When the Rules Move

Why does intelligence exist at all?

In biological systems, the answer appears straightforward. Organisms must continuously allocate limited resources-energy, time, attention-under conditions of uncertainty. Prediction allows those allocations to occur before consequences fully unfold. A creature that anticipates where food will be found, where predators may appear, or when weather conditions may change gains a measurable advantage. We do look in detail at the role of prediction in biological systems in Why Intelligence Exists.

From this perspective, intelligence is not primarily a capacity for abstract reasoning or pattern recognition. It is a process embedded in time. It enables an organism to form expectations about the near future and to act accordingly.

A first principle follows from this view: prediction is valuable only insofar as it remains coupled to consequences. Expectations must be continuously tested against the world, and internal representations must adjust when those expectations fail. Without this coupling, prediction becomes detached from reality.

For most of evolutionary history, this adjustment occurred gradually. The physical environment changed, but the underlying rules governing survival remained relatively stable across generations. This stability allowed organisms to accumulate structures-neural or behavioral-that remained useful for extended periods.

The modern economic environment presents a different situation.

The structure of prediction

Prediction relies on two elements. The first is representation: a system must construct internal structures that capture regularities in its environment. The second is coupling: those representations must remain connected to the consequences that follow from action.

Most predictive systems emphasize the first element. They attempt to identify statistical regularities in historical data and encode those regularities into models. When the structure of the environment remains stable, this approach works well.

In such contexts, historical data provides a reliable guide to future outcomes. The patterns embedded in past observations continue to describe the world with sufficient accuracy.

However, prediction becomes fundamentally more difficult when the rules governing a system change. In such cases, the patterns captured in historical data lose relevance more quickly. Representations built from the past must be revised at a pace that matches the shifting environment.

The difficulty is not that the world has become more complex in an abstract sense. Rather, the stability of the underlying rules has decreased.

A world of moving rules

Many of the most significant disruption in recent years share a common feature: the governing structures of the system changed abruptly.

Consider the early months of the COVID-19 pandemic. Entire sectors of economic activity were halted almost overnight as governments introduced lockdown measures. Historical demand curves, supply chains, and operational models-often refined over decades-became temporarily irrelevant.

Airlines grounded fleets built around long-term travel demand. Restaurants faced sudden closures. Energy consumption patterns shifted dramatically as office districts emptied and residential demand increased.

The challenge in these situations was not the absence of data. In many cases, organizations possessed large volumes of historical information. The difficulty lay in the fact that the structural assumptions embedded in that data no longer held.

A similar dynamic appears in energy markets. The rapid expansion of renewable generation has introduced new patterns of variability into power systems. Solar and wind output fluctuate with weather conditions, creating supply structures that differ significantly from the more predictable generation profiles of conventional power plants.

Historical price dynamics in electricity markets-once shaped primarily by fuel costs and demand cycles-now interact with intermittency, regulatory adjustments, and evolving grid management strategies. Models trained on earlier market structures often require constant revision as these conditions evolve.

Geopolitical events can produce comparable shifts. The disruption of European gas supply chains in 2022 altered the structure of energy pricing across the continent. Systems designed around stable supply assumptions were forced to operate under entirely different conditions.

In each of these examples, the challenge was not merely statistical uncertainty. The rules governing the system changed.

Representation and coupling

These developments reveal an important conceptual distinction.

Prediction systems can be evaluated not only by how well they represent historical patterns, but also by how tightly they remain coupled to ongoing consequences.

Representation concerns the internal structure of the model: the patterns it has learned from past observations. Coupling concerns the relationship between those patterns and the real-time outcomes that follow from decisions.

When environments remain stable, strong representations can compensate for relatively weak coupling. A well-trained model may continue to produce reliable predictions for long periods, even if it updates infrequently.

When rules shift rapidly, however, the balance changes. Representations built on past data may degrade quickly. The value of prediction then depends increasingly on the system’s ability to maintain alignment with the evolving environment.

In other words, prediction becomes less about static knowledge and more about continuous adjustment.

Where contemporary AI succeeds

Modern machine learning has achieved remarkable progress in domains where large volumes of stable data are available.

Language models provide a clear example. By training on vast corpora of text, these systems learn statistical structures that capture many patterns of human language. The resulting capabilities-translation, summarization, question answering-are substantial.

Similar advances appear in image recognition, speech processing, and recommendation systems. In these contexts, the structure of the underlying data distribution often changes slowly relative to the scale of training. The representations learned during training remain useful across many deployments.

It would be difficult to overstate the technical achievements that enabled these systems. Training methods, hardware infrastructure, and algorithmic innovations have expanded the practical reach of machine learning dramatically.

At the same time, these successes often occur in environments where the coupling between model and consequence is relatively indirect. A recommendation system may influence which video a viewer watches next, but the immediate consequences of that decision rarely reshape the underlying structure of the environment.

In settings where decisions alter the system itself, the requirements are different.

Structural limitations

Many contemporary AI systems operate within a clear separation between training and runtime.

During training, a model learns from historical data. Once deployed, it performs inference: applying the learned structure to new inputs. Updates typically occur through periodic retraining cycles, during which additional data is incorporated.

This architecture has proven effective across numerous domains. However, it assumes that the knowledge captured during training remains valid for some period of time.

When environments change rapidly, this assumption weakens. The lag between observed change and model retraining becomes operationally significant. Organizations must detect the shift, collect new data, retrain the model, and redeploy it. Each step introduces delay.

The limitation is therefore architectural rather than algorithmic. The system’s capacity to adjust is bounded by the rhythm of retraining cycles.

In many real-world systems, consequences unfold continuously. Prices change, supply chains adapt, regulations shift, and human behavior responds to incentives. A prediction system that remains partially detached from this flow may struggle to maintain alignment.

This does not invalidate the achievements of current approaches. Rather, it highlights a dimension of intelligence that has received less attention: persistent embedding in consequence.

Intelligence in time

Biological systems illustrate a different arrangement.

Organisms do not separate learning and deployment into distinct phases. The same processes that generate behavior also update internal expectations. Prediction and adjustment occur simultaneously.

This arrangement allows organisms to remain sensitive to small deviations between expectation and outcome. When those deviations accumulate, behavior changes accordingly.

Of course, biological intelligence has its own limitations. Evolutionary solutions are constrained by physical and metabolic factors, and biological learning can be slow or biased.

Yet the underlying architectural principle remains notable: intelligence is embedded in the temporal flow of interaction with the environment.

The question for artificial systems is whether similar coupling can be achieved in engineered forms.

The organizational dimension

For corporations operating in volatile environments, the implications extend beyond technical architecture.

Organizations themselves function as predictive systems. They allocate capital, plan production, design logistics networks, and manage risk based on expectations about future conditions.

When those expectations remain accurate, planning structures perform well. Long-term investments can be optimized against relatively stable assumptions.

When rules shift more frequently, the gap between expectation and outcome widens. Decision processes built for stable environments may struggle to adapt quickly enough.

This dynamic is visible across industries undergoing structural change: energy systems integrating renewable generation, supply chains adjusting to geopolitical tensions, and financial markets responding to evolving regulatory frameworks.

The problem is not necessarily that predictions become impossible. Rather, the horizon over which stable predictions remain reliable becomes shorter.

Organizations therefore face an architectural question similar to the one encountered in artificial intelligence: how tightly can predictive structures remain coupled to ongoing consequences?

Reframing the problem

Discussions about artificial intelligence often focus on the sophistication of algorithms or the scale of data. These are important factors, but they may not fully capture the central challenge posed by environments with moving rules.

The deeper issue concerns the relationship between prediction and time.

When the structure of the world remains stable, prediction can rely heavily on accumulated representations. When rules change more frequently, prediction becomes a process of continuous alignment between expectation and consequence.

This shift does not invalidate existing methods. Many domains will continue to benefit from large-scale models trained on extensive datasets. But in systems where structural change is common, additional architectural properties may become important.

The missing property is not greater complexity. It is persistent coupling: the ability of predictive systems to remain embedded in the stream of consequences they attempt to anticipate.

Developing such systems presents substantial difficulties. Continuous adjustment risks instability if not carefully controlled. Evaluating performance becomes more complex when models evolve during operation. And designing mechanisms that balance responsiveness with robustness remains an open challenge.

Yet the direction of inquiry may prove increasingly relevant as economic and technological systems continue to evolve.

If intelligence is ultimately a process embedded in time, then the question is not simply how accurately systems represent the past. It is how effectively they remain aligned with a future that may not follow the same rules.