Chapter 1 of 25
What Is Artificial Intelligence, Really?
Separating the hype from the actual idea
Motivation
Ask ten people what "artificial intelligence" means and you'll get ten different answers — a chatbot, a robot butler, a villain from a sci-fi movie, or maybe just "the thing that recommends videos I somehow always want to watch." That confusion isn't your fault. AI has been simultaneously overhyped by marketing and undersold by decades of research spread across wildly different techniques. Before this course can teach you how to build a neural network, it owes you an honest answer to a more basic question: what actually is this field, and where does the thing you're about to build fit inside it?
A Real-Life Scenario
Imagine you're a radiologist looking at a chest X-ray. You've spent years learning to spot a faint shadow that might indicate an early-stage tumor — a pattern so subtle that it took your brain thousands of training examples (other X-rays, textbooks, mentors pointing things out) to learn to recognize reliably. Now imagine trying to write that skill down as a step-by-step rulebook for a computer: "if there's a roughly circular shadow of density greater than X in the upper-left lobe, flag it." You'd fail almost immediately — the real pattern isn't a clean rule, it's a fuzzy statistical regularity your visual cortex absorbed from experience.
This is precisely the gap AI exists to close. Instead of a human trying to write down the rule, we show a system thousands of labeled X-rays and let it discover the pattern itself — the same way you did, just faster and at a scale no single radiologist could match.
Building the Intuition
Start with the plainest possible definition, then sharpen it.
Artificial Intelligence is the field of building machines that perform tasks which, when a human performs them, we'd say require intelligence — perception, reasoning, learning, planning, and decision-making.
Notice what this definition does not say. It doesn't say AI means "robots," "consciousness," or "human-like thinking." It's defined entirely by the tasks it tackles — recognizing a face, understanding a sentence, playing a strong game of chess, diagnosing a disease from an image, driving safely through traffic. Anything that convincingly does one of these counts as AI, regardless of how it works under the hood.
That "regardless of how it works" part matters a lot, because AI is not one algorithm or one technique. It's an umbrella spanning more than seven decades of very different approaches — some built from explicit, hand-written rules, others built entirely from statistical patterns discovered in data. The Multi-Layer Perceptron, the model this entire course builds toward, lives in the second camp, in a particular corner of it called deep learning, which the next two chapters define precisely.
Why does AI exist at all?
Human brains are strange machines. They're astonishingly good at things that feel effortless — recognizing a friend's face in a crowd, following a conversation in a noisy room, catching a ball on the run — yet those same tasks stumped computers for decades. Meanwhile, computers are effortlessly good at things brains find painfully slow: multiplying ten-digit numbers, or scanning a million database records in milliseconds.
AI exists to close that gap in both directions — to give machines some of the flexible, pattern-recognizing intelligence that comes naturally to us, while keeping the tireless precision that comes naturally to them. Get this right, and you can automate work that used to require a trained human, freeing people for the things that still need creativity, judgment, and ethical reasoning.
Why does it matter so much?
Once a human expert learns a skill, applying it to a new case still takes their time, one case at a time. Once a machine learns a pattern, applying it to the millionth new case costs almost nothing extra. That's the single biggest reason AI matters: it doesn't just replicate intelligence, it makes intelligence scale.
A model that reads medical scans can flag warning signs across a hospital system faster than any team of radiologists working around the clock. A translation model can remove language barriers for millions of people simultaneously. A predictive maintenance model can watch every machine on a factory floor at once, something no human inspector could do. This is the lever AI pulls: not more intelligence, but intelligence applied at a scale that was never available before.
A Brief Timeline (the short version)
The detailed history — with all its twists, winters, and revivals — gets its own chapter later. For now, here's just enough of the shape of it to understand why the field looks the way it does today.
The First Mathematical Neuron
McCulloch and Pitts show that simple binary units, wired together, can compute logical functions — the first hint that 'thinking' might be reducible to computation.
The Turing Test
Alan Turing proposes a practical criterion for machine intelligence: can a machine's conversation be mistaken for a human's?
The Field Gets a Name
The term 'Artificial Intelligence' is coined at the Dartmouth Conference, widely treated as the founding moment of the field.
The Perceptron
Frank Rosenblatt builds the first algorithm that learns its own weights from data instead of having them hand-set — the direct ancestor of everything in this course.
A Sobering Proof
Minsky and Papert prove the limits of a single-layer perceptron, triggering a long slowdown in neural network research.
Backpropagation Popularized
An efficient way to train multi-layer networks revives the field and makes the Multi-Layer Perceptron practical.
Deep Learning Takes Over
AlexNet's dominant win at the ImageNet competition kicks off the modern deep learning era in computer vision.
Large Language Models
The Transformer architecture extends deep learning's reach into natural language at massive scale.
Categories of AI
People classify AI along two mostly independent axes: what it can do (capability) and how it does it (approach). Both are useful mental filing cabinets.
By capability
| Category | Description | Exists Today? |
|---|---|---|
| Narrow AI (Weak AI) | Built to do one task, or a narrow family of related tasks — spam filtering, face recognition, chess. | Yes — every deployed AI system today is narrow AI. |
| General AI (Strong AI) | A hypothetical system that could learn and apply intelligence across any task a human can do. | No — purely hypothetical. |
| Super AI | A hypothetical system whose intelligence would exceed humans across virtually every domain. | No — a topic for philosophy, not engineering, right now. |
Every application you'll build in this course, including everything powered by an MLP, is narrow AI. It might be excellent at its one job, but it has no idea how to do anything else.
By approach
| Approach | Description |
|---|---|
| Symbolic AI (rule-based) | A human expert writes explicit rules and logical statements by hand. Nothing is learned from data automatically. |
| Machine Learning | Intelligence emerges from statistical patterns discovered automatically in data, rather than being hand-coded. Covered in depth next chapter. |
| Deep Learning | A subfield of machine learning that uses layered neural networks — including the MLP — to learn hierarchical representations directly from raw or lightly processed data. |
Deep Learning is a subset of Machine Learning, and Machine Learning is a subset of Artificial Intelligence. Every deep learning system is a machine learning system, and every machine learning system is an AI system — but the reverse doesn't hold. Plenty of AI (like a hand-written chess rulebook) involves no learning at all.
Where This Already Shows Up
Healthcare
Flagging abnormalities in X-rays, predicting patient risk from health records, recommending personalized treatment plans.
Finance
Credit scoring, fraud detection, algorithmic trading, and automated risk assessment.
Cybersecurity
Intrusion detection systems that learn what 'normal' network traffic looks like, so they can spot what isn't.
Recommendations
The product, movie, and content suggestions on nearly every e-commerce and streaming platform.
Manufacturing
Predictive maintenance that forecasts equipment failure before it happens.
Robotics
Perception and control systems that let robots navigate and manipulate the physical world.
Natural Language Processing
Machine translation, sentiment analysis, and conversational assistants.
Later chapters return to several of these applications with a specific focus on how an MLP, in particular, is used to solve them.
Key Takeaways
- AI is the field of building machines that perform tasks normally thought to require human intelligence — it's defined by the tasks, not by any particular technique. - AI matters because it lets intelligence scale: once a pattern is learned, applying it to new cases costs almost nothing extra. - AI can be classified by capability (Narrow, General, Super — only Narrow AI exists today) and by approach (Symbolic, Machine Learning, Deep Learning). - Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence — the Multi-Layer Perceptron sits at the innermost, most specialized level, while remaining one of the most foundational architectures in the whole field. - Key milestones — the McCulloch-Pitts neuron (1943), the Perceptron (1958), and backpropagation (1986) — form a direct lineage leading to the MLP you're about to learn in detail.
Common Mistakes
Popular media loves to conflate any AI system with human-like general intelligence. Every deployed system today — including every one you'll build in this course — is narrow AI: excellent at one job, clueless about everything else.
Symbolic, rule-based systems are AI too, even though they involve no learning at all. Machine learning — and deep learning within it — is only one branch of a much older tree.
Interview Corner
Where This Shows Up in Practice
Nearly every AI success story you've heard about — from medical imaging tools to translation apps to fraud detection systems — is narrow AI built on machine learning, and an enormous fraction of those are built on deep learning specifically. Understanding where the Multi-Layer Perceptron sits in this hierarchy is the first step toward understanding why it's worth an entire course: it's simple enough to fully understand end-to-end, yet it's the direct ancestor of the architectures running some of the most consequential software in the world today.