Level 1 Checkpoint Quiz

25 questions covering all of Level 1. Answer each one, read the explanation, and aim for 80% (20/25) or more before starting Level 2. Reload the page to retry.

Foundations (pages 01–02)

1. Which statement about the relationship between terms is correct?

Machine learning is a subset of LLMs Deep learning and AI are the same thing LLMs are a kind of generative AI built with deep learning Generative AI does not use machine learning

AI ⊃ ML ⊃ DL ⊃ GenAI ⊃ LLMs. See page 01.

2. How are LLMs primarily pre-trained?

Humans label every sentence as correct or incorrect Self-supervised: predicting the next token on huge amounts of text By reinforcement learning from game playing By querying a database of facts

Self-supervised next-token prediction allows training on trillions of tokens without manual labels. Post-training (SFT, preference/RL) comes after.

3. Roughly how many English words fit in 1,000 tokens?

100 250 4,000 750

Rule of thumb: 1 token ≈ 4 characters ≈ ¾ of a word in English. Code often uses more tokens.

4. You want deterministic-ish, focused output for code generation. Which setting?

Low temperature (e.g. 0–0.3) High temperature (e.g. 1.5) Larger context window More stop sequences

Low temperature makes the model pick high-probability tokens. Even at 0, outputs may vary slightly, so always validate.

5. Why do LLMs hallucinate?

They are programmed to lie Their database is corrupted They generate plausible continuations, not verified facts Only small models hallucinate

Plausible text is not necessarily true text. Ground with docs and tools, and verify.

6. Your chat assistant "remembers" your preference for Scala 3 in a new chat. What is most likely happening?

The model's weights were updated from your chats The product saved a memory and inserted it into the new prompt The context window persists forever Tokens are cached in your browser

Models are stateless between calls. "Memory" is context engineering by the product.

Prompting & context (pages 03–04)

7. Which is NOT one of the six prompt building blocks?

Constraints Output format Examples Threats or rewards

Role, task, context, constraints, format, examples. Calm, precise instructions beat emotional pressure.

8. Giving the model 3 input→output examples to copy a format is called…

Few-shot prompting Fine-tuning RAG Zero-shot prompting

Few-shot examples steer format and style without changing model weights.

9. A long chat keeps repeating the same wrong fix. Best move?

Type "try again" until it works Use ALL CAPS Start a fresh session with a better prompt that includes what you learned Paste the entire repository

The polluted context keeps steering the model wrong. A clean context plus the lessons learned usually fixes it.

10. Which are the four moves of context engineering?

Read, write, delete, repeat Write, select, compress, isolate Prompt, sample, decode, stop Plan, code, test, deploy

Write durable context, select relevant context, compress long context, isolate unrelated work. See page 04.

11. Why is a Scala type signature valuable context?

It reduces the price per token Models ignore prose It disables hallucinations It carries precise meaning in few tokens, and the compiler can check the result

Types are dense, precise specs, and the compiler provides immediate feedback.

Tools, IDEs & customisation (pages 05–07)

12. Which IDE mode suits "add an endpoint with service, repository and tests, and make the build pass"?

Agent mode Inline completion Ask mode only Spell-check

Multi-file changes plus running the build and tests is agent work. Plan first.

13. What do VS Code, JetBrains IDEs and Xcode increasingly have in common?

They all only support Swift They all ban external agents They connect agents to IDE capabilities through open protocols like MCP and ACP They require local models only

For example, Xcode exposes build/test/docs tools via MCP, and Metals exposes Scala compiler tools via MCP.

14. What belongs in AGENTS.md?

API keys so the agent can deploy Exact build/test commands, versions, conventions and boundaries A 2,000-line tutorial on Scala "Write clean code" and nothing else

Keep it concrete and short. No secrets. Move long procedures into skills.

15. In Agent Skills, what decides whether a skill gets loaded automatically?

The file size The folder's creation date The number of scripts The description in the SKILL.md front-matter

With progressive disclosure, only name + description are visible up front, so the description must say what the skill does and when to use it.

16. You need code to be formatted after every agent edit, no exceptions. Use…

A hook A polite request in the chat A higher temperature A custom persona

Instructions are requests. Hooks are deterministic guarantees.

Agents & MCP (pages 08–09)

17. In tool calling, who actually executes the tool?

The model, inside its weights The user, by hand, always The host application, after any permission checks The tokenizer

The model emits a structured call. The host executes it and returns the result into context. The host is the security boundary.

18. An agent "fixes" a failing test by changing the expected value. This is…

Ideal behaviour Test gaming, a classic agent failure mode Prompt injection Spaced repetition

Instruct "don't modify tests without asking" and review test diffs carefully.

19. Which problem does MCP mainly solve?

Making models smaller Training models faster Replacing HTTP The N×M problem of connecting many AI apps to many tools

One server works with every MCP-capable host: N+M instead of N×M.

20. Which three are MCP server primitives?

Tools, resources, prompts Sampling, elicitation, roots Tokens, embeddings, attention Hooks, skills, agents

Sampling, elicitation and roots are client features that servers can request.

21. A local MCP server launched by your IDE as a child process typically uses which transport?

SMTP WebRTC stdio FTP

stdio for local processes. Streamable HTTP for remote servers.

Scala, Python, IT, learning, safety (pages 10–14)

22. Which is a common AI mistake in Scala code?

Using exhaustive pattern matching on enums Mixing Scala 2 implicit syntax with Scala 3 given, and calling .get on Option Using opaque types for IDs Returning Either for recoverable errors

State the Scala version and conventions in AGENTS.md, and keep the compiler in the loop.

23. An AI suggests installing a Python package you have never heard of. What should you do first?

Install it globally with sudo Trust it, since AI checked it Rename it to something shorter Verify it exists, is maintained and is the intended package (slopsquatting risk)

Attackers register package names that models commonly hallucinate.

24. Which learning approach builds real skill?

Explain it in your own words, get critiqued, get quizzed, then review later Copy AI answers and re-read them Let AI do every exercise Avoid any difficulty

Active recall, Feynman, deliberate practice and spacing. See page 13.

25. Which combination makes prompt-injection data theft most likely?

Low temperature + short prompts + local model Read-only tools + no network + trusted files Access to private data + untrusted content + ability to send data externally Using AGENTS.md

The "lethal trifecta". Remove at least one leg, and keep approvals on for outbound or destructive tools.

What next?

Scored 20+? You are ready for Level 2 · Practitioner. Below 20? Revisit the linked pages, redo the related labs, and retake the quiz in a few days. Spacing helps.

16 · Level 1 Checkpoint Quiz

Foundations (pages 01–02)

Prompting & context (pages 03–04)

Tools, IDEs & customisation (pages 05–07)

Agents & MCP (pages 08–09)

Scala, Python, IT, learning, safety (pages 10–14)

What next?