Historical Precedents

The Roman Aqueducts

Roman engineers built aqueducts that supplied water to cities for centuries. After the empire fell, the knowledge to maintain them was lost. The aqueducts degraded over generations. Some cities lost running water for a thousand years.

The Antikythera Mechanism

The ancient Greeks built a mechanical computer to predict astronomical positions. The technology was lost. Nothing comparable was built again until the 14th century. Capability is not permanent.

Colonial Infrastructure

European colonizers built infrastructure in colonized nations but concentrated technical knowledge among colonizers. After independence, some nations struggled to maintain systems they hadn't built and didn't fully understand.

The Apollo Program

NASA sent humans to the moon in 1969. By 2024, doing it again required largely rebuilding the capability. The people who knew how retired. The documentation was incomplete. The institutional knowledge was gone.

Capability can be lost even within living memory.

The Modern Stack's Fragility

Current AI infrastructure has specific vulnerabilities:

Training Data Provenance

Modern models are trained on data whose provenance is often unclear. If something goes wrong—bias, security vulnerabilities, copyright violations—tracing the problem to its source may be impossible.

You can't maintain what you can't trace.

Hardware Dependencies

AI systems depend on specialized hardware (GPUs, TPUs) from a concentrated set of manufacturers. Disruption to that supply chain cascades through everything that depends on AI.

Who maintains the fab? What happens when the people who know how are gone?

Weight Opacity

Model weights encode learned behavior, but they're not interpretable. You can't read the weights to understand what the model knows. When problems arise, debugging requires experimentation rather than inspection.

Maintaining a system you can't read is fundamentally different from maintaining one you can.

API Dependencies

Modern applications depend on AI services accessed via API. The internals are hidden. If the provider changes the service, deprecates features, or goes out of business, dependent applications break.

You can't maintain a dependency you don't control.

The Knowledge Cliff

Technical knowledge is concentrated in a shrinking number of people:

The Age Pyramid

Many critical systems—mainframes, industrial control systems, legacy databases—are maintained by aging specialists. The young generation didn't learn these technologies because they seemed obsolete. Now the old generation is retiring.

The Expertise Bottleneck

Cutting-edge AI expertise is concentrated in a few thousand people worldwide. These people are oversubscribed. If the field expands faster than expertise can be transferred, the bottleneck worsens.

The Documentation Debt

No one documents as much as they should. AI systems are particularly bad—they change rapidly, and the relevant knowledge is often tacit rather than explicit.

Every undocumented system is a future maintenance emergency.

The Context Loss

Knowing what the code does isn't enough; you need to know why it was written that way. What constraints existed? What alternatives were considered? What assumptions were made?

This context is rarely preserved. When the original authors are gone, it's gone too.

Possible Futures

The Graceful Handoff

AI systems become sophisticated enough to maintain themselves and each other, with humans in a supervisory role. The transition is managed carefully, with redundancy and fallback.

This requires deliberate planning, which there's little evidence of.

The Fragile Steady State

Systems keep running because they're stable, but maintenance capability degrades. Small problems go unfixed. Technical debt accumulates. The systems work until they don't.

This is the current trajectory.

The Brittle Collapse

A major system fails—power, finance, communications—and the expertise to fix it doesn't exist in time. The failure cascades to dependent systems. Recovery takes years or decades.

This has happened locally. It could happen globally.

The Intentional Simplification

Society decides to reduce dependency on complex systems, rebuilding around more maintainable technologies. Some capabilities are sacrificed for resilience.

This requires collective choice that seems unlikely absent a catastrophe.

The AI Maintenance Class

A new profession emerges: people who specialize in maintaining AI systems, with AI assistance but ultimately human judgment. Society invests in training and retaining this class.

This is possible but requires recognizing the problem and acting on it.