The Coding Gopher Courses

1.Your team wants to ship a new prompt because '5 sample outputs look great'. What should gate the release?

A.A quick read of 20 more outputsB.An automated evaluation on a golden dataset, with deterministic checks, a validated model judge, and blocking thresholdsC.Whether the new prompt is shorterD.Turning temperature to zero

2.You switch to a newer embedding model. What happens to your existing vectors?

A.Normalize them againB.They still work; vectors are model-independentC.Every document must be re-embedded, ideally into a new index built in the background, then switched overD.Only the similarity cutoff needs changing

3.With structured outputs (constrained decoding), what is guaranteed?

A.Nothing: it's the same as asking for JSON in the promptB.The values are correctC.The output matches your schema (unless it's cut off or the model refuses)D.The model will never invent data

4.Some HR documents are managers-only. How should a RAG assistant protect them?

A.Retrieve them but instruct the model not to reveal themB.Filter by the user's identity at search time so restricted chunks are never retrievedC.Fine-tune a separate model for managersD.Put them in a separate system prompt

5.An agent's 10-step task has 95% reliability per step. Roughly how often does the whole task succeed?

A.About 60%B.About 90%C.About 99%D.95%

6.When is fine-tuning the right tool?

A.As the first step for any new featureB.To enforce who can see which dataC.When the model has the information but you need consistent behavior on a stable, high-volume task, or a smaller cheaper modelD.To teach the model facts that change every year

7.Why is generating each token (decode) usually limited by memory bandwidth rather than compute?

A.Because every token requires reading all the model's weights from GPU memoryB.Because the KV cache is recomputed every stepC.Because prompts are longD.Because of network latency

8.Prompt injection: what's the realistic defense?

A.A better system prompt that forbids following injected instructionsB.Assume it will sometimes succeed and limit what it can do: session-scoped data, narrow tools, limits in code, untrusted output, confirmationsC.Use a bigger modelD.Block any message containing the word 'ignore'

Where are you on the junior → senior path in AI engineering?