Roman concrete was not lost. Roman concrete remains, embedded in the Pantheon’s dome and in thousands of bridges and aqueducts that have stood for two thousand years. Vitruvius wrote down the recipe in the first century BCE. The texts have been continuously available. The materials still exist. By any reasonable test, the information about Roman concrete has been with us the entire time.
What was lost, for roughly 1,400 years, was the ability to make it.
That gap — between knowing what something is and being able to do it — is the most under-discussed fact in the history of human knowledge. It is the one fact that explains why every generation rediscovers techniques that were nominally available, why entire technologies disappear and reappear, and why a culture’s knowledge can be measured in libraries while its capability quietly collapses.
The substance metaphor
We talk about knowledge as if it were a substance. Knowledge “exists” in a book, “is preserved” in an archive, “is stored” in a database, “is held” in a person’s head. The vocabulary is industrial, mineral, water-tank. Knowledge is something with a location and a quantity.
This is the wrong metaphor. Knowledge is not a substance. It is a behavior. It happens, repeatedly, in transmission events between people who are doing something. It exists for as long as the transmission keeps happening and stops existing when the transmission stops.
The Roman concrete example is the cleanest case. Builders in 800 CE had access to most of the same materials that builders in 100 CE used. They had access to the texts. The ones who could read Latin had Vitruvius on the shelf. They could not, however, build with Roman concrete. What they could do was build with the inferior medieval mortar that replaced it, badly, until Renaissance Italian builders began experimentally re-deriving the technique from rubble.
What was lost was not the recipe. It was the network of practitioners who knew, embodied, in their hands and eyes, which volcanic ash to source, how to slake the lime, how to compact the aggregate, how the cure should look in the third week versus the first. The knowledge was a practice, distributed across a community of artisans and apprentices. When the community fragmented — when the western Empire’s labor markets collapsed, when the legions stopped commissioning bridges, when the apprenticeship lineages broke — the practice stopped happening. The substance that everyone now points to as “Roman concrete knowledge” was not a fact in the world. It was a stable pattern of behavior. The pattern stopped.
The pattern is everywhere
Roman concrete is the famous case, but it is one of dozens.
The keystone arch was known in Mesopotamia by 2500 BCE, refined by Roman engineers, and forgotten across most of post-Roman Europe except where a continuous masonry guild lineage preserved it. Where it survived, it survived because people kept building keystone arches and teaching each other to do it. Where it died, it died because nobody apprenticed anyone in the technique for a few generations.
Damascus steel — the high-carbon crucible steel of pre-modern Syria, with its watered patterns and unusual toughness — was lost in the late 1700s. The texts describing it survive. The archaeological samples survive. Modern metallurgists have spent decades trying to reproduce it and have only partially succeeded. The reason is straightforward: the practice depended on small impurities in specific ore sources that the smiths recognized by feel and color. When the ore source changed and the smiths could no longer get the right input, the practice could no longer be performed. The smiths who knew the difference died without successors who could compensate. The texts described the procedure but did not encode the embodied knowledge of which ore is the right ore.
Variolation — the deliberate exposure of children to mild smallpox to confer immunity — was practiced by Turkish women in the early 1700s as a folk medical tradition. It was observed and described by Lady Mary Wortley Montagu, brought to England, briefly practiced by physicians, then mostly abandoned for fifty years until Jenner reinvented a related technique with cowpox. The practice was known in England the entire time. It was not practiced because the network of practitioners who could do it competently had not formed.
The Antikythera mechanism — the bronze geared astronomical computer recovered from a first-century BCE shipwreck — represents a level of mechanical-computation craft that does not appear again in the Mediterranean for over a thousand years. The texts that describe its function (Cicero references devices like it) were continuously available. The metallurgy was continuously available. What was missing was the chain of clockmakers and astronomers who knew how to design and produce such a thing. When that chain broke, the device became impossible to build, regardless of what any text said.
The pattern is consistent. The texts survive. The materials survive. The artifacts survive. What is lost is the community of practice. Without the community of practice, none of the surviving substance reconstitutes the capability.
Why the substance metaphor persists
If the substance metaphor is wrong, why is it so dominant?
Because we built libraries. Beginning with Alexandria, continuing through monastic scriptoria, exploding with the printing press, formalized in the modern university, the substance metaphor became infrastructurally embedded. We built buildings to “hold” knowledge. We built professions to “transmit” knowledge in the sense of moving it from a teacher’s head to a student’s. We measured knowledge as a quantity — degrees, citations, page counts. We treated learning as the transfer of substance from a source to a destination.
This worked, for some kinds of knowledge, well enough. Theoretical physics is largely textual. Mathematical proofs are largely textual. Most analytic philosophy is largely textual. Domains where the proof or the reasoning is itself the object can survive in textual form, because the act of reading the text is approximately the same act as performing the knowledge.
The substance metaphor fails badly for domains where the knowledge is the action and the action requires embodied judgment that the text cannot encode. Surgery. Wine-making. Carpentry. Trial advocacy. Ceramics. Combat. Diplomacy. Distillation. Cooking. Welding. These domains have texts. The texts are mostly useless to someone who has not first apprenticed under a practitioner. The text is a mnemonic for someone who already knows how — not an instruction set for someone who doesn’t.
The list of domains where the substance metaphor fails is, on inspection, most of human productive knowledge. We have systematically over-counted what survives in libraries and under-counted what survives in workshops, kitchens, and operating theaters. When we try to estimate “how much knowledge has humanity accumulated,” we are mostly measuring the textual fraction, which is the easier-to-measure fraction, and underestimating the propagation-dependent fraction, which is the larger fraction.
Why texts can’t carry the propagation fraction
It is worth being precise about why texts fail at this.
A text describes a procedure. A procedure is a sequence of steps. The execution of any non-trivial procedure depends on judgment calls at each step that are not — and in many cases cannot be — fully specified. Compact the aggregate firmly says nothing about how firmly. Cure for several weeks says nothing about which week the color should change. Quench rapidly says nothing about the temperature differential or the medium. The competent practitioner fills the under-specification with embodied judgment built up across hundreds of trial-and-error reps with corrective feedback from someone who already knows.
This is not a flaw in particular texts. It is a structural property of the relationship between symbolic representation and embodied skill. A text could in principle specify every judgment call to arbitrary precision, but in practice no text does, because the writer is also relying on tacit assumptions that are obvious to fellow practitioners and invisible to outsiders. The text is, in effect, addressed to readers who already share most of the embodied knowledge it presupposes. Strip the readers of that shared substrate and the text becomes a hollow shell that describes the practice without enabling it.
The hardest part of any craft is not the part that gets written down. The hardest part is the part the master corrects on the apprentice’s third or thirtieth attempt — no, like this — where the like this is a demonstration that no text could capture. Lose the masters and the text becomes archaeology.
Implications for AI
This matters now because we are building systems that operate almost exclusively in the substance metaphor. A large language model is a textual artifact trained on textual artifacts. It can answer questions about how to make Roman concrete. It can recite Vitruvius. It can describe Damascus steel forging. It can list the steps of variolation.
It cannot, in any current architecture, apprentice anyone. It does not produce a stable community of practice. It does not maintain an embodied lineage of practitioners who can recognize the right ore by feel or the right cure by week three. It does not stand in a kitchen with a student and say no, like this.
This is an architectural fact, not a temporary limitation. A model trained on more texts is still a model trained on texts. The propagation problem is not solved by more textual coverage. It is a different problem in a different category.
The current AI debate operates as if the question is whether AI can substitute for the substance: can the model give the right answer? But for most knowledge — the propagation-dependent fraction — substitution is not what’s at stake. What’s at stake is whether the AI is part of the propagation network or parallel to it. An AI that documents Roman concrete is a substance contribution. An AI that mentors apprentices, watches their work in real time, corrects their hand position, recognizes when their cure is wrong, and certifies them as competent practitioners — that AI is a propagation-network contribution.
We have approximately none of the second kind. We have a great deal of the first. The press tends to confuse them, because the substance metaphor is the framing we already had.
The risk this creates is not “AI will replace experts.” The risk is more subtle, and worse: AI will substitute for the substance fraction of knowledge while leaving the propagation fraction unfunded. Students who would have apprenticed under masters will instead query the model. The model will give them the substance — the recipe, the procedure, the step list. They will believe they have learned the thing. The community of practice that would have formed if they had apprenticed — the lineage that recognizes the right ore, the right cure, the right grip on the chisel — will not form. The substance will be more accessible than ever. The capability will quietly stop propagating.
We have seen this movie before, with the textbook. Textbooks were supposed to substitute for masters. They didn’t. They substituted for the easy fraction of mastery, which is the substance fraction. The hard fraction — embodied judgment under uncertainty, recognition of the right input by feel, knowing when to break the rule the textbook gave you — kept being transmitted by the apprentice-master pattern, where it was transmitted at all. AI has the same shape, with the substance fraction much larger and faster than the textbook ever was, and the propagation fraction not addressed.
The honest version of the AI productivity story is not “AI replaces experts.” It is “AI handles the substance fraction so cheaply that we forget to fund the propagation fraction, and a generation later we discover, like the medieval builder staring at Vitruvius, that we cannot do the things the texts describe.”
What to do about it
I don’t have a complete answer. I have a list of partial ones.
First: name the distinction. Treat substance-knowledge and propagation-knowledge as different categories with different infrastructure needs. Stop describing AI tools that handle substance as if they were addressing propagation. The two are different problems and require different metrics.
Second: protect propagation infrastructure even when the substance is freely available. The fact that you can look up how to do a thing is not the same as the fact that there are people in the world who can do it competently. The latter is a separate, fragile, and chronically under-funded category of social infrastructure. Apprenticeships. Master-journeyman pipelines. Workshop residencies. Hospital teaching services. Trial advocacy clinics. These are not redundancies on top of the textual layer. They are the load-bearing structure underneath it.
Third: be specific about which fraction of which domain we are trying to preserve. Some domains are mostly textual and AI substitution is genuinely improving access. Some domains are mostly propagation-dependent and AI substitution is, paradoxically, accelerating their loss by hollowing out the demand for apprenticeship. The two cases require different policy, different investment, different success metrics.
Fourth: recognize that the question of whether a piece of knowledge “still exists” is empirical and behavior-based, not textual. If the practice has not happened in a generation, the knowledge has not survived, regardless of how many books describe it. Roman concrete did not survive between 600 and 1500 because nobody was making Roman concrete between 600 and 1500. The texts were beside the point.
This is uncomfortable because it implies that most of human knowledge has the lifespan of its practitioner network, not the lifespan of its archive. The archive is a souvenir. The network is the thing. The archive without the network is a beautifully preserved corpse.
The corollary
Roman concrete was not lost. The people who made it were lost. We have rebuilt the people, slowly, over the last two centuries, by reverse-engineering the substance and bootstrapping a small contemporary practitioner network. The substance was always there. The recovery was the network.
This is the right pattern to remember as we build the next generation of cognitive infrastructure. The question is not whether the AI can know what Vitruvius wrote. The question is whether anyone, in twenty years, will still know what to do with the cure on the third week — whether the network of people who can recognize no, like this will still be propagating, or whether we will have funded the substance into ubiquity while the propagation network atrophied beneath it.
The substance is the easy part to measure, which is why every public conversation about AI is about the substance. The propagation is the hard part to measure, which is why nobody is talking about it. By the time we notice it has thinned, it will already be thin enough that rebuilding it requires the kind of multi-generational reverse-engineering that brought Roman concrete back, and there is no guarantee we will do that work for every domain we lose.
We will, almost certainly, lose some domains and not notice for forty years. The texts will be there the entire time. The capability won’t.
That is the part of the AI conversation I cannot stop ruminating on. It is not a problem the substance metaphor can see, and the substance metaphor is the only metaphor most of the conversation is using.