Three ways to learn

That completes a set. All three of the main arrangements for teaching a machine have now appeared, and they differ on one question. Where does the feedback come from?

Supervised learning. A person supplies the answer. Millions of photographs, each with school bus or coffee mug written beside it by hand. The machine guesses, the answer is right there, and the gap between them is the correction. Accurate, and limited by how much labelling anyone can afford.

Reinforcement learning. The world supplies a score. Nobody says what the correct move was. Something went well or badly, possibly long after the action responsible, and the machine has to work out for itself what deserved the credit. This is how a system learned dozens of Atari games from nothing but pixels and a number going up.

Self-supervised learning. The data supplies the target. Hide part of a sentence and predict it, or predict the next token and compare it with the one already in the text. People do not have to label every example by hand.

That third arrangement made it practical to learn from far more text than people could label by hand. It is not free or limitless. Text must still be collected, filtered, stored, licensed or governed, and processed with substantial computation.

Existing writing could be turned into prediction examples automatically. Humans still made the writing, the dataset choices, and the training system.