Too Much to Know: They Complained About Information Overload 2,000 Years Before Your Inbox

The complaint predates your inbox by about two thousand years. Seneca was already telling a correspondent that reading many books scatters a person, and that the fix was to run over many thoughts and then take exactly one to digest that day.¹

The useful part is that the answers repeat. Strip two millennia of finding aids down to their mechanics and only two moves survive: shrink what a reader has to cross, and remove the requirement to cross it in order. Compression and addressing. Nothing else recurred.

That gives you a test you can run this afternoon on every tool, dashboard, digest and AI summary you currently pay for.

Why these three models

This piece runs the Wire Model: score the features of the decision, route to a small ensemble of formal models, then force the ensemble to produce dated actions. The scores that routed it:

• Factor-price divergence (0.9). The cost of producing text has collapsed repeatedly. The cost of deciding what a specific text is for has not moved at all.
• Retrieval topology (0.85). Every durable fix changed the shape of the path between a person and the sentence they needed, rather than the size of the library.
• Reusable judgment (0.8). The compilers did the selecting once and thousands of readers spent it. That is an aggregation asset, and it can be built badly.
• Cognitive load (0.5). Real, extensively documented, and scored too low to earn a card of its own here.

Those route to growth accounting (which input got cheap), network centrality (what shape the fix took), and crowd aggregation (who did the selecting and whether their errors were independent). Against the outcome-type map that is equilibrium, complex and random: three distinct types, so the models fail in different directions.

Behavioral and governance are folded rather than shipped as separate cards. The behavioral effect in play is that people prefer collecting to choosing, which produces the same prescription as the price argument, so a fourth card would have bought you no lever. Governance appears where it bites, inside the levers: an index with no named owner rots within a quarter.

The framework: three prices, one shape, and a borrowed judgment

1. The prices: three of the four operations got cheap

Ann Blair’s study of information management before the modern age is the reference work here, and it covers ancient and medieval Europe, the Islamic world and China before it reaches print.⁸ Writing in the Boston Globe, she set out what the tools of that period actually did. Printers, scholars and compilers built tools that listed, sorted under subject headings, summarized, and selected from books no one person could master. Her verdict on the present is blunt: our key methods of coping have not changed since the sixteenth century, because we still need to select, summarize and sort, and human judgment still has to guide the process.⁷

Growth accounting asks the next question. When one input to a process gets dramatically cheaper and the others do not, output reorganizes around whatever stayed expensive. So price the four operations.

• Storing. The price of a unit of computer storage has fallen by almost ten orders of magnitude in seventy years. A 256-gigabyte capacity, standard in a laptop today, would have cost roughly twenty billion dollars in the 1950s.¹⁰
• Sorting. Free, instant, and performed by software you did not configure.
• Summarizing. Now a few cents and a few seconds, on any text, on demand.
• Selecting. Unchanged since Seneca. Selecting requires knowing which decision you are about to make, and no external system can supply that.

Three of the four got cheap. The fourth never moved.

Meanwhile the corpus kept accelerating. Bornmann and Mutz measured the growth of science across three phases and found rates roughly tripling each time: under 1% a year to the middle of the eighteenth century, 2% to 3% up to the interwar period, then 8% to 9% through 2012.⁹ Growth compounds. Reading speed does not.

Growth accounting, the price lens

Assumes: the four operations are substitutable inputs and their relative prices drive where effort concentrates.

Fits because: three inputs fell by orders of magnitude while the fourth stayed at human speed.

Breaks when: a tool starts performing genuine selection against your stated criteria, which would finally move the fourth price.

Counteracts: the instinct to buy more storage, more feeds and more summarization capacity.

May reinforce: a false comfort that the answer is purely personal discipline rather than shared structure.

2. The shape: every surviving fix built a hub

Network centrality asks what the path from a person to a needed sentence looks like. Serial reading is a line: to reach item n you cross n items. An index is a star: one hop to the hub, one hop out. That topology change kept getting rebuilt, across centuries that shared no tools and mostly no texts.

Around 245 BCE, Callimachus compiled the Pinakes at Alexandria, widely treated as the first library catalogue in the West, organizing a collection of nearly 500,000 papyrus scrolls by author and subject. The catalogue itself ran to 120 volumes.³ A hub costing 120 volumes was worth building against 500,000 scrolls.

In AD 77, Pliny the Elder opened the Natural History with an explicit product spec. He had gathered 20,000 topics from about 2,000 volumes by 100 select authors, and he attached a contents list to the dedication so that Titus would not be obliged to read the books through, adding that the same aid would serve everyone else, “so that any one may search for what he wishes, and may know where to find it.”² Random access, stated as a feature, in the first century.

The sharpest case is medieval. The first concordance to the Vulgate was completed in 1230 under Hugh of Saint-Cher, assisted by a reported 500 fellow Dominicans. It contained no quotations at all. It was purely an index of where each word appeared, keyed to Stephen Langton’s recently invented chapter numbers, with each chapter split into seven lettered parts to sharpen the address.⁵ They spent 500 people to build a book that carries no content and only addresses. That is how valuable a hub is when the corpus is fixed and the questions are not.

Compression shows up in the same record, as a separate move. Photius produced the ninth-century Bibliotheca, 280 sections on works he had read, ranging from a single sentence to several pages, and Reynolds and Wilson call him the inventor of the book review.⁴ Vincent of Beauvais assembled the Speculum Maius, 3.25 million words across 80 books and 9,885 chapters.⁶ Blair’s summary of the printed era describes the same pair of moves at scale: detailed outlines and alphabetical indexes let readers consult books without reading them through.⁷

The operational rule falls straight out. An index earns its cost by holding addresses and one-line claims. The moment it also holds the content, it stops being a hub and becomes a second copy of the corpus, which is what most company wikis actually are.

Network centrality, the topology lens

Assumes: retrieval cost is a function of path length, and a hub converts a line into two hops.

Fits because: unconnected cultures across 2,000 years converged on catalogues, contents lists and concordances.

Breaks when: the corpus changes faster than the hub can be updated, at which point a stale index misroutes you with confidence.

Counteracts: duplicate storage, renamed documents, and knowledge bases that mirror rather than point.

May reinforce: centralization of authority in whoever controls the index and its vocabulary.

3. The selectors: someone else’s judgment, bought once and spent many times

Pliny read 2,000 volumes so his readers would not have to. Photius wrote up 280 works so his brother could read the account instead. The Dominicans pooled 500 people into one address layer. In every case the selecting was done once by a named party, and then reused.

Crowd aggregation prices that asset, and its warning is specific. Pooled judgment beats an individual only when the errors of the contributors are independent. Correlate them and the pool inherits one bias with extra confidence attached. The historical compilations had a real independence structure: different compilers, different orders, different patrons, different centuries, and a named person carrying the reputational cost of a bad selection.

An AI summary is a selector with no reputation exposure and no memory of what you decided last quarter. It feels like a crowd. It behaves like one correlated source.

The move that follows runs inside your company, and almost nobody makes it. Right now the same selection is being performed several times in parallel and thrown away each time. Four people read the same regulator circular, three of them reach a private conclusion in a chat thread, and none of it lands anywhere a fifth person can reach. Name one reader per domain: regulation, competitors, pricing, capital. That reader writes one line of what changed and which decision it moves, and posts it to the address. One selection, spent by everyone. That is Pliny’s arrangement, and it is the part of his stack you can stand up this week.

Crowd aggregation, the selector lens

Assumes: selection done once by a named party can be reused, and pooling helps only when contributors err independently.

Fits because: the durable compilations were built by identifiable people who carried the cost of a bad selection.

Breaks when: apparent diversity hides one shared upstream source, which is the normal state of a modern feed.

Counteracts: the same reading being done four times in parallel and discarded four times.

May reinforce: single points of failure once one reader owns a domain.

GEER: what to pull, starting with the reversible

Ranked by cost and by how easily you can undo it.

1. Name your recurring decisions. Most operating companies have six to nine: price, hire, fire, spend, ship, kill, raise, enter a market, extend credit. Write the list. Twenty minutes, fully reversible.
2. Give each one a permanent address. One page per decision, one stable name, never renamed and never duplicated. The address is the whole asset. Renaming it destroys the asset.
3. Move compression to the point of capture. The person on the distributor call writes the three lines, not the person who reads the transcript later. Pliny’s economics: read once, compress once, spend the compression many times.
4. Give each domain one named reader. Regulation, competitors, pricing, capital. One line posted to the address when something changes, with the decision it moves. Reversible in a week if the wrong person has it.
5. Strip content out of the index. Each entry holds a claim in one line, a link to where the evidence lives, and a date. WhatsApp order threads, the mobile money statement, the signed LPO, the distributor’s price list: the index says which and where, never a copy.
6. Assign one owner. An index with no name attached is stale in a quarter. This is the governance layer, folded in.

RADAR: dated, and split by side of the table

Relative to T, the day you read this.

DO NOW, reversible and dominant across scenarios.

• T+3. Write the recurring-decision list and create one page per decision. Put three fields at the top of each: what was decided, what evidence moved it, when it gets revisited.
• T+7. For each decision, record the single number or event that would reverse it, and the address where that number lives. A decision whose reversal trigger has no address is one you will silently make again.
• T+10. Name the reader for each domain and tell them what they own. One line, posted to the address, when something changes. Reversible the moment it is the wrong person.
• T+7, investor side. Replace the generic process question in diligence. Ask for the last three decisions the team reversed and the note that triggered each reversal. A team with an addressing layer produces them in minutes. A team without one produces a narrative.

HEDGE, cheap insurance against the tail.

• T+14. Log every AI-generated summary you acted on, with the source link, for two weeks. Cost: a line per item. Payoff: when one summary turns out to have dropped the qualifier that mattered, you can find every decision it touched.
• T+14. Keep one deliberately off-model input, a customer visit or a distributor you do not sell to, specifically to break the correlation in your selector set.
• T+21, investor side. Ask a portfolio company to reproduce, from its own records, the evidence behind a decision made two quarters ago. The time it takes is a cleaner read on operating maturity than any dashboard.

DEFER AND TRIGGER, irreversible, so pre-commit the observable.

• Do not buy a knowledge platform yet. Migration and lock-in make this hard to unwind. Trigger it on an observable, not a feeling: two people, working separately, cannot locate the same decision inside sixty seconds, twice in one month. Or headcount passes fifteen. Until then a shared folder with disciplined naming carries the same topology at zero switching cost.
• Do not put an AI agent in front of your decision record until the record exists and has been stable for a quarter. Automating retrieval over an unstructured pile reproduces the pile faster.

CHAIN: what the record says happens next

Build the comparison set on structure. The right peers are populations whose cost of producing text dropped by an order of magnitude while human reading capacity stayed flat. Five members qualify: Alexandria after the scroll boom, thirteenth-century Paris after the universities and the commercial book trade, Europe after the press, the web after the 1990s, and text generation after 2023.

Base rate across those five. Volume never came back down in any of them. In every case the durable fix was an addressing layer plus a compression layer, built by a small group and reused by many. In zero of the five did anyone resolve it by reading faster.

Three things about the present state change the arithmetic. Summarization is now near-free, which is genuinely new and works in your favour. The summarizer is anonymous and carries no reputational cost for a bad selection, which is new and works against you. And the addressing layer for most of what you read is owned by platforms optimizing for time spent rather than for your retrieval, which is the sharpest departure from every historical case, where the compiler and the reader wanted the same thing.

Now take out what would have happened anyway. Some founders who look calm about information are calm because their business has three inputs and one customer segment. Their calm is not evidence for any of this. Strip them out, and the remaining calm operators share one visible trait: they write decisions down somewhere with a fixed address.

Matrix-break flag. This whole ensemble assumes selection stays expensive. If a system reliably declines to answer because your decision criterion is undefined, and then selects correctly once you supply it, the fourth price has moved and the argument above needs rebuilding. Watch for that specific behaviour. Confident answers to underspecified questions are the opposite signal.

What this ensemble cannot see

Prices, topology and independence say nothing about taste. Two founders with identical indexes will still pick different things to care about, and that difference decides more outcomes than any of this. The ensemble also cannot see the political reason a company keeps five versions of the truth, which is usually that five people want to stay right. No index survives a room that does not want one.

And the historical record has a survivor problem. The finding aids we can study are the ones that were copied. Whatever was tried and abandoned left no trace, so “these two moves recurred” is partly a statement about what got preserved.

The action that survives all of that is small. By T+7, name your recurring decisions and give each one a permanent address. If everything else in this piece is wrong, you are left with a list of the decisions your company actually makes and a place to put them. That has been the winning move for two thousand years, and it costs you an afternoon.

Joshua Agonya Pi’Rwot, Founder.

Sources and notes

1. Seneca, Moral Letters to Lucilius, Letter 2, “On discursiveness in reading,” translated by Richard Mott Gummere. “And in reading of many books is distraction. Accordingly, since you cannot read all the books which you may possess, it is enough to possess only as many books as you can read.” Section 4 gives the remedy: “after you have run over many thoughts, select one to be thoroughly digested that day.” Wikisource full text.
2. Pliny the Elder, The Natural History, Preface, translated by John Bostock and H. T. Riley. “I have included in thirty-six books 20,000 topics, all worthy of attention … gained by the perusal of about 2000 volumes … procured by the careful perusal of 100 select authors.” And: “I have subjoined to this epistle the contents of each of the following books, and have used my best endeavours to prevent your being obliged to read them all through. And this, which was done for your benefit, will also serve the same purpose for others, so that any one may search for what he wishes, and may know where to find it.” Perseus Digital Library.
3. “Pinakes,” Wikipedia. Bibliographic work composed by Callimachus, popularly considered the first library catalogue in the West; he organized the library by authors and subjects about 245 BCE, and the work ran to 120 volumes. The collection at Alexandria contained nearly 500,000 papyrus scrolls, grouped by subject matter and stored in labelled bins. Article.
4. “Bibliotheca (Photius),” Wikipedia, quoting Reynolds and Wilson: “a fascinating production, in which Photius shows himself the inventor of the book-review,” with “280 sections … vary in length from a single sentence to several pages.” Article. The codex-by-codex contents list, in English, is at the Tertullian Project.
5. John Francis Fenlon, “Concordances of the Bible,” The Catholic Encyclopedia, Vol. 4, New York: Robert Appleton Company, 1908. “The first concordance, completed in 1230, was undertaken under the guidance of Hugo, or Hugues, de Saint-Cher … assisted, it is said, by 500 fellow-Dominicans. It contained no quotations, and was purely an index to passages where a word was found.” Passages were keyed to Stephen Langton’s chapter divisions, with each chapter divided into seven lettered parts. New Advent.
6. “Speculum Maius,” Wikipedia, on the thirteenth-century compilation by Vincent of Beauvais: “As a whole, the work totals 3.25 million words and 80 books and 9885 chapters.” Article.
7. Ann Blair, “Information overload, the early years,” The Boston Globe, 28 November 2010. Source of the four operations (“tools that listed, sorted under subject headings, summarized, and selected from all those books that no one person could master”), of “Detailed outlines and alphabetical indexes let readers consult books without reading them through,” and of the closing verdict: “In many ways, our key methods of coping with overload haven’t changed since the 16th century: We still need to select, summarize, and sort, and ultimately need human judgment and attention to guide the process.” Globe archive.
8. Ann M. Blair, Too Much to Know: Managing Scholarly Information before the Modern Age, Yale University Press, 2010. Cited here for the scope of the study rather than for a figure: it examines information management in ancient and medieval Europe as well as the Islamic world and China before focusing on printed Latin reference books in early modern Europe. Publisher page.
9. Lutz Bornmann and Rüdiger Mutz, “Growth rates of modern science: A bibliometric analysis based on the number of publications and cited references,” accepted in the Journal of the Association for Information Science and Technology. Segmented regression on cited references identifies three growth phases, each roughly tripling on the last: “from less than 1% up to the middle of the 18th century, to 2 to 3% up to the period between the two world wars and 8 to 9% to 2012.” Open access preprint: arXiv:1402.4578.
10. Edouard Mathieu, “The price of computer storage has fallen exponentially since the 1950s,” Our World in Data, 21 May 2024, using data collected by John C. McCallum. “In the last 70 years, the price for a unit of storage has fallen by almost ten orders of magnitude.” And: “A 256-gigabyte storage capacity, commonly found in standard laptops sold today, would have cost around 20 billion dollars in the 1950s.” Data insight.