Concept · Chapter 1: What Is Artificial Intelligence?
From Rules to Learning
Traditional programs turn rules and inputs into answers; machine learning turns inputs and answers into the rules — learning the knowledge from examples instead of having it written in.
The problem
Hand-writing the rules for perception, language or messy real-world judgement is impractical: the rules are too many, too subtle, and partly unknown even to experts.
The solution
Collect many examples with the right answers, choose a model with adjustable parameters, and let an algorithm tune those parameters until the model reproduces the answers — and generalizes to new cases.
The consequence
The bottleneck shifts from writing rules to gathering data and computing. With enough of both, learned models overtook hand-built systems in vision, speech, translation and eventually language generation.
You should understand first
The flip
In machine learning, the "rules" are parameters of a model — numbers such as the per-word weights of a spam filter, or the billions of weights of a neural network. A training algorithm adjusts them until the model's outputs match the examples.
Try it
Try it · toy model
Build a spam filter the 1980s way — by writing rules — then watch it meet new messages. Finally, let a learning algorithm work it out from examples instead.
What changed, and what it costs
Knowledge comes from data
The spam filter learned that reward, fee and loan signal spam without anyone writing a rule about them.Generalization is the goal
What matters is performance on messages the model has never seen — the central idea of Chapter 3.Data becomes the bottleneck
With 12 examples the learner was mediocre; with 40 it was far better. Modern AI's appetite for data starts here.Transparency drops
A rule says why it fired. A learned model's weights are harder to read — the reason interpretability research exists (Chapter 16).
Why should I care?
As a researcher
Every later chapter is a variation on this move: choose a model family, a training objective and data, and let optimization find the parameters.
As an engineer
Deciding whether a problem calls for explicit rules (clear, stable, auditable logic) or a learned model (fuzzy patterns, lots of data) is a daily engineering judgement.
Modern systems that depend on it
- All of machine learning
- Neural networks and deep learning
- LLM pretraining
Historical context
Before
Symbolic AI and expert systems: knowledge written by hand.
After
Statistical machine learning (1990s–2000s), deep learning (2012 onward) and large pretrained models.
Used today
Almost every AI product: spam filters, recommendations, speech recognition, translation, image recognition, and LLMs — all learned from data.
What to remember
- Programming: rules + data → answers. Learning: data + answers → rules.
- The 'rules' become model parameters, tuned to fit examples.
- The goal is generalization: correct answers on new, unseen cases.
- More (and better) data usually beats more hand-tuning.
- Learned models are harder to inspect than hand-written rules.
Key papers
The perceptron: A probabilistic model for information storage and organization in the brain.
F. Rosenblatt · 1958 · Psychological Review
Introduced the perceptron — a neuron model that learns its weights from examples. Every neural network descends from it.
Some Studies in Machine Learning Using the Game of Checkers
A. L. Samuel · 1959 · IBM Journal of Research and Development
One of the first programs that improved by learning from play — an early demonstration that learning can beat hand-tuning.