AI and LLMs Junior interview questions and answers

75 questions for a AI and LLMs (Junior) interview - with correct answers, explanations and links to the docs.

Take this quiz
  1. 1.What does a large language model (LLM) fundamentally do?
  2. 2.What is a token in the context of LLMs?
  3. 3.Why does the same text in Ukrainian often use more tokens than in English?
  4. 4.What is the context window of a model?
  5. 5.What does the temperature parameter control?
  6. 6.Which temperature suits extracting fields from an invoice into JSON?
  7. 7.What is an LLM hallucination?
  8. 8.What is a model's knowledge cutoff?
  9. 9.What is a system prompt?
  10. 10.Chat APIs are usually stateless. What does that mean for an app?
  11. 11.What is few-shot prompting?
  12. 12.Why ask a model to think step by step before answering hard problems?
  13. 13.How are LLM API calls usually priced?
  14. 14.Why do chat interfaces often stream the answer token by token?
  15. 15.What does it mean when a response stopped because it hit the max tokens limit?
  16. 16.What is an embedding?
  17. 17.How does semantic search with embeddings differ from keyword search?
  18. 18.What is retrieval-augmented generation (RAG)?
  19. 19.Why are documents split into chunks for RAG?
  20. 20.When is RAG usually a better choice than fine-tuning?
  21. 21.What is fine-tuning?
  22. 22.What is tool use (function calling)?
  23. 23.Who actually executes a tool the model asks to call?
  24. 24.What is an AI agent in simple terms?
  25. 25.What is the Model Context Protocol (MCP)?
  26. 26.Why ask for structured output like JSON from a model?
  27. 27.Where should the API key for an LLM provider live in a web app?
  28. 28.What should an app do when the LLM API returns 429?
  29. 29.What is prompt injection?
  30. 30.Why is indirect prompt injection dangerous?
  31. 31.What is a safe habit when using public AI chat tools at work?
  32. 32.Why must code from an AI assistant be reviewed?
  33. 33.Why is testing an LLM feature harder than testing normal code?
  34. 34.What is an eval for an LLM feature?
  35. 35.Which check is reliable for an LLM that must return a category from a fixed list?
  36. 36.What is a golden dataset in LLM testing?
  37. 37.What is LLM-as-a-judge?
  38. 38.Why version prompts like code?
  39. 39.Does temperature 0 guarantee identical answers every time?
  40. 40.What does top-p (nucleus sampling) do?
  41. 41.What is the role of the transformer's attention mechanism?
  42. 42.What is the difference between pre-training and instruction tuning?
  43. 43.What is RLHF?
  44. 44.What is a multimodal model?
  45. 45.What is a typical trade-off between a large and a small model?
  46. 46.Why put clear output instructions, like "answer in under 50 words", in prompts?
  47. 47.What is a guardrail in an LLM application?
  48. 48.Why should an AI feature show sources when it answers from documents?
  49. 49.What happens when a conversation grows longer than the context window?
  50. 50.What does prompt caching help with?
  51. 51.Why can very long prompts with many documents give worse answers?
  52. 52.What is bias in AI outputs?
  53. 53.Why should users be told when they're talking to an AI?
  54. 54.What is a reasoning or thinking model?
  55. 55.What should an AI chat do when it doesn't know the answer?
  56. 56.Why log prompts and outputs of an AI feature in production, with care?
  57. 57.What does it mean that LLM output is non-deterministic?
  58. 58.What is a simple automated check for a chatbot answer?
  59. 59.What is the role of a vector database in RAG?
  60. 60.Why might an AI feature refuse a harmless request?
  61. 61.What is red teaming for an AI application?
  62. 62.What is a jailbreak?
  63. 63.Why should an AI agent's tools follow least privilege?
  64. 64.What is "human in the loop" for AI features?
  65. 65.What does the AI SDK, e.g. Vercel's, help a web developer with?
  66. 66.Why should prompts separate instructions from user data, e.g. with tags?
  67. 67.How can you reduce hallucinations in an answer about company policy?
  68. 68.What does "model drift" mean for an AI feature?
  69. 69.What is the main risk of trusting AI-generated SQL without checks?
  70. 70.Why are evals run on many examples instead of one?
  71. 71.What is a chatbot's "context" made of in a typical app?
  72. 72.What does "grounded" mean for an AI answer?
  73. 73.Why do long answers take noticeably longer than long prompts?
  74. 74.Why give a model a role, like "You are a support agent for X"?
  75. 75.Which data should never go to a third-party LLM without approval?