Rubric Edu
Contents — domains, guide and mocks

AI, ML, deep learning, GenAI and agents — the vocabulary

AIF-C01 1.114 min read · checked 6 November 2026

Task statementExplain basic AI concepts and terminologies.

How the terms nest

  1. Artificial intelligencemachines doing tasks we associate with human intelligence
  2. Machine learninglearns patterns from data rather than explicit rules
  3. Deep learningML using multi-layer neural networks
  4. Generative AIcreates new text, images, audio, video, code
  5. Foundation models and LLMslarge models pre-trained on broad data
  6. Agentic AImodels given goals, tools and autonomy to act
Each layer is a subset of the one above. Agentic AI is shown last because it is a pattern built on top of models, not a new kind of model.

The terms the exam guide lists

The AIF-C01 guide names fifteen terms in this objective. Most of them pair up — a term and the one it is most often confused with — so learn them as pairs.

TermWhat it meansConfused with
ModelThe learned artefact: a function that turns an input into a prediction or outputAlgorithm — the recipe used to produce the model
AlgorithmThe procedure that learns from data (for example, linear regression, XGBoost, a neural network training method)Model
TrainingShowing the algorithm data so it adjusts its parametersInference
InferenceUsing a trained model to produce an output for new inputTraining
Neural networkLayers of connected nodes whose weights are learned from dataDeep learning — a neural network with many layers
Computer visionAI that interprets images and videoImage generation, which is generative AI
NLPAI that understands or produces human languageSpeech recognition, which turns audio into text first
BiasSystematic error — in data, labels or outputs — that skews results, often against a groupVariance, covered in lesson 4.1
FairnessOutcomes that do not disadvantage groups without justificationAccuracy — a model can be accurate overall and unfair to a subgroup
FitHow well the model captures the pattern: underfit (too simple), overfit (memorised the training data), good fitAccuracy on training data alone
LLMA foundation model trained mainly on text to understand and generate languageFoundation model — the wider class, which also covers image and multi-modal models
Generative AIModels that create new contentAny AI that outputs text — a sentiment classifier outputs a label, not new content
Agentic AIGoal-driven systems that reason, plan, call tools and act with some autonomyA chatbot that only answers questions

How AI, ML, GenAI, deep learning and agentic AI differ

AWS describes generative AI as a subset of AI that can create new content and ideas — conversations, stories, images, video and music — and foundation models as large generative models trained on a broad spectrum of data, with LLMs as one class of foundation model. Traditional machine learning, by contrast, usually predicts or classifies: is this transaction fraud, what will demand be next week, which segment does this customer belong to.

AWS Prescriptive Guidance draws the line to agentic AI clearly. Traditional AI is tool-centric and narrow, focused on prediction or classification. Traditional software agents add autonomy but are rule-bound. Agentic AI combines autonomy, asynchrony and agency: goal-driven systems that reason, act and adapt, with decisions that are contextual and often LLM-enhanced.

Same business, three kinds of AI

What does the system have to do?
  • Label or predict something from past data
    Traditional MLclassify a support ticket's priority
  • Produce new content
    Generative AIdraft the reply to the ticket
  • Reach a goal through several steps and tools
    Agentic AIlook up the order, issue the refund, reply, close the ticket

Types of data

The guide lists data types as pairs and families. Expect a scenario that describes the data and asks which kind it is, or which learning approach the data allows.

  • Labelled vs unlabelled. Labelled data carries the answer (this email *is* spam). Unlabelled data does not. Labels make supervised learning possible.
  • Structured vs unstructured. Structured data fits rows and columns with a fixed schema. Unstructured data — documents, images, audio, free text — does not.
  • Tabular. Rows and columns: customer records, transactions. The classic home of traditional ML.
  • Time-series. Values recorded in order over time: sensor readings, daily sales. Order matters, which is why forecasting is its own technique.
  • Image and text. Unstructured data handled by computer vision and NLP respectively, and the main inputs to foundation models.

Types of learning

TypeData it needsTypical taskOne-line test
SupervisedLabelled examplesClassification, regression“We have past examples with the right answer”
UnsupervisedUnlabelled dataClustering, anomaly detection, dimensionality reduction“Find structure we haven't defined”
ReinforcementRewards from an environmentSequential decisions: robotics, game play, tuning a policy“Learn by trial, reward and penalty”

Types of inferencing

The guide names four: batch, real-time, asynchronous and serverless. Amazon SageMaker AI offers all four as deployment options, with published limits that make the choice concrete.

OptionUse it whenLimits (SageMaker AI)
Real-timeYou need a fast answer per request and traffic is steadyPayload up to 25 MB; up to 60 s per request (8 min streaming); always-on endpoint
ServerlessTraffic is intermittent or unpredictable and you don't want to manage instancesPayload up to 4 MB; up to 60 s; pay for use, not idle time
AsynchronousRequests are large or slow and the caller can wait for a queued resultPayload up to 1 GB; up to 1 hour; can scale to zero
BatchA whole dataset is available up front and nobody is waitingLarge datasets; jobs can run for days; no persistent endpoint

Traps the wrong answers are built from

Tempting but wrongDo this instead
Calling any model that outputs text “generative AI”.Ask whether it creates new content. A sentiment model outputs a label — that is classification.
Treating agentic AI as a type of model.Agentic AI is a way of using models: a goal, a plan, tools and autonomy.
Choosing real-time inference for a scheduled job over a full dataset.If nobody is waiting and the data is available up front, use batch.
Judging fit by training accuracy.A model that scores near-perfectly on training data and poorly on new data is overfit.

You should now be able to

  • Define AI, ML, deep learning, neural networks, computer vision, NLP, model, algorithm, training, inference, bias, fairness, fit, LLM, GenAI and agentic AI.
  • Explain how AI, ML, deep learning, GenAI and agentic AI relate and differ.
  • Choose between batch, real-time, asynchronous and serverless inference from a scenario.
  • Identify labelled, unlabelled, tabular, time-series, image, text, structured and unstructured data.
  • Tell supervised, unsupervised and reinforcement learning apart by the data and feedback they use.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A bank wants a model that flags card transactions as fraudulent or legitimate. It has five years of transactions, each marked by investigators as fraud or not fraud.

    Which type of machine learning fits this data?

    1. ASupervised learning
    2. BUnsupervised learning
    3. CReinforcement learning
    4. DGenerative AI
    Show answer and reasoning
    1. ACorrect. Every example carries the right answer (the investigator's label), and the task is to predict that label for new transactions — the definition of supervised classification.
    2. BIncorrect. Unsupervised learning works on unlabelled data. Here the labels exist and are exactly what the model must learn.
    3. CIncorrect. Reinforcement learning learns from rewards in an environment through trial and error, not from a fixed labelled history.
    4. DIncorrect. The task is to classify, not to create new content.

  2. Question 2

    A media company runs a model each night that tags every video uploaded that day with topic labels. Nobody needs the tags until the next morning.

    Which inference approach is most cost-effective?

    1. ABatch inference
    2. BReal-time inference
    3. CServerless inference
    4. DStreaming inference
    Show answer and reasoning
    1. ACorrect. The full dataset is available up front, nobody is waiting, and there is no need for a persistent endpoint.
    2. BIncorrect. A real-time endpoint is always on and built for low-latency single requests — paying for it all day to run one nightly job wastes money.
    3. CIncorrect. Serverless suits intermittent single requests with small payloads (up to 4 MB in SageMaker AI); a nightly pass over a whole day of video is a batch job.
    4. DIncorrect. Not one of the four types the exam guide lists, and the scenario has no streaming requirement.

  3. Question 3

    Which statement correctly describes the relationship between these terms?

    1. ADeep learning is a subset of machine learning, which is a subset of AI.
    2. BMachine learning is a subset of deep learning.
    3. CGenerative AI and machine learning are unrelated fields.
    4. DEvery AI system uses deep learning.
    Show answer and reasoning
    1. ACorrect. Deep learning is machine learning with many-layered neural networks, and machine learning is one approach within AI.
    2. BIncorrect. The nesting is the other way round: deep learning is one family of machine learning methods.
    3. CIncorrect. Generative AI is built on machine learning — foundation models are very large ML models pre-trained on broad data.
    4. DIncorrect. Rule-based systems and many traditional ML models (for example, decision trees) are AI without deep learning.

  4. Question 4

    A model scores 99% accuracy on the data it was trained on but 62% on new customer data.

    What does this most likely indicate?

    1. AThe model is overfit.
    2. BThe model is underfit.
    3. CThe model has a good fit.
    4. DThe model needs real-time inference.
    Show answer and reasoning
    1. ACorrect. Overfitting means the model memorised the training data, including its noise, and fails to generalise to new data.
    2. BIncorrect. An underfit model performs poorly on training data too, because it is too simple to capture the pattern.
    3. CIncorrect. A good fit performs similarly on training and new data. A 37-point gap is the classic overfitting signature.
    4. DIncorrect. Inference type affects latency and cost, not how well the model generalises.

  5. Question 5

    An insurer is considering three systems: one that predicts claim cost, one that drafts letters to claimants, and one that receives a claim, checks the policy, requests missing documents, and schedules an assessor without a human stepping in.

    Which TWO statements are correct? (Select TWO.) (Select 2.)

    1. AThe letter-drafting system is generative AI.
    2. BThe end-to-end claims system is agentic AI.
    3. CThe claim-cost predictor is generative AI because it produces a number.
    4. DAll three systems are agentic because they run automatically.
    5. EThe letter-drafting system must be trained with reinforcement learning.
    Show answer and reasoning
    1. ACorrect. It creates new content — the definition of generative AI.
    2. BCorrect. It pursues a goal across several steps, using tools and acting with autonomy — the features AWS lists for agentic AI.
    3. CIncorrect. Predicting a value from historical data is regression, a traditional ML task. Producing an output is not the same as generating new content.
    4. DIncorrect. Running on a schedule is automation. Agentic AI means goal-driven reasoning and action across steps, not merely running without a click.
    5. EIncorrect. Nothing in the scenario requires it; a foundation model used with a prompt can draft letters without any training by the insurer.

Practise every question in Domain 1 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.