What is machine learning?

Rules you write versus rules you discover

Ordinary software is a set of instructions a person wrote down in advance: if the balance is below zero, decline the transaction; if the input matches this pattern, return that result. The programmer already understood the rule, and the code simply encodes it. Machine learning inverts this. Instead of writing the rule, you show a model many examples of inputs paired with the right outputs, and an algorithm searches for a rule that fits those examples well enough to generalise to new ones it has never seen.

This is the right frame for everything that follows in this roadmap. A spam filter does not consist of a list of banned words someone typed in; it consists of parameters fit to millions of real emails, discovered by an optimisation process rather than authored by a person. The upside is that it can catch patterns a person would never think to write down. The downside, which the rest of this roadmap spends considerable time on, is that nobody wrote the rule, so nobody can read it off directly — which is exactly why evaluating a model honestly matters so much more here than in software whose logic a person can simply read.

What to watch for

The video gives the core distinction cleanly: a system that improves at a task from experience rather than from explicit programming, illustrated with concrete, approachable examples. That one sentence is worth holding onto through every lesson that follows — a linear regression, a random forest and a neural network are all, underneath their very different mechanics, doing exactly this.

Where this differs from being an “AI engineer”

It is worth being explicit about a distinction this roadmap’s sibling, Upshard’s AI Engineer roadmap, is built entirely around. An AI engineer treats a large language model as a finished, pretrained dependency — someone else already ran the training, spent the compute, and shipped an API. The engineering problem there is what you build around that dependency: context, retrieval, tools, evaluation. This roadmap is about the other half of the picture: how a model gets trained from data in the first place, using classical methods that are often simpler, cheaper, and more interpretable than a large language model, and are frequently the right tool when the problem is narrower than “understand and generate human language.”

Neither skill replaces the other. A team shipping an LLM-powered product still needs someone who understands what a model actually is and how training works, if only to know when a much simpler trained model would do the job better, faster, and for a fraction of the cost.

The shape of what is ahead

This roadmap follows one coherent arc: preparing data a model can actually learn from, the classical models that learn from labelled examples, combining those models into ensembles, finding structure in data with no labels at all, honestly evaluating whether any of it worked, and finally where neural networks pick up from the classical foundation. Each group answers one question in that arc, and each assumes nothing from the ones that come after it — only from what came before.