AI Foundationspredict · compress · act
act I

The Core Question

Before we can build it, we should say what we mean by it. Start with the loop, not the brain.

01

What Is Intelligence, Really?

Everyone uses the word intelligence, and almost no one defines it the same way. So we will not start with a definition. We will start with a loop.

AGENTdecidesWORLDrespondsactionobservation + reward
the agent loop — the one shape that survives every era of AI

An agent is anything that takes in a situation and produces an action. An environment is everything the action lands on. The environment answers with a new observation and a reward — a number saying how well that went. Intelligence, for our purposes, is whatever lets the agent choose actions that raise the reward over time.

That is deliberately modest. It covers a thermostat, a chess engine, a foraging ant, and a large language model. It does not require consciousness, feelings, or a soul.

Three words, one loop

the vocabulary of every chapter that follows

The loop is the whole field in miniature. Almost every idea later in this guide is an answer to one of three questions: how does the agent represent the world, how does it choose, and how does it learn from what came back?

the frameAn AI is not a thing. It is a relationship between a decision-maker and a world that keeps score.
REPRESENTwhat does it know?
DECIDEwhat should it do?
LEARNhow does it get better?

Three levels people confuse

Most arguments about AI are people talking at different levels. Keep them apart.

task
the problem — win the game, answer the question, drive the car
model
the machinery — a function that maps inputs to outputs
agent
the loop — the model plus memory, tools, and a sense of purpose
system
the deployment — the agent plus the people, permissions, and costs around it

The one sentence to keep

Intelligence is the ability to take useful action under uncertainty.

Everything else in this guide is machinery for that sentence. We begin where the uncertainty lives: in information itself.

what is nextBefore we can talk about learning, we need a way to measure what the agent does not know. That measure was invented in 1948, and it is called the bit.
introduces →intelligenceagentenvironmentreward
next →Intelligence Is Compression