Practice precedes theory working comes before understanding

The claim—we see something that works, and then we understand it—inverts the textbook ordering of theory and practice. The textbook says: hypothesize, predict, test, confirm. Understanding leads, practice follows. The actual historical record usually runs the other way. The steam engine predates thermodynamics by a century. Chemists synthesized aspirin in 1897 and doctors prescribed it routinely for seventy years before anyone identified the prostaglandin mechanism. Bread predates microbiology by ten millennia. The Wright brothers flew before aerodynamicists could fully account for lift. Penicillin was a contaminant before it was a treatment, and a treatment before anyone had a coherent theory of antibiotic action. The working artifact arrives first; the explanation arrives, sometimes much later, to make sense of what already worked.

Selection—by markets, by evolution, by trial-and-error in workshops—produces working systems without ever understanding why. A blacksmith’s intuition for heat colors is competence without theory. A child’s grammar is competence without syntax. Natural selection has no theory of biochemistry; it keeps what survives. If understanding were prerequisite to working, the world would be far emptier than it is.

Contemporary AI systems exemplify the pattern in extreme form. Large language models translate, summarize, code, persuade—and interpretability research trails their deployment by years. We ship systems whose inner mechanisms we cannot read off the weights. The historical pattern has not been suspended; it has accelerated. We are once again in the position of the eighteenth-century engineer who could regulate a steam engine without being able to write down the gas laws.

The darker reading concerns what “works” silently smuggles in. Cargo cults work in the sense of being practiced; folk remedies work in the sense that the patient often improves. Survivorship bias erases the ten thousand bridges that fell so that the one that stood can be admired and copied. To say we understand something because we have seen it work risks calling regularity what is really luck, or calling mechanism what is really coincidence aligned with our hopes. The temporal sequence is reliable as description but treacherous as warrant. Working is necessary for understanding to have something to grip. It is not sufficient for the understanding to be correct.

A phenomenological layer sits underneath all of this. Polanyi’s observation that we know more than we can tell points to the priority of tacit competence over articulate explanation. The expert diagnostician pattern-matches the sick child before she could narrate what tipped her off; the articulation comes later, partial and lossy. Wittgenstein’s gesture toward meaning as use runs the same direction: we do not understand a word and then deploy it, we deploy it competently and call that understanding. Cognition is largely reverse-engineering of our own behavior. We watch ourselves succeed and theorize about how.

Something like an inverted pragmatism emerges. Theory is the residue of practice that has already been working. The honest scientist concedes this; the honest engineer never doubted it. It arrives as commentary on what has already happened, and the gap between the demonstrable fact of a thing working and any particular story told about why deserves appropriate humility.

What this asks of anyone trying to build well is a particular seriousness about practice. Respect the working artifact even when its principles are obscure. Trust accumulated craft and the people who carry it. Recognize that explanation, when it finally comes, is more often the recovery of what was already known by hand than the revelation of something that was never there. Treat theory as something owed to working practice, not a license to dismiss it—and, per the darker reading above, not a substitute for asking whether what works is working for the reason you think.