AI power: ownership, dependence and control

General capability matters because a reusable problem solver could coordinate work across domains and help improve the next generation of tools. We should look for growing ability to complete difficult tasks, not wait for a theatrical moment called AGI. The benefits could include stronger science, cheaper services and more capable public institutions.

Then measure the right thing. METR's time horizon describes task difficulty using human-expert completion time and a specified success rate. It is not how long an agent can roam independently, nor a percentage of an economy it can replace. The published task suite is mainly software-related and cleaner than much ordinary work.^[1]

Fair. Deployed autonomy also depends on permission. Anthropic's February study found longer uninterrupted runs in the extreme tail of Claude Code usage. The authors warn that duration mixes user choices, task ambition and product behavior. That shows a changing relationship between people and tools, not a machine escaping its operator.^[2]

And yet obedient systems can reorder society. Consider a firm that replaces labor while a few owners retain the productive assets. It might become less dependent on workers' cooperation without becoming independent of consumers, law or public infrastructure. The claim that ordinary people lose all leverage needs those additional dependencies examined. The linked analysis of production without mass employment follows that separate argument.

There is a strong counterargument to permanent concentration. Better algorithms and widely available models could reduce costs and let smaller organizations compete. States can preserve expertise and regulate access. A scenario in which one company effortlessly owns every useful capability assumes away competitors, adaptation and politics.

But access to a model is not access to unlimited productive capacity. The IEA's April 2026 report identifies constraints in grids, electrical equipment and computing supply chains.^[3] My inference is political: control of scarce capacity could confer leverage even when software is widely available. Bottlenecks can create bargaining power; they do not prove a permanent monopoly.

Nor do bottlenecks guarantee a comfortable adjustment period. AI-assisted research could shorten some development cycles. The International AI Safety Report identifies major uncertainty about that feedback loop, including oversight needs and slow large-scale experiments.^[4] A faster research loop is a reason to watch progress closely, not proof of an inevitable, instantaneous intelligence explosion.

The difficult case is strategic competition. A company or state may delegate more because it expects rivals to do so. Each actor can prefer a cautious world while fearing the cost of caution alone. The report discusses precisely these pressures on safeguards.^[4] A policy that assumes everyone will voluntarily accept an indefinite disadvantage has not resolved the incentive problem.

Then look for practical counterevidence as well as warning signs: independent evaluations, organizations demonstrating a usable fallback, and productivity gains that leave meaningful human review intact. A supplier's promise is weaker evidence than a customer actually exercising its right to stop, switch or reject an output.

Agreed. Watch who sets objectives, whose objections change decisions, and what happens when authorities try to intervene. Ownership concentration, dependency and autonomous failure call for different tests. A society could avoid rebellious machines and still surrender too much power; it could also overstate technical danger and use that fear to entrench incumbent control.