AIF-C01 fundamentals of GenAI explained
Fundamentals of GenAI is worth 24% of AIF-C01’s scored content — around 12 of the 50 scored questions. Together with applications of foundation models it makes generative AI 52% of the exam, which is the single most important planning fact about AIF-C01.
Foundation models
A foundation model is a large model trained on a broad dataset, adaptable to many tasks without being built for any one of them. That adaptability is the point: rather than training a separate model per task, you take a general one and steer it.
Terms to know cold:
- Large language model (LLM) — a foundation model for text.
- Multimodal model — handles more than one input type, typically text and images.
- Token — the unit models work in, roughly a word fragment. Cost and limits are counted in tokens.
- Context window — how much the model can consider at once, measured in tokens. Exceeding it means content is dropped, not summarised.
- Embedding — a numeric representation of meaning, letting systems find semantically similar content.
- Parameters — the learned values in a model; more generally means more capable and more expensive.
Why output varies
The same prompt can produce different answers. Models predict likely continuations, and settings control how much variation is allowed:
- Temperature — higher gives more varied and creative output, lower gives more predictable output. Lower is not “more accurate”, only more consistent.
- Top-p / top-k — alternative ways of constraining which next tokens are considered.
Exam framing: a requirement for consistent, repeatable output wants low temperature. A requirement for creative variety wants higher.
Prompt engineering
Named explicitly in the domain, and it is testable at a practical level:
- Zero-shot — ask directly, no examples.
- Few-shot — include examples of the input and output you want. The reliable fix when output format is inconsistent.
- System prompt or instructions — standing behaviour applied to every turn: tone, language, scope, refusals.
- Chain of thought — asking the model to reason step by step for multi-step problems.
The distinction that appears repeatedly: anything phrased as always or never belongs in the system instructions, not repeated in each user prompt.
Where generative AI fails
AWS expects a realistic view, and the failure modes are examinable:
| Failure | What it is |
|---|---|
| Hallucination | Confident output that is simply untrue |
| Non-determinism | The same input can give different output |
| Knowledge cutoff | The model does not know recent events |
| No source awareness | It cannot cite what it was not given |
| Cost and latency | Large models are slower and more expensive per call |
Hallucination is the most tested. The mitigation is grounding — giving the model the source content — not adjusting temperature and not choosing a bigger model.
Amazon Bedrock
The AWS service at the centre of this domain: managed access to foundation models from multiple providers through one interface, with no infrastructure to provision.
At foundational level you need to know what Bedrock is for, that it provides model choice, and that it removes the need to host models yourself. Contrast it with Amazon SageMaker AI, which is the platform for building, training and deploying your own models.
When generative AI is the wrong tool
An underrated part of the domain. Generative AI is not the answer to everything:
- A deterministic requirement — a calculation, a rule — wants code, not a model.
- A well-defined classification task at high volume is often cheaper and more reliable with a traditional ML model.
- A problem a pre-trained AI service already solves does not need a foundation model.
Sample questions
Question 1. A company's assistant returns answers in an inconsistent format despite clear instructions. Which prompt engineering technique most directly addresses this?
- A. Increase the temperature setting
- B. Few-shot prompting with examples of the desired output format
- C. Shorten the prompt to reduce confusion
- D. Switch to a model with more parameters
Show answer
Answer: B
Few-shot prompting includes examples of the desired input and output, which is the most direct way to make output format consistent. Raising temperature increases variation, shortening the prompt removes guidance, and switching to a larger model does not by itself fix formatting.
Question 2. A summarisation task must produce the same summary every time it runs on the same document, for audit reasons. Which setting change best supports this?
- A. Lower the temperature
- B. Raise the temperature
- C. Increase the context window
- D. Select a model with more parameters
Show answer
Answer: A
Lowering the temperature reduces randomness in token selection, making output more consistent for the same input. Raising temperature does the opposite, a larger context window changes how much is considered rather than how variable the output is, and more parameters does not imply determinism.
Question 3. A team needs access to foundation models from several providers through one managed interface, without running any infrastructure. Which AWS service is designed for this?
- A. Amazon SageMaker AI
- B. Amazon EC2 with GPU instances
- C. Amazon Bedrock
- D. Amazon S3
Show answer
Answer: C
Amazon Bedrock provides managed access to foundation models from multiple providers with no infrastructure to manage. SageMaker AI is for building and training your own models, EC2 requires managing servers, and S3 is storage.
What to practise
Open Bedrock, send the same prompt three times at high temperature and three times at low, and watch what changes. Then take a prompt that produces inconsistent formatting and fix it with two examples. Ten minutes of that teaches the temperature and few-shot questions permanently.