SKILLUMEN
Try for free

AI Interview Questions

The fundamentals

3 topics · 18 questions

Attention

 01In plain terms, what does self-attention let each token do? 02Walk through the three steps, scoring, softmax, and mixing, that produce a token's new vector. 03What are the Query, Key, and Value roles, and how are they made from a token's vector?

Embeddings

 01What is an embedding, and why is a raw token ID not enough? 02What does 'similar meanings sit close together' actually mean, and how do you measure it? 03Explain the king − man + woman ≈ queen example.

Tokenization

 01What does tokenization do to your text, and why is it necessary? 02Why split into word-pieces rather than whole words or single letters, and what does BPE buy you? 03Cost and length limits are counted in tokens. What surprising consequence does that have for a long word?

RAG and retrieval

4 topics · 24 questions

Advanced RAG

 01What does basic RAG get wrong often enough that advanced RAG exists to fix? 02How do query rewriting, multi-query, and HyDE each improve the search before retrieval happens? 03Why does HyDE search with a made-up answer instead of the user's raw question?

Hallucination

 01What exactly is a hallucination, and what makes it dangerous to a reader? 02Why does a model trained to predict plausible next tokens produce confident falsehoods when it lacks knowledge? 03Which mitigations reduce hallucination, and why is grounding with RAG the biggest lever?

RAG

 01What problem does RAG solve about a model's knowledge, and how does it reduce hallucination? 02Walk through the RAG pipeline from documents at rest to an answer, naming the retriever's role. 03Why does RAG let a model use private or very recent information it was never trained on?

RAG vs Fine-tune vs Long-Context

 01What are the three ways to give a model knowledge, and the one-line pitch for each? 02When is RAG the right choice, and what properties of the knowledge push you toward it? 03Why is fine-tuning for behaviour rather than facts, and how do you pair it with RAG?

Agents

4 topics · 24 questions

Agent Evaluation

 01What's the difference between agent observability and agent evaluation, and which do most teams skip? 02How does trajectory evaluation score an agent, and what does it look at beyond the final answer? 03Why does a dashboard that logs every step still leave you exposed, and what does an eval gate catch that logging can't?

Agents & Tool Calling

 01What is the repeating loop an agent runs, and when does it stop? 02In tool calling, what does the model emit, who runs the function, and how does the result get back? 03What do the Thought, Action, and Observation steps of ReAct each correspond to?

Model Context Protocol (MCP)

 01What problem does MCP solve? 02What are the two roles in MCP and how do they talk? 03What three things does an MCP server expose?

Multi-Agent Orchestration

 01What's the core idea of multi-agent orchestration versus one agent doing everything? 02What roles do a planner, workers, and an orchestrator play? 03Distinguish the planner-executor, supervisor/router, and debate/critique patterns.

Tuning and serving

3 topics · 18 questions

Fine-tuning

 01Name the three stages of building a model and what each one teaches. 02In RLHF, how do human rankings get turned into something that tunes the model, and what algorithm is often used? 03What behaviour does SFT add that pretraining alone leaves missing?

KV-Aware Serving

 01How does the KV cache shape production serving, and what two generation phases does it split into? 02How do prefill and decode differ in what they bottleneck on, and why does mixing them on one GPU cause problems? 03What problem does disaggregation solve by running prefill and decode on separate GPU pools linked by a fast connection?

Quantization

 01What does quantization change about how a model's numbers are stored, and what do you trade for it? 02Going from fp16 to int4, roughly how much memory does a 70B model save, and why does that matter? 03Why does using fewer bits hurt quality, and how do GPTQ and AWQ keep the damage tiny?
600+ more interview questions live inside each concept card in Skillumen
Practice them free