Introduction

Artificial intelligence is easiest to misunderstand when we begin with impressive examples: a chatbot answering questions, a model generating an image, a phone recognizing speech, a car detecting pedestrians, a fraud system flagging unusual transactions. These examples are real, but they are not the foundation. They are outcomes of a deeper idea.

At its core, artificial intelligence is the study and engineering of systems that perform tasks we associate with intelligence: perceiving information, representing situations, making predictions, choosing actions, learning from experience, and adapting under uncertainty. This broad view is close to the way modern AI textbooks frame the field: AI is not only about imitating humans, but also about building systems that act effectively or rationally in their environments (Russell & Norvig, 2021).

The purpose of this book is to build that idea from first principles.

A first principle is a basic idea that we do not merely memorize as a rule of thumb. We use it as a starting point for reasoning. For example, instead of beginning with “neural networks are powerful,” we will ask: What is a function? What does it mean to approximate a function from data? What is error? How can error be measured? How can a system adjust itself to reduce error? Only after those ideas are clear does a neural network become less mysterious.

This book is not a catalog of tools. Tools change. Libraries change. Model names change. What lasts longer is the structure underneath: data, representation, probability, optimization, generalization, evaluation, deployment, and human judgment.

Why artificial intelligence matters

AI matters because many important decisions and workflows can be expressed as problems of using information well.

A doctor may want help detecting signs of disease in an image. A bank may want to detect fraudulent transactions. A factory may want to predict equipment failure. A researcher may want to search thousands of documents. A teacher may want to adapt practice questions to a learner’s progress. A software team may want assistance writing, testing, and reviewing code.

In all these cases, the system must connect input to output.

An input is information given to a system. An image, a sentence, a table row, a sound recording, and a sensor reading can all be inputs. An output is what the system produces: a class label, a probability, a generated paragraph, a recommendation, a control action, or a warning.

For example:

• Input: an email message.
  Output: “spam” or “not spam.”

• Input: a customer’s past purchases.
  Output: product recommendations.

• Input: a chest X-ray image.
  Output: a probability that a certain abnormality is present.

• Input: a user’s written question.
  Output: a helpful answer, perhaps with citations or tool calls.

The simple input-output view is powerful, but incomplete. AI systems do not merely transform inputs into outputs in isolation. They are built by people, trained on data, evaluated by metrics, deployed into organizations, and used by humans with goals, incentives, and constraints. An AI system that is technically impressive but unsafe, unfair, unreliable, or economically inappropriate is not a successful system.

That is why this book treats AI as both a scientific and engineering discipline: we study what is possible, then learn how to build systems that behave reliably enough for real use.

A short orientation: intelligence as information processing

The phrase “artificial intelligence” became established as a research ambition in the 1950s, especially around the Dartmouth proposal organized by John McCarthy and colleagues, which described the goal of studying intelligence in a way that could be simulated by machines (McCarthy et al., 1955). Even before that, Alan Turing asked whether machines could exhibit behavior that would count as intelligent in conversation, opening a practical way to discuss machine intelligence without requiring a perfect definition of human thought (Turing, 1950).

In this book, we will use a practical working idea:

An AI system is a computational system designed to perceive information, represent relevant structure, learn or reason from that information, and produce useful outputs or actions under uncertainty.

This sentence contains several important terms.

Computational means that the system operates through explicit procedures that can be carried out by a computer. These procedures may be hand-written rules, learned numerical parameters, search algorithms, statistical estimators, or combinations of these.

Perceive means to receive and process information from the world or from a data source. A camera-equipped robot perceives through images. A speech recognition system perceives through audio. A document search system perceives through text.

Represent means to encode information in a form the system can use. For example, a word may be represented as a number, a vector, or a token. A customer may be represented by features such as age range, purchase history, and account activity. A chess position may be represented as the arrangement of pieces on a board.

Learn means to improve performance by using experience or data. Tom Mitchell’s influential textbook definition says that a computer program learns from experience with respect to a class of tasks and a performance measure if its performance at those tasks improves with experience (Mitchell, 1997). This definition is useful because it forces us to name three things: the task, the experience, and the performance measure.

Reason means to draw conclusions or choose steps based on available information. Reasoning may be symbolic, as in logic and planning, or statistical, as in updating beliefs from evidence.

Uncertainty means that the system does not have perfect information. Most real AI problems involve uncertainty. A medical image may be ambiguous. A user’s request may be underspecified. A sensor may be noisy. A future market price is unknown. AI systems often work not by knowing with certainty, but by estimating, ranking, predicting, and deciding despite incomplete information.

This is one of the central themes of the book: modern AI is not magic certainty. It is structured uncertainty management.

Programmed behavior and learned behavior

To understand AI clearly, we need to distinguish programmed behavior from learned behavior.

In programmed behavior, a human writes explicit instructions. For example:

If the transaction amount is greater than $10,000 and the country is outside the customer’s usual region, flag the transaction for review.

This rule may be useful, but it depends on a person anticipating the pattern. If fraud patterns change, the rule may fail.

In learned behavior, the system adjusts itself using data. Suppose we have many examples of past transactions labeled as fraudulent or legitimate. A machine learning model can search for patterns that help predict the label. It may discover that fraud risk depends not only on amount and country, but also on time of day, merchant category, device fingerprint, account age, and interaction among these factors.

This does not mean learned systems are automatically better. A learned model can learn misleading patterns, reproduce historical bias, fail on new conditions, or perform well on average while failing badly for a small group. The point is narrower and more precise: learned systems can infer patterns from examples rather than relying only on manually written rules.

This distinction will return throughout the book. It is the difference between telling a machine exactly what to do and designing a process by which the machine adjusts its behavior from evidence.

The practical shape of an AI problem

Most AI projects can be understood through a small set of questions.

First, what is the task? A task is the job we want the system to perform. Examples include classification, regression, ranking, generation, detection, clustering, and control.

Second, what is the data? Data are recorded observations. Data may be numbers, images, text, sound, video, logs, human labels, or interaction histories. Data are never neutral raw truth. They are produced by measurement processes, business processes, sensors, people, and historical choices.

Third, what is the model? A model is a simplified computational structure that maps inputs to outputs. A linear regression model, a decision tree, a neural network, and a large language model are all models, though they differ greatly in complexity.

Fourth, what is the objective? An objective is what the training process tries to improve. In machine learning, this is often expressed through a loss function, which assigns a numerical penalty to wrong or undesirable outputs. Training means adjusting the model to reduce that loss.

Fifth, how do we evaluate the system? Evaluation asks whether the system performs well on cases it has not already seen. This matters because memorizing past examples is not the same as learning a pattern that works in the future.

Sixth, what happens when the system is used in the real world? Deployment introduces latency, cost, monitoring, security, user behavior, legal duties, and social consequences. A model that looks good in a notebook can still fail as a product, service, or organizational process.

These questions form the practical backbone of the book.

AI is not one thing

The term “AI” names a family of methods and systems, not a single technique.

Machine learning is the part of AI concerned with systems that improve from data or experience. A spam filter trained from labeled examples is a machine learning system.

Deep learning is a subfield of machine learning that uses neural networks with many layers to learn representations from data. Deep learning has become especially important in vision, speech, language, and generative systems; standard treatments describe it as representation learning through layered models trained by optimization methods such as gradient-based learning (Goodfellow et al., 2016).

Generative AI refers to systems that create new content, such as text, images, audio, video, molecules, or code. A language model that writes a paragraph and an image model that produces a picture from a prompt are generative systems.

Reinforcement learning studies agents that learn by interacting with an environment and receiving rewards. A game-playing agent, for example, may improve by trying actions, observing outcomes, and adjusting its policy; this agent-environment-reward framing is central to reinforcement learning (Sutton & Barto, 2018).

An agent is a system that observes an environment and chooses actions. A thermostat is a simple agent: it senses temperature and turns heating or cooling on or off. A warehouse robot is a richer agent: it perceives surroundings, plans movement, avoids obstacles, and acts physically. A tool-using language model can also be treated as an agent when it observes a user request, calls tools, receives results, and decides the next step.

These categories overlap. A modern AI assistant may use deep learning for language, retrieval to access documents, tools to perform actions, rules for safety, and human feedback for improvement.

What AI can and cannot do

AI systems can be remarkably effective when a task has patterns that can be captured from data, when success can be measured, and when errors can be tolerated or controlled. They can classify millions of examples, search large spaces, detect subtle statistical patterns, translate text, generate drafts, summarize documents, and support human decision-making.

But AI systems do not automatically understand the world in the way humans do. A model may produce fluent language without having reliable knowledge of whether its statement is true. A classifier may achieve high average accuracy while failing on rare but important cases. A recommendation system may optimize clicks while harming user well-being. A planning agent may find a technically valid path that violates unstated human expectations.

This is why evaluation, monitoring, and governance are not optional. The National Institute of Standards and Technology emphasizes that AI risk management requires attention to validity, reliability, safety, security, resilience, accountability, transparency, explainability, privacy, and fairness across the system lifecycle (NIST, 2023).

A mature view of AI avoids both extremes. AI is not mere hype, and it is not an all-purpose replacement for human judgment. It is a powerful set of methods for building systems that use information to produce outputs or actions. Whether that power is useful depends on problem framing, data quality, modeling choices, evaluation, deployment discipline, and responsibility.

How this book will build understanding

The chapters ahead follow a deliberate path.

We begin by defining AI and the computational view of intelligence. Then we build the mathematical foundation: functions, vectors, matrices, probability, statistics, optimization, gradients, and loss functions. These are not included for decoration. They are the language in which modern AI systems are designed and evaluated.

Next we study data, features, measurement, and generalization. This is where many practical AI projects succeed or fail. A sophisticated algorithm trained on poor data can produce poor results. A simple baseline trained and evaluated carefully can be more valuable than a complex model used carelessly.

We then examine supervised learning, unsupervised learning, optimization, neural networks, deep learning architectures, natural language processing, large language models, and generative AI. These chapters explain not only what the methods are, but why they work, when they fail, and how to reason about trade-offs.

Later chapters turn toward building complete AI systems: retrieval, tools, agents, reinforcement learning, end-to-end development, MLOps, safety, security, robustness, ethics, law, and social impact. This order is intentional. It is not enough to know how to train a model. You must also know how to decide whether a model should be built, how to evaluate it, how to deploy it safely, and how to keep learning as the field changes.

The attitude to bring

The best attitude for learning AI is disciplined curiosity.

Be curious enough to ask: What can this system do? What new product, service, analysis, or scientific tool becomes possible?

Be disciplined enough to ask: What data trained it? What does the metric measure? What are the failure modes? Who is affected by errors? What assumptions are hidden? What would count as evidence that this system works?

Whenever you encounter a new AI claim, practice translating it into first-principles questions:

• What are the inputs and outputs?
• What pattern is the system trying to learn?
• What data provide the evidence?
• What objective is being optimized?
• What uncertainty remains?
• How is performance measured?
• What happens when the environment changes?
• What are the costs of being wrong?

If you can ask these questions clearly, you are already thinking like an AI practitioner.

The goal of this book is not to make AI seem simple. It is to make it understandable. Some ideas will require patience. Some equations will need rereading. Some methods will seem abstract before they become practical. That is normal. The field is large, but its foundations are learnable.

We now begin with the first question: what, exactly, is artificial intelligence?

References

Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.

McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.

Mitchell, T. M. (1997). Machine Learning. McGraw-Hill.

National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1.

Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.

Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction (2nd ed.). MIT Press.

Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.