Level 5 Checkpoint Quiz

15 questions covering Level 5. Aim for 12/15 (80%). This is the last checkpoint — the real final exam is your published capstone. Reload to retry; answer order is shuffled each time.

Research & interpretability (pages 01–02)

1. The first pass of a paper should take about…

An hour 5–10 minutes — title, abstract, headings, figures, conclusion A full day As long as it takes to read every word

Most papers deserve only pass 1. See page 01.

2. A paper reports "+3% over baseline". Before crediting it, check…

Nothing — published means proven Only that the code exists The baseline's strength, whether the benchmark leaked, and whether error bars overlap The authors' reputation

Weak baselines, contamination and noise all fake "improvements".

3. An ablation tells you…

Which parts of a method actually cause the improvement, by removing them one at a time How fast the model runs The model's size The training loss

A result that survives ablations is robust; one needing all tricks is fragile.

4. "Superposition" in interpretability means…

Models are quantum Each neuron stores exactly one concept Models have too many parameters Concepts are stored as overlapping directions, not one-per-neuron

That's why there's no neat "the X neuron". See page 02.

5. A sparse autoencoder (SAE) is used to…

Compress the training data Expand dense activations into many interpretable features Speed up inference Train the model

The current leading tool for decomposing superposed features.

6. "Reward hacking" is when a model…

Refuses to answer Trains faster Optimises the measured metric in a way that defeats the actual goal Gets jailbroken

E.g. hardcoding the expected test output instead of solving the task.

On-device & multimodal (pages 03–04)

7. The biggest hardware advantage of Apple silicon for local AI is…

Unified memory — CPU/GPU/ANE share it with no copying, so large models fit The most CPU cores The biggest GPU The fastest network

See page 03.

8. On Apple platforms, the right tool for shipping a model inside an app is…

MLX llama.cpp vLLM Core ML (converts to .mlpackage, routes to ANE/GPU)

MLX is for dev/research; Core ML is the production path for apps.

9. The hybrid on-device pattern means…

Always use the cloud Small model on-device for most queries; escalate hard ones to a server model Never connect to the internet Run the biggest model possible on-device

Privacy + instant latency for most, frontier power only when needed.

10. A vision-language model fails most on…

Describing a photo's overall content Reading a clear headline Fine print, dense tables, precise counting and exact positions Detecting whether an image is a photo

See page 04.

11. A voice-agent "cascade" architecture is…

ASR → LLM → TTS: debuggable but higher latency Audio straight to audio with no text Several microphones A waterfall chart

Speech-to-speech models cut latency but are newer and capture tone.

Adoption (page 05)

12. The first step in an AI rollout should be…

Buying licences for everyone Measuring a baseline of task time and quality before AI Writing the policy A company-wide announcement

No baseline = no way to prove any gain. See page 05.

13. The most reliable productivity metric is…

Lines of code written Acceptance rate of suggestions "I feel faster" surveys Time + quality + cost on real tasks, measured against your own baseline

LoC and acceptance don't equal correct work; feelings mislead.

14. An AI usage policy works best when it…

Is short and positive: what's allowed, what's not, and the review duty — with easy compliant paths Bans everything Is a 50-page document Says nothing

A policy that blocks work drives shadow AI on personal devices.

15. "Skill atrophy" as a team risk means…

Tools get slower Models get worse over time If AI does all the writing, people never build the skill to review it — keep deliberate no-AI practice The licence expires

Judgement and review are the durable skills; they need exercising.

You made it

All five levels complete Score 12+ and finish your published capstone and you've climbed the whole curriculum — from "what is a token" to building agents, reading research, and leading adoption. The field keeps moving; keep the learning-loop habit from Level 1 and you'll keep pace with it.

L5 · 08 · Level 5 Checkpoint Quiz

Research & interpretability (pages 01–02)

On-device & multimodal (pages 03–04)

Adoption (page 05)

You made it