A compressed argument about what compression costs.
Every major forecasting failure of the past two decades – in energy, economics, epidemiology, technology – shares a common structure. Not bad data. Not insufficient computing power. Not even ideological bias, though that plays a role. The deeper failure is conceptual: treating an open problem as a closed one.
The distinction matters more than most analysts acknowledge.
Some complex systems approach what we might call the closed ideal. Metabolic networks, power grids, digital protocols, supply chains, cascade within a bounded possibility space. The components are largely known. The governing rules are relatively stable. The interactions may be astronomically large in number, yet enumerable in principle. Uncertainty is real, but it is epistemic: we may not know everything, but we know the space where things can happen. Better models, more data, faster computation these genuinely help.
Go is a closed system in exactly this sense. With some 10^170 legal board positions, it dwarfs chess by orders of magnitude — yet the rules are fixed, the board finite, the win condition unambiguous. When AlphaGo played Move 37 against world champion Lee Sedol in 2016 a move so unexpected that Sedol briefly left the room — it was a genuine achievement: a system navigating an astronomical but bounded space, finding patterns no human had seen. The board had not changed in two thousand years.
Many real problems behave as if the board could expand mid-game.
Open systems economies, wars, pandemics, organizations, societies don’t merely evolve in time. They evolve in structure. As events unfold, new actors emerge, new couplings form, cascades of second-order effects alter the conditions for everything that follows. The crucial variables are not simply unknown. They do not yet exist.
This is the domain of contingency: the recognition that multiple futures remain genuinely possible, and that the future is not merely discovered but partially created through action, accident, and the interactions of agents who are themselves trying to anticipate each other.
The history of forecasting is largely a history of underestimating this.
Who imagined, at the invention of the automobile, that America would end up with more cars than drivers or that this would generate not just new industries and infrastructures but entirely new ways of organizing cities, families, and daily life? The automobile did not merely satisfy existing demand for transportation. It restructured the space of what was possible.
Who imagined, after two decades of essentially flat per-capita electricity demand, that AI data centers consuming the power of mid-sized cities would suddenly rewrite the energy equation straining grids, reviving mothballed power plants, and forcing utilities to abandon forecasts they had built their capital plans around? The transition was not a deviation from a trend. It was a transformation of the system generating the trend.
The 2008 financial crisis offers perhaps the starkest case. The models that failed were not naive. They were sophisticated, well-funded, and staffed by some of the most quantitatively capable people in the world. What they assumed implicitly, structurally was that the correlations governing normal times would hold under stress. They treated a system capable of endogenous transformation as if it were merely complex. When the structure itself shifted, the models had no language for what was happening. They were, in a precise sense, solving the wrong problem.
The forecasters who missed these shifts were not careless or naive. They were applying closed-system thinking to an open-system world treating the future as computable within a known space, when what was actually unfolding was a transformation of the space itself.
Hindsight hides this from us. Once events occur, we construct narratives that make outcomes appear inevitable. The unrealized alternatives disappear because they leave no traces. We forget how open the world once was, how many paths were available, how different the future looked to intelligent observers at the time.
The philosopher Michael Oakeshott called this the “abridgment” of history: the reduction of a contingent unfolding into a tidy causal sequence. What gets lost may be what mattered most, the genuine openness of the situation, the real possibility of other outcomes. Backshadowing describes the tendency of historical narratives as if the past were inevitable.
And we do the same thing to the future. We project from the present as if the structure of the present were fixed — as if we were navigating a known board rather than one that might expand, contract, or change its rules entirely. Clocks and clouds capture these narratives.
Clock thinking and cloud thinking are two fundamentally different orientations toward complexity.
Clocks are complicated. They have many parts, precisely engineered, interacting in predictable ways. A skilled clockmaker can take one apart and reassemble it. Given the current state, the future state is in principle calculable. The metaphor suits much of classical physics, most of engineering, and a surprising amount of economic modeling.
Clouds are complex. They are composed of innumerable interacting elements, but the interactions are nonlinear, sensitive to initial conditions, and generative of emergent properties that cannot be read off from the parts. You cannot take a cloud apart and reassemble it. The future state is not merely unknown — it is, in a meaningful sense, undetermined. Small perturbations ramify. Structure evolves. What the cloud “is” at any moment is inseparable from how it is changing.
Most of the problems that matter — economic systems, political orders, technological transitions, ecological dynamics — are clouds that we persist in treating as clocks. The persistence is not irrational. Clock thinking offers something that cloud thinking cannot: precision, tractability, the comfort of a definite answer. The question is whether precision attached to the wrong model is worth having.
The appropriate response to a clock problem is a better forecast. The appropriate response to a cloud problem is something harder to name — call it orientation: a stance toward uncertainty that preserves optionality, that attends to weak signals, and that remains capable of revision when the structure shifts.
Which brings us back to the hook.
“The hook,” “the thesis,” “the three takeaways”, this is clock thinking applied to ideas. It assumes that arguments can be compressed without loss: that the essential content can be separated from the form, that the structure is merely packaging.
Sometimes it can. Some arguments are genuinely summarizable. The periodic table is not diminished by a summary. A well-specified model can be stated precisely in fewer words than the paper that derives it.
But some arguments are not summaries of themselves. The structure is part of the content. The unresolved tensions are not failures of compression: They are the argument. Strip them away, and what remains is not the same idea made accessible. It is a different idea, reshaped to fit the available space and today’s attentional economy.
This is the paradox of compression in the age of open systems — the ideas we most need the ones that preserve contingency, resist premature closure, keep the future genuinely open — are precisely the ones least suited to the form we have built to carry them. The medium selects against the message.
Compression itself can erase contingency.
This is an argument for calibration: knowing which kind of problem you are facing before deciding how to approach it. Not all problems reward the same cognitive tools. Not all arguments reward the same editorial instincts.
In a closed system, the right response may be a better forecast.
In an open one, the deeper question is whether forecasting is the right response at all.
That question does not compress into a takeaway.
It asks for orientation, not certainty.
