← Writing

MBA · STO

Via negativa: the discipline of subtraction

You know with near certainty that smoking harms you. Nobody knows with comparable certainty which supplement helps you. Via negativa is that asymmetry turned into policy: negative knowledge is firmer than positive knowledge, so removal is a safer default than addition. Nassim Taleb gave the old idea its modern form in Antifragile. For a founder it has an awkward corollary: the highest-yield item on your roadmap is probably a deletion, and nothing in your incentives will ever surface it.

A rule older than its author

Taleb names the principle in Antifragile (2012). In domains where errors are rare but severe, act by removal before you act by addition. His reasoning is epistemic. We understand what we remove, because it has a track record inside our system. We do not understand what we add, because its interactions have not happened yet. Subtraction operates on known material; addition imports unknowns. Taleb files everything from medicine to state policy under this heading, and he takes the family name from theology.

Apophatic theology, the branch he borrows from, is a method for describing an object too complex for positive language. Pseudo-Dionysius the Areopagite argued in The Mystical Theology (5th–6th century) that every positive statement about God fails. The honest path is negation: say what the thing is not, and stop there. Strip away the theology and a general method remains. When an object exceeds your descriptive power, negative statements stay reliable long after positive ones collapse. A company is a smaller mystery, and still too complex for confident positive claims about what will work.

Sculpture supplies the folk version. A line attributed to Michelangelo — no reliable source records it, so treat it as a fable — has him explain that the statue already stands inside the block, and his work is to remove everything that is not the statue. Note what the image smuggles in: the marble had to be quarried and paid for first. Someone performed the additive step. Keep that debt in mind; it returns in the objections.

One boundary line, to keep the tools sharp. Inversion, Munger's habit of asking what would guarantee failure, is a direction of analysis: it produces the list of hazards. Via negativa is a policy of intervention: it says the strongest available fix is usually to delete something, and it applies even when nothing threatens failure yet. I wrote about inversion separately on this blog. This essay is about the second tool — the claim that of the two moves you can make on any system, the subtractive one is more trustworthy.

The asymmetry of negative knowledge

Karl Popper built a philosophy of science on one observation (The Logic of Scientific Discovery, 1934/1959). No count of white swans proves that all swans are white. One black swan settles the question. Confirmation and refutation are asymmetric operations. A theory can only be corroborated for now; a refutation is permanent. The firmest stock of knowledge we hold is therefore negative: the catalogue of things shown to be wrong.

Taleb compresses this into a phrase, subtractive epistemology: knowledge grows more by removing error than by adding truth. The "more" is arguable. The reliability ranking is hard to argue with. Disconfirmed claims stay disconfirmed, while confirmed claims sit on probation forever.

Now move from science to a system you operate. Benefit in a running business hides in counterfactuals. You rarely observe what the new feature would have earned, because the world where you skipped it is gone. Harm announces itself: the incident report, the support queue, the churned account, the meeting everyone dreads. The evidence in front of an operator is mostly evidence about damage. Removal acts on the documented side of the ledger; addition acts on the speculative side.

There is also arithmetic. A system with twelve components has sixty-six pairwise interactions. Add a component and you have seventy-eight: one visible new part, twelve invisible new relationships. Remove one and fifty-five remain. Additions grow the space of things you do not understand combinatorially; removals shrink it. This is why a "small" addition is never small, and why a codebase feels harder to reason about every quarter even when every change was reviewed.

Iatrogenics: the harm column nobody fills in

Medicine has a name for harm caused by the healer: iatrogenics. It earned the concept because it kept the receipts. Doctors bled patients for over two thousand years on a theory of humours. George Washington's physicians bled him repeatedly on the day he died in 1799, by their own account. The intervention looked like action, was billed as expertise, and had negative expected value. The lesson took another century to absorb: in some states of knowledge, the best treatment is the one you decline to administer.

Taleb's move in Antifragile is to generalize iatrogenics beyond medicine. Wherever an agent gets credit for visible action and does not pay for delayed damage, intervention will be oversupplied. Consultants add frameworks. Product teams add features. Each addition books a small, immediate benefit to the intervener and a diffuse, delayed cost to the system. Nobody logs the second entry, so the ledger always says the intervention worked.

The reverse asymmetry explains why subtraction is undersupplied. An added initiative has a sponsor whose name is attached to it. A removed initiative has no constituency: its beneficiaries are everyone, slightly, later. No one gets promoted for the meeting they deleted or the feature they refused to ship. So organizations accrete by default, and the accretion is nobody's fault and everybody's cost.

Lindy: time as a filter

Via negativa needs a selection rule: remove what, keep what? Taleb's answer leans on the Lindy effect. The name comes from Albert Goldman's 1964 essay "Lindy's Law", about comedians trading career forecasts in a New York deli; Taleb generalized it in Antifragile. For a perishable thing, age predicts less remaining life. For a non-perishable — an idea, a practice, a book, a protocol — each year survived is evidence of durability, so expected remaining life grows with age.

The mechanism is subtractive, which is why it belongs here. Time does not certify quality directly. Time runs removals: every year deletes some fraction of the practices that failed their users. What remains after five centuries has passed trials nobody could list in advance. So when you must bet between a practice that survived generations and an addition invented last quarter, the old one carries survival evidence and the new one carries a pitch deck. That is a prior, and only a prior. Its failure modes get their own section below.

Why nobody subtracts by default

If removal is this reliable, why is it rare? In 2021 a University of Virginia team published an answer in Nature. Adams, Converse, Hales and Klotz ran a series of experiments with one design: give people something to improve, then count whether they add or remove. People overwhelmingly added. Asked to stabilize a Lego structure, participants added bricks even when each added brick cost money and removal was free. Asked to improve a piece of writing, they lengthened it.

The paper's title states the finding plainly: people systematically overlook subtractive changes. Additive ideas arrive in mind first; subtractive ones require a deliberate search. Under time pressure or cognitive load, the search never happens. Prompts worked in the experiments — participants reminded that removal was allowed subtracted more. The deficit sits in attention, and attention is exactly what a procedure can redirect.

The finding scales uncomfortably well from Lego to management. The grammar of corporate planning is additive: objectives launch things, roadmaps ship things, budgets fund things, CVs list things. No standard artifact exists in which a quarter's deletions are planned and rewarded. The bias is not a personal weakness; it is a missing checklist item. That is the most hopeful reading available: install the prompt, and the subtraction shows up.

Four ledgers where removal pays

Doctrine earns nothing until it changes a decision, so here are the four ledgers where a founder can run via negativa this month. The table gives the shape; the paragraphs give the reasons.

LedgerAdditive reflexSubtractive moveCost to reverse
StrategyAnother segment, another partnershipA written list of what the firm refuses to doLow: amend the note
ProductA feature per loud requestRank by usage; sunset the bottom decile with noticeMedium: staged, announced
CalendarA standing meeting per problemAsk per series: what breaks if this stops?Near zero: recreate in thirty seconds
InformationAnother source, another alertUnsubscribe; sample weekly instead of hourlyZero

Strategy. Michael Porter located the essence of strategy in deciding what not to do ("What Is Strategy?", 1996). A strategy that only lists intentions is a wish. The binding part is the exclusions, because exclusions are testable: either you declined the tempting segment or you did not. Peter Drucker supplied the operating version decades earlier in The Effective Executive (1967). For every activity, ask what would happen if it were not done at all; where the answer is "nothing", abandonment is the default and continuation carries the burden of proof.

Product. Product is where addition compounds fastest, because every yes is permanent. Fried and Heinemeier Hansson wrote the durable version in Getting Real (2006): make "no" the default answer and make each feature fight hard for inclusion. A shipped feature is code to maintain, states to test, UI to explain, and a promise to support — the interaction arithmetic from earlier, wearing a roadmap. The subtractive move is a scheduled feature audit: rank by usage, then sunset the bottom decile with notice. Deleted surface cannot break, so removal is the cheapest quality work available.

Calendar. The calendar is the founder's ledger of default additions. Marcus Aurelius, passing on Democritus, gives the test in Meditations IV.24: do few things, and ask of each act whether it belongs to the necessary. Most standing meetings fail that question quietly. The stoic framing prices the resource correctly: what a meeting spends is a share of a finite life, and the spend recurs weekly without further consent. The worked example below runs this ledger end to end.

Information. The information diet responds to subtraction faster than any other ledger. Taleb's argument in Antifragile is statistical: sample the news hourly and you read mostly noise; sample weekly and the signal share rises. Adding filters and dashboards fights volume with more volume. Unsubscribing fights it with zero. For an attention system tuned to novelty — mine is — every deleted input is a filter that costs no willpower to run. Lindy gives the replacement rule: read old books, because they have survived more readers than this morning's feed.

A worked example: the calendar, minus eighteen percent

Procedures beat exhortations, so here is the audit run once, with arithmetic. The numbers describe one concrete run; the procedure is the deliverable. Export four weeks of calendar into a sheet. Say it shows 96 scheduled events totalling 116 hours. Separate recurring series from one-offs: 11 recurring series account for 68 of those hours, or 59 percent. Recurrence is where subtraction pays, because a recurring event is a standing tax. A weekly hour is roughly fifty hours a year, decided once and never re-decided.

Apply one question to each of the eleven series — Drucker's: what breaks if this stops entirely? Grade the honest answers into three classes. Class A, nothing breaks: a weekly status whose numbers already sit in a dashboard (four hours a month), a biweekly sync whose decisions moved into a written doc (two hours), a weekly half-hour call that had drifted into pleasantries (two hours). Class B, something breaks but less would do: four series totalling 24 hours a month. Class C, real damage: four series totalling 36 hours a month — sales, hiring, governance, the weekly review. The classes sum back to 68.

Now the moves. Class A: delete all three series today; eight hours a month return for the cost of three cancellation emails. Class B: shrink. Default lengths drop from sixty minutes to thirty, two weekly series become biweekly, one meeting becomes a written update with comments. The 24 hours become 11, and 13 come back. Class C stays untouched. Via negativa is a filter, and a filter that removes everything is broken.

Every deletion runs as a four-week experiment, and this step makes the method safe. Recreating a calendar series costs thirty seconds, so the worst case of a wrong deletion is one month of mild friction. The asymmetry is the point: additions create constituencies and become hard to unwind, while deletions stay reversible for weeks. In this run, one deletion failed the experiment. The biweekly sync turned out to carry an unwritten escalation channel, and it returned as a monthly half hour. Cost of the error: half an hour a month, against twenty-one recovered.

Net result: 20.5 hours a month recovered — 17.7 percent of scheduled time, and the title rounds up — from about ninety minutes of audit work and no new tool. The second-order gains are unpriced and real. Every removed event also deletes its preparation and the attention switch on either side of it, and the freed slots merge into blocks long enough for actual work. One guard keeps the gain: a one-in-one-out rule, under which any new recurring meeting must name the recurring meeting it replaces. Accretion is the default; the rule turns it back into a decision.

Where via negativa fails

Every essay here owes its model an honest bill of failure modes. Via negativa has four.

One: subtraction cannot create. The statue was latent in the marble, and the marble was quarried and paid for by an additive act. Before product-market fit there is no stock to refine. The search is additive by nature: you must build things that do not yet exist to learn what the market wants. James March's distinction between exploration and exploitation (1991) marks the boundary cleanly. Via negativa is an exploitation-phase tool: it sharpens a system that already produces value. Applied to an exploration phase, it starves the search and refines you into a smaller version of the wrong thing.

Two: the additive column contains most of civilization. Anesthesia, vaccines, antibiotics and the transistor were additions, and each one beat every subtraction of its century. Taleb himself concedes the case where intervention wins: severe conditions, large and proven benefit. The honest generalization is that via negativa free-rides on additive risk-takers. Somebody accepted iatrogenic risk, ran the additive experiments and produced the stock of practices that Lindy later filters. A world of pure subtractors would be safe and small.

Three: Lindy inherits survivorship bias. Old is evidence of durability, and durability is a different property from benefit. Bloodletting survived two millennia; astrology has outlived every science that refuted it. Practices persist for reasons other than serving their users: because they serve whoever holds power, because exit costs are high, because ritual defends itself, because nobody measured. Lindy also assumes a stable environment in silence. A practice durable under old conditions carries no warranty for new ones; ask any industry that met the internet with five centuries of confidence. Treat Lindy as a prior with modest weight, and pair it with a live check on whether the environment moved.

Four: removal is also an intervention, with iatrogenics of its own. Chesterton's fence (The Thing, 1929) is the standing warning: if you cannot state why the fence was built, you have not earned the right to remove it. The calendar audit above encodes the defence twice — the function question comes before any deletion, and every deletion stays reversible for a month. Subtract a control or a buffer whose purpose you never understood, and you are the naive interventionist again, this time holding a delete key. Via negativa without the fence rule collapses into vandalism with a philosophy.

Perfection is finally attained not when there is no longer anything to add, but when there is no longer anything to take away.Antoine de Saint-Exupéry, Terre des hommes (1939)

Saint-Exupéry wrote that as a working pilot, about aircraft, where every part is also a failure mode. The sentence survives because it is an operating instruction. The finished state of a system is discovered by removal, and much of what you added along the way was scaffolding.

The stoics ran this program on the self long before Taleb priced it for systems; Marcus's morning question — is this necessary? — is via negativa in four words. Your next quarter will offer you additions by the dozen, and every vendor and every framework is one. Nobody will ever sell you a removal. Schedule the subtraction yourself, run it with a reversibility window, and count the hours it returns. Then, before you add anything, ask the sculptor's question: is this part of the statue, or part of the block?

Sources

  1. primaryNassim N. Taleb, Antifragile: Things That Gain from Disorder (2012) — via negativa, iatrogenics, subtractive epistemology, the Lindy effect.
  2. primaryPseudo-Dionysius the Areopagite, The Mystical Theology (5th–6th c.) — apophatic method: description by negation.
  3. primaryKarl R. Popper, The Logic of Scientific Discovery (1934/1959) — the asymmetry of confirmation and refutation.
  4. primaryMarcus Aurelius, Meditations, IV.24 (c. 170–180) — "do few things": the necessity test.
  5. primaryAntoine de Saint-Exupéry, Terre des hommes (1939) — perfection as the end state of removal.
  6. secondaryGabrielle S. Adams, Benjamin A. Converse, Andrew H. Hales, Leidy E. Klotz, "People systematically overlook subtractive changes", Nature 592 (2021).
  7. secondaryPeter F. Drucker, The Effective Executive (1967) — systematic abandonment; "what would happen if this were not done at all?"
  8. secondaryMichael E. Porter, "What Is Strategy?", Harvard Business Review (1996) — strategy as the choice of what not to do.
  9. secondaryAlbert Goldman, "Lindy's Law", The New Republic (1964) — origin of the Lindy observation.
  10. secondaryJason Fried, David Heinemeier Hansson, Getting Real (2006) — "build less"; no as the default answer.
  11. secondaryG. K. Chesterton, The Thing (1929) — the fence parable: understand before you remove.
  12. secondaryJames G. March, "Exploration and Exploitation in Organizational Learning", Organization Science 2(1) (1991).