Task 1.1: Explain basic AI concepts and terminologies.

• AAsynchronous inference

  Correct

  Asynchronous inference queues requests, accepts payloads up to 1 GB with processing times up to one hour, and can scale the endpoint down to zero when there is nothing to process.

• BReal-time inference on a persistent endpoint

  Incorrect

  Real-time inference is for low-latency online requests and supports payloads up to 25 MB with 60-second processing for regular responses, far below a 500 MB file that runs for 20 minutes.

• CServerless inference

  Incorrect

  Serverless inference suits intermittent traffic, but it supports payloads only up to 4 MB and processing times up to 60 seconds.

• DBatch transform jobs over the full dataset

  Incorrect

  Batch transform is for offline processing when a large dataset is available upfront. Here each file arrives on its own and must be queued the moment it is uploaded.

Large payloads, long processing times and bursty traffic with a wish to scale to zero are the profile SageMaker AI documents for asynchronous inference.

• ADeep learning is the broadest of the four fields, and both AI and ML are specialized techniques that sit inside deep learning.

  Incorrect

  It is the other way round: AI is the umbrella term, and deep learning is a subset of machine learning.

• BML is one branch of AI, deep learning is a subset of ML, and generative AI is a type of AI that creates new content.

  Correct

  AWS describes ML as one of many branches of AI, deep learning as a subset of ML, and generative AI as a subset of AI that creates new content such as conversations, stories, images and music.

• CAI and ML are two names for the same thing, because every AI system has to learn its behavior from data.

  Incorrect

  Not all AI is ML. The field of AI also includes methods such as rule-based systems and search algorithms.

• DGenerative AI is separate from ML, because generative models are written as fixed rules and are not trained on any data.

  Incorrect

  Generative AI grew out of advances in deep learning, and foundation models are ML models trained on a broad spectrum of data.

Think of nested circles: AI contains ML, ML contains deep learning, and generative AI is the part of AI, built on deep learning, that produces new content.

• ASupervised learning

  Incorrect

  Supervised learning trains on input data paired with labeled outputs, and there are no labeled ideal routes to learn from here.

• BUnsupervised learning

  Incorrect

  Unsupervised learning finds patterns and groups in unlabeled data. It does not learn from a score given to the actions it tries.

• CReinforcement learning

  Correct

  Reinforcement learning learns by trial and error: actions that move toward the goal earn rewards, and the model learns the actions that maximize the cumulative reward.

• DSemi-supervised learning

  Incorrect

  Semi-supervised learning combines a small labeled dataset with a large unlabeled one. There is no labeled data at all in this scenario.

A reward signal for actions taken in an environment, with no labeled answers, is the defining setup of reinforcement learning.

• ARecorded audio of customer service calls

  Correct

  Audio files do not fit neatly into rows and columns, so AWS lists them as unstructured data.

• BAn inventory listing stored in a relational database table

  Incorrect

  Inventory listings with short text and numeric fields are a typical example of structured data, and relational databases are structured data stores.

• CVideo footage from in-store security cameras

  Correct

  Video monitoring is one of the examples AWS gives of unstructured data.

• DHR records with fields for name, start date and salary

  Incorrect

  Records made of short text, dates and numbers fit a table with one column per attribute, which makes them structured data.

• EMonthly sales figures per store kept in a spreadsheet table

  Incorrect

  Sales figures are discrete numeric data in rows and columns, the kind of data AWS describes as structured.

Structured data fits a table of rows and columns. Audio, video, images and long free text do not, and they are where machine learning is most often applied.

• AA workflow engine that runs a fixed, predefined sequence of steps

  Incorrect

  Traditional software follows predefined rules. The requirement is for a system that chooses its own next actions toward a goal.

• BA generative AI model that writes text only in reply to each prompt

  Incorrect

  That is what the existing assistant already does. AWS contrasts this prompt-by-prompt guidance with agentic AI, which acts on its own.

• CAn unsupervised model that groups past onboarding tickets

  Incorrect

  Grouping unlabeled records finds patterns in data. It does not plan steps or act on other systems.

• DAgentic AI that plans steps and uses tools to reach the goal

  Correct

  Agentic AI is an autonomous system that acts independently to achieve goals; AI agents choose their next actions themselves and use tools and systems to complete tasks without constant human oversight.

The difference between a generative AI assistant and agentic AI is agency: an agent is given a goal, plans its own steps and acts through tools.

• AReal-time inference

  Incorrect

  Real-time inference provides a persistent endpoint for low-latency online requests and sustained traffic, which this monthly job does not need.

• BServerless inference

  Incorrect

  Serverless inference targets intermittent request traffic with payloads up to 4 MB and 60-second processing, not one large offline dataset.

• CAsynchronous inference

  Incorrect

  Asynchronous inference queues individual requests with large payloads. The whole dataset is available upfront, which is the batch case.

• DBatch transform

  Correct

  Batch transform is for offline processing when large amounts of data are available upfront and you do not need a persistent endpoint; it handles datasets that are gigabytes in size.

All the data at once, no one waiting, no endpoint wanted: that is batch inference, which SageMaker AI calls batch transform.

• AServerless inference

  Correct

  Serverless inference suits intermittent or unpredictable traffic that can tolerate cold starts; it scales to zero when idle so you pay only for what you use, with payloads up to 4 MB.

• BReal-time inference on an always-on instance

  Incorrect

  A real-time endpoint is persistent and backed by the instance type you choose, so it is billed while idle between bursts.

• CBatch transform

  Incorrect

  Batch transform runs offline over a dataset that is available upfront; it does not serve individual API requests.

• DAsynchronous inference

  Incorrect

  Asynchronous inference is for queued requests with large payloads and long processing times, not small requests answered in under a second.

Small payloads, short processing, spiky traffic and no budget for idle time is the profile SageMaker AI describes for serverless inference.

• AUnsupervised learning

  Incorrect

  Unsupervised learning works on unlabeled data to discover patterns. Here every email already has a label.

• BSupervised learning

  Correct

  Supervised learning trains on inputs paired with labeled outputs, and AWS gives email spam classification as an example of it.

• CReinforcement learning

  Incorrect

  Reinforcement learning learns from rewards for actions taken in an environment, not from a labeled dataset.

• DGenerative pre-training

  Incorrect

  Foundation models are pre-trained on broad, generalized and unlabeled data. Learning to reproduce known labels is supervised learning.

Known answers attached to every example means supervised learning; tagging into two categories makes it binary classification.

• ATraining and inference are the same step run on different hardware

  Incorrect

  They are different steps: training adjusts parameters, and inference uses the trained model to produce outputs.

• BInference adjusts the model's parameters, while training is the step that produces the outputs

  Incorrect

  It is the other way round. Training adjusts parameters to reduce error; inference generates an output from an input.

• CTraining adjusts parameters from data; inference uses the trained model on new input

  Correct

  During training a model adjusts its parameters to minimize the gap between predictions and known outcomes. Inference is getting predictions from the trained model for a given input.

• DInference happens only once, before the model is deployed

  Incorrect

  Inference is how a deployed model is used; SageMaker AI provides deployment options to get predictions, or inferences, from trained models.

Training is the learning phase; inference is the using phase, and it runs every time the deployed model answers a request.

• ADetecting whether product reviews express positive or negative feelings

  Correct

  NLP lets computers interpret human language, including understanding the intent or sentiment hidden in it.

• BIdentifying cars and pedestrians in traffic camera images

  Incorrect

  Deriving information from images and video is computer vision.

• CTranslating support articles from English into German

  Correct

  Machine translation is an NLP task; NLP research began with experiments in translating sentences between languages.

• DPredicting next month's electricity use from past meter readings

  Incorrect

  Predicting a numeric value from past data is a regression or forecasting problem, not language processing.

• EGrouping network traffic into types to spot security incidents

  Incorrect

  Grouping unlabeled records by similarity is clustering, an unsupervised learning technique.

NLP works on human language, written or spoken; images belong to computer vision, and numeric prediction or grouping are general ML techniques.