This module followed another chain of old ideas becoming practical together.

• In 1988, Richard Sutton showed how later estimates could correct earlier ones.
• In the 1990s, TD-Gammon joined that learning rule to a neural network.
• In 2012, AlexNet joined deep networks, GPUs, ImageNet, and improved training methods.
• From 2013 onward, DeepMind joined deep networks to reinforcement learning, search, and self-play.

Deep learning describes the network machinery. Reinforcement learning describes feedback from consequences. Either can exist without the other. Their combination helped systems learn Atari games and Go, while AlphaFold used deep learning for a different scientific prediction task.

The pattern matters. Progress came from work built on earlier work, sometimes after the first idea had waited years for enough data, compute, or another missing method.

These systems do not all receive the same kind of raw material. An Atari screen begins as measured pixel values. A Go board can be encoded as a grid of stone positions. A protein sequence begins as symbols for amino acids, so AlphaFold must turn those symbols and their relationships into numerical representations.

The next problem begins with symbols.

The word dog carries no built-in number saying how it relates to puppy, animal, or bark. Before the machinery in this module can process language, the symbols need numerical representations that preserve useful relationships.

That is where Module 6 begins.