What is machine learning?
Patterns, tested against reality.

Instead of writing every rule, learn a useful pattern from data.

In machine learning, a developer chooses a task, a model family, and a way to measure performance. A training process uses data to adjust the model’s parameters. The resulting model can then make predictions or produce other outputs for new inputs. For the wider picture, read what artificial intelligence includes ↗.

This does not mean software has stopped following instructions. People still choose what the system is for, what data it receives, how training is run, how success is measured, and what happens to the output. Most real products combine learned models with ordinary software, curated data, and explicit rules.

“Learning” describes changes to a model through a training process. It does not imply human understanding, a reliable memory of every example, or an ability to judge whether a prediction should be acted on.

Teach a spam filter from labeled messages.

Imagine a collection of emails marked “spam” or “not spam.” A training example may include the message text and other permitted signals as input, plus the label as the expected answer. During training, the model’s predictions are compared with those labels, and its parameters are adjusted to reduce a chosen loss.

After training, the filter can score a new message it has not seen before. That step is inference. It is not learning from that message unless the product has a separate process that records feedback and later uses it to update or retrain the model.

The examples define the lesson. If labels are inconsistent, important kinds of messages are missing, or the training set reflects old patterns, the model can inherit those limits. The score should also leave room for a person to appeal or correct an error when the decision matters.

The feedback tells the model what kind of task it is solving.

These are common learning setups, not four interchangeable buttons. The right one depends on the available data, the feedback that can be collected, and the consequences of a wrong output.

Machine learning includes a wide range of model families.

Linear models, decision trees, nearest-neighbor methods, clustering algorithms, and neural networks solve different kinds of problems. Deep learning refers to machine-learning methods based on neural networks with multiple layers. It is a powerful family, not a synonym for all machine learning. For how these layers are trained, see the deep-learning guide ↗.

Large language models are one example of deep learning. Their text generation is often trained as next-token prediction. Read how language models work ↗, then follow how image generation uses learned models ↗ to see how a different output is produced.

Test on unfamiliar cases, and inspect the cost of errors.

A model can fit its training examples without learning a pattern that holds in new situations. This is overfitting. A basic evaluation keeps examples separate from the data used to fit the model, then measures performance against a task-specific target. That first check is useful, but a real system also needs testing across the users, conditions, and failure cases it will face.

1. Define the decision. State the task, intended user, and what counts as a useful result.
2. Check the examples. Review coverage, labels, data rights, privacy, and how the data was collected.
3. Choose the error measures. Accuracy alone may hide the specific mistakes that matter most in your setting.
4. Keep evaluation separate. Test on examples not used to train or tune the final model, and keep looking for data leakage.
5. Test the whole product. Include interfaces, retrieval, tools, review steps, and what happens when the model is uncertain or wrong.

For language models and agents, a benchmark score is only one piece of evidence. See how to read AI benchmarks ↗, how to evaluate an AI agent ↗, and how synthetic training examples need their own checks ↗.

Follow the model into the system around it.

Start with the AI field guide ↗ for a map of models, context, memory, retrieval, tools, and evaluation. For hands-on foundations, Google’s Machine Learning Crash Course ↗ has lessons and exercises.