domain_1

Core AI Terms

Artificial Intelligence, or AI, is the big umbrella. It covers any technology that enables machines to do things we usually think of as needing human intelligence—like understanding speech, recognizing images, making decisions, or planning steps.

Machine Learning, or ML — a branch of AI where systems learn patterns from data instead of being explicitly programmed rule by rule. For example, learning to predict house prices or detect fraud from past examples, rather than following a list of if-then instructions written by a human.

How AI differs from ML. AI is the entire field of making machines smart. ML is one approach within that field—the data-driven approach where the machine improves by seeing lots of examples. Not all AI uses machine learning, but today most AI systems you hear about do.

Deep Learning, or DL, is a subset of Machine Learning. It uses neural networks with many hidden layers. Think of it as ML with a much deeper architecture, which makes it incredibly powerful for messy, complex data like pictures, voice clips, and long text. It typically needs more data and much more compute power than traditional ML.

How ML differs from Deep Learning. Traditional ML often works well on structured data, like rows and columns in a spreadsheet, with simpler algorithms. Deep Learning shines on unstructured data—images, audio, freeform text—by learning through many layers of abstraction. It’s a more powerful but more demanding type of ML.

A neural network is the engine under Deep Learning’s hood. It has three basic parts. The input layer receives the raw data. The hidden layers—often many of them—learn to detect useful patterns. The output layer gives the final result, like a classification or a prediction.

Computer Vision, or CV — the branch of AI that lets machines interpret and understand the visual world. Typical tasks include classifying what’s in an image, detecting objects and faces, reading text from pictures, and moderating content.

Natural Language Processing, or NLP — the branch of AI that enables machines to understand, analyze, and generate human language. This covers sentiment analysis, entity recognition, translation, summarization, question answering, and powering chatbots.

Amazon Comprehend is a fully managed service that provides NLP capabilities. But NLP tasks can also be handled by foundation models, which are larger, more general models we’ll talk about shortly.

Model — a trained mathematical representation that turns inputs into outputs. For instance, a fraud detection model takes in transaction details and produces a prediction about whether the transaction looks fraudulent.

Algorithm — a step-by-step procedure for solving a problem or training a model. The algorithm is the recipe; the trained model is the dish that comes out of following that recipe with data.

How a model differs from an algorithm. The algorithm is the learning procedure—the set of steps. The model is the thing you get after running that algorithm on data. You use the algorithm only during training. You use the model afterward for predictions or actions.

Training is the process where a model learns patterns from data. During training, the model’s internal parameters are adjusted so it gets better at its task. Inference is when you take that trained model and feed it new, unseen data to get a prediction or output. Training builds the model; inference uses it.

How training differs from inference. Training is the learning phase—it can take hours or days and uses labeled or unlabeled data to tune the model. Inference is the real-time or batch application phase—it uses the already-trained model to generate answers on new data.

Model fit describes how well a model captures the real patterns in data. A good fit generalizes well to new data. Overfitting means the model learned the training data too perfectly, including its noise, so it stumbles on data it hasn’t seen before. Underfitting means the model failed to learn enough useful patterns even from the training data.

Bias in AI is a systematic error or unfair behavior that shows up in data, models, or decisions. Fairness means the system treats relevant individuals or groups appropriately and avoids unjust outcomes. Bias can creep in from unrepresentative training data, historical discrimination baked into the data, measurement error, biased sampling, or choosing the wrong features.

Generative AI, or GenAI — a category of AI that creates new content based on patterns learned during training. It can generate text, images, audio, video, code, and even synthetic data. Unlike traditional ML, which often predicts a category or a number, Generative AI produces something new that didn’t exist before.

How traditional ML differs from Generative AI. Traditional ML typically classifies or predicts: is this email spam? What will the temperature be? Generative AI makes something new: write an email reply; create a picture of a mountain at sunset. One discriminates or estimates; the other creates.

Foundation model, or FM — a large, pre-trained model that you can adapt to many different tasks. Think of it as a general-purpose brain that you can customize. Techniques to adapt it include prompt engineering, giving in-context examples, fine-tuning on your own data, retrieval-augmented generation, and distillation.

A Large Language Model, or LLM, is a foundation model specifically trained to understand and generate language. LLMs power tasks like question answering, summarization, translation, classification, information extraction, and code generation.

How a foundation model differs from a Large Language Model. All LLMs are foundation models, but not all foundation models are LLMs. A foundation model is a general term for any large pre-trained model—it could be for vision, audio, or language. An LLM is a foundation model built specifically for language tasks.

Agentic AI — a system that combines a model with tools, memory, instructions, and a workflow logic to accomplish multi-step goals. It doesn’t just respond to a prompt; it can decide what steps to take, call APIs, check databases, and orchestrate a process to finish a task. For example, an agent might handle a furniture return by checking the order database, verifying eligibility, creating a shipping label, and sending a confirmation.

How Agentic AI differs from a foundation model or LLM. A foundation model produces an output given an input. An agent uses a model as its brain but adds the ability to act: it can chain multiple steps, make decisions, remember context, and use external tools to get things done. It’s the difference between asking a smart assistant a question, and having that assistant go and complete an entire workflow for you.

Relationship Between AI, ML, Deep Learning, GenAI, and Agentic AI

Artificial Intelligence — the broadest field, creating systems that can perform tasks normally associated with human intelligence, like understanding language, recognizing images, making predictions, or automating decisions.

Machine Learning — a subset of AI where systems learn patterns from data instead of being explicitly programmed, enabling tasks like fraud detection, product recommendations, or predicting customer churn.

Deep Learning — a subset of Machine Learning that uses neural networks with many hidden layers, especially powerful for large amounts of unstructured data like text, images, audio, and video.

Generative AI — a category that creates new content based on learned patterns, such as generating text, images, code, or summaries, rather than just predicting a label or number.

Agentic AI — a system that combines a model with tools, memory, instructions, and workflow logic to accomplish a goal and perform multi-step actions, like ordering a product and scheduling a delivery.

These concepts form a nested hierarchy. Artificial Intelligence is the widest umbrella. Inside it, Machine Learning learns from data. Deep Learning is a specialized branch of ML using deep neural networks. Generative AI and Agentic AI are different capabilities that can be built on top of these layers. For example, an agent might use a generative foundation model trained with deep learning to plan and execute tasks, so the categories overlap in practice.

How Generative AI differs from Agentic AI: Generative AI is about creating new content, such as poems, images, or dialogue. Agentic AI is about taking actions — it uses tools, checks databases, calls APIs, and follows workflows to achieve a goal. An agent may generate content as part of its plan, but its core identity is action-oriented.

Types of Machine Learning

Machine learning is about teaching computers to learn from data. There are several major approaches, each suited to different kinds of problems and data.

Supervised learning — a type of machine learning where the training data includes both the input and the correct answer. The model learns to map inputs to outputs by seeing many labeled examples.

Supervised learning divides into two main tasks. Classification predicts a category, like whether an email is spam or not spam. Regression predicts a numerical value, like tomorrow's temperature or the price of a house.

Think of classification like sorting mail into labeled bins. Think of regression like guessing a number on a scale. Both need a teacher who has already provided the right answers for many examples.

Unsupervised learning — a type of machine learning that uses data without any labels. The model finds hidden patterns, groupings, or unusual points all on its own.

Unsupervised learning includes clustering, where you group similar items together, like segmenting customers by behavior. It also includes anomaly detection, where you flag observations that are very different from the rest, like unusual credit card charges.

Clustering is like organizing a messy bookshelf by topic without any pre-made labels. Anomaly detection is like spotting a single red apple in a basket of green ones.

Reinforcement learning — an agent learns by interacting with an environment. It receives rewards for good actions and penalties for bad ones, gradually improving its strategy to maximize long-term reward.

Think of training a dog: you give a treat when it sits, and a gentle correction when it doesn't. Over time, the dog learns the best behaviors. In reinforcement learning, the agent is the dog, the environment is the world, and the reward is the treat.

Self-supervised learning — a clever approach where the model creates its own training signals from the data itself, without needing humans to label everything. For example, a language model might learn by predicting missing words in a sentence.

This is how large language models are pre-trained. They read vast amounts of text, hide some words, and learn to fill in the blanks. The data becomes its own teacher.

Data Types

Data comes in different shapes and forms, and knowing the type of data you're working with helps you choose the right AI approach. Let's walk through the four main categories.

Structured data, also called tabular data — data organized into a fixed schema. Think rows and columns, like a spreadsheet or a relational database table. Examples include customer records, financial transactions, and inventory lists.

Semi-structured data — data that has some organization but doesn't follow a strict table structure. It uses tags or key-value pairs, like JSON or XML files. Application logs and API responses are common examples.

Unstructured data — data without a fixed schema or model. It doesn't fit into rows and columns. Text documents, emails, images, audio, and video are all unstructured.

Time-series data — data where time is a critical part of each record. Every data point has a timestamp and the sequence matters. Sensor readings over time and website traffic logs are typical examples.

How structured data differs from semi-structured and unstructured data: structured data has a rigid schema with clearly defined fields, like a database table. Semi-structured data has some organization, like tags or key-value pairs, but the schema can vary. Unstructured data has no fixed format at all — it's freeform and needs more processing before an AI model can use it.

Inference Methods

Inference methods describe how you deploy a trained model to get predictions. Choosing the right method depends on how quickly you need results, how much data you have, and how steady your traffic is.

Real-time inference — This method gives you an immediate response with low latency. It runs through a persistent endpoint that is always ready to handle requests. Think of fraud detection during a payment, product recommendations while browsing, or real-time stock trading.

Serverless inference — The platform automatically provisions and scales compute resources. It is great for intermittent or unpredictable traffic and can save costs. However, be aware that it can have cold-start latency, meaning a slight delay when the service is first invoked after a quiet period.

Asynchronous inference — Use this for large inputs or long-running tasks. The request is processed, and the result is stored or returned later. It is ideal when you don't need an instant answer, like processing high-resolution videos or large documents.

Batch inference — This method processes many records together, all at once, without maintaining a persistent endpoint. It is perfect for bulk processing where immediate responses are not required, such as overnight customer segmentation or generating daily product recommendations.

Real-time inference versus batch inference. Real-time gives you an immediate answer, like a friend instantly replying to a text. Batch is like sending a long email to a whole team and waiting for the summary later. Real-time needs a live endpoint; batch crunches many records in one go.

To match the method to the use case, remember: if low latency is crucial, choose real-time. If your traffic varies and you want to save money, go serverless. If you have large payloads or lengthy processing, use asynchronous. If you can process everything in bulk without an immediate response, batch is best.

Practical AI Use Cases

AI can deliver value in several distinct ways. Think of these as the jobs you can hire AI to do. It can support human decisions, automate repetitive tasks, score or rank items, personalize experiences for individual users, predict future outcomes, generate new content like text or images, and even orchestrate multi-step actions through agentic automation.

A critical skill is mapping a business need to the right AI technique. This mapping is a major exam focus. For example, if you need to predict a number, like a house price, you reach for regression. If you need to predict a category, like spam or not spam, you use classification.

To group similar records together, like segmenting customers by behavior, use clustering. To find unusual behavior, like a fraudulent transaction that stands out from the norm, use anomaly detection. For understanding text, use natural language processing. For understanding images, use computer vision.

For recommending products, you use a recommendation system. To predict future values over time, like next month's sales, use forecasting or time-series modeling. To answer questions from a set of documents, use a retrieval-augmented knowledge base. To generate new content, use a foundation model or generative model. To automate a multi-step process, like handling a return, use an AI agent.

AI is powerful, but it is not always the right answer. You must recognize when a simpler solution is better. A good rule of thumb: if a simple deterministic rule can solve the problem perfectly, you do not need AI. If the exact same outcome is required every single time, AI's probabilistic nature makes it a poor fit.

AI is also inappropriate when you lack enough reliable data to train a model, when the cost of building and running the system outweighs the expected business value, or when the model cannot meet strict latency or availability requirements. If errors could cause unacceptable harm, or if the business demands a guaranteed outcome rather than a prediction, AI is not the right tool.

Traditional ML Versus Foundation Models

Let's compare traditional machine learning and foundation models. Think of traditional ML as a specialist, while a foundation model is more of a generalist.

Traditional ML — a model built and trained to do one specific task really well, like classifying spam or predicting a house price.

Foundation Model — a large, pre-trained model that can be adapted to many different tasks, like understanding language, generating text, or describing images.

How they differ. Traditional ML is task-specific. You train one model for fraud detection and a separate one for product recommendations. A foundation model is multi-task. The same underlying model can summarize a document, answer a question, or translate a sentence.

Another key difference: data strength. Traditional ML shines on structured data, the kind you find in spreadsheets and databases. Foundation models are built for unstructured data — language, images, audio, and video.

Cost and size also set them apart. Traditional ML models are often smaller and cheaper to train and run. Foundation models are typically much larger and more expensive, but they bring pre-trained capabilities that would be impossible to build from scratch.

When should you choose traditional ML? Go this route when the task is narrow and well-defined. You have a small dataset or specific domain expertise. You need predictable, explainable results, or the cost of a large model just isn't justified.

When is a foundation model the better choice? Pick a foundation model when you need deep language understanding or generation. This includes conversational interfaces, summarization, translation, or content creation. Also, when you want to use the same pre-trained model for multiple business needs, a foundation model is a strong fit.

For the exam, always match your model choice to the task, the data you have, your budget, and the requirements for predictability and speed. If the scenario mentions structured data, a small budget, or strict need for explainable results, lean toward traditional ML. If it highlights natural conversation, multiple language tasks, or leveraging pre-trained knowledge, think foundation model.

AI and ML Development Lifecycle: Define the Business Goal

Every AI or machine learning project starts with a clear business goal. This first phase is all about understanding why you're building the system and what success looks like.

Business goal definition — the initial phase where you identify the problem, expected business value, stakeholders, success metrics, cost, return on investment, and risk.

You need to know the problem you're solving, the value it will bring, who cares about the outcome, how you'll measure success, and whether the benefits outweigh the costs and risks.

MLOps Fundamentals

MLOps brings the kind of reliable, repeatable engineering practices used in software to the machine learning lifecycle. It’s about making the whole process of building, deploying, and maintaining models more robust and automated.

Experimentation — systematically tracking every run: what dataset was used, which model version, the hyperparameters chosen, and the resulting metrics. This ensures your work is reproducible and comparable.

Repeatable Processes — automating the steps in your pipeline, such as data preparation, training, evaluation, and deployment. This replaces error-prone manual handoffs.

Scalable Systems — designing your training and inference infrastructure to grow gracefully with your data volume and request traffic.

A model isn't production-ready just because it’s accurate. Truly production-ready systems also require reliable deployment, comprehensive monitoring, robust security, detailed logging, graceful failure handling, cost controls, version management, and a governance or approval process.

Technical Debt — the accumulation of hidden costs that occurs when machine learning systems rely on undocumented data, manual processes, untracked model versions, or fragile, tangled pipelines. You reduce this debt through automation, documentation, and rigorous testing.

Monitoring and Retraining — the practice of continuously watching for data drift, model drift, prediction quality, and bias. When the model’s performance degrades below requirements, it triggers a retraining cycle.

Several key SageMaker tools enable MLOps. Think of them as a toolkit, each with a specific job.

SageMaker Pipelines — a service for creating automated, repeatable machine learning workflows. It orchestrates every step from data processing to model deployment.

SageMaker Model Registry — a central catalog for tracking model versions, their metadata, and their approval status. It’s the single source of truth for which model version should be deployed.

MLflow with SageMaker — a capability for meticulously tracking experiments. It records the parameters you used, the metrics you observed, and the artifacts you created so nothing gets lost.

SageMaker Model Monitor — a service that continuously watches your deployed models. It detects deviations in data quality, model prediction quality, and feature importance over time.

SageMaker Clarify — a tool for detecting statistical bias in your data and models, and for generating explanations that make your model’s predictions more interpretable.

Traditional ML Evaluation Metrics

To understand these metrics, you need to know four basic counts. A true positive is a case your model correctly identifies as positive. A true negative is correctly identified as negative. A false positive is when the model says positive but the truth is negative. A false negative is when the model says negative but the truth is positive.

Accuracy — the share of all predictions that are correct. You look at the number of correct predictions, both positive and negative, divided by the total number of predictions. Use accuracy when your classes are fairly balanced, meaning there are roughly equal numbers of positive and negative examples.

Precision — out of all the items the model labeled as positive, how many really were positive. High precision means you have very few false positives. Prioritize precision when a false positive is more costly than a false negative.

A classic example where false positives hurt most is a spam filter. If a legitimate email is marked as spam, you lose an important message. That false positive is far worse than a spam message sneaking into your inbox.

Recall — out of all the actual positive items in your data, how many your model managed to find. High recall means you have very few false negatives. Prioritize recall when missing a positive case is expensive.

Think of disease screening. If the test is negative for a patient who actually has the disease, that false negative could be life-threatening. Here, recall is the metric that matters most.

How precision differs from recall. Precision cares about false positives; recall cares about false negatives. If wrongly flagging something as positive costs you more, choose precision. If overlooking a real positive costs you more, choose recall.

F1 score — a single number that balances precision and recall. It is the harmonic mean of the two, which means it will be low if either precision or recall is low. Use the F1 score when both metrics matter and your classes are imbalanced.