AIF-C01 Study Platform
← All lessons
Lesson 1 · Domain 1: Fundamentals of AI and ML (20%)

The Core Vocabulary: AI, ML, Deep Learning, GenAI, and Agentic AI

Task 1.1 — Explain basic AI concepts and terminologies
Take the quiz for this lesson →
How AI, ML, Deep Learning, GenAI, and Agentic AI relate Four nested circles: Artificial Intelligence is the largest outer circle, containing Machine Learning, which contains Deep Learning, which contains Generative AI. A separate connected box shows Agentic AI as systems built on top of generative AI models plus tools, memory, and planning. Where each term sits Artificial Intelligence (AI) Machine Learning (ML) Deep Learning (DL) Generative AI (GenAI) Agentic AI GenAI model + tools, memory, planning Each ring is a subset of the one around it. Agentic AI is built using GenAI, not a separate ring.
Evolution of handling a customer support email, from rule-based AI to agentic AI A five-step horizontal timeline showing the same task, handling a customer support email, solved by five successive approaches: rule-based AI, machine learning classifier, deep learning, generative AI drafting a full reply, and agentic AI acting end to end. Same task, five eras: "handle this customer email" AI hand-written if/then rules "if subject has 'refund', route to X" ML trained spam classifier learns from 1000s of labeled emails Deep Learning neural net reads full email nuance catches sarcasm, tone, context GenAI LLM drafts a full personalized reply writes new text, not just a label Agentic AI checks order DB, drafts + sends reply plans multiple steps, acts without asking
Linear regression versus gradient boosting Left panel shows linear regression as a scatter plot of data points with a single straight best-fit line through them. Right panel shows gradient boosting as a sequence of three small decision trees, each correcting the errors left by the previous tree, combining into one final prediction. Linear Regression one straight line, fit once House size (sq ft) → Price Gradient Boosting many small trees, fit in sequence Tree 1 rough guess Tree 2 fixes errors Tree 3 fixes what's left Combined final prediction
End-to-end pipeline for building and using an AI model, from raw data to inferencing A seven-step vertical pipeline: Collect data, Pick a learning type, Choose an algorithm, Train to produce a model, Evaluate fit and fairness, Deploy, Run inferencing. From raw data to a working AI system 1 Collect data Data types: structured / unstructured labeled / unlabeled, time-series 2 Pick a learning type Learning types: supervised / unsupervised / reinforcement 3 Choose an algorithm Examples: linear regression, gradient boosting, neural network 4 Train → produces a model Recall: algorithm = the method model = the trained result 5 Evaluate fit & fairness Check for: overfitting, underfitting, bias / fairness issues 6 Deploy the model e.g. Bedrock, SageMaker AI endpoint, managed API service 7 Run inferencing Inferencing types: batch / real-time / asynchronous / serverless Loop back: monitor performance in production, retrain periodically (MLOps).

The hierarchy: AI, ML, DL, GenAI, Agentic AI

Artificial Intelligence (AI) is the broadest term: any technique that lets a computer mimic human-like intelligence. AI is a category, not a specific technology — hand-coded rule systems and modern chatbots both count.

Machine Learning (ML) is a subset of AI: instead of hand-written rules, the system learns patterns from examples. The output is a model — a function shaped by data. Shaping it is training; using the finished model on new data is inferencing.

An algorithm is the general method used to train a model (e.g., linear regression, gradient boosting). The model is the specific trained artifact produced by applying an algorithm to data.

Deep Learning (DL) is a subset of ML using neural networks — layered structures loosely inspired by neurons. It powers computer vision (interpreting images/video) and NLP — Natural Language Processing (understanding/generating human language).

Generative AI (GenAI) is a subset of deep learning: models sophisticated enough to generate new content rather than just classify or predict a number. The engines behind most GenAI are Large Language Models (LLMs) — deep learning models trained on huge text corpora that predict 'what comes next' well enough to write, answer, and converse.

Agentic AI is NOT nested inside GenAI — it's drawn as a connected box. It's a GenAI model wrapped with tools, memory, and multi-step planning: given a goal, it plans, acts, observes, and adjusts across multiple steps autonomously.

Bias, fairness, fit

  • Bias — systematic skew in predictions, often from unrepresentative training data.
  • Fairness — the broader goal that outcomes don't unfairly disadvantage particular groups.
  • Overfitting — model memorized training noise; great on training data, poor on new data.
  • Underfitting — model never learned the pattern; poor on both.

The three learning types

  • Supervised learning — trained on labeled data (every example has the correct answer attached).
  • Unsupervised learning — no labels at all; finds structure/groups on its own (e.g., clustering).
  • Reinforcement learning — no fixed dataset; an agent takes actions and learns from reward/penalty feedback over time (later refined via RLHF — Reinforcement Learning from Human Feedback for LLMs).

Data types

Labeled vs. unlabeled; structured/tabular (rows and columns) vs. unstructured (images, audio, free text); time-series (sequential, order matters — e.g., IoT sensor data, GPS pings).

Data type, learning type, technique, and inferencing type are independent axes the exam tests separately — don't assume one implies another.

Types of inferencing

  • Batch — many inputs processed together on a schedule, nobody waiting live.
  • Real-time — single request, instant response, someone waiting live.
  • Asynchronous — single request submitted, response delivered later.
  • Serverless — an infrastructure/billing classification (no servers to manage, pay per use, scales to zero when idle) — independent from the timing classifications above.

Algorithm examples: Linear Regression vs. Gradient Boosting

Linear regression fits one straight line through a scatter of data points, minimizing overall distance to all of them. Simple, fast, highly explainable — used for regression problems (predicting a continuous number). AWS SageMaker built-in: Linear Learner.

Gradient boosting builds many small decision trees in sequence — each new tree corrects the residual errors of the previous ones; the final prediction combines them all. More accurate on complex tabular data, less explainable. AWS SageMaker built-in: XGBoost.

The customer-support email story (same task, five eras)

  1. AI — hand-written if/then rules, no learning at all.
  2. ML — a spam classifier trained on thousands of labeled emails (supervised learning).
  3. Deep Learning — a neural net reads the full email and catches nuance/tone a shallow model misses.
  4. GenAI — an LLM drafts a full personalized reply from scratch (creating text, not picking a label).
  5. Agentic AI — the system checks the order database, drafts, and sends the reply autonomously, planning multiple steps without being told each one.

Ask the AI tutor about this lesson

Ask a follow-up question about this lesson — the tutor answers using this exact lesson's content as context.