0
← Journal index

AI / Wanderings 2026

Symmetry

Every increase in AI capability needs a matching increase in evaluation, oversight, cybersecurity, governance and safeguards. That is defensive co-scaling.

By Martin Uetz2 min read
Two glass-and-metal towers, one glowing amber and one blue, rise side by side from linked bases.

AI capability is scaling quickly: more compute, better models, longer context, stronger reasoning and more autonomy.

Our ability to understand, control and defend those systems has to follow the same curve. That is defensive co-scaling.

Every increase in capability needs a matching increase in evaluation, monitoring, cybersecurity, interpretability, governance, human oversight and safeguards against misuse. Protection belongs in the release plan, at the same time as the new capability.

Capability needs a progression

Human alignment offers a useful analogy. We arrive with intelligence, instincts and an impressive ability to create chaos. Over many years, parents, education, values, laws, social norms, accountability and trust form the surrounding system.

As a child becomes more capable, society expands freedom alongside responsibility. A child can ride a bicycle before driving a car and manage pocket money before controlling a company budget. Judgement and accountability develop with the authority to act.

AI needs a similar progression. Alignment is a continuous and imperfect process supported by institutions, rather than a one-time training exercise that produces a perfectly obedient machine.

The model remains one part of the system. A highly capable AI with weak oversight resembles a brilliant teenager given unlimited money, a private jet and no parents. Intelligence offers no automatic supply of wisdom or responsibility.

Protection makes capability usable

Defensive co-scaling provides the infrastructure that allows more innovation to be used safely. Cars spread because roads, seat belts, traffic laws and insurance developed around them. Aviation grew because navigation, maintenance and air-traffic control grew with it.

AI will follow the same pattern. The strongest systems will combine powerful models with controls that remain dependable as that power increases.

For every capability you add, name the evaluation, safeguard, accountable person and stopping condition that grow with it. Scale intelligence and our ability to live with it at the same speed.