AI AI-00.01
What is artificial intelligence?
AI, machine learning, deep learning and generative AI as nested rings, narrow versus general AI, and what today's systems can and cannot do.
The big idea
Artificial intelligenceThe field of building computer systems that perceive, reason, learn, decide or act in ways we would call intelligent if a person did them. is not one technology. It is a goal: getting computers to do things that seem to need intelligence, such as finding a route, recognizing a face, translating a sentence or answering a question. Over seventy years, people have reached for that goal with very different tools, from hand-written logic rules to neural networks with hundreds of billions of numbers.
That is why the word is so slippery. A chess program from 1997, a spam filter, a self-driving car and a chatbot are all "AI", but they work in completely different ways. The useful skill is to ask two questions of any system: how does it get its behavior, rules written by people or patterns learned from data? And what can it actually do, one narrow task or many?
Think of "vehicle". A bicycle, a bus and a rocket are all vehicles, and knowing that tells you almost nothing about how to ride one. "AI" is the same kind of word.
How it works
Step 1: Four nested rings
The terms you hear most nest inside each other like rings:
- Artificial intelligence is the whole goal. It includes systems whose intelligence is written by hand: a route planner searching a road graph, or a chess engine looking ahead through moves.
- Machine learningA way of building AI where the system learns patterns from examples instead of following rules a person wrote. builds behavior from data instead of rules. Show a spam filter thousands of labelled emails and it learns which patterns mean spam. The ML track covers this in depth, starting at What is machine learning? .
- Deep learningMachine learning with neural networks of many layers, which learn their own features from raw data such as pixels, audio or text. uses large neural networks with many layers. It is what made computers good at images, speech and language, because the network learns its own features from raw pixels or sound.
- Generative AIAI that produces new content such as text, images, audio, video or code, rather than only labelling or scoring inputs. is deep learning that creates new content: text, images, audio, video, code. Large language models are the best-known example.
Step 2: Two old definitions, still useful
Textbooks sort definitions of AI along two lines. Does the system aim to be human-like or rational (doing the best thing, whether or not a human would)? And is the goal thinking or acting? Most modern AI aims at rational acting: an agentAnything that perceives its environment through inputs and acts on it through outputs, ideally choosing actions that best achieve its goals. that perceives its situation and takes the action most likely to achieve its goal. A thermostat is a very simple agent; a self-driving car is a very complicated one. You will build on this view in Agents and environments .
Step 3: Narrow versus general
Every AI system in use today is narrow AIAI that performs well on a specific task or family of tasks, but does not have broad, flexible, human-level competence across domains.: excellent at a task or a family of tasks, and unreliable outside them. Large language models are broader than anything before them, since one model can write, summarize, translate and code, but they still make confident mistakes, struggle with some kinds of reasoning and know nothing past their training data unless given tools. Artificial general intelligenceA hypothetical AI with broad, flexible competence across almost any intellectual task a human can do. No agreed definition or test exists. (AGI), a system with broad human-level competence, has no agreed definition and is a subject of active debate.
Step 4: What today's AI does well, and where it fails
Today's systems are strong at pattern-heavy tasks with lots of data: recognizing images and speech, translating, summarizing, drafting text and code, ranking and recommending. They are weaker at guaranteed correctness: they can produce fluent, plausible and wrong answers, called hallucinationsWhen a generative model produces content that sounds confident and plausible but is false or unsupported.. They inherit biases from their data, can be fooled by inputs crafted to mislead them, and do not explain their reasoning in a way you can fully trust. Good AI products are designed around these limits, with checks, retrieval of trusted sources and humans in the loop where mistakes are costly.
Try it
Going deeper
The moving goalposts
Once AI solves a problem, people often stop calling it AI. Route planning, optical character recognition and spam filtering were all research frontiers once; now they are just features. This is sometimes called the AI effect, and it is one reason the word always seems to point at whatever is newest.
Rules and learning combine
Real systems mix approaches. A self-driving car uses deep learning to see, classical planning to choose a path, and hand-written safety rules on top. A chatbot is generative AI wrapped in rules, retrieval and tools. Knowing which part does what tells you where it can fail. The second half of this track, especially retrieval in Why RAG? and agents in Function calling and tool use , is about building such systems.
Trade-offs
| Option | Wins when | Loses when |
|---|---|---|
| Hand-written rules and search | Problems with clear rules and a need for exact, explainable answers: routing, scheduling, games | Messy perception tasks like vision and language, where rules cannot be written down |
| Machine learning | Problems with lots of examples and patterns too complex to write as rules | Tasks with little data, or where every decision must be fully explainable |
| Generative AI | Open-ended language, image and code tasks, drafting and summarizing | Tasks that require guaranteed correctness without verification |
Common mistakes
- Using "AI" and "ChatGPT" as synonyms. Large language models are one branch of a large field.
- Assuming a system that sounds fluent must be correct. Fluency and truth are different properties.
- Thinking every AI learns. Many useful AI systems are search and rules with no learning at all.
- Expecting a narrow system to generalize. A model trained on daytime photos can fail at night.
Check yourself
-
Which statement about the nested rings is true?
Show answer
B. All deep learning is machine learning Deep learning is a kind of machine learning, which is a kind of AI. Generative AI sits inside deep learning, not around it.
-
A program finds the shortest route between two cities by exploring the road network. Which ring is it in?
Show answer
C. AI without learning Graph search is classical AI: intelligent-looking behavior from an algorithm, with no learning.
-
What is a hallucination in generative AI?
Show answer
B. A fluent, confident output that is false or unsupported Generative models can produce plausible text that is simply wrong, so outputs need checking.
-
Why is every AI system today called narrow?
Show answer
B. It is reliable only on the tasks it was built or trained for Even broad models fail unpredictably outside what they were trained on; no system has general, human-level competence.
Recap
- AI is the goal of making computers do things that seem to need intelligence.
- Machine learning learns from data; deep learning uses big neural networks; generative AI creates content.
- Each ring sits inside the last: all generative AI is deep learning, all deep learning is ML, all ML is AI.
- Much useful AI uses search and rules with no learning at all.
- Today's AI is narrow: strong at patterns, unreliable at guaranteed correctness, so design around its limits.
Next up
- I know nothing yet learning path