AIF-C01 Fundamentals of Generative AI Fundamentals of Generative AI Practice Questions

Generative AI creates new content, and producing photorealistic images from text descriptions for a marketing campaign is content generation, a classic generative AI use case.

Intrusion detection is a security and classification task. Forecasting stock trends from historical data is predictive analytics. Optimizing database indexing is performance tuning. None of these create new content, so they are not generative AI use cases.

Generative pre-trained transformers are large language models trained to generate text, and they can translate a natural-language request into a structured SQL query, which lets non-technical staff retrieve data without writing SQL themselves, exactly as the requirement describes.

A residual neural network (ResNet) is a deep architecture built for image recognition and computer vision, not language generation. A support vector machine (SVM) is a classic supervised algorithm for classification or regression and cannot generate SQL from text. WaveNet is a generative model for producing realistic audio and speech, so it is the wrong modality for converting text into queries.

A generative adversarial network (GAN) is built from two competing neural networks, a generator that creates synthetic samples resembling the real data and a discriminator that tries to tell real from generated, and through this competition it learns to produce realistic synthetic data based on an existing dataset, which is exactly what the company needs.

XGBoost is a gradient-boosting algorithm for supervised classification and regression, not for generating new data. A residual neural network (ResNet) is a deep convolutional architecture used mainly for image recognition and classification. WaveNet is a generative model, but it is specialized for synthesising raw audio such as speech rather than producing general synthetic data from an arbitrary existing dataset.

Inference cost is driven by the number of tokens consumed, because LLM pricing charges for the input and output tokens processed by each request.

The temperature value only shapes output randomness and has no cost impact. The amount of training data and the total training time affect the one-time cost of building the model, not the cost of running inference.