Jason Karpeles · Companion

Systems Thinking with AI, illustrated

Each reader takes one chapter's own model and lets you change one or two values at a time. Every picture was computed in advance from the chapter's code pack.

Every example in these readers is a constructed teaching example built from the chapter's own numbers. Chapters 36 to 39 start from committed public records; their fitted values are inferred from those records, and nothing here is a forecast.

Part I

Why systems surprise us

3 readers

  1. 03Chapter 3

    Events Are Not Explanations

    A reference mode is a behavior target, and the shapes it names depend on a few parameters.

    1. Growth looks like S-shaped growth until the ceiling bites
    2. Decay is goal seeking with the goal at zero
    3. One erosion rate turns a ceiling into a collapse
    4. Noise hides a slow approach first

    Open reader

  2. 04Chapter 4

    The Bathtub Most People Misread

    A level and a rate have different units, and reading one off the other is the error.

    1. A falling inflow can still fill the tub
    2. The step length turns a rate into an amount
    3. Read every flow, then write every stock
    4. A floor is a claim, and conservation can see it

    Open reader

  3. 05Chapter 5

    The Beer Game and the Cost of Local Rationality

    Four sensible ordering desks, joined by delays, make a swing no desk intended.

    1. One station's order, with and without the supply line
    2. Expected demand lags a step in orders
    3. A four-case step grows at every station upstream
    4. The supply-line fix lands three stations away

    Open reader

Part II

From a messy problem to a dynamic hypothesis

2 readers

  1. 08Chapter 8

    Causal-Loop Diagrams Without Causal Theater

    A diagram is a list of claims, and each claim can be counted, audited and removed.

    1. Two negative links make a reinforcing loop
    2. An audit counts what the picture rests on
    3. One more arrow closes loops nobody counted
    4. Remove the weakest arrow

    Open reader

  2. 09Chapter 9

    System Archetypes as Hypothesis Templates

    A template accepts almost any situation; the boundary and the observation decide what it predicts.

    1. An early limit is easy to dismiss
    2. Same engine, two boundaries
    3. The window decides what the record can show
    4. One parameter turns B into A

    Open reader

Part III

The formal machinery

8 readers

  1. 12Chapter 12

    Stocks, Flows, Sources, and Sinks

    Every flow names two endpoints and a unit with a time base, and the accounting can then check itself.

    1. A tub with no drain can only grow
    2. Conservation as an executable claim
    3. A clamp hides a mechanism; a limited flow names it
    4. Units as a type system: a rate times a time

    Open reader

  2. 13Chapter 13

    Auxiliaries, Parameters, and Decision Rules

    Four kinds of quantity share one notation, and each needs its own evidence and its own test.

    1. A decision rule that never goes negative
    2. A belief that lags the truth
    3. Dependencies fall out of the parse
    4. An auxiliary is tested by arithmetic

    Open reader

  3. 14Chapter 14

    Feedback and Loop Dominance

    A model with two loops has a dominant loop at a moment, and the moment is part of the answer.

    1. Two loops through one stock
    2. Knockout at one state
    3. The handover
    4. The same intervention in two regimes

    Open reader

  4. 15Chapter 15

    Nonlinearity and Lookup Functions

    A lookup shows where its evidence ends; a fitted formula keeps answering after it has none.

    1. Inside the data, a lookup and a fit agree
    2. Outside the data, the fit answers and the lookup refuses
    3. Monotonic and bounded are claims, so test them
    4. Does the conclusion survive a change of shape?

    Open reader

  5. 16Chapter 16

    Material and Information Delays

    Two delays with the same mean behave differently, and where a delay sits matters as much as how long it is.

    1. The conveyor and the tank
    2. Chaining tanks walks toward the conveyor
    3. A tank stepped at its own mean is a conveyor
    4. The same delay inside a correction loop

    Open reader

  6. 17Chapter 17

    Cohorts, Aging Chains, and Coflows

    People are not interchangeable units, and the experience they hold moves when they move.

    1. Experience cannot be left behind
    2. What the surge does
    3. Read headcount and average experience together
    4. The recommendation that reverses

    Open reader

  7. 18Chapter 18

    Events, Queues, Agents, and Hybrid Models

    When an aggregate model and a queue are joined, the seam between them is part of the model.

    1. Exchange frequency is a modeling decision
    2. Nine patients per period become nine arrival events
    3. The staffing rule alone has an equilibrium
    4. Does the frequency finding survive another seed?

    Open reader

  8. 19Chapter 19

    Numerical Integration Is Part of the Model

    The same equations give different answers under different solvers, steps and update orders.

    1. Three answers from one model
    2. What Euler assumes about the rate
    3. Sequential update is a different model
    4. The refinement test

    Open reader

Part IV

Models as software

1 reader

  1. 25Chapter 25

    Turn Models into Management Flight Simulators

    A scenario runner is worth what its record can replay and what its constraints can catch.

    1. Three scenarios against a declared ceiling
    2. The settings belong in the record
    3. Averaging scenarios shows a path nobody ran
    4. A narrative may cite only the record

    Open reader

Part V

The AI roles

2 readers

  1. 29Chapter 29

    The AI as Experiment Designer

    Reduce the uncertainty that most changes the decision, at the least cost, without breaking anything.

    1. Rank uncertainties by their swing
    2. Rank by what it costs to find out
    3. One at a time against a joint sample
    4. Endpoints can miss a reversal

    Open reader

  2. 30Chapter 30

    The AI as Policy Searcher

    Rank by the worst case among the policies that satisfy every constraint, and say what was excluded.

    1. The best average is excluded
    2. Every draw, not the average
    3. Same draws, every policy
    4. When no policy is admissible

    Open reader

Part VI

Cases, calibration, and capstone

8 readers

  1. 32Chapter 32

    Growth Can Destroy the Engine of Growth

    A reinforcing loop can consume the capacity that sustains it, and the warning is a ratio nobody plots.

    1. Same growth policy, two hiring speeds
    2. Throttling intake makes the business smaller
    3. Which parameter decides the outcome
    4. The load ratio crosses one while the dashboard is green

    Open reader

  2. 33Chapter 33

    Technical Debt as a Stock

    Debt is a stock nobody can count, and the review window decides which policy looks wise.

    1. The review window picks the winner
    2. Capacity is not what the plan says
    3. Rework share moves before delivery does
    4. Test the conclusion against the proxy's range

    Open reader

  3. 34Chapter 34

    Hybrid Case: Hospital Capacity and Patient Flow

    A policy can improve the average and fail one group, and only a per-group table shows it.

    1. The policy that improves the average
    2. The staffing rule swings and does not settle
    3. A mean can hide the tail
    4. Stress the arrivals until a group disappears

    Open reader

  4. 35Chapter 35

    Fitting Is Not Confirming

    A fit is the answer to a search, and the holdout and the choice of knob decide what it is worth.

    1. Scoring a cycle by point-by-point distance
    2. How many knobs the record allows, and what they cost
    3. The tank, fitted by a grid
    4. The wrong knob passes the window and fails the holdout

    Open reader

  5. 36Chapter 36

    Elective Backlogs as a Stock

    A waiting list is a level, and the chapter fits one to a public record without forecasting it.

    1. A gap held open is the growth
    2. A fit that passes and a holdout that fails
    3. Which parameter decides the long-wait stock
    4. A recovery date, or the honest None

    Open reader

  6. 37Chapter 37

    Hiring Is a Pipeline, Not a Number

    A hiring target buys heads, and capability follows a ramp the record barely constrains.

    1. Heads rise and capability does not
    2. Two surveys, one identity
    3. A fit within tolerance, a holdout outside it
    4. Which parameter decides capability

    Open reader

  7. 38Chapter 38

    Capacity Arrives When the Price Has Gone

    A loop with two delays and nothing outside it can swing like a public record, which proves little.

    1. Period and amplitude, not the path
    2. Sensible at every step, and still a cycle
    3. The construction delay carries the period
    4. A period is not a date

    Open reader

  8. 39Chapter 39

    Delay Is a Curve, Not a Line

    Taxi-out rises with load while the quoted delay falls, and a cap can only be priced on the curve.

    1. Two measures, two slopes
    2. A convex curve fitted to one year
    3. Pricing a cap needs two points on the curve
    4. Padding hides the delay it responds to

    Open reader

No reader matches that search.