question is used to search a knowledge source, the most relevant passages are inserted into the prompt, and the model answers from them. Nothing about the model changes. Training the model on documents is fine-tuning, building a smaller imitation model is distillation, and changing sampling settings is an inference parameter, so none of those retrieve anything.

1. Recall that 'retrieval' happens at question time, not training time.
2. Recognise that the retrieved passages travel to the model inside the prompt.
3. Rule out descriptions of fine-tuning, distillation and temperature.

Exam tip: RAG = retrieve relevant passages, add them to the prompt, generate; the model's weights never change.

Veracity (truthfulness) is the responsible-AI feature about factual accuracy of outputs.

A model stating false information describes a truthfulness failure, which maps to veracity. Robustness (reliability under bad inputs), inclusivity (serving diverse people), and fairness (equitable treatment) are separate features that do not describe factual accuracy.

1. Identify the symptom: the model produces false statements presented as fact.
2. Match the symptom to the feature about accurate, factually grounded output.
3. That feature is veracity, also called truthfulness.

Exam tip: Veracity, or truthfulness, is the responsible-AI feature about outputs being factually accurate.

The shared responsibility model splits security into AWS's part (security of the cloud) and the customer's part (security in the cloud); using a managed AI service does not move the customer's data off the customer's side.

The retailer always owns its data: what goes into prompts, how sensitive it is and how it is protected. Patching model hosts, the service's operating system and physical data center security are all AWS's job for a managed service like Amazon Bedrock. The more managed the service, the more layers AWS runs, but data, access and configuration stay with the customer.

1. Place each task on one side: infrastructure and service operation, or customer data and configuration.
2. Hosts, operating systems and buildings belong to AWS for a managed service.
3. The data the customer puts into prompts belongs to the customer.

Exam tip: Managed AI service: AWS runs the infrastructure; you own your data, access and configuration.