Intelligence is becoming a utility

Most conversations about AI assume a race: a handful of labs, a finish line called AGI, and a single winner who takes everything. Billions of dollars, export controls and a lot of policy now rest on that assumption.

Everybody got electricity. Intelligence is heading the same way, and decisions made as if it were a race with one prize put money, talent and safety in the wrong places.

Intelligence is becoming ambient

Software wants to spread. Open-weight models that rival the closed frontier already run on a single desktop machine instead of racks of servers. A few months later someone compresses them to run on a laptop, and a few months after that on a phone. The intelligence in those small systems is already more than most people use most of the time.

Something that is everywhere and nearly free cannot be hoarded. A few companies would like to monopolise it, but they are working against the economics. When intelligence is a utility, the interesting question stops being who has it and becomes what you do with it.

The useful model is usually the small one

The race picture also assumes the biggest model wins. In practice, a lot of real progress comes from small models trained for one job. A drug discovery team doesn't need a model that can write sonnets; it needs one trained deeply on its disease, small enough to run on a laptop and cheap enough to iterate on every day. The model most likely to help cure a disease is a specialist, not the largest generalist.

That matters for where the money goes. If efficiency keeps improving the way it has, some of the capital pouring into ever bigger data centres would do more good in hospitals, power plants and roads, and the concentration of that spending makes the wider economy more fragile, not less.

A race reframes a technology as a weapon

Treat AI as an arms race and it becomes a weapon by default. Labs rush, safety becomes a cost to cut, and every step is justified by the fear that the other side will get there first. It is a familiar playbook: find an enemy, make it big and frightening, and position yourself as the answer. If a product is as transformative as claimed, it is fair to ask why it needs quite so much lobbying.

The race also points at the wrong threat. States already live with a hard-won balance: using a decisive weapon invites escalation that destroys everyone, which is why nuclear arms haven't been used in eighty years. A super-capable AI doesn't change that logic. The new risk is small groups and individuals. When capable models fit on a laptop, the ability to do serious damage, in cyber, chemistry or biology, moves from states to people.

That kind of risk is only manageable together. We already have working examples: know-your-customer rules for chemical precursors and pharmaceuticals, and DNA synthesis providers who screen orders against shared lists of dangerous sequences. They work because the systems talk to each other. Split the world into two technology stacks that refuse to share warnings, and the gaps between them become the places bad actors hide. Even at the height of the Cold War, rivals agreed limits and kept a hotline open, because a phone call is cheaper than a mistake. Shared safety standards and incident reporting for AI are the obvious place to start.

Value that doesn't show up in GDP

General purpose technologies are hard to measure. AI pushes the cost of both thinking work and physical work down: design, testing and manufacturing get automated, and something that cost a thousand becomes a hundred. A contract review that once meant hours of expensive legal time is now a question you ask before you sign. A second opinion on a sports injury, a tax question, a medical letter: a few years ago that kind of expertise on demand wasn't available to the richest people in the world.

Almost none of that registers as output. Prices fall, so measured value can even shrink while people are plainly better off. If we keep judging progress only by what is produced and sold, we will misread one of the biggest improvements in everyday life in a generation.

What to tell people starting out

The short-term picture is harder. Early-career people in jobs most exposed to AI are already finding it tougher to get hired, while new businesses are being started faster than ever. Both are true, and the advice that holds up is the same either way:

• Go wide. Read history, psychology, philosophy, management. Breadth is what lets you connect ideas and judge an answer instead of accepting it.
• Go deep on one thing you actually built. Take something through its whole life: design, build, launch, run, and retire. That is the experience a model can't hand you.
• Don't become a wrapper. If your job is to approve whatever the AI produced, someone will eventually notice they are paying for a rubber stamp.

Abundance is already here

The darker forecasts are easy to find: runaway systems, wars, or a few companies running the world. The more likely bumpy part is economic. Bubbles burst, and that hurts, but it is recoverable, and a shock like that has a way of changing minds. When people are focused on their own survival, fighting someone else becomes much less appealing.

The hopeful path follows from that. Spend less on racing and more on people. Bring the cost of chips and memory back down, and help the parts of the world that are behind catch up instead of being left out, the way rebuilding after a war once created decades of shared growth. Use the time AI gives back to spend more of it with other people.

Much of the world already lives with astonishing abundance; it just isn't evenly distributed. Treat intelligence as a race and we build weapons. Treat it like electricity and the question becomes what we all build with it.