../

Models from other fields

A catalog of the big ideas from physics, engineering, chemistry, biology, economics, mathematics, military strategy, computer science and social life that transfer to business, engineering and everyday decisions. Each table gives the model, the idea, the question it prompts and an example; the most useful models get a worked section. General tools are in thinking tools, probability and statistics in probability and risk, psychology in depth in cognitive biases. Feedback loops, stocks and flows are in systems thinking.

How to use this catalog

The value of borrowing from another field is a different question. A product problem viewed through physics ("where is the friction?"), biology ("what niche does this fit?") and economics ("who is paid to do what?") produces three different diagnoses. Munger's argument for a multidisciplinary "latticework of models" is summarized in thinking tools.

StepDo
1state the problem in one sentence, without jargon
2run three models from three different fields; write the question each one asks
3note where they agree (probably true) and disagree (where to dig)
4check each model's domain: does the analogy actually hold here?
5act on the diagnosis that is cheapest to test

Physics and engineering

ModelIdeaUse it to askExample
leveragea small force at the right point moves a large load"Where does one unit of effort produce the most output?"code, media and capital are leverage: one engineer's library used by 1,000 teams
inertia and momentumobjects keep doing what they're doing; changing direction takes force proportional to mass"How much force will it take to change this, and what's the cost of stopping what's already moving?"a big company's roadmap; a habit; a growing product with word-of-mouth
frictionresistance that wastes energy at every step"What small obstacles stand between the user (or team) and the action?"removing one sign-up field; a one-command dev setup
activation energya reaction needs an initial push before it proceeds by itself"What's the up-front hump that stops people starting?"a new tool nobody tries because setup takes a day
critical massbelow a threshold a chain reaction dies out; above it, it sustains itself"How many participants before this is self-sustaining?"a marketplace needs enough sellers per city before buyers stay
entropyclosed systems drift toward disorder; order takes continuous energy"What maintenance does this need, or it decays?"docs rot, dependencies age, codebases and cultures degrade without upkeep
equilibriumsystems settle where forces balance; disturbed, they tend back (Le Chatelier's principle in chemistry)"What forces hold the current state in place?"a price war ends at a new equilibrium; a reorg drifts back to old habits
redundancy and backupsspare components so one failure doesn't stop the system"What is the single point of failure?"N+1 servers; two people who can deploy; a second supplier
margin of safetydesign for loads well above the expected maximum"Do we survive being wrong by 30–50%?"capacity headroom; runway buffer (see thinking tools)
bottlenecksthroughput is set by the slowest step"Where does work queue up?"see theory of constraints below
feedback and controlmeasure output, compare to target, adjust input"What's the sensor, what's the setpoint, how long is the delay?"autoscaling; OKRs; see systems thinking
breakpoints and non-linearitybehavior changes abruptly at thresholds (phase transitions: water at 0 °C and 100 °C)"Where does more of the same stop working, or suddenly work?"a database fine at 1M rows falls over at 100M; a team of 8 vs 15
scale and the square-cube lawscale a shape up and volume grows with the cube of size, area with the square (Galileo, Two New Sciences, 1638)"What grows faster than what as we get bigger?"why ants can't be elephant-sized; why coordination costs outrun headcount

Bottlenecks and the theory of constraints

Eliyahu Goldratt's business novel The Goal (1984, with Jeff Cox) argues that any system has one constraint that limits its output, and improving anything else is wasted effort. His five focusing steps:

  1. Identify the constraint.
  2. Exploit it: get the most out of it as it is (never let it sit idle).
  3. Subordinate everything else to it: other steps work at its pace, not their own.
  4. Elevate it: add capacity only after the steps above.
  5. Repeat: once it's no longer the constraint, find the new one.

Worked example. A team's PRs wait on average two days for review and 20 minutes for CI. Buying faster CI runners (non-constraint) changes nothing. Exploit: reserve review blocks each day. Subordinate: cap work in progress so people review before starting new work. Elevate: train two more reviewers for the core module. The constraint then moves, perhaps to QA or product decisions.

Where it misleads. Knowledge work has shifting, invisible constraints (attention, decisions, unclear requirements), not a fixed machine. And local utilization targets ("keep every engineer 100% busy") create queues; see Little's law on system design.

Scale: the square-cube law for organizations

Communication channels in a group of nn people grow as n(n−1)2\frac{n(n-1)}{2}: 10 channels for 5 people, 45 for 10, 1,225 for 50. Output grows roughly linearly with headcount; coordination grows roughly quadratically. Fred Brooks made the software version famous in The Mythical Man-Month (1975): adding people to a late software project makes it later (Brooks's law, paraphrased). The fix at scale is structure: small teams, clear interfaces between them, fewer required channels.

Chemistry

ModelIdeaUse it to askExample
catalystsa substance that lowers activation energy and speeds a reaction without being used up"What person, tool or event would make this happen faster without being consumed?"a respected engineer adopting a tool first; a deadline; a template that makes the first draft trivial
alloyingcombining elements gives properties neither has alone (steel: iron plus a little carbon is far stronger than iron)"What combination produces a property no single element has?"engineer + domain expert co-founders; design + data in one team
activation energysee physics above; chemistry's original model"How do we lower the hump?"free trial, sample data, one-click import

Biology

ModelIdeaUse it to askExample
evolution by natural selectionvariation + selection + inheritance → adaptation, without a designer"What varies, what selects, and what gets copied forward?"A/B testing; startups as variants selected by markets; practices that spread through a company
adaptation and the Red Queen effectcompetitors co-evolve, so you must keep improving just to hold position (Leigh Van Valen, "A New Evolutionary Law", 1973)"If we stand still for a year, where do we end up relative to competitors?"security arms races; ad-targeting vs ad-blocking; feature parity treadmills
ecosystems and nichesspecies survive by fitting a niche; ecosystems have interdependencies"What niche do we fill that nobody else fills as well? Who depends on whom?"a vertical SaaS for dental clinics vs a generic CRM; platform partners
hormesisa low dose of a stressor triggers adaptation that makes the organism stronger; a high dose harms"What small, controlled stress would make this system stronger?"exercise; chaos engineering; game days; small code-review critiques early in a career
homeostasisorganisms actively hold key variables (temperature, blood sugar) within a range"What is this system defending, and how will it push back if I change it?"a team re-creates the old process after a reorg; prices that snap back
incentives in biologybehavior follows payoffs; costly signals are credible because they're costly (Amotz Zahavi's handicap principle, 1975)"What does this behavior pay off for the actor? Is the signal costly enough to be honest?"a free "commitment" is cheap talk; a signed LOI with a deposit is a costly signal
replicationwhat copies itself with enough fidelity and selection spreads"Does this idea or feature copy itself? What's its copying fidelity?"viral loops; memes (Richard Dawkins coined the word in The Selfish Gene, 1976)

Evolution, worked. A growth team runs 40 experiments a quarter. Evolution suggests: maximize variation (many cheap, different tests, not 40 button colors), make selection honest (pre-registered metrics, adequate sample sizes), and ensure inheritance (winning changes get merged and documented, not lost when the experimenter leaves).

Where biology misleads. Evolution optimizes for survival and reproduction, not for anyone's goals, and it's slow and wasteful: most variants die. "Survival of the fittest" means fittest to the current environment, which can change overnight. And borrowing biology for social questions has an ugly history (social Darwinism); keep it to mechanisms, not moral conclusions.

Economics

Theory and worked examples are on microeconomics and macroeconomics; this is the portable version.

ModelIdeaUse it to askExample
supply and demandprice moves to balance quantity wanted and quantity offered"What happens to price if demand or supply shifts?"senior ML engineers' salaries rising when every company wants AI features
incentivespeople respond to rewards and costs, including unintended ones"Who gets rewarded for what? What would I do in their shoes?"sales paid on bookings, not retention, sell to customers who churn
comparative advantagespecialize in what you give up least to do, even if someone else is better at everything (David Ricardo, 1817)"What is my opportunity cost of doing this myself?"the CTO who codes best should still delegate code and do hiring
opportunity costthe cost of a choice is the best alternative forgone"Compared to what?"see decision-making
diminishing returnseach extra unit of an input adds less output"Is the next unit of effort worth as much as the last?"the fifth round of design polish; the 12th engineer on one service
economies of scaleaverage cost falls as volume grows"Does unit cost fall with volume here, and do we have the volume?"cloud providers; a startup can't win on cost against incumbents
network effectseach user makes the product more valuable to others (Metcalfe's law says value grows with n2n^2; the exponent is disputed)"Does one more user make it better for the others, and for whom exactly?"messaging apps, marketplaces; a to-do app has none
switching coststhe cost of moving to a competitor, in money, time, data or retraining"What would it cost a customer to leave?"data lock-in, integrations, trained staff
principal–agent probleman agent acting for a principal has different incentives and more information (formalised by Michael Jensen and William Meckling, 1976)"Whose interests does this person actually serve?"an agency paid by the hour; managers optimizing their own metric; a founder vs investors
tragedy of the commonsshared resources get overused when each user bears only a fraction of the cost (Garrett Hardin, Science, 1968; Elinor Ostrom's Governing the Commons, 1990, shows communities often manage commons well with the right rules)"Who pays for the shared thing, and who uses it?"a shared staging environment; a monorepo's CI minutes; everyone's calendar
creative destructioninnovation destroys old industries and firms as it creates new ones (Joseph Schumpeter, Capitalism, Socialism and Democracy, 1942)"What does our success destroy, and what will destroy us?"streaming vs DVD rental; LLMs vs parts of search
Goodhart's lawa measure turned into a target stops measuring well (Charles Goodhart, 1975; the popular wording is Marilyn Strathern's, 1997)"How would a smart person game this metric?"lines-of-code targets; tickets closed; test coverage %

Incentives

If you remember one model from economics, make it this. Munger's talk "The Psychology of Human Misjudgment" (1995; revised text 2005) puts reward and punishment first in its list of tendencies; the self-deception version, incentive-caused bias (people come to believe what they're paid to believe), is on cognitive biases.

Worked example. A support team is measured on average handle time. First-order effect: shorter calls. Second-order: agents close tickets before the problem is solved; customers call back; total contacts rise and satisfaction falls. Better design: measure first-contact resolution with a counter-metric (repeat contacts within 7 days), and read a sample of tickets.

INCENTIVE AUDIT
For each actor (employee, team, customer, partner, investor):
  What are they paid / praised / promoted for?
  What are they punished for?
  What do they know that I don't?
  If I were them, what would I do - including gaming it?
  Does that match what we want? If not, change the
  incentive, not the speech.

Mathematics

ModelIdeaUse it to askExample
compoundinggrowth on growth: FV=PV(1+r)nFV = PV(1+r)^n"What small rate, sustained for years, beats a large one-off?"skills, relationships, reputation, code quality, money
power lawsa few items account for most of the total; no "typical" value"Is this distribution dominated by its extremes?"startup returns, book sales, customer revenue, bug impact
multiplicative systemswhen factors multiply, a zero anywhere zeroes the result"Is this a product of factors? Which one is near zero?"a conversion funnel; team × market × product
regression to the meanextreme results tend to be followed by more average ones"Was the extreme result partly luck?"see probability and risk
permutations and combinatoricsthe number of combinations explodes with the number of parts"How many states or paths are there really?"feature flags (210=1,0242^{10} = 1{,}024 combinations), test matrices, coordination channels
asymmetry and convexitya convex payoff gains more from upside than it loses from downside; Nassim Taleb's Antifragile (2012) builds a philosophy on seeking convex bets"Is my downside capped and my upside open?"cheap experiments with large possible payoffs; options; open-source side projects

Compounding

FV=PV(1+r)n,doubling time≈72r (%)FV = PV(1 + r)^n, \qquad \text{doubling time} \approx \frac{72}{r\,(\%)}

Worked example. $10,000 at 7% a year for 30 years grows to about $76,000 (1.0730≈7.611.07^{30} \approx 7.61). The rule of 72 gives a doubling time of about 10 years (72/7≈10.372/7 \approx 10.3), so roughly three doublings. The same shape applies to anything whose growth feeds on its current size: an audience, a codebase's tech debt interest, a team's reputation for shipping.

Where it misleads. Real growth rates don't hold: compounding curves become S-curves when they hit a market size, a physical limit or competition. "Get 1% better every day" (1.01365≈37.81.01^{365} \approx 37.8) is arithmetic, not a finding about skills. And compounding works in both directions: small persistent losses (churn, a slow regression) compound too. Money specifics are on investing.

Multiplicative systems

10,000 visitors
 x 3%  sign up        =    300
 x 40% activate       =    120
 x 10% pay            =     12 customers
 
Double any one step   ->  24 customers
Any step at 0%        ->   0 customers

In an additive system you can compensate for a weak part; in a multiplicative one you can't. Early-stage companies are multiplicative (team × market × product × distribution), which is why fixing the worst factor beats polishing the best one.

Power laws

In a power-law world the average is misleading and the outliers are everything. Venture returns are the classic case: a small fraction of investments return most of the fund. Consequences: in power-law domains, optimize for the size of the upside, not the hit rate; in normal-distribution domains (heights, most operational metrics), averages and consistency are fine. Peter Thiel's version of this argument is on Peter Thiel's advice.

Military and strategy

ModelIdeaUse it to askExample
OODA loopobserve, orient, decide, act, faster than the opponent (John Boyd)"Whose loop is faster, and where is ours slow?"see the OODA loop
Schwerpunkt / concentration of forceSchwerpunkt ("point of main effort", German doctrine): concentrate strength at the decisive point instead of spreading it evenly"Where is our one main effort this quarter?"one launch, one segment, one metric; not ten priorities at 10% each
asymmetric warfarea weaker side avoids the stronger side's strengths and attacks where it's weak"Where can't the incumbent follow us without hurting itself?"a startup targets a segment too small for the incumbent, or a model that cannibalises its revenue
seeing the frontleaders who go to where the work happens see what reports and dashboards miss"When did I last see this with my own eyes?"founders doing support shifts; managers sitting in on sales calls; see extreme ownership

Where military models mislead. Business is mostly positive-sum; framing every customer as terrain and every competitor as an enemy produces paranoid, zero-sum decisions. Borrow the concepts about speed, focus and information; leave the rhetoric.

Psychology basics

Human behavior runs through every model on this page. The biases are cataloged in cognitive biases; the three that most often break borrowed models:

ModelIdeaUse it to askExample
incentive-caused biaspeople sincerely come to believe what they're rewarded for believing"What would this person believe if they were paid differently?"a vendor's benchmark; a team defending its own project
confirmation biaswe seek and remember evidence for what we already think"Where's the evidence against?"a founder who only hears the positive customer calls
man-with-a-hammerthe tool you know best becomes the lens for everything"Am I reaching for this model because it fits, or because it's mine?"the engineer who sees every problem as a tooling problem

Computer science

Several of these are drawn from Brian Christian and Tom Griffiths, Algorithms to Live By (2016).

ModelIdeaUse it to askExample
explore/exploittrade off trying new options (explore) against using the best known (exploit); the multi-armed bandit problem"How much time is left? Explore more when the horizon is long."try new restaurants when you've just moved; exploit favorites before you leave; early-stage vs late-stage product bets
optimal stopping (37% rule)with nn options seen one at a time and no going back, look at the first n/e≈37%n/e \approx 37\% without committing, then take the first one better than all so far; this picks the best about 37% of the time"Am I in a sequential, no-recall search? Have I set my look-then-leap point?"a 30-day flat search: look for ~11 days, then take the first flat better than everything seen
cachingkeep what you'll need soon close at hand; evict the least recently used"What do I access most, and is it within reach?"a desk with today's work; a CDN; a FAQ for the ten most common support questions
batchinggroup similar tasks to amortize the switching cost"What set-up cost am I paying repeatedly?"email twice a day; deploy trains; batch invoice processing
sortingordering costs work up front; only worth it if you'll search later"Will I ever search this? If not, don't sort it."inbox zero vs search; tagging every document nobody will look up
abstractionhide detail behind a simpler concept so you can reason at a higher level"What level of detail does this decision need?"a finance model at monthly granularity; an API that hides a vendor
interfacesagree on the contract between parts so each can change independently"What's the contract between these teams or systems?"team APIs, SLAs, a clear RACI; see decision-making
big-O intuitionhow cost grows with input size matters more than cost at today's size"What happens to this process at 10× and 100×?"a manual onboarding step that is O(customers); a meeting that is O(team²)
idempotencedoing it twice has the same effect as doing it once"Is it safe to retry?"payment APIs with idempotency keys; runbooks you can re-run; a policy that's fine if applied twice
tech debtshortcuts are a loan: you ship sooner and pay interest until you repay the principal (Ward Cunningham, OOPSLA 1992)"What's the interest rate on this shortcut, and when do we repay?"a hard-coded config fine for one customer, painful at twenty

Where CS models mislead. Algorithms assume a well-defined problem: the 37% rule needs a known number of options, no recall and rank-only information, and real searches often let you go back or learn values. People are not caches, and a "batch" of one-to-ones delays feedback. The metaphors are useful because they are precise; check the preconditions.

Tech debt as a model

Cunningham's framing (1992) is a financial metaphor: some debt is a sensible loan taken to learn faster, and the danger is not the borrowing but never repaying. It generalizes beyond code:

KindLoanInterest
codecopy-paste, missing tests, hard-coded valuesslower changes, more bugs
process"just DM me" instead of a documented processevery new hire asks the same questions
organisationala reporting line that "works for now"unclear ownership, slow decisions
producta one-off feature for one big customerevery later feature has to work around it

Take debt deliberately, write it down, and schedule repayment when the interest (measured in slowed work) exceeds the cost of paying it off.

Human nature and social

ModelIdeaUse it to askExample
trusttrust lowers transaction costs: fewer checks, contracts and approvals"What would we stop doing if we trusted each other more? What earns it?"high-trust teams skip approval layers; trust is built by keeping small promises
reciprocitypeople feel obliged to return favors (Robert Cialdini, Influence, 1984)"What have I given before asking?"open-source contributions before asking a maintainer for help; useful content before a sales call
social proofunder uncertainty people copy what others do (Cialdini)"What will people see others doing?"logos and case studies; the first 10 users who make a community look alive
statuspeople care intensely about relative standing, often more than absolute rewards"How does this change who looks good or bad?"a title change that costs nothing but causes resignations; public praise
reputation as a stockreputation accumulates slowly from many actions and can drain quickly from one (a stock, in systems thinking terms)"Is this action a deposit or a withdrawal, and how large?"a data breach handled badly; years of reliable on-time delivery

Reputation, worked. Every on-time delivery, honest post-mortem and kept promise is a small inflow. A hidden outage or a broken commitment is a large outflow. Because the stock builds slowly and drains fast, the asymmetric move is to protect it: disclose problems early, under-promise and deliver.

Where social models mislead. Influence techniques used cynically work until they're noticed, then destroy trust; reciprocity manufactured through small "gifts" is manipulation. Social proof also spreads mistakes (information cascades).

Building your own latticework

  1. Pick a small core

    Start with about ten models from at least five fields: compounding, incentives, opportunity cost, bottlenecks, feedback loops, margin of safety, inversion, evolution, power laws, map vs territory.

  2. One model a week, applied deliberately

    Each week pick one model and apply it on purpose to three real situations: a work decision, a news story, a personal choice. Write two lines per application: what the model predicted and what it missed.

  3. Learn the source, not the summary

    Read one primary text per field (a textbook chapter, the original paper, the book that named the idea). Summaries flatten the conditions under which a model holds.

  4. Collect failures

    Keep a list of times a model misled you. The boundaries of a model are more valuable than its definition.

  5. Cross-check with a second field

    For any important decision, run at least two models from different disciplines. Agreement raises confidence; disagreement shows where to look.

  6. Explain it to someone

    If you can't explain a model plainly with a real example, you don't own it yet (see learning by explaining in thinking tools).

MODEL CARD (one per model, keep them in a notes file)
Name:
Field of origin / source:
One-line idea:
Question it makes me ask:
Preconditions (when it holds):
Worked example from my own work:
Where it misled me:
Related models:

Pitfalls

PitfallWhat it looks likeFix
model-hoppinga different model for every meeting, none applied properlydepth over breadth; finish the analysis with one model before adding another
fitting the world to a favorite modeleverything is network effects; everything is incentivesname two alternative models and ask what each would predict
vocabulary without mechanism"we need more momentum" with no idea what mass or force means herestate the mechanism in plain words; if you can't, drop the metaphor
precision theatrephysics equations applied to fuzzy social quantitiesuse models for direction and magnitude, not decimals
ignoring preconditionsthe 37% rule for a decision where you can go backlist a model's assumptions before using its conclusion
retrofittingpicking the model after deciding, to justify the choicewrite down which model you're using before the conclusion
collecting instead of usinga list of 200 models, none habitualfewer models, more reps
ignoring the local expertan outsider's analogy overrides someone who knows the domainanalogies generate questions; domain knowledge answers them

References

  • Charles Munger, "A Lesson on Elementary, Worldly Wisdom" (USC, 1994), Farnam Street transcript (opens in a new tab): the case for multidisciplinary models
  • Eliyahu M. Goldratt and Jeff Cox, The Goal: A Process of Ongoing Improvement (North River Press, 1984): theory of constraints, five focusing steps
  • Galileo Galilei, Dialogues Concerning Two New Sciences (1638): the square-cube law
  • Frederick P. Brooks Jr., The Mythical Man-Month (Addison-Wesley, 1975): Brooks's law, communication overhead
  • Leigh Van Valen, "A New Evolutionary Law", Evolutionary Theory 1 (1973): the Red Queen hypothesis; overview (Wikipedia) (opens in a new tab)
  • Amotz Zahavi, "Mate Selection: A Selection for a Handicap", Journal of Theoretical Biology 53 (1975): the handicap principle
  • Richard Dawkins, The Selfish Gene (Oxford University Press, 1976): replicators and memes
  • David Ricardo, On the Principles of Political Economy and Taxation (1817): comparative advantage
  • Michael C. Jensen and William H. Meckling, "Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure", Journal of Financial Economics 3(4), 1976: principal–agent costs
  • Garrett Hardin, "The Tragedy of the Commons", Science 162 (1968) (opens in a new tab)
  • Elinor Ostrom, Governing the Commons (Cambridge University Press, 1990): how communities manage shared resources
  • Joseph A. Schumpeter, Capitalism, Socialism and Democracy (1942): creative destruction
  • Goodhart's law (Wikipedia) (opens in a new tab): Goodhart 1975, Strathern 1997, Campbell's law
  • Nassim Nicholas Taleb, Antifragile: Things That Gain from Disorder (Random House, 2012): convexity and asymmetric payoffs
  • Brian Christian and Tom Griffiths, Algorithms to Live By: The Computer Science of Human Decisions (Henry Holt, 2016): explore/exploit, optimal stopping, caching, sorting
  • Secretary problem (Wikipedia) (opens in a new tab): the 37% rule and its assumptions
  • Ward Cunningham, "The WyCash Portfolio Management System", OOPSLA '92 experience report (1992): origin of the technical debt metaphor
  • Robert B. Cialdini, Influence: The Psychology of Persuasion (1984; revised 2021): reciprocity, social proof
  • Microeconomics: the full treatment of supply and demand, elasticity, market failure