The Exponential Divergence

Executive Overview

Executive Overview

For nearly all of human history, we have been the beneficiaries of technological progress. The technologies themselves looked very different — printing presses, railroads, broadband internet — but they all grew on remarkably similar growth curves: an S-shaped logistic curve. Slow start, rapid acceleration, then a natural saturation point. Institutions lagged for a season, then caught up. The bound on the technology was also the bound on the gap.

Advanced computational intelligence is different. For the first time in history, a major technology appears to be growing on a fundamentally different curve — exponential, with no theoretical ceiling. There is no saturation point to aim for. There is no natural moment at which the gap closes on its own.

The practical challenge is that a logistic curve and an exponential curve look identical in the early going. Only after the inflection point do they separate — and by then, the trajectory is already set. Our institutions, our regulations, our habits of mind were all built for a world of logistic growth. They have never encountered anything like what may be unfolding now.

Section 1

Logistic vs. Exponential Growth

Technology has always looked different on the surface — the printing press bears little resemblance in form to the railroad, and the railroad bears little resemblance in form to broadband internet. Beneath the surface, the growth dynamics have been consistent. Nearly every major technology in human history has followed the same shape: a slow start, a period of rapid acceleration, then a natural plateau. The S-curve, or logistic curve, is not a coincidence — it is a reflection of how physical constraints work in the real world.

The constraint usually comes from the very thing the technology was designed to serve. Once every home has a dishwasher, every town has a railroad terminal, and every city has broadband internet access, there is no natural need for more of it. Demand saturates. Growth flattens. The industry shifts into replacement cycles, add-on services, or margin compression — but the underlying growth curve has completed its arc. Institutions, regulators, and society have time to catch up, because the technology itself eventually stops moving faster than they can.

AI is different. Unlike every prior wave, advanced computational intelligence does not appear to have an obvious physical saturation point. Intelligence compounds: better models produce better training data, more capable systems attract more investment, each improvement enables the next. The world has not seen this kind of compounding in a major technology wave before. Practically, it will likely encounter some physical boundary — energy constraints, compute limits, architectural ceilings. But theoretically, there is no wall. The curve does not know where to flatten.

The thing being scaled in the AI moment is intelligence itself. Intelligence is not limited by geography, by number of homes, by number of attention spans. The constraint that bounded every prior wave — the saturation of the very thing the technology was designed to serve — may not bind this one.

Section 2

This Will Look the Same Until It's Too Late

A logistic curve and an exponential curve are visually identical for their first half. The point of divergence is also the point at which it becomes too late to choose differently.

This is the most uncomfortable property of exponential growth. In the early phase — when the absolute numbers are still small — the two trajectories look the same. The same data points fit both models. Two observers can look at the same chart and tell two different stories, both defensible. Only past the inflection do the curves separate; once they have, the distance opens fast.

This matters because the policy and personal questions about powerful technology have to be answered in the early phase, when the curves still look the same. If you wait until you can see for sure that the trajectory is exponential, the moment for shaping the trajectory has already passed.

This is the structural reason that prudent action in this domain looks different from prudent action in domains where the curve shape is known. In a known-logistic world, "wait and see" is rational — by the time the wave matures, you will have learned what kind of wave it is. In a possibly-exponential world, "wait and see" is the same as "decide too late." The cost of being wrong in one direction is much larger than the cost of being wrong in the other.

Section 3

Our Institutions Are Not Built to Handle Exponential Growth

Losing the Grace of Logistic Growth

Every past technology eventually saturated — and the institutions that lagged it eventually caught up.

The chart below is a stylized model of what the past eight thousand years actually looked like. The grey diagonal line is the slow, roughly linear rise of institutional capacity — laws, norms, schools, regulators, professional training, accumulated common sense about how to live with a given tool. The blue bumps are technological waves: each one rises rapidly above the line for a season (the capability gap C), then is matched by the institutional line catching up (the time to catch up T).

This is the deep reason institutions have been able to lag behind technology for centuries without society falling apart. The printing press took roughly a century to be matched by the institutions of public education, copyright, and journalistic practice. The automobile took roughly fifty years to be matched by licensing, traffic law, urban planning, and insurance. The internet has taken twenty-five years and is still being matched — but it is still matchable, because the internet itself eventually saturates (you can only put so many people online).

The key feature of every past wave is this: the capability gap had a maximum size. C was bounded because the technology was bounded. When the technology saturated, the institutions had a fixed target to aim at. Eventually they hit it.

The rightmost wave on the chart shows what happens when a technology does not saturate. The blue line keeps going, the gold gap keeps widening, and the grey institutional line — moving at its historically normal pace — falls further and further behind with no asymptote to reach.

Section 4

This Time May Actually Be Different

The other studies in EconFaithAI describe specific consequences — what's happening to children, to attention, to capital, to religious life, to the moral character of invention itself. This framework piece is the answer to a question those studies provoke but don't fully answer: why is this happening now, and why doesn't the usual reassurance apply?

The usual reassurance is some form of: societies have always adapted to new technologies; we'll adapt to this one. That reassurance is built on the historical pattern in Section 3 — bounded technology, bounded gap, eventual catch-up. It is a true statement about every prior wave.

It may not be a true statement about this one. If AI capability is on the exponential curve — and the early evidence is at least consistent with that — then the institutional adaptation strategies that worked for railroads, broadcast media, and the early internet are structurally inadequate to the moment. The same amount of institutional energy directed at a non-saturating target produces a falling-further-behind, not a catching-up.

None of this proves the curve is exponential. It might saturate at superhuman-but-still-bounded levels. It might bend for reasons we don't yet see. The honest position is that we cannot tell, in the early phase, which curve we are on. The honest action that follows is to behave as if we are on the more demanding one, and be pleasantly surprised if we are not.

No one knows what will happen. But the chance that something genuinely different is unfolding — different from any prior point in history — is more material now than it has ever been. That demands something different from us in return.

Not from the technologists, who have every incentive to keep building. What is required is activation from moral institutions, from religious communities who have long held the questions of what it means to be human, and — most urgently — from policymakers willing to act before the curve makes the decision for them.

Our reactive systems — laws written after harm is done, regulations drafted after industries mature — were designed for logistic change. They may not be fast enough here. If we get in front of this growth, we can still shape it. If we wait until the divergence is undeniable, we will find ourselves trying to steer something that has already left us behind.

The curve does not wait for us to make up our minds.

Section 5

Conclusion and What We Should Do

The framework in this article is the framework for the rest of the project. What every other EconFaithAI study describes — the children, the attention, the capital, the religious life — looks different if the underlying curve is exponential rather than logistic. The downstream consequences compound. The window for response narrows.

We cannot tell yet which curve we are on. The risk of getting it wrong, though, is deeply asymmetric. If we prepare as though the curve is exponential and it turns out to be logistic, we will have over-prepared — costly but recoverable. If we prepare as though it is logistic and it turns out to be exponential, we will have under-prepared at the exact moment under-preparation cannot be undone. That asymmetry is the whole argument.

What is required now is not certainty about the future. It is the willingness to act before certainty arrives. Logistic technologies gave us time to be sure. The exponential — if that is what this is — will not.

The urgency is not uniform across roles, but the direction is. Select your role below.

Appendix A

Methodology

The logistic and exponential curves shown in this article are illustrative models, not empirical fits to specific datasets. The logistic curve uses the standard Verhulst formula with parameters chosen to represent a generic S-curve adoption pattern. The exponential curve is calibrated to match the logistic at its inflection point, then allowed to grow unconstrained. The divergence between the two is the core argument of the article — not a prediction of any specific technology's trajectory.

The institutions chart uses a simplified piecewise model: linear growth in human institutional capacity, and a logistic-to-exponential transition for AI capability. The crossover point and scaling parameters are illustrative. The intent is to show the structural shape of the divergence, not to forecast a specific date.

Appendix B

Limitations

1. Models are illustrative, not predictive. The curves in this article represent structural arguments about growth regimes, not empirical forecasts. Real technological trajectories are messier, and the timing of any specific transition is unknown.
2. Institutions are not monolithic. Some institutions adapt quickly; others are deliberately slow by design. The institutional lag argument applies unevenly across sectors and regions.
3. Exponential growth always eventually saturates. The article does not claim AI capability grows exponentially forever — only that the current period of growth is better described by an exponential than a logistic, and that this distinction has practical consequences while it holds.

Appendix C

References

On Growth Curves and Technological Diffusion

• Verhulst, P. F. (1838). Notice sur la loi que la population suit dans son accroissement. Correspondance mathématique et physique, 10. The original logistic model of constrained population growth.
• Rogers, E. M. (2003). Diffusion of Innovations (5th ed.). Free Press. The standard text on S-curve technology adoption.
• Modis, T. (2002). Forecasting the Growth of Complexity and Change. Technological Forecasting and Social Change, 69(4). Logistic-curve modeling of technological evolution.

On AI Scaling and Exponential Capability

• Kaplan, J., et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361. Foundational paper on the relationship between model size, data, compute, and capability.
• Hoffmann, J., et al. (2022). Training Compute-Optimal Large Language Models (Chinchilla). arXiv:2203.15556.
• Sevilla, J., et al. (2022). Compute Trends Across Three Eras of Machine Learning. arXiv:2202.05924. Documents the post-2012 exponential growth in training compute.
• Karnofsky, H. (2021). This Can't Go On. Cold Takes. Argues that exponential trends in compute and capability cannot continue indefinitely under any normal physical assumption — but the current rate has not yet hit a wall.

On Institutional Lag

• Acemoglu, D., & Robinson, J. A. (2012). Why Nations Fail. Crown Publishers. On the slow adaptive pace of institutions relative to technology and economy.
• Allen, R. C. (2009). The British Industrial Revolution in Global Perspective. Cambridge University Press. Documents the multi-generational lag between industrial technology and institutional matching.
• Mokyr, J. (2002). The Gifts of Athena: Historical Origins of the Knowledge Economy. Princeton University Press.

Appendix D

Project Files