What machine learning actually is

In every program you've written so far, you write explicit rules, and the computer applies them to produce output:

Machine learning inverts this. You provide examples of inputs paired with the correct answers, and the training process works out the rules (formally, a mathematical model) that map one to the other:

Once trained, that model can be given a brand-new input it's never seen and produce a plausible answer — because it learned a general pattern from the examples, not because anyone hand-wrote a rule for that specific case.

Some problems are genuinely too complex, or too fuzzy, for anyone to write explicit rules for. Nobody can write an if/else chain that reliably recognizes a cat in a photo — there's no finite set of pixel-value rules that covers every possible cat, angle, and lighting condition. But if you show a model tens of thousands of labeled photos ("this is a cat," "this is not a cat"), it can learn a general pattern that generalizes to photos it's never seen. This is the exact class of problem machine learning was built to address: pattern recognition from examples, where the rules themselves aren't obvious enough for a human to write down directly.

It's tempting to imagine a trained model as an enormous table that memorized every example it saw — but that's not what actually happens, and it's not what makes ML useful. A model is a mathematical function with internal numeric parameters (sometimes billions of them, in large language models) that get adjusted during training until the function's output closely matches the correct answers for the training examples. Once tuned, that same function is applied to new inputs it was never shown — the whole point is generalizing beyond exactly what it saw, not recalling it.

A spam filter that checks a fixed list of banned words is traditional rule-based programming, not ML — the rules were written by a person, not learned from data. A spam filter that was trained on millions of labeled emails, and learned on its own which patterns correlate with spam, is ML. The distinction matters practically: rule-based systems are predictable and easy to debug line by line; ML systems are probabilistic, can be wrong in ways that are hard to predict in advance, and need to be evaluated on real data rather than reasoned about from the code alone — a theme this whole domain returns to.

The rest of this domain builds directly on this one idea: what different kinds of learning-from-examples exist (next lesson), what a neural network — the specific kind of model behind nearly everything usable today, including every large language model — actually computes, and then a deep dive specifically into how LLMs turn text into predictions, since that's the technology the AI Agents & LLM Integration domain builds real, production features on top of.