Level 4 Checkpoint Quiz

20 questions covering all of Level 4. Aim for 16/20 (80%) or more, and finish the capstone, before Level 5. Reload to retry; answer order is shuffled each time.

Maths & internals (pages 01–02)

1. The dot product of [1,2,3] and [4,5,6] is…

[4,10,18] 32 12 [5,7,9]

1·4 + 2·5 + 3·6 = 32. See page 01.

2. Softmax turns logits into…

Integers Logits in a wider range A probability distribution (all positive, summing to 1) Binary 0/1 values

Exponentials make them positive and amplify the top score.

3. Cross-entropy loss heavily punishes…

Confident wrong answers (low p on the correct token) Confident right answers Long outputs Large vocabularies

-log(p): p=0.01 → loss 4.6; p=1 → 0.

4. In attention, Q·K computes…

The final answer Positional encodings The loss How relevant each key is to each query

Then softmax → weights → weighted sum of values. See page 02.

5. The causal mask exists so…

The model runs faster Heads don't collide A token can't attend to future tokens, keeping next-token prediction honest Attention uses less memory

Without it the model "cheats" by looking ahead.

6. In a transformer block, the feed-forward network…

Mixes information across tokens Processes each token independently (most parameters live here) Computes attention scores Does nothing important

Attention mixes; the MLP transforms per-token.

Training & inference (pages 03–04)

7. A "base model" is…

A next-token text completer before post-training The smallest model size A model with no weights A model trained only on code

Post-training (SFT + preferences) makes it an assistant. See page 03.

8. DPO vs RLHF: DPO is…

Always better quality More expensive Needs a reward model Simpler — optimises on preference pairs without a separate reward model

Fewer moving parts; usually the default now.

9. RLVR produces reasoning because…

It reads reasoning textbooks The prompts say "think step by step" Rewarding verifiable answers makes long reasoning chains improve reward — self-checking emerges It uses a bigger model

Verifiable rewards (maths, tests) let reasoning be learned, not scripted.

10. The KV cache…

Caches model weights on disk Stores past tokens' keys and values so generation doesn't recompute them — and grows with context × users Is a type of quantisation Only matters for training

See page 04.

11. Decode (generation) is bottlenecked by…

Memory bandwidth (reading weights per token) Disk speed Network latency only The tokenizer

That's why bandwidth and quantisation dominate generation speed.

12. Speculative decoding speeds up generation by…

Using a bigger model Skipping the softmax Caching the prompt A small draft model proposes tokens; the big model verifies them in one parallel pass

Same output, ~2–3× faster.

Reasoning & architecture (pages 05–06)

13. Test-time compute means…

A bigger training set More GPUs Letting the model generate more reasoning tokens per answer to get better results Faster hardware

A second scaling axis beyond model size. See page 05.

14. Reasoning models help most on…

Tasks with a checkable right answer (maths, code, logic) Quick lookups and classification Casual chat Everything equally

Elsewhere the extra thinking buys little and costs a lot.

15. Hidden "thinking" tokens…

Are free Are billed as output tokens and can dwarf the visible answer Are always faithful reasoning Don't affect latency

Monitor them in production.

16. An "AI gateway" is…

A faster model A brand of GPU A firewall A single layer all LLM calls pass through: routing, budgets, guardrails, caching, tracing, fallback

See page 06.

17. The strongest reason to keep models swappable behind the gateway is…

It's tidier Providers like it Build-vs-buy drifts: models improve, prices change, a second provider avoids lock-in and single points of failure It saves memory

Vendor risk is real; make a swap a config change.

Governance (page 07)

18. Under the EU AI Act, a customer-support chatbot is most likely…

Unacceptable (banned) Limited risk — transparency duty (disclose it's AI) High risk Unregulated entirely

High risk covers hiring, credit, critical decisions — not a normal chatbot. See page 07.

19. Model risk management starts with…

An inventory/register of every model feature in production Buying insurance A bigger model More prompts

You can't govern what you don't list.

20. The most effective governance approach is…

Committee approval for every prompt change Banning all AI tools Doing nothing Paved paths: make the compliant option (approved models, evals in CI, templates) the easiest option

Governance that blocks work pushes people to shadow AI.

What next?

Level 4 complete? Scored 16+ and finished the capstone? You're ready for Level 5 · Master & Specialist: reading research, interpretability, on-device AI, multimodal systems, and leading AI adoption.

L4 · 10 · Level 4 Checkpoint Quiz

Maths & internals (pages 01–02)

Training & inference (pages 03–04)

Reasoning & architecture (pages 05–06)

Governance (page 07)

What next?