AI Fundamentals

Artificial Intelligence refers to the branch of computer science focused on building systems that perform tasks normally associated with human thinking. This includes recognizing speech, identifying objects in images, making decisions from incomplete information, and generating language. The field traces back to a foundational workshop at Dartmouth College, where researchers including John McCarthy coined the term itself and proposed that every aspect of learning could in principle be simulated by a machine^[1]. That founding ambition, building a machine that reasons the way a person does, still shapes how the public imagines AI today, even though most real systems fall far short of it.

A useful way to separate the hype from the reality is to distinguish between two categories: Narrow AI and General AI. Narrow AI describes any system built to do one job, and only that job, extremely well. A chess engine that can defeat any human grandmaster cannot fold laundry or hold a conversation, because it was never built to do those things. Virtually every AI system in commercial use today, from spam filters to recommendation engines, falls into this narrow category.

General AI, sometimes called artificial general intelligence, describes a hypothetical system capable of understanding and learning any intellectual task a human can, transferring knowledge fluidly between domains. No such system exists yet, despite frequent claims in popular media. Keeping this distinction clear prevents a common confusion: a chatbot that writes convincing sentences is still narrow, tuned for language generation, not a general reasoning mind that understands the world the way a person does.

The callout below captures this distinction in a form you can return to whenever a news headline overstates what an AI system can do.

Consider three familiar examples to see the narrow category in action. A voice assistant that sets a timer and answers trivia questions seems flexible, but each of its skills was individually built and trained by engineers, then stitched together behind one interface. A fraud-detection system at a bank flags suspicious transactions by finding patterns in past data, but it has no concept of money, trust, or crime beyond the numbers it was shown. A photo app that automatically tags faces recognizes patterns of pixels associated with a name, without any understanding of who that person actually is. Each system looks intelligent within its lane, and useless outside it.

This narrow-but-effective quality explains why AI has spread so quickly through industries in the last decade. Companies do not need a machine that understands everything, they need a machine that reliably automates one costly, repetitive task. A narrow system trained to read chest X-rays for signs of pneumonia can process thousands of images a day, catching patterns a tired radiologist might miss late in a shift. That narrowness is a feature, not a limitation, when the goal is efficiency rather than philosophical achievement.

It also explains why so many AI failures happen at the edges of a system's training. A self-driving car trained mostly on sunny California roads can struggle badly with snow it has rarely seen, because its skill was built from specific examples rather than general understanding. This fragility connects directly to Bias in AI, a concept explored further in the vocabulary section, where the data a system learns from directly shapes and limits what it can do well.

A second useful lens for understanding artificial intelligence is to ask what problem a given system is solving, rather than how "smart" it sounds. Some systems classify things into categories, such as sorting an email as spam or not spam. Others predict a number, such as estimating a house price from its size and location. Still others generate new content, such as producing an image or a paragraph of text, a category known as Generative AI that later chapters explore through language and vision applications. Naming the task type quickly cuts through marketing language.

The distinction between artificial intelligence as a broad field and Machine Learning as a specific approach within that field trips up many newcomers. Early AI research in the 1950s and 1960s often relied on hand-written logical rules, where programmers explicitly coded "if this, then that" instructions for every scenario. Machine learning, developed and popularized through work by early researchers and expanded dramatically since, took a different route, letting systems find their own patterns from examples rather than following rules a human wrote out in advance^[2]. Nearly all AI systems built today use this learning-based approach rather than the older rule-based style.

Understanding artificial intelligence as a field, then, means holding two ideas at once. First, AI is the overall ambition of building machines that perform intelligent tasks, an ambition that stretches back nearly seventy years and is still far from its most ambitious form. Second, almost everything called AI in practice today is a narrow, task-specific system trained through machine learning on examples rather than programmed with explicit rules. Keeping both ideas in view prevents the two most common mistakes: dismissing AI as "just math" with nothing impressive about it, and overestimating a chatbot or image generator as something approaching human understanding.

Machine Learning is the technique that makes most modern AI possible, and it deserves a precise definition before anything else. It is a method of building systems that improve their performance on a task by being exposed to examples, rather than by following instructions written line by line. A spam filter built with machine learning does not contain a rule that says "block emails with the word free in all caps." Instead, it has seen thousands of labeled emails and learned, on its own, which patterns of words and formatting tend to appear in spam versus legitimate mail.

That example points directly to the next essential term: Training Data. This is the collection of examples a machine learning system studies in order to learn its task, and its quality shapes everything the system can later do. If a spam filter's training data only ever included English-language emails, it would likely perform poorly on spam written in French or Japanese, because it never saw examples in those languages. This dependency on training data is not a minor footnote, it is arguably the single most important limiting factor in how well any AI system performs in the real world.

Once a machine learning system has been trained, the result is called a Model. A model is the trained artifact itself, a set of learned patterns stored in a form the computer can use to make decisions on new data it has never seen before. Think of a model as a recipe distilled from thousands of cooking attempts. It is not the ingredients (the training data) and it is not the act of cooking (the training process), it is the finished, reusable set of instructions that resulted from that process.

Using a trained model to make a decision on new data is called Inference. If training is the process of a student studying for years, inference is the moment that student sits down and answers an actual exam question. When a photo app instantly recognizes a face in a picture just taken, that instant recognition is inference, the model applying what it learned during training to something brand new. Training typically happens once, often taking hours, days, or even weeks on powerful hardware, while inference happens repeatedly and quickly, often in a fraction of a second, every time the system is used.

Machine learning itself splits into three major learning styles, and distinguishing them clearly resolves a huge amount of confusion found in casual AI discussions. The table below lays out how each style differs in terms of what data it needs and what kind of problem it solves.

Supervised Learning is the most common style in commercial use, and it works by showing a model pairs of inputs and correct outputs until it can predict the output for a new input on its own. A supervised system for predicting house prices is trained on past sales, each one paired with its actual sale price, so the model learns the relationship between features like square footage and the final number. Nearly every classification and prediction task discussed later in this book, including much of computer vision and sentiment analysis, is built using supervised learning at its foundation.

Unsupervised Learning takes the opposite approach, working with data that has no correct answer attached at all. Instead of predicting a known label, an unsupervised system looks for hidden structure or grouping within the data itself. A retailer might feed years of purchase history into an unsupervised system to discover that customers naturally cluster into groups, such as bulk buyers versus occasional browsers, without ever telling the system in advance what those groups should be. This makes unsupervised learning especially useful for exploration, when the goal is to discover unknown patterns rather than confirm a known outcome.

Reinforcement Learning works differently still, training a system through trial and error using rewards and penalties rather than a fixed set of labeled examples. An agent takes an action, observes the result, and receives a signal indicating whether that action was good or bad, gradually learning a strategy that maximizes reward over time. This approach became famous through DeepMind's AlphaGo system, which learned to play the board game Go at a superhuman level partly through reinforcement learning against itself, defeating a world champion player^[3]. Reinforcement learning suits problems involving sequences of decisions, such as robotics or game-playing, where the "right answer" only becomes clear after a series of actions plays out.

Deep Learning is a specific and highly influential subset of machine learning, built around structures called neural networks that stack many processing layers on top of each other. The word "deep" refers to that stacking, dozens or even hundreds of layers, each one transforming the data a little further before passing it to the next. This layered structure is what allows deep learning systems to handle messy, high-dimensional data like images, audio, and raw text far more effectively than older machine learning techniques could. The internal mechanics of how these layers actually function, including concepts like weights and activation functions, belong to the discussion of neural networks and algorithms in a later chapter.

It is worth pausing on why deep learning became so dominant only in the last fifteen years or so, despite the underlying mathematical ideas dating back decades earlier. Two ingredients had to arrive together: enormous amounts of digital training data, generated by the internet and smartphones, and enormous amounts of cheap computing power, largely enabled by graphics processing hardware originally built for video games. Without both pieces in place at once, even a brilliant algorithm cannot perform well, which is why many older AI ideas remained theoretical curiosities until relatively recently.

Bias in AI describes a systematic and often unintended skew in a model's behavior, usually traced back to patterns or gaps in its training data. If a hiring model is trained mostly on résumés from past successful employees at a company with a historically homogeneous workforce, it can end up favoring candidates who resemble that existing group, regardless of actual qualifications. This is not a matter of a computer holding conscious prejudice, it is a mathematical reflection of imbalances baked into the examples it was shown. Recognizing bias early requires actively auditing training data and model outputs, rather than assuming a system is neutral simply because it is built from numbers rather than opinions.

A well-known real-world instance of this problem occurred when a facial analysis system was found to perform significantly worse at identifying darker-skinned faces compared to lighter-skinned ones, a gap traced to training datasets that overwhelmingly featured lighter-skinned individuals^[4]. This example illustrates why training data quality is not a side issue but a central ethical and technical concern running through every application discussed later in this book. A model can only reflect the world it was shown, and if that world was incomplete or skewed, the model's decisions will carry that same skew forward into every prediction it makes.

Bringing these terms together with a concrete scenario helps cement them. Picture a company building a system to recognize handwritten digits on mailed checks. It gathers thousands of scanned digit images, each labeled with the correct number, that collection is the training data. It chooses supervised learning, since each example comes with a known correct answer attached. It builds a deep learning system, layering multiple processing stages to handle the visual complexity of handwriting. After training finishes, the resulting model is deployed at a bank, where it performs inference on every new check that arrives, reading digits in a fraction of a second. If the training images mostly came from one style of handwriting, bias could creep in, causing the model to misread checks written in an unfamiliar style.

The checklist below offers a quick way to test your grasp of these nine core terms before moving into how the underlying neural networks that power deep learning actually function.

These nine terms, artificial intelligence, machine learning, deep learning, training data, model, inference, supervised learning, unsupervised learning, and reinforcement learning, along with narrow AI, general AI, and bias in AI, form a vocabulary set that will resurface constantly. When the next chapter opens up the internal mechanics of neural networks, layers, weights, and the mathematics of learning, these words will already feel familiar rather than foreign. That familiarity is the entire purpose of building this foundation first, so the more technical material ahead has something solid to stand on.