What Is Artificial Intelligence?

A field, not a single technology

"Artificial intelligence" is the umbrella discipline: building systems that perform tasks which, when humans do them, require intelligence — perceiving, reasoning, planning, learning, communicating, acting. Machine learning is one (currently dominant) toolbox inside it; deep learning is a toolbox inside that. The nesting matters: chess programs, route planners, theorem provers and expert systems are all AI with little or no learning; conversely, a linear regression is ML but nobody calls it "an AI".

A useful map of approaches:

Modern systems mix them: AlphaGo = deep networks (evaluate positions) + tree search (plan ahead); an LLM agent = a transformer + tool use + symbolic scaffolding.

The agent framework

AI's unifying abstraction — you'll recognize it from the RL preview: an agent perceives its environment through sensors and acts on it through actuators, choosing actions to achieve goals. A thermostat, a chess engine, a self-driving car, ChatGPT-with-tools: all agents differing in the richness of percepts, actions and goals. A rational agent picks actions that maximize its expected goal achievement given what it knows — "expected" and "given what it knows" doing heavy lifting: real environments are uncertain (probability path!) and partially observable.

Narrow vs. general

Everything deployed today is narrow AI: superhuman inside its lane (Go, protein folding, image labeling), helpless one step outside it. A chess engine cannot play checkers; an image classifier can't explain a joke. Artificial general intelligence (AGI) — one system matching human flexibility across arbitrary domains — remains a research goal, though modern LLMs' breadth has genuinely blurred the line and reignited the debate about how far scaling can go.

The Turing test (1950) proposed conversation indistinguishable from a human as a benchmark. It shaped decades of imagination but measures imitation, not intelligence — chatbots can pass restricted versions via tricks while lacking any world model. Modern evaluation prefers task batteries: math, coding, reasoning, factuality, safety.

A recurring cultural pattern worth knowing — the AI effect: once something works reliably, people stop calling it AI ("that's just autocomplete / just search"). AI is perpetually redefined as "whatever machines can't do yet".

Key takeaways

• AI ⊃ ML ⊃ DL; multiple traditions (symbolic, search, learning) — modern systems hybridize them.
• The agent lens — perceive, decide, act toward goals under uncertainty — unifies the whole field.
• Today's AI is narrow; AGI is the open frontier, and the goalposts have a habit of moving.