Every big change is bumpy. This one is faster.
AI is spreading faster than any earlier technology and its costs are arriving before most of its benefits, so we need to prepare sensibly and make sure more people get the chance to build with it.
A colleague started a call with me the other day by asking whether AI was going to kill us all in ten years.
I have no idea. I've been thinking about a much less dramatic question: what happens when something we rely on stops working for a while?
I live in the Pacific Northwest. We get windstorms and power outages, and we live with earthquake risk, so I've always known it makes sense to keep enough food, water and other basics at home to manage for a few days. Like a lot of people with busy lives, though, I had a gap between knowing what I was supposed to have and checking that I had it.
Lately I've added cyber disruption to the list. I'd thought about an earthquake cutting power or water, and about a storm taking down the electricity. I hadn't thought much about a cyberattack disrupting a utility, a payment system or part of the food supply chain, or what that would mean for the things I assume will work at home.
Then in October, attackers got into the internet-connected refrigeration controls of hundreds of French supermarkets. France's finance minister put the number at about 700. Temperatures in fridges and freezers reached as high as 50°C, and stores had to throw away food. The attack reminded me that most emergency-preparedness advice isn't new. Governments have told people for years to keep enough to manage for several days if normal services stop, and I just needed to get around to doing it. Getting ready was a small step, not a return to homesteading, which is a good thing because I would make a terrible farmer.
AI didn't cause the supermarket attack, and cyberattacks have been around for a long time. AI changes how much one person or a small group can do, and how fast. In October, Reuters reported that one person had used inexpensive AI tools to attack more than 1,200 websites and break into the websites of more than 100 companies. AI is helping defenders too, as models find vulnerabilities and help people respond faster. So cyber is now on my list next to windstorms and earthquakes.
The larger changes from AI are bumpy too. Every big technological transition has been, and this one is moving faster than any before it.
Earlier technologies couldn't spread much faster than the infrastructure built for them. Electricity needed power plants and wiring in every home, and cars needed roads and gas stations. Even the personal computer meant buying a new piece of hardware. Each generation also inherited more of what came before. Smartphones arrived on top of cellular networks and the internet, and generative AI arrived on top of almost all of it. Hundreds of millions of people already had a phone or computer and an internet connection, and nobody had to run a new wire to their house before they could try it.
The adoption numbers show the difference. Electricity took about 40 years to go from 10 percent of American homes to 90 percent, the telephone took 66 years and the car 74. The cell phone took 19. Less than two years after ChatGPT appeared, 39 percent of working-age Americans had used generative AI, a faster start than either the personal computer or the internet had. The measures aren't directly comparable across a century, but the direction is hard to miss.
It's a little like a roller coaster that keeps getting faster. None of the turns has to be big to make the ride harder. The next one just arrives sooner, with less time to adjust in between.
There are good reasons to slow or stop particular uses of AI. Some uses need more testing or firm limits, and in some cases the consequences justify saying no. Slowing down won't take us back to a world where society has a generation or two to adapt to a major new technology, though, because technology doesn't spread at that speed anymore.
AI reached hundreds of millions of people through infrastructure we already had, but it still needs new infrastructure of its own. Its rapid adoption is driving a race to build data centers and chips, and the electricity to run them. The International Energy Agency expects electricity use by data centers worldwide to roughly double by 2030, and in its 2025 analysis it projected that data centers would account for almost half of the growth in US electricity demand over the same period. With earlier technologies, infrastructure limited how fast adoption could spread. With generative AI, adoption came first and we're now building the infrastructure behind it, which is one reason so many of the bumps are arriving close together.
Communities are asking reasonable questions about the data centers being built near them: what resources they'll use, and what the community gets from the bargain.
Work is changing at the same time. Companies are reorganizing jobs around AI, and through October employers had cited AI in about one in five announced US job cuts this year. That figure doesn't tell us what AI will do to total employment. Earlier technologies eliminated jobs and also created work that didn't exist before. People live the transition through what happens to their own jobs, though, and a long-run statistic is little comfort to someone whose role is changing this year.
Many of the benefits people hear about are still ahead. AI could help us discover new treatments and make work much more productive, and I expect it will do both. For someone whose work is changing now, or whose community is being asked to host a large new data center, that return is still a forecast.
Finance has taught me to pay attention when costs and benefits arrive at different times or in different places. An investment can make sense overall and still be hard on the people who absorb the costs first. We need to work out how the value AI creates reaches those people and communities, because their question is a simple one: how does this turn out well for me? Telling them that earlier technological revolutions eventually created new jobs isn't much of an answer.
Anyone who has led a large transformation knows you don't wait until the change is finished to think about the people affected by it. People whose work is changing need the chance to learn new skills before the old ones lose their value. Reskilling people for somebody else's version of the future isn't enough, though. More people should get the chance to build it, and that makes access to models and computing important.
We don't need the largest model for every task, and not every AI request has to travel to a large cloud data center. Smaller models can already do useful work, and more of them run on devices and equipment organizations already own. That doesn't remove the need for large data centers. Some work needs huge amounts of computing, and making AI cheaper may lead us to use more of it overall. Local models still give us choices. Some information can stay on the device, some applications can work without a constant connection, and organizations can decide which tasks belong on their own equipment and which need the largest cloud models.
The more places useful models can run, the more people can experiment and build with them. I want students, researchers, entrepreneurs and small businesses to have that chance. I look at AI the way I look at any critical dependency. When too much runs through too few places, the opportunity concentrates there, and so does the risk.
My colleague's question is one of the largest questions about AI. Some of those questions matter a great deal, and most of us have little ability to answer them. We can do more with the changes closer to us.
At home, the step can be as mundane as having water, food, medications, cash and batteries on hand if something doesn't work for a few days. Preparing for a disruption isn't predicting one.
At work, learn the tools and try them on real problems, so you know where they make you better and where they don't. If someone around you is seeing their work change, help them learn too. Waiting until a job has disappeared is a terrible time to start preparing for the next one.
And look for things to build. The people who created value in earlier technological revolutions mostly learned what the new tools could do and found useful things to do with them, whether or not they predicted where the technology would lead. Smaller and local models can put the ability to build with AI into more hands, so more people can try ideas, and companies can find uses nobody designing today's models had in mind.
We should also watch where the costs and benefits land. Communities hosting new infrastructure should have a voice in the bargain, people whose work is changing need a path to new skills before they need them, and the chance to build with AI should reach beyond the organizations that own the largest computing systems.
There will be bumps we don't see coming. We can prepare for the ones we can see, and while everyone debates where AI eventually takes us, we can learn to use it and build useful things along the way.
Continue the conversation. I read and reply to comments on Substack. The sharpest responses usually come from readers.
Get the essays in your inbox
Clear Calls publishes on Substack. No promotion. Unsubscribe anytime.