We are taught to expect straight lines. Work twice as hard and get twice the result. Spend twice as much and get twice the growth. Learn twice as much and become twice as capable. It is an intuitive way to understand the world because linear relationships are easy to see, easy to measure, and easy to predict.
The problem is that many of the systems shaping our lives do not behave that way.
Money compounds. Networks multiply. Ideas spread. Markets tip. Skills stack. Attention concentrates. Technology accelerates. Small advantages accumulate until they stop looking small. Sometimes nothing appears to be happening for a very long time, and then everything changes.
The mistake is assuming the curve is a line simply because we can only see the beginning of it.
- Straight-line thinking works until a system compounds, crosses a threshold, forms a network or loops back on itself.
- Each shape calls for a different habit: consistency, persistence, connection, direction.
- Before you forecast, ask what shape the system is.
01 · The line we expect
Linear thinking is useful. Until the system bends.
The simplest relationship in mathematics looks like this: increase the input by one, and the output increases by one. Predictable. Comfortable. A large amount of human planning quietly depends on this assumption. We estimate projects this way, forecast careers this way, set goals this way, and often judge progress this way.
But real systems contain interactions, dependencies, thresholds, bottlenecks, delays, and feedback loops. Once those appear, the straight line begins to bend. And when the line bends, intuition becomes less reliable.
Consider a relationship such as y = x². At the beginning, it may not look dramatically different from a straight line. Keep moving, though, and the distance between the two grows quickly. Exponential growth is even more deceptive. The early curve can look almost boring before it suddenly becomes impossible to ignore.
This is what makes nonlinear systems difficult. We tend to look at the first few steps and assume the next few will resemble them. Often, they do not.
02 · Compounding
Small differences do not stay small.
Compounding is one of the clearest examples of nonlinearity. The important feature is not simply that something grows. It is that the result of one period becomes part of the next calculation. Output becomes input.
Learning can behave this way. One concept helps you understand another. That second concept makes the third easier to recognize. Eventually you are not simply collecting information. You are building a network of ideas that changes how quickly you can understand what comes next.
Reputation can behave similarly. One opportunity creates evidence. Evidence creates trust. Trust creates larger opportunities. Larger opportunities create stronger evidence. A loop begins to form, and the system starts creating some of its own momentum.
Tiny differences in architecture can become enormous differences in outcome.
Imagine two teams that begin in roughly the same place. One gets feedback slightly faster. One has slightly better data. One removes a small amount of friction from every decision. After one cycle, the difference may be invisible. After one hundred cycles, they may no longer look like the same kind of organization.
03 · Thresholds
Sometimes progress is invisible until the rules change.
Some systems behave differently again. Increasing the input produces little visible change until a threshold is crossed. Then the state of the system changes.
A product can exist for years before distribution becomes dense enough for adoption to accelerate. A creator can publish for months before the body of work becomes large enough to generate discovery. A team can experiment with AI repeatedly before enough workflows become connected that the operating model itself changes.
Before the threshold, progress can look disappointing. After the threshold, the same system can appear obvious in hindsight.
This is why “nothing is happening” can be a dangerous conclusion. Something may be happening beneath the surface. You simply have not reached the point where the system reveals it yet.
The sharpest demonstration is a thought experiment by the sociologist Mark Granovetter. Picture a crowd of one hundred people deciding whether to join a riot. One person will act alone, another will act once one person has, another once two have, and so on up to someone who will only act when ninety-nine others already have. The crowd erupts completely. Now swap the person who needed one other for someone who needs two. No one is ready to follow the instigator, and the riot stops at one. The two crowds have almost the same average threshold and opposite outcomes.1
04 · Networks
The value is not always in the node. Sometimes it is in the connection.
If you add one person to an isolated task, you add one person. If you add one person to a network, you may add many new possible relationships. The node matters, but so do the connections between nodes.
A hundred disconnected people and a hundred deeply connected people are not the same system. The difference is not headcount. It is structure. Who can communicate with whom? Where does information travel? Which nodes connect different groups? Where do bottlenecks form? Where does knowledge disappear?
A famous rule of thumb, Metcalfe's law, says a network's value grows with the square of its users, because the number of possible pairs does. In 2006 Briscoe, Odlyzko and Tilly argued that most of those pairs are never worth making, and that value grows closer to n log n.2 The argument is about how steep the curve is. Nobody in it thinks the curve is a line.
These sound like organizational questions, but they are mathematical questions too. Increasingly, they are AI questions.
A collection of isolated AI tools is not the same thing as a coordinated intelligent system. A library of prompts is not the same thing as an architecture. The intelligence may live in the nodes, but the leverage often lives between them.
05 · Feedback
Direction matters before scale.
Feedback is another force that bends the line. In a simple system, output leaves the system. In a feedback system, output returns as input.
Imagine a product customers love. More people use it. More usage creates more data. More data improves the product. A better product attracts more customers. The loop strengthens itself.
Now reverse the direction. Reliability falls. Trust declines. Usage decreases. Less usage creates weaker feedback. Investment becomes harder to justify. The product weakens further.
Both systems contain feedback. One compounds upward and one compounds downward. This is why speed by itself is not enough. You do not want to accelerate a broken loop.
Systems thinkers call these reinforcing and balancing loops, and Donella Meadows' Thinking in Systems is still the clearest introduction to them.3
06 · Power laws
The average can hide the real story.
Many important outcomes are not distributed evenly. A small number of products generate a large share of revenue. A small number of creators capture enormous attention. A handful of nodes carry much of a network’s traffic. A few decisions shape years of downstream consequences.
This feels strange because human intuition likes averages. We expect most things to sit somewhere near the middle. Power-law environments do not cooperate. The largest outcome may not be twice the average. It may be one hundred times larger.
Mark Newman's survey of power laws lists word frequencies, city populations and the wealth of the richest among them. The usual first check is to plot the data on logarithmic axes and look for a straight line, though Newman warns that this alone is not proof.4
This matters in the age of AI because the cost of experimentation is falling. If a person can test thirty ideas instead of three, the goal may not be to make all thirty equally successful. The goal may be to create enough intelligent experiments to discover the few that produce disproportionate value.
That is a different strategy. It rewards experimentation, fast feedback, and the ability to recognize outliers early.
07 · AI
AI does not only make the old loop faster. It can create a tighter loop.
AI is often described as a productivity tool. Write faster. Research faster. Code faster. Analyze faster. That framing is useful, but incomplete.
If AI simply made every task twenty percent faster, its impact would be relatively easy to model. It would mostly be a linear productivity story.
What is more interesting is that AI changes the cost of experimentation. A person who once had time to explore three ideas may now be able to explore thirty. A prototype that once required several specialists can sometimes be created by one person with enough domain knowledge and AI fluency. Research, analysis, simulation, writing, and iteration can move much closer together.
More experiments create more feedback. More feedback improves judgment. Better judgment improves future experiments. The loop gets tighter.
And tighter loops compound.
08 · Skill stacks
Some capabilities multiply each other.
A person who understands customers has value. A person who understands data has value. A person who can prototype has value. A person who understands AI systems has value.
But someone who combines customer understanding, data, product thinking, systems design, and AI may not simply possess the sum of those abilities. The capabilities can increase the usefulness of one another.
Customer insight changes what gets built. Technical fluency changes what feels possible. Data improves prioritization. AI increases iteration speed. Product judgment determines which experiments deserve to survive.
This is one reason unusual combinations of skills can suddenly become powerful when technology changes. Their value is not always additive. Sometimes it is multiplicative.
09 · Data
The data can be correct while the model is wrong.
We love averages because they simplify reality. Average revenue. Average response time. Average productivity. Average growth. But averages can conceal the structure underneath.
Two datasets can have the same average and describe completely different worlds. One may be tightly clustered. Another may contain extreme outliers. One may be stable. Another may contain two distinct populations being blended into a single number.
So the question cannot only be, “What is the number?” It has to become, “What is the distribution?” Where does behaviour change? Which observations are pulling the average? Are there clusters? Is there a threshold? Is the relationship changing over time?
Averages describe. Structures explain.
The textbook case is Anscombe's quartet. In 1973 the statistician Francis Anscombe published four small datasets with the same mean, the same variance, the same correlation and the same fitted line. Plotted, one is a rough line, one is a smooth curve, one is a line with a single outlier, and one is a vertical stack with a single point far to the side. The summary numbers could not tell them apart. The pictures could.5
The same is true of prediction. A perfectly accurate dataset fed into the wrong model can still produce terrible conclusions. If you assume the future is linear when the underlying system is exponential, your forecast will fail. If you assume variables are independent when they interact, your forecast will fail.
The data can be right. The model can still be wrong.
10 · The signal
The future often looks small at first.
The earliest evidence of nonlinear change is usually easy to dismiss. A strange customer request. A tiny group behaving differently. An unexpected spike in a metric. A new workflow forming around technology that still looks primitive.
The important question is not always how large the signal is today. Sometimes the better question is: what happens if this compounds?
What happens if the cost keeps falling? What happens if adoption doubles? What happens if the feedback loop gets faster? What happens when two previously separate systems connect? What happens after the threshold?
This is the shift from looking at a moment to studying a curve.
11 · In practice
How to tell the shapes apart.
Knowing the shapes helps only if you can spot them in a real system. For each one there is a question to ask and a pattern to look for in the numbers.
| Shape | Ask | Look for |
|---|---|---|
| Compounding | Does this period's output feed the next period's input? | Constant percentage growth. On a log axis, a straight line. |
| Thresholds | Is there a level below which nothing visible happens? | A long flat stretch, then a sudden change of state. |
| Networks | Does value depend on who is connected to whom? | Value per user rising as users are added. Clusters and bridges. |
| Feedback | Does output return as input, and does it push the system further or back? | Runaway growth or collapse, or a steady return to a set point. |
| Power laws | Do a few cases carry most of the total? | The mean far above the median. A straight line on a log-log plot. |
The twelve-number test.
Take any number you track and write down its last twelve values. Work out the change between each pair. If the change in units is roughly constant, the system is behaving linearly. If the percentage change is roughly constant, it is compounding. If the percentage change was rising and has started to fall, you may be on an S-curve and past the steepest part. It takes five minutes, and it turns "it feels like it is accelerating" into something you can check.
Two mistakes worth naming.
Reading the early curve as a line. Something that grows ten percent a month looks like steady progress for a quarter: 100, 110, 121, 133. After a year it stands at 314. After two years, 985. The first three months gave no hint of the last twenty-one.
Assuming every curve stays exponential. Most do not. Resources run out, markets saturate and people stop adopting, so the curve flattens into an S. A line fitted to the first half of an S will overshoot badly. The skill is holding two ideas at once: it is not a line, and it is not forever.
The takeaway
Think in curves, not moments.
The point is not to assume everything becomes exponential. Most things do not. Some systems plateau. Some decay. Some follow S-curves. Some oscillate. Some collapse. Some remain stubbornly linear.
The point is to stop assuming the line before examining the system.
If the system compounds, consistency matters. If it contains a threshold, persistence matters. If it follows a power law, experimentation matters. If it depends on networks, connections matter. If it contains feedback loops, direction matters. If variables multiply, balance matters.
You cannot understand a changing world by looking only at where things are today. You have to understand the shape.
By the time exponential change looks exponential, the most important part of the curve may already be behind you. By the time a signal becomes obvious, it is no longer an early signal.
If this is not a line, what is it?
The world is not linear. Our thinking cannot afford to be either.
Sources
- Granovetter, M. (1978). Threshold models of collective behavior. American Journal of Sociology, 83(6), 1420 to 1443. Stanford Sociology record
- Briscoe, B., Odlyzko, A., and Tilly, B. (2006). Metcalfe's law is wrong. IEEE Spectrum. spectrum.ieee.org
- Meadows, D. H. (2008). Thinking in Systems: A Primer. Chelsea Green Publishing.
- Newman, M. E. J. (2005). Power laws, Pareto distributions and Zipf's law. Contemporary Physics, 46(5), 323 to 351. arXiv:cond-mat/0412004
- Anscombe, F. J. (1973). Graphs in statistical analysis. The American Statistician, 27(1), 17 to 21.
Writes The Visionary, field notes from the AI-native frontier. Building in public: practical builds, named methodology, receipts, no fluff.