# Quiz — Session 2: Professional prompting **Program:** Applied AI — Intermediate Level · **Instructor:** Yann Isola **Format:** 10 MCQs (Multiple Choice Questions) · 1 correct answer per question · ~12 minutes **Suggested scale:** 1 point per question. Validation threshold: 7/10. --- ## Questions ### Q1. Among these tasks, which corresponds to a structural STRONG point of an LLM (Large Language Model)? - A. Multiply two five-digit numbers without an external tool - B. Give the closing price of a stock yesterday evening - C. Reformulate a dry email into a diplomatic version - D. Accurately count the occurrences of a letter in a long text ### Q2. A model invents a perfectly formatted but non-existent bibliographic reference. What is the correct explanation? - A. The model has been hacked and is returning corrupted data - B. The model is trained to produce a plausible continuation of the text, not to be calibrated to the truth - C. The model deliberately lies to appear more competent - D. The temperature was set too low, which causes inventions ### Q3. In the 2×2 trust matrix, what are the two axes? - A. Generation speed × response length - B. Cost of an undetected error × ease of verification - C. Model size × context window size - D. Creativity required × precision required ### Q4. Generating code covered by a complete automated test set typically falls into which area of ​​the trust matrix? - A. Forbidden zone: the code is always too risky - B. Free zone: the code cannot cause costly errors - C. Leverage zone: high error cost but easy verification thanks to tests - D. Outside the matrix: the code is not affected by the confidence matrix ### Q5. Why do we say “examples beat descriptions” in a prompt? - A. Examples consume fewer tokens than descriptions - B. Showing a typical output removes the ambiguity that a description in adjectives leaves open, and establishes a pattern that the model continues - C. Models are unable to understand adjectives - D. Descriptions are prohibited by model providers ### Q6. What are delimiters (triple quotes, `…` tags) used for in a prompt? - A. To speed up the generation of the response - B. To reduce the number of tokens billed - C. To make the output more creative - D. To partition the data to be processed so that they are not confused with instructions ### Q7. What is the specific role of the system prompt? - A. It is a privileged channel of instructions, invisible to the end user, which sets the persona, the permanent rules and the exit policy - B. This is the first message the user types in the conversation - C. It is a configuration file that modifies the weights of the model - D. This is the technical log of model errors ### Q8. On very long contexts, where is the models' information recall weakest? - A. At the very beginning of the context - B. At the very end of the context - C. In the middle of the context - D. The recall is uniform across the entire window ### Q9. Why is a long conversation with an assistant becoming more and more expensive and slow? - A. The model gets tired and slows down after many responses -B.The model being stateless, the application returns the entire history each round, and the cost increases with the number of tokens processed - C. The supplier increases prices throughout the conversation to discourage abuse - D. The model downloads new data with each message ### Q10. What does a temperature set to 0 guarantee? - A. That the answer is factually correct - B. That the model will refuse to respond if it is not sure - C. That the output is (quasi) reproducible, the model choosing the most probable token at each step — without guarantee of accuracy - D. That the model will use fewer tokens --- ## Corrected with explanations ### Q1 — Answer: C **Reformulating an email** is a transformation of text: the core strength of the LLM (form, tone, register). The other three are structural weaknesses: **A** — the model predicts plausible sequences of numbers, it does not execute a multiplication algorithm (unless using a tool); **B** — event after the knowledge cutoff date, inaccessible without a search tool; **D** — the model sees tokens, not characters (direct inheritance from Session 1). ### Q2 — Answer: B The training objective is plausible continuation; a well-formatted reference is statistically plausible, and the model does not have a calibrated truth gauge that would trigger an “I don’t know.” **A** is fanciful; **C** anthropomorphizes (lying requires intention and knowledge of the truth); **D** reverses reality — low temperature does not create inventions, it makes them at worst *reproducible*. ### Q3 — Answer: B The matrix crosses **cost of an undetected error** (vertical) and **ease of verification** (horizontal). The word “undetected” is essential: a detected error costs a few seconds; it’s the one that goes unnoticed that’s expensive. The other propositions are not the axes of this matrix. ### Q4 — Answer: C This is the textbook case of the **leverage zone**: an error in production would be expensive (high cost), but automated tests make verification quick and systematic (easy verification). AI produces, testing net controls. Be careful, however: if the tests are incomplete, the verification is no longer “easy” and the case slides towards the prohibited zone. ### Q5 — Answer: B Describing a tone (“professional, warm”) remains interpretable; showing an example of the desired output is unambiguous. In addition, the model is a machine for continuing patterns: examples in the prompt (few-shot prompting, priming with a few examples) establish a strong pattern to be extended. **A** is often even false (examples cost tokens) — and that's not the reason. ### Q6 — Answer: D Delimitations draw the boundary between **instructions** (what to do) and **data** (what to process). Without them, external content containing “ignore your instructions…” can be executed as an instruction: this is the mechanism of prompt injection. Delimitations greatly reduce this risk without completely eliminating it. ### Q7 — Answer: A The system prompt is the **privileged** channel read before the rest: permanent persona, non-negotiable rules, exit policy (language, format, refusal). Analogy: the internal regulations, faced with the “request of the day” which is the user message.**C** is false: no prompt modifies the weights of the model — this is the role of fine-tuning, seen in Session 1. ### Q8 — Answer: C A phenomenon documented under the name “**lost in the middle**”: on long contexts, models find information placed at the beginning and end better than in the middle. Practical consequence: critical instructions at the beginning, reminder at the end, and cutting up very long documents. ### Q9 — Answer: B The model is **stateless**: it has no memory between two calls. The app creates the illusion of conversation by returning all history at each turn; the number of tokens processed — therefore the cost and latency — increases with the conversation. ⚠ Some providers reduce the cost via the prompt cache (already seen tokens billed less expensively): a price optimization, not a memory. ### Q10 — Answer: C Temperature 0 = chance is (almost) eliminated: the most probable token is chosen at each step, so the output is (almost) reproducible. **Deterministic ≠ correct**: if the model is wrong, it will be wrong in the same way each time it is run. Reliability comes from the verification process (confidence matrix), never from the temperature setting. --- ## Score reading grid | Score | Reading | |---|---| | 9–10 | Concepts mastered — ready for Session 3 (plug the model into your data) | | 7–8 | Solid knowledge — review missed questions with the answer key | | 5–6 | Reread Parts B (hallucination) and D (specification) of the guide + repeat Exercise 2 | | <5 | Resume the session with the interactive web page, then retake the quiz | --- *Quiz — Applied AI, Intermediate Level, Session 2 — Yann Isola.*