Agriculture is usually told as an "upgrade" for humanity, but the skeletal evidence points the other way: early farmers were shorter, sicker and worked more than the foragers of their time. Rather than liberating us, farming locked us into a ratchet we could never reverse—it was a one-way bet, not a free choice.
Settlement → surplus → storage → more children → more mouths to feed → the need to farm still more. This is a positive-feedback ratchet: once a population swells on agricultural surplus, there is no route back to foraging—the population density that gathering can sustain is far lower than farming's. People become trapped by the very population pressure they created, forced to crop the same land ever more intensively, working ever-longer hours to sustain a group that can no longer shrink.
Compare human bones from before and after early Neolithic settlement: as grain became the staple, average height fell, dental caries surged, and marks of iron-deficiency anemia and infectious disease multiplied—the price of a monotonous carbohydrate diet and dense settlements shared with livestock. Anthropologists have estimated that foragers "worked" only around twenty hours a week, the rest being leisure. In other words, agriculture raised the number of people a plot of land could feed, not the quality of each life: it made humans more numerous, not better off. Some have called it "the worst mistake in the history of the human race."
In economics this is the classic Malthusian trap and path dependence—technical progress eaten by population growth, locked into a low-welfare equilibrium. In complexity science it is a phase transition with hysteresis: cross a certain density and you cannot retrace your steps. In software engineering it maps precisely onto "lock-in": a throughput-boosting technology, once its capacity is filled by the business, can no longer be removed—only fed. Anything that "raises output while lifting the irreversible floor" is an agricultural bet.
When adopting an AI tool stack, the real question is not "how much will it speed us up" but "once it does, will my scope expand and never contract again." A team that doubles delivery with AI often goes on to take twice the demand and hire more people to feed the pipeline—capacity rises, but the exit is paved over too. Whether a productivity gain is liberation or a new ratchet depends on whether you hold onto the floor of "we could always go back."
Which tool or process you recently adopted has grown to the point where ripping it out hurts more than maintaining it? It promised to save effort—what is it actually locking you into now?
Writing was not invented for poetry or philosophy—it was invented for accounting. The earliest known texts are overwhelmingly receipts, inventories, tax rolls, ration accounts. Writing is first a technology of the state and of control: it freezes fleeting memory and externalizes it, making large-scale coordination and cross-generational accumulation possible. Literature is a byproduct that arrived centuries late.
Mesopotamian cuneiform began as clay tokens used for counting: one lump for a sheep, one for a basket of grain, sealed inside a clay envelope. People eventually pressed the tokens' shapes onto the surface of the tablet, then abstracted them into signs—accounting gave birth to writing. The essence of writing is lifting language out of "the air of the present moment" and fixing it as a portable, checkable, accumulable object. Once memory is externalized it is no longer bound by the capacity and lifespan of a single brain, and only then does the layered edifice of law, bureaucracy, contract and science have a foundation.
The earliest batch of tablets from Uruk is overwhelmingly administrative: so much grain, so many head of livestock, rations issued to so many workers. Narrative literature appears only some five hundred years later. More telling is an ancient doubt handed down from Greece—the worry that writing would weaken memory: since anything could be written down, people would no longer bother to remember, outsourcing their recall to symbols. This is the ancient version of today's "extended mind" debate: every time we hand cognition to an external medium, we bargain over the same gains and losses.
In neuroscience, literacy reshapes the brain: reading "recruits" regions of visual cortex originally used to recognize objects and faces, forming a dedicated word-processing area (the neuronal recycling hypothesis), so literate and illiterate brains handle the same image differently. In information theory, writing is a storage-and-error-correction medium resistant to the noise of time. In distributed systems, it is essentially the write-ahead log (WAL) and the tamper-evident ledger—ledgers, double-entry bookkeeping, blockchains all solve the same problem: how a group of mutually distrustful people can share one checkable record.
Large language models are externalizing a new layer of cognition, just as writing once externalized memory. Today's prompts and context windows are the new "clay tablets"—first used to automate, compute metrics and fill out forms, in service of control and efficiency, which is exactly writing's "ledger phase." The question worth asking: what will AI's late-arriving "literature phase" be? Only when it stops merely doing our accounting and begins to expand what we can think, and how, does that layer truly change cognition.
Are you mainly using AI to "keep the books" (automate, summarize, speed up), or to widen the frontier of your thinking? If writing took five hundred years to travel from ledger to epic, which phase is the AI in your hands stuck at?
Societies solve problems by adding complexity—more bureaucracy, infrastructure, specialization. But complexity obeys diminishing returns: each increment costs more and yields less, until a society can no longer afford its own complexity. At that point collapse is not pure catastrophe but can be a rational "downsizing"—a retreat to a level it can actually sustain.
The first investments in complexity are extraordinarily cheap: the first irrigation canal, the first cadre of officials—small input, large return. But the low-hanging fruit is picked first, and the remaining problems grow harder, demanding ever more expensive solutions—redundant layers, larger armies, more elaborate coordination. When the marginal return nears zero or turns negative, the system becomes brittle: any drought, plague or invasion can make "keeping the complexity running" less worthwhile than "simplifying on purpose." Collapse is the system's rapid fall back to lower complexity under stress.
The Roman Empire paid for its own complexity through expansion and plunder; once expansion stopped, it could only prolong its life by debasing the currency, raising taxes and swelling the bureaucracy—costs climbing, returns falling, until it was fragile enough to shatter. More counterintuitively: after the empire dissolved, the living standards of some provincial peasants actually rose, because the heavy taxes and vast machine bearing down on them vanished too. For most of those living through it, "collapse" is sometimes the lifting of a burden, not a plunge into the abyss.
In economics this is the law of diminishing marginal returns, blown up to civilizational scale. In energy research it corresponds to energy return on investment (EROI): every system needs a net-positive energy surplus, and when complexity devours that surplus the system can no longer sustain itself. In software and distributed systems it is technical debt and microservice sprawl—each added service or layer raises coordination cost, and one day maintaining it costs more than the value it creates, with "collapse" appearing as a tear-down rewrite. Complexity is not free; it charges interest.
As organizations and architectures grow, people habitually solve each new problem by "adding one more service, one more team, one more middle layer." Tainter urges you to do the marginal math: is this unit of complexity still net-positive? Sometimes the optimal move is to simplify deliberately before collapse is forced on you. And AI is an intriguing variable—it may lower the marginal cost of complexity and push the wall of diminishing returns outward; but beware that it only postpones the turning point rather than abolishing it, letting the system sprint on at an even higher, more brittle level of complexity.
Which of your systems or processes has already crossed the point of diminishing returns—where the effort to maintain it is outrunning the value it delivers? What would it take to simplify it on purpose, and why have you been putting it off?
What we call "collapse" is often an illusion manufactured by the archaeological record and by an elite-centric view of history. Cities abandoned, monuments no longer built—it looks like the end of a civilization, yet it is frequently just decentralization, migration or reorganization: the people did not vanish, they changed how they lived. The word "collapse" often says more about our bias (no megalithic temple, no civilization) than about what actually happened.
Archaeology reads collapse through indirect proxies: deforestation in pollen layers, breaks in pottery styles, the halting of monumental construction, abandoned settlements, marks of nutritional stress on bone. But the disappearance of elite markers does not equal the disappearance of people. The absence of palaces and inscriptions may only mean that the power structure dissolved and society flattened—not that everyone died. The record itself is a biased sample: those who left durable ruins are mostly the ruling class, while the daily life of ordinary people rarely enters the strata.
The Maya "collapse" is often imagined as the total extinction of a civilization, but although the cities were abandoned on a huge scale, millions of Maya descendants are alive today—it was a political and demographic reorganization, not the extinction of a people. Even more telling is Easter Island: the popular "ecocide" story (islanders felling all their trees and destroying themselves) has come under serious challenge in recent years, with new evidence suggesting the population did not crash before Europeans arrived, and that the real catastrophe came from imported slavery and disease. The moral parable we tell is largely a modern lesson projected onto ambiguous data.
In statistics this is survivorship and sampling bias: we study the part that "left traces," not everything that happened—just as counting bullet holes only on returning warplanes yields a fatally wrong conclusion. In machine learning it is the missing-data problem, where training on biased or survivor data necessarily leads to false conclusions. In history it is the deep-seated elite bias of "history written by monuments." Signal must always be painstakingly reconstructed from a lossy, filtered record.
A postmortem is a kind of archaeology. When a system "collapses," the logs and metrics you hold are a heavily skewed sample—you can only see where you happened to instrument, only the parts that happened to be recorded. Reconstructing an incident's causality from partial telemetry is the same craft as reconstructing a city's fall from half a stratum. Beware: how much of the "collapse story" you reconstruct is the truth, and how much is an artifact of what you happened to measure?
Look back at your most recent postmortem: is the widely accepted "root cause" narrative pointed to by the data itself, or written for you by your observability blind spots? With a different set of instrumentation, would the same incident be told as a completely different story?