Generative AI Concepts and Technologies practice questions

PaLM 2

Why wrong: PaLM 2 is a large language model for text, not image generation.

Gemini Pro Vision

Why wrong: Gemini Pro Vision can process images but is not optimized for generating high-quality images from scratch.

Codey

Why wrong: Codey is specialized for code generation, not images.

Use prompt engineering to instruct the model to ignore outdated information

Why wrong: Prompt engineering cannot inject new knowledge that the model was not trained on.

Increase the model's context window and include all documents in the prompt

Why wrong: Context windows are limited; including all documents is impractical and expensive.

Fine-tune the model weekly on the updated knowledge base

Why wrong: Weekly fine-tuning is costly and slow; knowledge cutoff remains until the next fine-tuning cycle.

Fine-tune Gemini on redacted transcripts and rely on the model to not generate PII

Why wrong: Fine-tuning reduces but does not eliminate PII generation; the model might still produce PII in novel contexts.

Use a smaller model that has never seen PII in training

Why wrong: Smaller models can still generate PII if the training data contained it; also, they may not perform well on the summarization task.

Only use prompt instructions telling the model to avoid PII

Why wrong: Prompt instructions alone are not reliably enforced; the model may still leak PII.

The top-p sampling is too low, limiting the model's vocabulary

Why wrong: Top-p affects token selection, not the recency of knowledge.

The temperature setting is too high, causing creative but incorrect outputs

Why wrong: Temperature affects randomness, not knowledge of recent APIs.

The context window is too short to include relevant API documentation

Why wrong: Even with a longer context, the model cannot generate information it never learned.

PaLM 2

Why wrong: PaLM 2 is text-only; it cannot process images.

Gemini 1.0 Pro

Why wrong: Gemini 1.0 Pro supports text and images but has a smaller context window and is less advanced than 1.5 Pro.

Gemini Nano

Why wrong: Gemini Nano is for on-device applications, not enterprise invoice processing.

Imagen

Why wrong: Imagen is for image generation.

Codey

Why wrong: Codey generates code, not speech.

Gemini

Why wrong: Gemini can process audio but is not primarily a text-to-speech service; Chirp is dedicated to speech generation.

The learning rate is too high, causing the loss to oscillate

Why wrong: Oscillating loss would be visible in the training curves, not just lack of improvement; low rank is a more fundamental limitation.

The LoRA adapter is applied to all layers, which slows convergence

Why wrong: Applying to all layers is common and generally helps convergence, not hinders it.

The base model is too large for the dataset, causing overfitting

Why wrong: Overfitting would show good training performance but poor validation; the scenario says performance is not improving at all.

Set temperature to 0 and top-k to 1

Why wrong: This makes the output deterministic and repetitive, the opposite of creative.

Decrease temperature and increase top-k

Why wrong: Decreasing temperature reduces creativity, making outputs less varied.

Increase the context window length

Why wrong: Context window affects how much preceding text the model considers, not the randomness of token selection.

They improve the model's ability to generate code

Why wrong: Code generation uses specialized models like Codey, not directly related to embeddings.

They reduce the size of the model by compressing weights

Why wrong: Embeddings are for representing data, not compressing model weights.

They allow the model to process images directly

Why wrong: Image processing requires multimodal models; embeddings are for text (or other modalities) but not a direct image processor.

Extract frames using Video Intelligence API and then send them as images to Gemini

Why wrong: This is possible but less efficient than sending the video directly; Gemini 1.5 Pro handles video natively.

Transcribe the video using Chirp, then analyze the text with Gemini

Why wrong: Chirp is for speech generation, not transcription; also, this loses visual information.

Use a custom model fine-tuned on video understanding tasks

Why wrong: Unnecessary; Gemini 1.5 Pro already has strong video understanding capabilities.

Chirp

Why wrong: Chirp generates speech, not video.

Imagen

Why wrong: Imagen generates images, not videos.

Gemini

Why wrong: Gemini can process video but is not primarily a video generation service.

Supervised fine-tuning and RLHF achieve the same results, but RLHF is faster

Why wrong: They have different goals and outcomes; RLHF is typically more time-consuming due to human annotation cycles.

Supervised fine-tuning is better for reducing toxicity because it uses labeled safe examples

Why wrong: Supervised fine-tuning can teach safe outputs but may not generalize well to unseen toxic inputs; RLHF directly optimizes for human preference.

In-context learning is always superior to both for safety, as it can adapt on the fly

Why wrong: In-context learning relies on prompt engineering and can be brittle; it does not permanently change model behavior.

Using a zero-shot prompt

Why wrong: Zero-shot prompting does not incorporate external knowledge and is prone to hallucination.

Using only the largest Gemini model with no retrieval

Why wrong: Even the largest model has a knowledge cutoff and may hallucinate.

Increasing the model's temperature to 1.5

Why wrong: Higher temperature increases randomness, likely increasing hallucinations.

Use prompt engineering to instruct the model not to output personal data

Why wrong: Prompt engineering is not a reliable compliance control; it may still output personal data.

Set temperature to 0 to ensure deterministic outputs

Why wrong: Deterministic outputs do not prevent personal data leakage.

Chirp

Why wrong: Chirp is for speech generation, not multimodal Q&A.

Imagen

Why wrong: Imagen generates images, not answers questions.

Codey

Why wrong: Codey specializes in code generation, not multimodal Q&A.

Fine-tune a base LLM on the policy documents monthly

Why wrong: Fine-tuning is expensive and time-consuming; monthly cycles are impractical and fine-tuned knowledge becomes stale immediately after cutoff.

Use a larger foundation model with a longer context window and paste all documents into each prompt

Why wrong: Pasting all documents into every prompt is expensive, hits context limits for large document sets, and does not scale as the document library grows.

Train a custom model from scratch on the policy documents each month

Why wrong: Training from scratch requires massive compute resources and weeks of time — entirely disproportionate for monthly document updates.

It sets the maximum number of tokens the model can generate in a single response

Why wrong: That is controlled by 'max output tokens', not temperature.

It determines the number of candidate tokens considered at each step

Why wrong: That is controlled by top-k, not temperature.

It adjusts the similarity threshold for vector search retrieval

Why wrong: Temperature is not related to vector search; it is a generation parameter.

Decrease the top-k value to 10

Why wrong: While reducing top-k may have a minor effect on sampling speed, the impact on latency is much smaller than reducing output token count.

Increase the temperature to 1.5

Why wrong: Temperature does not affect latency; it only changes sampling randomness.

Increase the top-p value to 1.0

Why wrong: Top-p does not significantly affect latency; it controls nucleus sampling.

Vertex AI Gemini API + Cloud Key Management Service (KMS) + Secret Manager

Why wrong: KMS and Secret Manager manage encryption keys and secrets, but do not inspect or redact sensitive content in model outputs.

Vertex AI Gemini API + Cloud Data Loss Prevention (DLP) standalone + Cloud NAT

Why wrong: Cloud NAT provides internet access for private instances, not data protection; DLP standalone requires separate integration.

Vertex AI Gemini API + Cloud DLP + Cloud Armor

Why wrong: Cloud Armor is for network security (DDoS, WAF) and does not inspect generative AI outputs for PII.

Gemini 1.5 Pro with image input

Why wrong: Gemini can process images but is not designed for text-to-image generation or inpainting.

Codey on Vertex AI

Why wrong: Codey generates code, not images.

Veo on Vertex AI

Why wrong: Veo generates videos, not still images; it does not offer image inpainting.