The Four Causes
"Why?" has four answers. Science kept one and a half — and the rest never actually left.
Aristotle held that you do not understand a thing until you can say why it is as it is, and that "why" is not one question. It is four, and answering only one of them and calling it understanding is a specific kind of mistake.
A translation problem gets in the way immediately. His word is aitia, standardly rendered "cause," but the modern word has narrowed to mean roughly "the event that made it happen." Aitia is closer to "what is responsible for," or simply "because." Read "the four becauses" and the doctrine stops sounding like bad physics.
The four
Take a bridge.
The material cause is what it is made of. The formal cause is its structure or defining arrangement — what makes this steel a bridge and not a heap. The efficient cause is what brought it into being, the agent or process. The final cause is what it is for, the end it exists to serve.
The important structural claim is that these do not compete. "Because it is made of steel" and "because they needed to cross the river" are not rival answers where one must be wrong. They are answers to different questions that both happen to be asked with the word "why."
What science dropped
The scientific revolution was, in significant part, a decision about which of these to keep. Francis Bacon put it bluntly: enquiry into final causes is sterile, and like a virgin consecrated to God it produces nothing. The programme that replaced Aristotelian science kept efficient causes — the pushes, the forces, the antecedent events — and treated material composition as a constraint. Formal cause survived quietly as structure. Final cause was ejected.
And the ejection worked. Asking what a falling stone is for is a dead end, and centuries of physics had been slowed by treating natural motions as strivings toward proper places. When Newton described how bodies move without asking what motion aims at, the gain was enormous.
Where it came back
Purpose did not stay ejected, because there is a class of things about which "what is it for" is a real question with a determinate answer.
Biology never stopped using functional language, and cannot. The heart is for pumping blood; that is not a poetic flourish, it is the fact that distinguishes the heart's pumping from the heart's making a thumping sound. Both are things the heart does; only one is what it is there for. Before Darwin this was embarrassing, since function seemed to imply a designer. After Darwin it was licensed: a trait's function is what it was selected for, which is a fact about a causal history and requires nobody's intentions. Final cause returned wearing a mechanism.
In anything made, it never left at all. The purpose of an artifact is not a subjective gloss on it — it is the fact that determines whether the artifact is defective. A bridge that cannot be crossed is a broken bridge, and no amount of material and structural description delivers that verdict. You need the final cause to have the concept of a fault.
The four causes are a checklist for the question "do I actually understand this?" — and the usual failure is answering one of the four thoroughly and mistaking that for the whole. A complete account of what a system is made of and how it was built can leave you with no idea what it is for, and therefore no way to tell working from broken.
Debugging is a four-cause procedure
Watch what a good diagnosis actually asks, in a domain Aristotle never saw.
Something is failing. What is it made of — which versions, which dependencies, what data is actually in there? What is its structure — how are the parts arranged, what talks to what, where are the boundaries? What brought this about — which change, which deployment, which event preceded the failure? And what is it for — what was this supposed to do, such that we are calling this outcome a bug at all?
Investigations that stall are usually stuck in one cause. A hunt through the efficient chain — which commit, which release — with no account of intended function will identify what changed and be unable to say what is wrong. A purely structural analysis will produce a beautiful diagram of a system that is nevertheless failing for a reason in the data. The four questions are cheap, and asking all of them explicitly is the entire technique.
The same lens explains why "why did the model produce this?" is such a confused question in practice: it is asked as one question and answered from four different registers at once — the weights, the architecture, the training run, and what the system is supposed to be doing. Those are four answers, and people talk past each other by giving different ones.
"Why" has four senses: what it is made of, what it is arranged as, what brought it about, and what it is for. Physics was right to drop the fourth for stones and wrong to think that generalised. For anything with a function — organisms, artifacts, software — the final cause is what makes "broken" a meaningful word.