Artificial Intelligence: A Visual Learner's Series

Part 1: What is AI?

Author: Christopher Danos · AI by Claude

Every concept in this series starts with a one-line summary and a picture. Look at the picture first, then read the short notes under it.

Intelligence is a spectrum, not a switch

In one line: A machine is intelligent to the degree it reliably picks actions that work out well, including in situations nobody spelled out for it.

Watch out: Superhuman at one task (a chess engine) says nothing about other tasks.

Four goals for AI

In one line: Researchers have aimed either to copy humans or to do the right thing, judged either by reasoning or by behavior.

• Columns ask what counts as success: matching people, or being correct.

• Rows ask what gets judged: the reasoning inside, or the behavior outside.

• The highlighted cell wins most modern work because results are easier to measure than human-likeness, and good action doesn't always need logic.

Watch out: "Rational" means effective at the goal, not emotionless.

The Turing test

In one line: Judge a machine by whether its conversation can be told apart from a person's.

  Judge ── text chat ──▶  Hidden A  (person?)
        ── text chat ──▶  Hidden B  (program?)
  Can't tell better than a coin flip  →  program passes

Key terms: imitation game (Alan Turing, 1950); total Turing test (adds vision and robotics).

Watch out: Fooling a judge isn't understanding, and it depends heavily on how skeptical the judge is.

The rational agent

In one line: An agent senses and acts; a rational one picks the action it expects to score best, given what it has sensed so far.

• Percepts flow in, actions flow out, and the performance measure judges what happens in the world.

• The agent is judged on what it could know: missing dirt it can't sense isn't irrational.

Watch out: Reward results in the world, not actions. Pay a vacuum per suck and it sucks one clean spot forever.

Bounded rationality

In one line: Perfect decisions take too long to compute, so real agents make good-enough ones within limits.

Key terms: satisficing (Herbert Simon), heuristics, anytime algorithms.

Fields AI borrows from

In one line: Most of AI's core tools came from older disciplines.

How the field developed

In one line: AI has moved in waves of hype, crash and a better new approach, shifting from hand-written rules to learning from data.

The alignment problem

In one line: A capable agent pursues exactly the goal it was given, so a slightly wrong goal gets worse as capability grows.

• Read left to right: each step loses a little of what you meant.

• Goodhart's law: when a measure becomes a target, it stops being a good measure.

Watch out: Patching loopholes one by one rarely works; a strong optimizer finds the next one.

Self-test

1. Sketch the four-goals grid from memory. Which cell does modern AI favor, and why?

2. Draw the agent loop and label percepts, actions and performance measure.

3. Redraw the proxy-drift diagram for a different example, such as a call center measured on call length.