06 · Conclusion

The central claim#copy

Model capability matters: it begins the analysis, and the deployed system supplies the evidence of dependable operation.

Four transitions matter: capability to dependable operation; operation to valid learning; contained execution to delegated action; and demonstrated control to boundary expansion. Each requires its own evidence. Broad economic transformation also requires durable value beyond local cases.

Supercycle and Helix are hypotheses about broad economic diffusion and repeated boundary expansion. Flywheel describes how a system can learn through operation, with a testable claim about local compounding. Agentic defines delegated action and control flow through an agency envelope. Whether a particular deployment works remains an empirical question.

The connections are conditional. A team may learn and improve a system while keeping its scope fixed. It may delegate more work than it can monitor or control. Controls that support expansion in one organization may remain costly to adapt elsewhere, limiting their contribution to a Supercycle. Capability may improve throughout while dependable value remains narrow.

What seems robust#copy

Across reasonable futures, several claims remain useful:

• The deployed system is the unit of analysis and operational accountability. Claims about value should name the level at which value is measured: task, person, organization, market, or society.
• Capability and reliability diverge. A demonstration or benchmark can justify further testing; dependable performance must be observed in live operation.
• Measurement and monitorability are part of the control surface. Observations must be representative, attributable enough to act on, valid for the decision, and legible within the available intervention time.
• Scope is multidimensional. Permissions, horizon, state, delegation, reversibility, and oversight capacity can vary separately; their effects and tradeoffs must be evaluated together.
• Compounding is conditional. Repetition creates momentum only when learning is retained, controls are reusable, and value survives the full cost of operation.

What remains uncertain#copy

The hardest questions concern those transitions:

• whether reliability, monitoring, and control can keep pace as the agency envelope expands;
• whether controlled evaluations remain predictive in live, changing environments;
• whether reusable controls materially reduce the cost of dependable outcomes across deployments;
• whether task-level gains produce net organizational value after full costs;
• and whether local boundary expansion becomes common enough to alter workflows, interfaces, labor, or market structure.

These uncertainties determine whether the future resembles a broad compounding cycle, a set of valuable but bounded application waves, or repeated expansion followed by contraction.

How to use this model#copy

Use the model to judge how far a claim reaches and what evidence would take it further. Compare systems doing specified work, identify their limits, and design observations that could change the assessment.

Three limits matter:

• This paper identifies conditions and mechanisms. A forecast would also need an explicit horizon, assumptions, and uncertainty estimates.
• Model rankings require system and workflow context.
• Claims of transformation require evidence beyond adoption, activity, or task speed; claims that broader agency is desirable require explicit value judgments.

The practical work of designing bounded workflows, controls, evaluations, and incident processes belongs in the AI Operators Handbook. This paper supplies the analytical frame.

What should change the assessment#copy

Evidence should strengthen the compounding hypotheses when:

• live reliability holds under stated error limits as the agency envelope expands, and supervision per comparable outcome falls with total burden visible;
• incidents become durable evaluations, controls, and process changes that reduce the cost of later deployments;
• organizations change workflow boundaries or commitments on the basis of measured performance;
• and added scope produces net value after full system costs across more than isolated cases.

Evidence should weaken or narrow those hypotheses when:

• evaluation gains do not survive live operation or incidents exceed error limits specified for the task mix and exposure as scope expands;
• monitoring and human correction remain binding or hidden costs;
• each workflow requires largely bespoke integration and governance;
• or local productivity gains produce neither durable value at the level claimed nor a change in what organizations are willing to delegate.

The evidence reviewed for this revision includes measured capability gains on selected tasks. A field study estimates productivity gains in assisted customer support during a 2020–21 rollout; an early-2025 coding experiment found a slowdown in its setting. I still regard reliable agency at scale, dependable net economics across deployments, and repeated boundary expansion as open empirical questions. The evidence is stronger for some parts of this argument than for others.

I will update this paper as those transitions become observable—or repeatedly fail. I want it to help people judge what they can build and depend on. People operating these systems and relying on their decisions live with the consequences. That is why conditions and failure modes deserve this attention.