The Thing in the Room
You have probably used artificial intelligence today. Maybe you asked a question and got an answer that surprised you with its fluency. Maybe a streaming service recommended something you actually wanted to watch. Maybe your email filtered out spam before you saw it, or your phone unlocked when it recognized your face.
All of that is AI. But those four things work in completely different ways, built on completely different principles, solving completely different problems. The word "artificial intelligence" has become a container that holds almost everything, which means it increasingly explains almost nothing.
This article is an attempt to change that — to give you a clear, honest map of what AI actually is, how the most powerful current form of it works under the hood, and what the different types of AI are and what each one is actually good for. Not to impress you with terminology. To give you a framework that makes the next ten years of news about AI genuinely legible.
What Artificial Intelligence Actually Means
The simplest definition of artificial intelligence is a system that performs tasks that would normally require human intelligence to perform. Recognizing a face. Translating a language. Recommending a product. Writing a sentence. Diagnosing an image.
That definition is useful but incomplete, because it describes a moving target. Fifty years ago, a program that could play checkers was considered AI. Today, checkers-playing software is so trivial it ships inside wristwatches. The field has a long-running joke about this: once a problem gets solved, it stops being called AI and starts being called software. What we call AI is always the frontier — the things machines can almost but not quite do as well as humans.
What makes modern AI different from the software of previous decades is not that it follows more rules. It's that it learns its own rules from data. Traditional software is a set of instructions written by humans: if this happens, do that. AI systems, at least the kind that matter most today, are trained on examples and learn to recognize patterns well enough to generalize — to handle situations nobody explicitly programmed them for.
That distinction — learned patterns versus written rules — is the conceptual key to understanding almost everything that follows.
How Large Language Models Work
The most consequential AI systems right now are large language models — the technology behind ChatGPT, Claude, Gemini, and others. Understanding how they work, even roughly, changes how you think about what they can and cannot do.
Start with the training data. A large language model begins as a system with no knowledge of anything. It is then trained on an enormous quantity of text — articles, books, websites, code, conversations, scientific papers — spanning hundreds of billions or trillions of words. The model reads this text not to memorize it, but to learn the statistical patterns of how language works. Which words tend to follow which other words. Which concepts tend to appear together. Which sentence structures are used to express which kinds of ideas.
Then come the tokens. The model doesn't read text the way you do, word by word. It breaks text into chunks called tokens — roughly, pieces of words or short common words. The word "understanding" might be one token. The word "uncharacteristically" might be broken into three or four. Every token gets represented as a point in a vast mathematical space, and the model learns how tokens relate to each other — which ones cluster together, which ones imply which others, how the meaning of a token shifts depending on what surrounds it.
The transformer architecture. The specific design that makes modern language models work is called a transformer. The key innovation of the transformer is something called attention — the ability to weigh which parts of a sentence, or a paragraph, or an entire conversation, are most relevant to understanding the current word or token being processed. When you write a long, complex prompt, the model isn't just reading the last few words. It is attending to the entire context, weighing relevance across the whole thing, and generating a response shaped by all of it simultaneously. This is what makes modern AI feel qualitatively different from earlier chatbots, which could only remember the last sentence or two.
What the model is actually doing when it responds. When you type a question and a language model answers, it is not retrieving a stored answer from a database. It is predicting, one token at a time, what the most plausible next token is given everything that came before — your question, the conversation history, its training. Each token it generates becomes part of the context for the next one. The result is text that is statistically coherent with the input and with the patterns of language the model absorbed during training.
This is also why language models sometimes say things that are fluent but wrong. They are generating plausible continuations, not verifying facts. They have no access to ground truth. If the training data contained a pattern that produces a wrong answer with high confidence, the model will produce that wrong answer with high confidence. The output sounds certain because language models are not built to hedge — they are built to complete.
Understanding this doesn't diminish what these systems can do. It contextualizes it. A tool that generates highly plausible, coherent text based on patterns in human knowledge is genuinely extraordinary, and genuinely useful, when you understand what it is.
The Different Types of Artificial Intelligence
AI is not one thing. It is a family of related approaches, each suited to different kinds of problems. Here is a clear-eyed map of the major types.
Narrow AI (also called Weak AI)
This is the AI that actually exists today — all of it. Every AI system currently deployed, no matter how impressive, is narrow. It is optimized for a specific domain or task and cannot transfer that competence outside it. A model that plays chess at superhuman levels cannot play checkers. A language model that writes brilliant prose cannot drive a car. A facial recognition system cannot recommend music.
Narrow AI has produced genuinely transformative results — medical imaging analysis that matches or exceeds specialist radiologists, translation systems that handle dozens of languages with remarkable accuracy, recommendation engines that shape what billions of people read and watch. But it is always bounded. The boundary is invisible until you cross it, which is when these systems fail in ways that surprise people who thought of them as generally intelligent.
Machine Learning
Machine learning is the methodology underlying most modern AI. Rather than being programmed with explicit rules, a machine learning system is trained on data and learns to make predictions or decisions based on patterns it identifies. It is less a type of AI than a technique — the dominant technique of the current era.
There are several varieties. Supervised learning trains a model on labeled examples — thousands of images labeled "cat" or "not cat," until the model learns to classify new images. Unsupervised learning looks for structure in unlabeled data — grouping customers by behavior patterns nobody explicitly defined. Reinforcement learning trains a model through a reward system, letting it explore actions and learn which ones produce better outcomes — the method behind game-playing AI and increasingly behind robotic systems.
Deep Learning
Deep learning is a subset of machine learning that uses neural networks — systems loosely inspired by the structure of the human brain, with layers of interconnected nodes that transform input data progressively until they produce an output. "Deep" refers to the many layers in these networks. Large language models are deep learning systems. So are the image recognition systems in your phone, the voice recognition in smart speakers, and most of the AI that has made headlines in the past decade.
Deep learning is particularly powerful because it can learn to extract its own features from raw data — it doesn't need humans to specify what to look for. It figures that out. This makes it applicable to complex, high-dimensional data like images, audio, and text, where the relevant patterns are too subtle and numerous for humans to specify by hand.
Generative AI
Generative AI refers to systems that produce new content — text, images, audio, video, code — rather than classifying or predicting from existing content. Large language models are generative AI. So are image generation systems like Midjourney and DALL-E, music generation tools, and video synthesis systems.
What distinguishes generative AI is not just that it creates, but that it creates in open-ended ways. You don't choose from a menu of outputs — you describe what you want in natural language and the system generates something new. This is a qualitative shift in what AI can be used for, and it is the source of both the current excitement and the current concern about AI's cultural and economic effects.
Symbolic AI (also called Rule-Based AI or Expert Systems)
Before machine learning dominated, AI was largely symbolic — systems built from explicit rules and logical representations of knowledge. An expert system for medical diagnosis might contain thousands of if-then rules encoded by human experts: if the patient has these symptoms and these lab results, consider these diagnoses. Symbolic AI is transparent — you can trace exactly why it reached a conclusion — but it is brittle. It only knows what it was explicitly told, and it fails when reality doesn't match the rules.
Symbolic AI has not disappeared. It is still used in legal systems, financial compliance, and anywhere that decisions need to be explainable and auditable. It is also experiencing a revival as researchers try to combine symbolic reasoning with the pattern-matching power of neural networks — a hybrid approach that might address some of the weaknesses of both.
Robotics and Embodied AI
Robotics applies AI to physical systems that interact with the world — factory arms, surgical robots, autonomous vehicles, delivery drones. The AI challenges here are distinct from language or image tasks, because the physical world is continuous, unpredictable, and unforgiving of errors. A language model that gets something slightly wrong produces a mildly inaccurate sentence. A robot that gets something slightly wrong can break equipment or injure someone.
Embodied AI is making rapid progress, particularly in industrial settings and in research labs, but it remains significantly harder than the AI problems that have been most publicly visible in recent years.
Artificial General Intelligence (AGI)
AGI refers to a hypothetical AI system that can perform any intellectual task a human can perform — not just one task, but all of them, with the ability to transfer learning from one domain to another the way humans do. AGI does not currently exist. There is genuine, unresolved scientific disagreement about whether current approaches could lead to it, how soon it might arrive, and whether it would be safe if it did.
The concept matters because it is the implicit reference point for most public conversation about AI's long-term trajectory. When people debate whether AI will take all the jobs, or whether it poses existential risks, or whether it will solve climate change — they are mostly talking about AGI, even if they don't use the term. Understanding that we are currently in the era of narrow AI, and that AGI is a separate and still-theoretical category, helps separate what is actually happening now from what might happen later.
Artificial Superintelligence (ASI)
Artificial superintelligence is the further hypothetical of an AI system that exceeds human capability not just in one domain but across all domains — smarter than any human at everything, including the design of AI systems themselves. This is the subject of serious philosophical and technical work by researchers who believe it is a real possibility, and the source of the more dramatic existential risk arguments in public discourse.
Like AGI, it does not exist. Unlike AGI, there is no credible near-term roadmap to it. It belongs in this list because it is part of the map — and because understanding where it sits relative to what exists today clarifies how much distance remains between current AI and the scenarios that generate the most alarm.
Why the Distinctions Matter
A technology that everyone calls by one name but that encompasses a dozen fundamentally different approaches will produce confused policy, confused investment, confused fear, and confused optimism in roughly equal measure. That is where we are now.
The image recognition system that helps radiologists catch cancer earlier is narrow AI, built on deep learning, trained on millions of labeled medical images. The chatbot you used this morning is a large language model, a generative AI system, also narrow but in a very different way. The spam filter in your email is machine learning that has been running quietly and effectively for twenty years. The robot assembling cars in a factory is embodied AI operating under constraints that make it look nothing like the language models that dominate headlines.
Each of these deserves to be understood on its own terms — what it actually does, how it actually works, where it actually fails. The umbrella term "AI" is useful shorthand. It is a poor substitute for the actual map.
In the articles that follow, we will take each type apart in depth — how it works, where it is being used today, what its limitations are, and what it might mean for the people and industries it touches.
Every technology that changed the world looked like magic until someone explained the mechanism — and then it looked, in retrospect, like the only thing that could have happened.