Building with AI — Module Assessment

Question 1

Which statement best captures the shift from using AI to building with it?

Accepted Answer

You treat the model as a reusable component inside a tool rather than a chatbot you talk to.

Answer

You must train a brand-new model from scratch instead of chatting.

Answer

You stop paying for AI because building tools is always free.

Answer

You must learn to program before you can build anything useful.

Question 2

What does Hugging Face primarily provide to builders?

Accepted Answer

A large public hub of openly shared models and datasets to browse and download.

Answer

A single proprietary chatbot that replaces all other AI tools.

Answer

Free unlimited computing power to run any model at no cost.

Answer

A legal service that removes all vendor lock-in automatically.

Question 3

In an API call to a model, what does a 'request' consist of?

Accepted Answer

A prompt (your instructions and context) plus parameters that shape the output.

Answer

Only the model's final answer, with no input from you.

Answer

A payment receipt and nothing else.

Answer

A full copy of the model's internal weights.

Question 4

Why might an organisation self-host an open-weight model like Meta's Llama instead of calling a closed API?

Accepted Answer

For more control and privacy and to avoid per-token fees, accepting that it must run the hardware itself.

Answer

Because open-weight models never require any computing hardware to run.

Answer

Because closed APIs are illegal to use for businesses.

Answer

Because open-weight models are always more capable than any closed model.

Question 5

What is the defining promise of no-code and low-code platforms?

Accepted Answer

They let people build working tools with little or no programming, using visual blocks and connectors.

Answer

They make AI models more accurate than custom-built ones.

Answer

They eliminate all ongoing costs once a tool is built.

Answer

They remove the need to understand your data or check outputs.

Question 6

Which situation is the strongest reason to choose custom code over a no-code platform?

Accepted Answer

Bespoke, high-stakes logic with unusual edge cases that the platform's blocks cannot bend to fit.

Answer

A common task like sorting incoming emails and drafting replies.

Answer

A quick internal form that classifies entries automatically.

Answer

Any task, because no-code tools cannot use AI at all.

Question 7

In retrieval-augmented generation, what happens before the model generates its answer?

Accepted Answer

Relevant passages are retrieved from a document collection and given to the model.

Answer

The model is retrained overnight on the user's documents.

Answer

The model erases its training data to avoid bias.

Answer

The user manually rewrites the model's memory.

Question 8

Which statement about grounding AI in your own data is accurate?

Accepted Answer

Grounding reduces hallucination, but wrong retrieval or vague prompts can still produce wrong answers.

Answer

Grounding completely eliminates hallucination in every case.

Answer

A citation guarantees the answer is correct without any human check.

Answer

Grounding requires retraining the underlying model on your files.

Question 9

What most distinguishes an AI agent from a single prompt-and-response?

Accepted Answer

It wraps the model in a loop that plans and takes multi-step actions toward a goal

Answer

It uses a larger and more expensive language model

Answer

It always produces the same output for the same input

Answer

It removes the need for any human oversight

Question 10

Why are long autonomous agent chains often unreliable even when each single step is fairly good?

Accepted Answer

Because errors compound across steps, so many steps in a row are far less reliable than any one step

Answer

Because models cannot use tools at all

Answer

Because agents always run out of memory after two steps

Answer

Because a single prompt is always more capable than an agent

Question 11

What makes evaluating AI outputs harder than testing ordinary deterministic software?

Accepted Answer

AI outputs can vary run to run and may confidently hallucinate, so there is often no single correct string to match

Answer

AI features never produce any errors worth testing

Answer

AI outputs are always identical, so one test covers everything

Answer

Ordinary software cannot be tested at all

Question 12

What best describes red-teaming an AI system?

Accepted Answer

Deliberately trying to make the system fail or produce harmful output before release

Answer

Publishing the model's benchmark scores to compare it with rivals

Answer

Letting the model grade its own outputs automatically

Answer

Running the same prompt many times to average the answers

Question 13

Why is a polished AI demo a poor indicator that you have a finished product?

Accepted Answer

A demo only has to work once, on chosen inputs, for a friendly audience, while a product must handle messy real-world inputs reliably.

Answer

Demos always use worse models than production systems do.

Answer

Products never need the features that demos show.

Answer

A demo proves nothing at all about whether the idea can work.

Question 14

A beginner wants to ship an AI feature responsibly for the first time. Which approach best reflects this lesson?

Accepted Answer

Validate a real user's need, ship a narrow MVP with a human in the loop and an 'I'm not sure' guardrail, be honest about limits, and iterate from feedback.

Answer

Launch the broadest possible version at once so it looks impressive, and hide the AI's limitations from users.

Answer

Wait until the model is perfect before releasing anything, since guardrails and human review won't be needed then.

Answer

Skip evaluation because a strong enough model makes testing unnecessary.