What Is Intelligence, Really? (From Zero to Agents, Episode 1)

Views: 6

Ask most people to define intelligence and you’ll get some version of: “it can figure out what’s needed and produce the right solution.” It’s a reasonable first instinct. It’s also wrong in an instructive way — and figuring out exactly how it’s wrong is the fastest route to understanding what modern AI systems are actually being built to do.

The definition that breaks in two places at once

Try it against a calculator: given “2+2,” it analyzes the input and produces the exactly correct output, every time. By a spec-satisfaction definition, that’s intelligent. Almost nobody agrees.

Try it against a thermostat: it continuously satisfies its one requirement — hold the room at 21°C — perfectly. Same problem.

Now flip it: a manager tells an employee “make the customer happy,” with no spec at all. The employee has to figure out what the requirement even is before doing anything resembling “solving.” That feels like a more intelligent act than the first two — and it doesn’t fit the original definition at all, because there was no clear requirement to analyze.

So the common-sense definition is both too generous (it lets in a thermostat) and too strict (it excludes the thing that actually feels central to intelligence: handling ambiguity and novelty).

What AI research actually settled on

A few decades of attempts, compressed:

• Turing (1950) sidestepped defining intelligence internally and tested it externally — can a human judge differentiate a machine’s responses from a human’s? Elegant, but a good enough mimic can pass without any real generalization underneath — the core critique now aimed at large language models.
• Russell & Norvig’s rational-agent definition, the backbone of modern AI, defines an intelligent agent as one that selects actions to maximize expected performance given uncertainty and past experience. This is what nearly all of reinforcement learning — and every agentic system built today — is formally built on.
• François Chollet’s reframing, in On the Measure of Intelligence, is the piece most casual definitions miss entirely: stop measuring skill at a task, and measure skill-acquisition efficiency — how well a system performs on tasks it wasn’t specifically prepared for, relative to how much prior knowledge and experience it needed to get there. A chess engine is enormously skilled but generalizes to nothing outside chess. A toddler who’s seen a handful of stacking examples can apply “stacking” somewhere entirely new. By raw skill, the engine wins. By intelligence in this sense, the toddler does something more impressive.

The definition that survives

Putting the pieces that hold up together:

Intelligence is the capacity to select actions that perform well against a goal, under uncertainty, generalizing to situations beyond what was explicitly prepared for — efficiently relative to the prior knowledge and experience available.

Three components of that are worth remembering, because they each become their own arc later in the course: acting well under uncertainty (the reinforcement-learning formalism behind agentic systems), generalizing beyond training (the entire reason learning beats hard-coded rules), and doing it efficiently (why bigger models aren’t automatically smarter — a live research problem, not a solved one).

Seeing the difference in code

Two small systems make the distinction concrete. One is a lookup table — it “solves” every case it’s tested on because it was only ever tested on cases it already memorized. The other is a simple nearest-neighbor classifier trained on a handful of labeled points, then tested on a point it’s never seen. The first fails completely outside its table. The second generalizes — imperfectly, but genuinely — from structure it learned rather than answers it memorized. That gap, between memorized specification and generalized understanding, is the entire subject of this episode, worked through in full (with runnable code) on GitHub.

Next up

Every example so far has quietly only ever dealt with numbers. Real requirements arrive as language. Episode 2 asks the more basic question underneath everything else in this course: how does a word become a number in the first place — and what’s lost, or gained, in that translation.