FREE AI INTERVIEW QUESTIONS & ANSWERS
Prepare for AI engineer and Forward Deployed Engineer (FDE) interviews with free topic questions, matched source answers, foundation practice, and customer scenarios. All questions and answers are in English, with original source numbers preserved.
Explain the concept aloud, reveal the reference answer, and use the linked lab to try it yourself.
15 questions · Showing 1–10
Specify the task, relevant context, output requirements, and what to do when information is missing. Include an example when it resolves an ambiguity that words alone leave open.
Watch for: Extra instruction length is not itself evidence of better results.
The reader or model can more easily distinguish the requested operation from text to analyze. Named fields or consistent delimiters make that intent explicit.
Watch for: A delimiter does not make malicious source content trustworthy.
No. Use limited tool permissions, validation, separation of trust levels, and checks appropriate to each action, then test failures. Prompt wording is one layer.
Watch for: Treating model compliance as the only security boundary leaves actions exposed.
Use them to demonstrate a tricky format or decision boundary. Include relevant edge cases and compare results with a simpler baseline.
Watch for: Examples with irrelevant details can teach accidental patterns.
Hold the model settings and evaluation cases fixed. Compare correctness, formatting, refusal or abstention behavior, and cost with a predefined acceptance rule.
Watch for: Selecting only successful outputs can hide regressions.
Include fixed instructions, tool definitions, messages, retrieved evidence, and space for output. Verify model-specific handling of other inputs and reasoning with the provider.
Watch for: Counting only the most recent user message understates the request.
When older material contains useful decisions but too much detail to resend. Preserve constraints, unresolved issues, and retrievable source references, then check the summary for omissions.
Watch for: Summarization has a cost and can remove information needed later.
Keep identifiers, user constraints, critical numbers, code that must run, and evidence whose wording matters. Compress surrounding explanation instead.
Watch for: An approximate paraphrase can change a permission or a requirement.
Caching may reuse processing of eligible repeated input; trimming changes which input is sent. Their billing and quality effects differ and must be measured separately.
Watch for: A cached request can still include a large context and paid output.
Narrow the task, retrieve a better subset, process it in stages, or choose a suitable model after checking its limits. Make missing evidence explicit.
Watch for: Silently dropping the exception to a rule may produce a confident wrong answer.
