What is Machine Learning?

Machine learning is how computers got good at things nobody could write rules for. Try writing explicit rules that recognise a cat in any photo, or that spot every fraudulent transaction. You can’t — there are too many cases. Instead, you show a machine learning system thousands or millions of examples and let it work out the patterns itself.

How it works

1. Data. Collect examples relevant to the task — emails labelled spam or not, house sales with prices, photos with captions. This is the training data.
2. Model. Choose a model type: anything from a simple decision tree to a giant neural network.
3. Training. The model adjusts its internal parameters to make its predictions match the examples as closely as possible.
4. Evaluation. Test on examples it hasn’t seen to check it generalises rather than memorises.
5. Prediction. Use the trained model on new data — that step is called inference.

The three classic types

• Supervised learning: learn from labelled examples (this email is spam).
• Unsupervised learning: find structure in unlabelled data (group these customers into segments).
• Reinforcement learning: learn by trial and error, rewarded for good outcomes (game-playing systems, and part of how chatbots are tuned with human feedback).

An analogy for the meatbags

Teaching a child what a dog is doesn’t involve a rulebook (“four legs, fur, tail, barks…” — which also describes a cross cat). You point at dogs until they get it. Machine learning is pointing at examples, at enormous scale, with a student that never gets bored.

Why it matters for the Chosen

Machine learning has quietly run the economy for years before ChatGPT made AI fashionable: recommendation engines, ad targeting, credit scoring, fraud detection, demand forecasting, search ranking. Deep learning — machine learning with very large neural networks — then produced generative AI.

For the Chosen, the practical lessons:

• Data is the moat. Models are increasingly available off the shelf; unique, clean data about your customers or niche is not.
• ML predicts; it doesn’t guarantee. Every output is a probability. Design processes that handle being wrong.
• Not everything needs a model. If a simple rule or a spreadsheet works, use it. Use ML where patterns are too complex for rules.

Examples

Netflix and YouTube recommendations, Gmail’s spam filter, bank fraud alerts, Google Maps’ arrival-time estimates, product recommendations in online stores, and the large language models behind every AI assistant.