How to Hire a Generative AI Engineer

• aUse larger parent chunks or parent-child retrieval with overlap so matched context includes neighboring textCorrect
• bSwitch to a higher-dimensional embedding model
• cIncrease top-k from 5 to 50 without changing the chunking
• dLower the similarity threshold so more distant chunks are included

Explanation: Small chunks retrieve precisely but strip context; parent-child retrieval or overlapping larger chunks restores surrounding conditions. Bigger embeddings add no missing text, and flooding top-k with tiny chunks adds noise and cost rather than coherent context.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the mid-level experience tier.

• aAsk the judge to be objective and unbiased in the prompt
• bAlways show the longer answer second to cancel the bias
• cUse a rubric with explicit criteria, randomize answer order, and calibrate the judge against human-labeled pairsCorrect
• dReplace the judge with exact string matching against reference answers

Explanation: Position and verbosity biases are well documented in LLM judges; rubrics, order randomization, and human calibration measurably reduce them. Politeness instructions barely help, systematic ordering just relocates the bias, and exact matching fails on open-ended answers.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the mid-level experience tier.

• aIncrease temperature so tokens generate faster
• bAdd few-shot examples to improve answer quality
• cSwitch from streaming to returning the full answer at once
• dShorten and cache the shared prompt prefix and stream tokens as they generateCorrect

Explanation: TTFT is dominated by prefill over input tokens; shrinking and caching the stable prefix reduces prefill work, and streaming shows tokens as they arrive. Temperature does not affect speed, few-shot examples lengthen prefill, and non-streaming worsens perceived latency.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the mid-level experience tier.

• aDiffusion models need no training data
• bDiffusion training is more stable and covers more of the data distribution, while GANs are prone to mode collapseCorrect
• cDiffusion models generate images in a single forward pass
• dGANs fundamentally cannot produce images larger than 256 pixels

Explanation: Diffusion's iterative denoising objective trains stably and yields better diversity and coverage than adversarial training, which suffers mode collapse. Diffusion actually needs many sampling steps and plenty of data, and GAN resolution limits were engineering hurdles, not fundamentals.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the mid-level experience tier.

• aFine-tune on the catalog hourly and prompt for style
• bRetrieve the style guide at query time and fine-tune on the catalog
• cFine-tune or use strong style examples for voice and format, and use retrieval for the hourly-changing catalogCorrect
• dPut the full catalog and style guide in the system prompt

Explanation: Fine-tuning is good at durable behavior and format; retrieval is good at fresh, changing facts. Reversing them means hourly fine-tunes, which is operationally unworkable, and a giant system prompt is costly, stale between edits, and crowds the context window.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the mid-level experience tier.

• aThe LLM silently changed its system prompt
• bNew documents were embedded with the new model while old vectors remain from the old model, mixing incompatible vector spaces in one indexCorrect
• cThe vector database is compressing old vectors to save space
• dUsers suddenly started asking worse questions

Explanation: Different embedding models produce incomparable vector spaces; mixing old and new embeddings corrupts similarity ranking and degrades gradually as new content accrues. The fix is re-embedding the full corpus or versioning namespaces per model. The other options do not produce this gradual asymmetric pattern.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the senior experience tier.

• aRaise the temperature so the model explores the middle more
• bDuplicate the entire context twice so the clause appears more often
• cSwitch to the cheapest model so you can afford an even longer window
• dRetrieve and rerank so only the most relevant chunks are included, placing critical content near the start or end of the contextCorrect

Explanation: Models show lost-in-the-middle recall: attention is strongest at the beginning and end of long contexts. Tighter retrieval, reranking, and position-aware ordering beat brute-force stuffing. Temperature does not change attention patterns, and duplication wastes tokens and can confuse the model.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the senior experience tier.

• aA small draft model proposes tokens that the target model verifies; accepted tokens match the target model's own distribution, so quality is unchanged while latency dropsCorrect
• bIt approximates the target model, so quality drops slightly but speed doubles
• cIt caches previous answers, so repeated questions are instant
• dIt quantizes the model to 4 bits, trading a little quality for speed

Explanation: Speculative decoding accepts draft tokens only when they pass verification against the target model's probabilities, preserving the output distribution (identical outputs under greedy decoding). It is not approximation, answer caching, or quantization.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the senior experience tier.

• aThe base model's tokenizer changed during RLHF
• bRLHF always increases honesty as a side effect
• cThe reward model is an imperfect proxy, so the policy exploits features humans rate well (length, agreement) instead of true helpfulnessCorrect
• dThe KL penalty was set too high, freezing the policy

Explanation: Reward models capture correlates of human preference, and policies Goodhart them: longer, agreeable answers score higher without being more correct, a classic reward-hacking pattern. A high KL penalty would prevent drift, not cause it, and tokenization is untouched by RLHF.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the senior experience tier.

• aThe model's weights get larger with more users; shard the weights
• bKV cache memory grows with concurrent sequences times context length; use continuous batching with paged KV cache, cap context lengths, and right-size the max batchCorrect
• cTokenization becomes the bottleneck; move it to the GPU
• dNetwork bandwidth between GPUs saturates; add NVLink

Explanation: Autoregressive serving stores per-token keys and values for every active sequence, so memory scales with concurrency times prompt length; paged attention, continuous batching, and context caps are the standard remedies. Weights are constant regardless of traffic, and tokenization is rarely the bottleneck.

Why it matters: Evaluates practical understanding of Generative AI concepts and problem solving for Generative AI Engineer at the senior experience tier.