Artificial Intelligence

I. Where we start

Today, AI has already left the laboratory and entered the office, the school, and the phone. Models generate acceptable text, code, image, and video. Use inside organisations is broad; deployment of agents that own a whole business process is still rare. That gap matters. The cost of inference falls fast while the cost of frontier training and infrastructure rises. There is job panic, productivity euphoria, and a race between the United States and China, with Europe trying to regulate and the rest of the world trying to sign an API contract. There is not yet, in demonstrated and stable form, a system that replaces an entire company without a human in the loop.

II. The decade in between

Cognition becomes a utility: cheap enough for firms to rebuild whole processes around agents. The housing boom supplies a stubborn test. Models reconcile plans, quantities, schedules, and code checks; licensed people sign off on safety and appeals. A design approved once becomes easier to build again, but a confidently wrong drawing cannot be allowed to reproduce unchecked. Firms create paid swarm-residencies because someone must learn why the exception matters. The expensive part is accountable judgment, not another generated floor plan.

Energy and capital set the ceiling — and the ceiling rises all decade, because this is the same ten years in which electricity gets abundant. Data centres become anchor tenants that finance grids instead of straining them. "AI sovereignty" matures from a minister's slogan into a workable middle path: the commodity tier gets so good that most countries run real capability on their own soil, and the frontier clubs lease what remains scarce. The gap that matters is not chatbot access. It is whether your agents can act inside the payment rails, the hospital, the factory, and the ministry — and each year the list of countries where they can gets longer.

III. Ten years from now

What is normalised ten years from now:

• The default firm in software, finance, law, media, logistics coordination, and professional services is a handful of humans supervising a swarm. The swarm reads, files, codes, books, invoices, hires other swarms, and closes the books. A three-person company with forty agents is not a stunt; it is a common capital structure — and founding one is what ambitious twenty-two-year-olds now do instead of applying for the junior job the swarm absorbed.
• Every person gets an Aristotle. The tutor, the clinician's first read, the legal first draft, the tax filing, the translation — capabilities that used to be an hourly rate are now a subscription, and the cheap tier is close enough to the expensive one that daily life runs on it. The great equalising fact of the decade is that the best explainer on earth is in a billion pockets.
• Consumers rarely reach a human by default — and mostly stop wanting to, because the agent resolves in minutes what the queue used to take weeks to mishandle. Human attention becomes a premium people choose for the moments that deserve it, not a toll they pay for everything.
• Education, first-line medicine, accounting, programming, customer service, and most analysis are machine-native with a human exception desk. The exception desk is well paid — and growing it is what the new swarm-residencies are for.
• Agents hold wallets, negotiate prices, place orders, file taxes, and settle with other agents. A non-trivial share of GDP is machines paying machines — audited by regulators' own swarms, which turn out to be better at catching collusion than any inspector with a briefcase ever was.

What is still genuinely human turns out to be what humans actually wanted to keep: legal responsibility (someone must be suable), taste that other humans will pay for, care that requires a body against a body, politics, and the unscripted exception. A model drafts the petition and often files it. A model proposes the diagnosis and initiates the protocol. A human still gets the call when the stakes are a life — and, freed from the drudgery, has time to take that call properly for the first time in a generation.

Frontier versus commodity. Cheap, local, "good enough" models run on the device and in bargain cloud — and "good enough" now covers most of daily economic life. Expensive frontier systems sit inside a few laboratories and states and matter where they should: research, defence, and the design of the next model. The hierarchy did not collapse; it stopped deciding who eats. Countries with only the commodity tier can run schools, clinics, courts, and firms on it — which, a decade ago, was the utopian scenario.

Society. AI delivers the beginning — the beginning — of the noble leisure it always promised. Essentials deflate: diagnosis, tutoring, drafting, design, translation. Machine surplus funds real income floors across the rich world and the first ones elsewhere. The generation that skipped the old cognitive apprenticeship gets a new one — running swarms — and its first cohort, raised with a tutor in the pocket, arrives at adulthood measurably harder to fool. Those who own or steer swarms still take the largest share; the difference from every previous automation wave is that the floor rises fast enough for the majority to feel it in the bill, the clinic, and the school.

IV. Uncertainties

This scenario includes hospital errors, financial fraud, and coordinated agent failures. Cheap capable models benefit attackers too; the defence is bounded authority, independent checks, recoverable transactions, and people who can halt a process. Housing makes correlated failure physical: one mistaken specification can enter a thousand homes. The machine runs the paperwork and the schedule. It cannot certify its own safety by asking another copy of itself.

The open questions that could bend the line: whether the frontier keeps compounding or plateaus into a very useful commodity; whether the swarm-residencies scale as fast as the old ladder burned; whether one public catastrophe produces a real pause rather than a press conference; and whether the surplus keeps reaching the pavement once the novelty wears off the politics.