Skip to content
Back to AI principles

AI principles · Chapter 1 of 8

What AI is, and what it is not

Most AI systems in use today learn patterns from examples and use them to infer an output for something new. Such a system is not a mind, a database or a search engine, and its examples shape every answer.

12 minutes0 of 4 goals done

Goals

  • Place examples of two kinds and let the classifier learn them.To do
  • Test a new fruit and see which examples decide.To do
  • Add a skewed batch of examples and watch the boundary move.To do
  • Pass the check.To do

Step 1 of 4: The read

Learning from examples

Most AI systems in use today are programs that learn patterns from examples, then use those patterns to produce an output for something new. The examples are called data. The output can be a label such as “spam”, a number such as next week’s sales, or the next word of a sentence.

Working out an output this way is called inference: the answer comes from patterns, not from looking it up.

Training, then use

Such a system has two phases. In training, it is shown many examples and adjusts itself to fit them. In use, it applies what it learned to inputs it has never seen. Using it does not change what it learned; that takes another round of training.

What it is not

It is not a mind: it computes outputs from patterns and wants nothing. It is not a database: a database gives back only what was stored, while an AI system answers for inputs nobody stored, and can be wrong. It is not a search engine: it does not fetch pages, although some tools connect it to one.

The examples decide

A system learns only from its examples, so they shape every answer. If the examples are skewed, meaning one kind or one source has far more of them than the rest, the answers lean the same way. A system that learned mostly from photos taken in one country tends to do worse on photos from elsewhere.

In the visual

You train a classifier, a program that sorts things into kinds, to tell lemons from oranges by size and colour. Each fruit is a dot on a map. To judge a new fruit, it finds the 3 examples closest to it on the map, its nearest neighbours, and answers with the kind most of them are. The line where its answer switches from one kind to the other is its boundary.

For this simple classifier, training means keeping the examples. Large models instead adjust a huge number of internal values while they train, and keep those values rather than the examples.

Step 2 of 4: Try it: Teach a tiny classifier

Teach a tiny classifier

Place lemons and oranges on the map, test a fruit it has not seen, then add a skewed batch and watch the boundary move.

Loading the visual…

Mode:RealReal: the classifier runs in this page and learns only from the fruit on the map. In the last step the page places the new lemon and the batch, and it puts the batch where it changes that lemon’s answer.

Step 3 of 4: Check

Check

Question 1 of 3A system was trained on thousands of photos labelled “cat” or “dog”. Now it labels a photo it has never seen. What is it doing?

Next: Question 2 of 3

Question 2 of 3A tool that sorts a support inbox learned from many billing messages and only a few delivery messages. A new message is unclear: it could be either. What is the likely result?
Question 3 of 3In the visual, why did the classifier start calling a typical lemon an orange?

Step 4 of 4: What you can do now

What you can do now

  • Explain in plain words that an AI system infers outputs from patterns in its examples.
  • Tell training (when a system learns from its examples) apart from use (when it applies what it learned).
  • Say why an AI system is not a mind, a database or a search engine.
  • Explain why skewed examples lead to skewed answers.