Fundamentals of gen AI practice questions

A large language model (LLM) is a foundation model trained on vast text data to understand, process, and generate natural language.

Option (a) gives the standard definition: massive text training plus language understanding and generation. Option (b) understates LLMs as narrow bilingual translators. Option (c) confuses language modeling with deterministic numeric computation. Option (d) describes a text-to-image model like Imagen, not a language model — a common mix-up between model modalities.

1. Note the training data: enormous volumes of text.
2. Note the capability: understanding and generating human language across many tasks.
3. Distinguish from image/video models (Imagen, Veo), whose outputs are visual rather than textual.

Exam tip: LLM = foundation model trained on massive text data for understanding and generating language.

Multimodality is a model's ability to understand and combine multiple types of input and output — text, images, audio, video — in one model. Gemini is Google's flagship multimodal foundation model.

Option (d) is the exact term for handling multiple data modalities together. Option (a) is an adaptation technique, orthogonal to modality. Option (b) is a learning paradigm, not an input capability. Option (c) is a lifecycle stage for moving data into pipelines. Only 'multimodality' names the described capability.

1. Spot the clue: multiple input types (text, images, audio) handled together.
2. Match the term: 'multi' (many) + 'modal' (types of data) = multimodal.
3. Connect to Google Cloud: Gemini is natively multimodal, a key selection factor when a use case involves mixed media.

Exam tip: Multimodal models process multiple data types in one model — Gemini is the canonical Google example.

Diffusion models generate content (especially images) by learning to reverse a progressive noise-adding process: they start from random noise and denoise step by step until a coherent image emerges.

Option (a) matches the described mechanism exactly — the noising/denoising process is the defining trait of diffusion models, which power modern text-to-image systems like Imagen. Options (b) and (c) are classical predictive ML methods for structured data, not generative techniques. Option (d) is not machine learning at all.

1. Recall the training idea: gradually add noise to real images so the model learns the reverse (denoising) path.
2. Recall generation: start from pure noise and iteratively denoise, guided by the text prompt.
3. Associate the technique with image and video generation models such as Imagen and Veo.

Exam tip: Diffusion models create images by learning to reverse a noising process — the technology behind text-to-image generation.

Prompt engineering is the practice of designing effective natural-language prompts (instructions, context, examples) to get better outputs with no training. Prompt tuning is a parameter-efficient adaptation technique that learns a small 'soft prompt' from training examples while leaving the foundation model's weights frozen.

Option (a) correctly separates the two: wording craft versus learned parameters. Option (b) reverses them and confuses prompt engineering with full fine-tuning, which does retrain model weights. Option (c) is wrong because one requires no data or training while the other does. Option (d) invents a modality restriction that exists for neither technique.

1. Ask: does the technique involve training? Prompt engineering — no; prompt tuning — yes (a small learned component).
2. Ask: are model weights changed? Neither changes the base model's weights; full fine-tuning does.
3. For business framing: prompt engineering is the cheapest, fastest customization; prompt tuning needs example data but is far lighter than fine-tuning.

Exam tip: Prompt engineering = writing better prompts (no training); prompt tuning = learning a small soft prompt from data while the model stays frozen.

The three classic ML approaches: supervised learning uses labeled examples (input → known output), unsupervised learning finds patterns in unlabeled data, and reinforcement learning learns behavior from rewards through trial and error.

The scenario's giveaway is the label attached to every record ('repaid or not'). Learning from input-label pairs is supervised learning, so (a) is right. Option (b) fails because unsupervised learning requires no labels. Option (c) fails because there is no agent interacting with an environment for rewards. Option (d) is not a training approach at all.

1. Check the data: does each example carry a known outcome label? Yes → supervised.
2. If the data had no labels and the goal were to discover groupings, it would be unsupervised.
3. If the system learned by receiving rewards for actions over time, it would be reinforcement learning.

Exam tip: Labeled outcomes in the training data signal supervised learning.

Unsupervised learning finds patterns and structure in data that has no labels — clustering, anomaly detection, and dimensionality reduction are typical applications.

The scenario explicitly says there are no predefined categories, ruling out supervised learning (a), which requires labeled targets. There is no reward signal or sequential decision-making, ruling out reinforcement learning (b). Prompt tuning (c) is a model-adaptation technique, not a data-analysis approach. Clustering unlabeled behavioral data is exactly what unsupervised learning (d) does.

1. Note the absence of labels or predefined categories in the data.
2. Note the goal: discover natural groupings the business has not defined in advance.
3. Match: no labels + pattern discovery = unsupervised learning (e.g., clustering).

Exam tip: No labels plus a goal of discovering structure means unsupervised learning.