Examworthy
Open the appSign inStart free

AIF-C01 domain - 24% of the exam

Generative AI fundamentals

Generative AI fundamentals is 24% of the AIF-C01 exam. These are the objectives it covers, each with practice questions, with every answer explained.

The domain in numbers

78
Practice questions
3
Objectives
21
Easy
57
Medium

Where people go wrong

  • Reversing the effect of temperature

    The wrong answers move temperature the wrong way. To make wording consistent they raise it, and to widen variety they lower it. Lowering temperature sharpens the distribution so the most probable tokens dominate, converging towards greedy decoding at zero. Raising it flattens the distribution so lower-probability tokens are sampled more often. Consistent, reproducible output needs a low temperature; a wider spread of ideas needs a higher one.

    See it in a sample question

  • Confusing temperature, top-p and top-k

    The wrong answers swap what each control does. Temperature rescales the probability of every token, making the distribution flatter or sharper. Top-p truncates, keeping the smallest set of top tokens whose cumulative probability reaches a threshold, so the candidate count varies with the distribution. Top-k keeps a fixed count of the most probable tokens. Describing top-p as a fixed count, or temperature as a cutoff, misattributes the mechanism.

    See it in a sample question

  • Expecting a model setting to supply current facts

    The wrong answers reach for a setting to fix stale or inaccurate answers: a different temperature, a larger context window or a longer output limit. None of these puts the current document in front of the model, so a lower temperature only repeats an outdated figure more consistently. Grounding retrieves the verified source text and supplies it in the prompt at inference time, so the answer is anchored to that material rather than to what the weights encoded in training.

    See it in a sample question

  • Swapping Trainium and Inferentia

    The wrong answers give each AWS accelerator the other's job. Trainium is the purpose-built chip for training deep learning and generative AI models, so it suits a lengthy training run. Inferentia is built to run trained models in production with high throughput and low cost per inference. Offering Inferentia for training, Trainium for serving, or Graviton as a training chip reverses or ignores that split.

    See it in a sample question

  • Pricing a request by anything other than tokens

    The wrong answers price a request by something other than tokens: the word count of the visible answer, processing time, a flat charge, or retraining. Hosted models bill the input and output tokens processed. A large supplied document or retrieved reference block raises cost even when the reply is short, and a request for five full versions multiplies the output tokens. Inference calls never retrain the model.

    See it in a sample question

Objectives in this domain

What to study

The official documents this domain's practice questions cite most.

  • Foundation modelscited by 4 questions
  • Top-p (nucleus) samplingcited by 4 questions
  • Amazon Bedrock overviewcited by 3 questions
  • Amazon SageMaker JumpStart documentationcited by 3 questions
  • AWS documentation on embeddings and vector representationscited by 3 questions
  • Embeddingscited by 3 questions
  • Tokenscited by 3 questions
  • Amazon Bedrock service overviewcited by 2 questions

Sample questions from this domain

Free sampleGenerative AI fundamentalsmedium

A support team sends the same prompt to a hosted LLM many times a day and gets noticeably different wording each time. They want answers to stay closely on-message and vary far less, without demanding byte-for-byte identical output. Which change best achieves this?

  • ALower the temperature toward 0.2 so the model concentrates on the highest-probability tokens. Correct
  • BRaise the temperature toward 1.0 so the model considers a wider spread of next tokens.
  • CIncrease the maximum output-token limit so each answer has room to finish completely.
  • DRaise top-p toward 1.0 so the model samples from the full probability distribution.
Lowering temperature concentrates output on the most-probable tokens, reducing run-to-run variation. Temperature scales how sharply the model favours its most-probable tokens. Lowering it toward 0 concentrates probability mass on the top tokens, so the same prompt produces closely similar, on-message wording each time without forcing fully identical output.

Why A is correct: Correct. A low temperature sharpens the distribution toward the likeliest tokens, so wording stays consistent run to run while still allowing slight variation.

Why B is wrong: Higher temperature flattens the token distribution and increases variation, which moves away from the on-message consistency the team asked for.

Why C is wrong: The token limit governs how long an answer can be, not how much its wording varies, so consistency is unchanged.

Why D is wrong: A top-p of 1.0 keeps the entire distribution in play, which preserves or increases variation rather than reducing it.

Free sampleGenerative AI fundamentalsmedium

A developer is tuning a model's top-p (nucleus) sampling value and wants to be sure they understand its effect. Which statement best describes how top-p selects the tokens the model samples from?

  • AIt caps the total number of tokens the model is allowed to generate in the response.
  • BIt restricts sampling to the smallest set of top tokens whose cumulative probability reaches p. Correct
  • CIt sets a fixed random seed so the same prompt always returns the same response.
  • DIt scales the model's confidence by reweighting every token's probability before sampling.
Top-p sampling draws from the smallest set of tokens whose cumulative probability reaches p. Top-p, or nucleus sampling, ranks tokens by probability and keeps the smallest set whose cumulative probability reaches the threshold p, discarding the long tail. The model then samples only from that nucleus, so a lower p narrows the candidate set.

Why A is wrong: That describes a maximum-output-length setting, not top-p. Top-p shapes the candidate token set, not the response length.

Why B is correct: Correct. Nucleus sampling keeps the smallest group of most-likely tokens whose probabilities sum to p, then samples only from that group.

Why C is wrong: A seed governs reproducibility of the random draw, not which tokens are eligible, so it is a different control from top-p.

Why D is wrong: Reweighting the whole distribution by a factor describes temperature, not the cumulative-probability cutoff that top-p applies.

Free sampleGenerative AI fundamentalsmedium

An architect is explaining grounding to colleagues. Which statement most accurately distinguishes a grounded generative AI response from an ungrounded one?

  • AA grounded response is always shorter than an ungrounded one because the source material constrains how much text the model can produce.
  • BA grounded response is produced at a lower temperature, while an ungrounded response uses a higher temperature to encourage creativity.
  • CA grounded response is anchored to specific external source material supplied to the model, whereas an ungrounded response relies only on the model's internal parameters. Correct
  • DA grounded response is generated entirely without the foundation model, using a database lookup that returns stored answers directly to the user.
Grounding anchors a model's output to supplied external source material rather than to its internal parameters alone. Grounding provides the model with relevant external evidence at generation time so its response can be tied to that material. An ungrounded response depends only on what the weights encoded during training, which makes it harder to verify.

Why A is wrong: This is tempting because grounding adds constraints, but length is unrelated; a grounded answer can be longer or shorter, and grounding concerns sourcing, not output size.

Why B is wrong: This is tempting because both relate to output control, but grounding supplies external evidence rather than adjusting sampling; temperature changes randomness, not whether output is anchored to sources.

Why C is correct: Correct: grounding feeds verified external context to the model so its answer can be traced to that material, while an ungrounded answer draws solely on knowledge encoded in the model's weights.

Why D is wrong: This is tempting because grounding involves retrieval, but the model still generates the answer; retrieved material informs generation rather than replacing the model with a raw lookup.

Free sampleGenerative AI fundamentalsmedium

An organisation is training a large foundation model from scratch and wants to reduce the cost of the lengthy training run by using a purpose-built AWS accelerator chip designed specifically for high-performance, cost-effective deep learning training. Given this objective, which AWS hardware should the team target for the training workload?

  • AAWS Inferentia, the chip family AWS built to lower the cost of running model training jobs at scale
  • BAWS Graviton, the general-purpose Arm-based processor AWS recommends for accelerating model training
  • CAWS Outposts, the on-premises rack AWS sells as a dedicated training accelerator for foundation models
  • DAWS Trainium, the chip family AWS built to deliver high-performance, lower-cost training for deep learning and generative AI models Correct
AWS Trainium is the purpose-built accelerator for cost-effective, high-performance deep learning and generative AI model training. Trainium chips are designed by AWS specifically to accelerate the compute-intensive training of deep learning and generative AI models at lower cost than general-purpose hardware, which is why they fit a from-scratch training run.

Why A is wrong: Inferentia is AWS silicon optimised for inference, not training, so selecting it for the training run would misuse the chip's intended purpose even though the names are easily confused.

Why B is wrong: Graviton is a general-purpose CPU for everyday compute workloads and is not a deep learning training accelerator, so it does not meet the performance need for training a large model.

Why C is wrong: Outposts brings AWS infrastructure on-premises but is not itself a training accelerator chip, so it does not address the requirement for purpose-built training silicon.

Why D is correct: Trainium is AWS's purpose-built accelerator for cost-effective, high-performance model training, which is precisely the workload described.

Free sampleGenerative AI fundamentalsmedium

A finance lead wants to understand why the bill for a hosted large language model can vary widely between two requests that each return a short final answer. Which factor most directly drives the cost of a single request under typical token-based pricing?

  • AThe number of distinct words in the final answer alone, since providers bill only for the visible text the user receives.
  • BThe wall-clock time the request spends being processed, billed per second of compute regardless of how much text is involved.
  • CA flat fee charged once per request, so identical pricing applies whether the prompt is a sentence or a long report.
  • DThe combined number of input and output tokens processed, so a long supplied document raises cost even when the returned answer is brief. Correct
Token-based pricing scales with combined input and output tokens, so large inputs raise cost even with short answers. Hosted models bill per token across both the prompt and the generated response. A lengthy input document is read and charged in full, which means cost can be high even when the model returns only a short answer.

Why A is wrong: This is tempting because the answer is what the user sees, but billing counts tokens not words and includes the input as well, so a brief answer does not guarantee a low charge.

Why B is wrong: This is tempting because latency feels like a cost driver, but standard hosted pricing is per token rather than per second, so processing time is not the direct billing unit.

Why C is wrong: This is tempting because flat pricing is simple, but token-based models scale with text volume, so a long report costs far more than a single sentence.

Why D is correct: Correct: token-based pricing charges for both the prompt tokens read and the tokens generated, so a large input inflates cost regardless of how short the final answer is.

Other domains in this exam

See also the AIF-C01 cert hub, the study guide, and the cheat sheet.

Examworthy is not affiliated with or endorsed by Amazon Web Services. Original, blueprint-aligned practice material only.