Generative AI Interview Questions and Answers | 25 Questions, Basics to System Design

Twenty-five generative AI interview questions, answered the way you would actually answer them out loud. Made by JobScanner, where you can search technical roles read straight from 82,080 company career pages — free, no account. Every question opens with the short version — the two or three sentences you say first, before anyone asks a follow-up. Then the depth, in case they do. And for most of them, what the interviewer is actually checking, which is usually not what the question appears to be about. The order escalates: foundations, then how these models are trained and controlled, then retrieval and agents, then the production questions that separate mid-level from senior, and finally three scenario questions with no single right answer. Nothing here is tied to a specific model version or a benchmark score. Those change faster than you can memorise them, and interviewers know it — what they test is whether you understand the mechanism. So these answers stay correct. If you are short on time, the three worth rehearsing out loud are Q9, Q20 and Q23. CHAPTERS 0:00 Intro 1:01 Q1 — What is generative AI, and how is it different from traditional machine learning? 2:11 Q2 — What is a foundation model, and why does the term exist? 3:18 Q3 — How does a large language model actually produce text? 4:23 Q4 — What is a transformer, and why did attention replace what came before? 5:27 Q5 — What are embeddings, and why does everything seem to use them? 6:22 Q6 — What are diffusion models, and how do they differ from LLMs? 7:28 Q7 — Walk me through how a modern LLM gets trained. 8:41 Q8 — What is RLHF, and what problem does it solve? 9:54 Q9 — Prompt engineering, RAG, or fine-tuning — how do you choose? 10:56 Q10 — What is LoRA, and why is parameter-efficient fine-tuning everywhere? 11:58 Q11 — What is the context window, and what breaks as you fill it? 13:00 Q12 — Explain temperature, top-p and top-k, and when you would change them. 14:19 Q13 — Explain RAG end to end. 15:33 Q14 — How do you chunk documents, and why does it matter so much? 16:51 Q15 — Your RAG system returns irrelevant results. How do you debug it? 18:07 Q16 — What is a vector database, and do you always need one? 19:17 Q17 — What is an AI agent, and when do you actually need one? 20:36 Q18 — How does tool calling actually work under the hood? 21:46 Q19 — What causes hallucination, and how do you reduce it? 23:13 Q20 — How do you evaluate a system with no single correct answer? 24:40 Q21 — What is prompt injection, and how do you defend against it? 26:06 Q22 — How do you control cost and latency in an LLM application? 27:29 Q23 — Design an internal assistant over 200,000 company documents. 29:05 Q24 — It works in the demo and fails for 20% of real users. Diagnose it. 30:24 Q25 — Explain generative AI to a non-technical stakeholder. 32:06 What to rehearse Covered along the way: foundation models, large language models, transformers and attention, embeddings, diffusion models, pretraining and supervised fine-tuning, RLHF and DPO, LoRA and parameter-efficient fine-tuning, context windows, temperature and nucleus sampling, RAG, chunking, reranking, vector databases and hybrid search, agents, tool calling, hallucination, evaluation, prompt injection, cost and latency, and two production scenarios. ───────────── ABOUT JOBSCANNER Most job boards are inventory businesses: employers pay to put a listing in front of you, and the board has no incentive to take it down. We are the opposite shape — nobody can pay to be here, so nothing has to be protected from being removed. · Read from 82,080 company career pages across 57 hiring platforms, never copied from another board · Every listing re-checked every 48 hours and removed when the employer removes it · No promoted listings, no recruiter reposts, no paid placement · Free to search, no account, nothing gated Search at jobscanner.co #GenerativeAI #AIInterview #LLM #RAG #AIEngineer #MachineLearning #InterviewPreparation