AIF-C01 - Generative AI fundamentals (24% of the exam) - Section 2.1
Explain the basic concepts of generative AI.
Describe tokens, embeddings, prompts, and foundation models, and how a large language model predicts the next token. Distinguish generative models from the discriminative models used for classification.
TokensEmbeddingsFoundation models
Practice question for this objective
lock_openFree sampleGenerative AI fundamentalseasy
An architect is explaining to a stakeholder how a generative foundation model differs from a traditional discriminative classifier that outputs a single label. Which two statements accurately capture these concepts? Select TWO.
AA foundation model is pre-trained on broad data and can be adapted to many downstream tasks rather than just one.check_circle Correct
BA discriminative classifier generates novel samples that resemble its training data instead of labelling inputs.
CA generative model produces new content such as text or images, while a discriminative model assigns inputs to categories.check_circle Correct
DA foundation model is purpose-built for one narrow task and must be retrained from scratch for any other use.
EA generative language model selects its output by returning one fixed class label drawn from a closed list.
Foundation models are broadly pre-trained and reusable across tasks, and generative models create new content whereas discriminative models assign categories. A foundation model gains its breadth from pre-training on large general data, which lets one model serve many tasks. Generative and discriminative models differ in output: a generative model synthesises new content, while a discriminative model maps an input to a category rather than producing fresh samples.
Why A is correct: Broad pre-training and adaptability across tasks are the defining traits of a foundation model, so this statement is correct.
Why B is wrong: Generating novel samples is a generative behaviour; a discriminative classifier predicts a label and does not synthesise new data.
Why C is correct: Generating fresh content versus drawing a boundary between classes is the core distinction between the two model types.
Why D is wrong: This inverts the concept; the value of a foundation model is reuse across tasks, not single-task lock-in requiring retraining.
Why E is wrong: Returning a label from a fixed list describes classification; a generative model builds open-ended output token by token instead.
Answers that look right on this material and are not. Each one is a distractor from a different question in the AIF-C01 bank for this domain.
When the input text is split into tokens and each token is mapped to a numerical ID by the tokeniser.
errorWhy it is wrong: Tokenisation is a pre-processing step that happens before the model runs. The model does not generate tokens until the tokenised input has been processed through all transformer layers.
A generative model returns a fixed probability for one of a predefined set of labels, while a classification model creates entirely new content from scratch.
errorWhy it is wrong: This is tempting because it names both behaviours, but it reverses them; producing a label probability describes classification, and creating new content describes generation.
It is a small model trained on one narrow labelled dataset for a single task, which keeps its behaviour predictable and easy to certify.
errorWhy it is wrong: This describes a conventional task-specific model, the opposite of a foundation model, so although it sounds reassuringly controlled it is incorrect.