Choosing an LLM surface: chat, search, code, and voice

10 questions, one correct answer each. The goal is to check whether you can predict what a surface does to the context window and which verification it demands.

1. What does "choosing a surface" actually decide?

A) Which model architecture answers the question
B) What content occupies the context window
C) How many parameters the model has
D) The tokenizer used for the prompt

Show answer

B — The same underlying model is often reachable through several surfaces. What changes is the material injected into the window: retrieved passages, repository files, or an uploaded corpus.

2. A plain chat turn with no tools answers a question about last week's news. What is the dominant risk?

A) The context window overflows
B) Stale or fabricated recall, because the answer relies on training data with a cutoff
C) The tokenizer mis-splits the words
D) The answer costs too much to generate

Show answer

B — Without retrieval the model can only draw on parameters fixed at training time, so recent events are either missing or confabulated.

3. Enabling a retrieval tool changes the failure mode to…

A) No failure mode; retrieval makes answers correct
B) Source quality, passage selection, and unfaithful summarization
C) Slower tokenization only
D) Loss of the knowledge cutoff

Show answer

B — Retrieval reduces staleness but introduces new dependencies: which sources were found, which passages were selected, and whether the summary is faithful to them.

4. What is the minimum acceptable way to verify a retrieval-backed factual claim?

A) Read the snippet quoted in the answer
B) Check that at least three sources are cited
C) Open the cited source and confirm the specific claim appears there
D) Ask the model whether it is confident

Show answer

C — The snippet and the citation list are both generated text. Only opening the source turns a citation into evidence.

5. Why prefer a code-executing surface for arithmetic over a plain chat turn?

A) It uses a larger model
B) The computation runs in an interpreter and leaves an inspectable program
C) It never makes mistakes
D) It is always cheaper

Show answer

B — The arithmetic leaves the token stream and runs deterministically, and the emitted program becomes the artifact a human can review.

6. A code-executing surface prints a plausible average. What should you review first?

A) The formatting of the number
B) The emitted program: which rows it selected and which it silently dropped
C) The model version
D) The response latency

Show answer

B — A correct-looking number can come from wrong code — a bad filter, a dropped null, a wrong column. The program is the evidence.

7. What is the operating cost of pasting a very long document into the context window?

A) None; context is free once the model is loaded
B) Higher price per turn and slower decoding, since attention and decoding scale with tokens held in the window
C) Only a one-off upload fee
D) It reduces cost by avoiding retrieval

Show answer

B — Attention and decoding both scale with the tokens held in the window, so a long context is metered on every turn it stays in.

8. A colleague says "extrapolate this trend to 2030" and the surface returns a confident figure. What is the honest reading?

A) The figure is validated by the computation
B) The computation is exact, so the extrapolation is exact
C) The arithmetic may be exact while the extrapolation beyond observed data remains an assumption
D) Extrapolation is impossible for a language model

Show answer

C — Exact arithmetic on the observed range says nothing about whether the trend continues. The assumption sits in the model of the world, not in the calculation.

9. Persistent memory and custom instructions are best understood as…

A) A pure convenience feature
B) A convenience feature and a privacy surface that should be reviewed
C) A way to reduce token cost
D) A replacement for retrieval

Show answer

B — They silently carry state across sessions, which is convenient but also accumulates personal information that deserves periodic review.

10. Which selection rule best summarises the session?

A) Always use the most capable surface available
B) Always enable every tool, since more context is always better
C) Choose the cheapest surface that can answer correctly, then verify what that surface makes verifiable
D) Choose the surface with the largest context window

Show answer

C — Capability and cost are traded per question. The discipline is matching the surface to the question, then performing the verification that surface demands.

Scoring

8-10 correct: you can select a surface and justify the verification it requires. 5-7: revisit the failure-mode section; you likely still treat retrieval as a correctness guarantee. Below 5: re-run Lab A and Lab B before moving on, focusing on what actually entered the context window.

Learner course · Go to the exercises