Hard AI Engineer (LLM & Generative AI) questions
32 questions · 20 points each for a correct first answer.
- 1.Injected instructions in a web pageguardrails, security, agentshard
- 2.Why long contexts eat GPU memoryllm-serving, transformershard
- 3.Stuffing 100 chunks into the promptrag, context-managementhard
- 4.What LoRA trainsfine-tuning, lorahard
- 5.Shipping a prompt change safelyevaluation, prompt-designhard
- 6.Agent stuck in a loopagents, reliabilityhard
- 7.Upgrading the embedding modelembeddings, vector-searchhard
- 8.What streaming improveslatency, llm-servinghard
- 9.Sizing the KV cachekv-cache, transformers, deploymenthard
- 10.Why attention divides by sqrt(d_k)attention, transformershard
- 11.What the causal mask preventsattention, transformershard
- 12.Counting LoRA parameterslora, fine-tuninghard
- 13.Where generation time goeslatency, deploymenthard
- 14.Estimating response timelatencyhard
- 15.Temperature 0 is not deterministicsampling, llm-basicshard
- 16.Counting letters in a wordtokenization, tool-callinghard
- 17.Data leaking through image linksprompt-injection, securityhard
- 18.A judge with 90% agreementllm-as-judge, evaluationhard
- 19.Computing nDCG at 3retrieval-metrics, evaluationhard
- 20.Is a 4-point gain realevaluation, statisticshard
- 21.When a tool call failsagents, tool-callinghard
- 22.A suspicious third-party MCP servermcp, prompt-injection, securityhard
- 23.Budgeting the context windowcontext-window, raghard
- 24.How speculative decoding helpsinference, latencyhard
- 25.Continuous batching in servingdeployment, inferencehard
- 26.Serving a mixture-of-experts modelinference, deploymenthard
- 27.How many results to rerankreranking, raghard
- 28.Adding BM25 and cosine scoreshybrid-search, bm25hard
- 29.Fine-tuned model scores too wellfine-tuning, evaluationhard
- 30.Field order in an output schemastructured-output, prompt-designhard
- 31.One failure in Promise.alljavascript, reliabilityhard
- 32.Jailbreaks spread across turnsjailbreaks, guardrailshard