Suddenly
In the summer of 2003, a few weeks before he moved to Taiwan, Ben Thompson installed WordPress. The software promised a five-minute install. It took him considerably longer, because he first had to learn how to create a MySQL database and configure PHP. Running a public website with little idea of what sat underneath it seemed reasonable then.
He recalled that summer on Monday, prompted by two critical flaws in PHP, together known as PHP2Shell, that researchers at Calif, OpenAI, and Wiz found with AI assistance in code paths dating back to 2012. PHP still runs roughly 70% of the world’s websites. Stratechery still runs on WordPress, but a managed host patched it before the disclosure. Doing it any other way today, he writes, would be “crazy.” The internet he is mourning is the one where anyone could stand up a server, and that server counted as much as any other.
In the same article, though, he describes making software for himself and his colleagues that would previously have been too costly or troublesome to build. He expects far more people to do the same. In his account, the freedom to put something online is becoming harder to sustain just as the freedom to make something useful for yourself is expanding.
In The Sun Also Rises, Mike Campbell is asked how he went bankrupt. “Two ways,” he says. “Gradually and then suddenly.” The arrival of AI has the same shape. Decades of accumulation became consequential to a great many people within a much shorter period, and the recognition arrived all at once.
AI makes the accumulated technology of the information age usable in ways that can feel like a historical break. But the internet whose passing Thompson mourns helped make those new capabilities possible in the first place. Decades of computers and networks, and the digital record they accumulated, supplied both the foundations on which AI was developed and a world already prepared for its use.
AI’s arrival is therefore both the culmination of one era and the beginning of another: the achievements of the preceding era produced the capabilities that now unsettle its arrangements. Whether the maturity of that inheritance tells us how much remains possible, now that people can use it this way, is a separate question.
Gradually
Nicolas Colin places AI within the mature phase of the computing-and-networks revolution. Drawing on Carlota Perez, whose historical framework I discussed recently, he sees AI as an efficiency breakthrough that extends the productive life of the technological system dating back to the microprocessor.
He argues that accumulated organizational knowledge becomes increasingly valuable as markets mature and innovation offers startups less protection from established competitors. He also suggests that application saturation may have set in: computing and networks have already reached most sectors where they can plausibly work, and many of the remaining opportunities are apparent to well-funded incumbents.
In The Commanding Heights, published this week, Will Manidis and Joe Roberts describe AI as still in its “adolescence,” with most of its diffusion into the economy ahead of it. Colin’s response calls it the “closing chapter” of the computing-and-networks revolution. He also points out that late-cycle buildouts can reward users more than the people who finance them. People could get a great deal out of AI while investors struggle to get their money back.
I agree that AI’s immediate usefulness owes a great deal to this inheritance. But the maturity of the technology underneath it tells us only so much about what people will find to do with it. We’re still finding out.
Decades of people writing, publishing, conducting research, and building software had produced an enormous digital record. We created it for our own purposes, then used it to train systems capable of working across many of the same domains. Software and mathematics offered another advantage: many results could be checked, giving researchers a way to reward correct answers during training.
By the time most people encountered a useful language model, cloud services could deliver it through devices they already owned. An improvement could feel abrupt because the infrastructure required to distribute it had taken shape over decades.
People were also conducting much of commerce and administration through software. As agents become the new users of software and the internet, they can help carry out requests through the same tools, provided they have the appropriate access.
Those tools also connect to valuable information and consequential decisions. They contain weaknesses and assumptions about who can act. The world was prepared for AI to become useful within it, and for mistakes or misuse to matter immediately. How far either the usefulness or the harm can go depends on the access granted and the checks placed around the system.
Within all that preparation, the models themselves are the unfamiliar part. In a recent Hidden Forces discussion, Paul Kedrosky described language models as “grown, not engineered.” Engineers design the architecture and training process, but training develops internal strategies they haven’t individually specified. Anthropic’s interpretability researchers have found strategies they didn’t expect, while acknowledging that their methods explain only a fraction of the computation they examine. We can demonstrate usefulness before we understand every mechanism producing it.
The transformer was a substantial advance, and understanding what training produces remains a scientific problem. We spent decades preparing the conditions for a technology whose form we didn’t anticipate and whose behavior we still don’t fully understand.
For most people, trying it requires something reassuringly familiar. Google spent decades getting us to type requests into a box, although we often had to reduce what we wanted to a few keywords the system could handle. Now we can explain more of the request in ordinary language. That familiarity makes it easy to begin, even though learning to direct the system and judge what comes back is a much larger undertaking.
Ask, and It Shall Be Given
People often know what they want to accomplish long before it becomes practical to try. Learning to use the relevant software can require more effort than the result is worth, particularly when it matters to only one person. AI changes that calculation by allowing people to direct more of the process in ordinary language. Many of its immediate applications come from purposes people already understand and now have a way to pursue.
Kedrosky’s “complexity discount” describes this constraint on the use of existing technology. Cloudflare offers services that many potential users find too difficult to operate. When an AI assistant handles some of those interactions, people can use infrastructure that was already there for purposes they already had. The platform’s technical maturity tells us relatively little about how many people could benefit from it under these new conditions.
A useful result also gives someone evidence about what their next attempt might require. A limit they had understood as personal can turn out to have depended partly on the tools available. That gives them a reason to reconsider ambitions they’d previously dismissed, while experimenting can reveal possibilities they hadn’t considered at all. This process still depends on their ability to judge the result. A disappointing attempt can show that further effort isn’t worthwhile, just as a successful one can encourage them to continue.
Venkatesh Rao argues that people may become more ambitious as intelligence becomes cheaper, and Thompson draws on this idea in his own account of making software. When sustained attention produces results that were previously inaccessible, people have a reason to invest more of it. Someone may think harder about a problem because they can now do something useful with the answer. Their interests can develop through that effort, so demand for AI can grow from ambitions people haven’t yet formed. Estimating its uses from the tasks people already perform would miss part of what the technology makes possible.
But bad actors benefit from these capabilities too. The expertise and effort that made a worthwhile project impractical may also have discouraged someone from attempting an attack. As AI supplies both, the person making the request can get exactly what they want while creating problems for everyone else.
Future Shock
Even if the models stopped improving tomorrow, we’d still have to adjust to what people are learning to do with them. Ethan Mollick calls the gap between what models can do and what people currently use them for a “capability overhang.” The consequences emerge unevenly as people discover applications and technology diffuses irregularly and unpredictably into society.
Kedrosky, for example, describes how someone who wants an image may obtain it directly from a model, bypassing software such as Photoshop that they would previously have paid to use. The same reduction in difficulty can also allow more people to produce software, increasing competition without necessarily creating enough paying demand to sustain its makers. A user can benefit long before a company finds a viable response. Successful uses can be economically disruptive even when everyone involved acts legitimately.
Even the person initiating an action may not anticipate how far its effects will extend. Kedrosky points to the Morris worm, deliberately released in November 1988. It exploited weaknesses in Unix software and network services, spreading across thousands of computers and repeatedly infecting machines until they became too overloaded to function. The disruption exceeded what Morris said he’d intended. His program consisted of explicitly written instructions, but its destructive reach depended on how those instructions interacted with the network it entered.
With AI, we’re connecting systems whose learned behavior we understand only partly to an environment where ordinary software could already produce consequences beyond its author’s intentions. Observing a model perform reliably in one setting tells us only part of what will happen when it interacts with other applications. The permissions it receives determine how far it can act, while the connections among those applications can extend the effects beyond the immediate task. Understanding the model is one part of understanding what happens when someone puts it to work.
The Electronic Frontier Foundation estimates that three percent of all software vulnerabilities ever recorded were found in October alone, and nearly a quarter of them in this year, because frontier models have made zero-days “incredibly cheap and easy to find.”
But people defending these systems can use AI to find and correct weaknesses too, and Thompson expects software security to improve as they do. Their results still have to be put into practice, which can require coordination among organizations with different responsibilities even when the technical solution is available. Someone may put the technology to a new use before everyone affected has adjusted.
That need to adjust to other people’s choices helps explain how AI can intensify the institutional strain Colin describes. People can accomplish more within systems whose rules and economics developed around earlier limits, leaving the organizations responsible for them to revise how they operate while continuing to serve existing users. Proceeding carefully can be reasonable, even as decisions made elsewhere make delay costly. The historical break may become apparent when people who have never used AI find they can no longer proceed on the old assumptions.
What’s Past Is Prologue
Agents could change familiar financial relationships by removing the friction that keeps people from exercising options they already have. As I wrote recently, Morgan Stanley’s research suggests they could make refinancing easier for borrowers. Businesses would have to reconsider assumptions about how much effort customers will make to obtain better terms.
Manidis and Roberts make a similar argument about bank deposits. Customers leave money in accounts paying very little because finding a better rate and moving it is a hassle. That inconvenience is valuable to the bank. An agent that handles the comparison and transfer could force banks to pay more to retain deposits they previously took for granted. The posts circulating on Twitter/X about agents moving savings into Treasury bonds describe another version of the same problem for banks.
We make similar assumptions about effort when we judge someone’s attention from a detailed, polished response. As AI makes that response easier to produce, we can infer less about the sender’s investment from its form alone. Someone may be expressing a sincere thought more effectively, or producing the appearance of consideration with very little of it. We’ll have to develop other grounds for trust as people become more capable of producing the signals on which we used to rely.
But people may also begin organizing their lives around things they hadn’t previously wanted. Drawing on Joyce Appleby’s history of capitalism, Rao describes how ordinary Britons developed tastes for imported goods such as tea and Indian cottons, then found it worthwhile to work more to obtain them. Their reasons for making an effort changed as new possibilities became available. A forecast based only on how efficiently they could meet their earlier needs would have missed the changes underway in what they wanted and how they chose to live.
That history is one reason I hesitate to follow Colin from technological maturity to application saturation. Companies can struggle to find growth within established markets while people discover valuable uses that would never have justified a commercial product. Those projects still have to be worth the time and resources people devote to them, and their usefulness doesn’t guarantee attractive investment returns. But the prospects of existing businesses give us an incomplete account of what people might do with the technology those businesses helped develop.
Whether we eventually describe this as a new era will depend partly on how people reorganize their lives around these capabilities. The first uses tell us something about what is possible, but much less about what people will come to consider worth doing.
In the small groups Thompson describes, for example, people can increasingly decide what software they need and make it for themselves. They still depend on models and infrastructure that others provide, but their purposes need only justify the effort of building for their own use. Thompson and his colleagues are already creating applications they have no intention of releasing publicly. He describes discovering “software-shaped holes in my life that suddenly I can fill without asking permission.”
The World Was All Before Them
An era can end as a victim of the success of its own technology. We spent decades making knowledge accessible and conducting more of our lives through software. Along the way, we organized businesses and careers around the expertise required to use it. The difficulty of getting something done helped determine whose services we needed, what we paid them, and which ambitions we considered realistic. Much of what became familiar was an accommodation to those difficulties.
AI depends on that accumulated development, and its usefulness gives us reasons to reconsider some of the accommodations. The same infrastructure can support a different division of labor. A person can continue using the same computer while revising their understanding of what they can undertake without assembling a team or persuading a company to serve them. When people act on those discoveries, others have to reconsider arrangements they had counted on. Technical continuity can produce a substantial break in how people live.
That is why I can accept AI as a culmination and still experience it as a beginning. The information age has produced capabilities that make some of its own organizing principles less reliable. We now have to decide which arrangements remain valuable, who gets to change them, and what they owe the people affected. Those decisions involve interests and obligations that greater intelligence alone cannot reconcile. For me, the sense of entering another era begins when things we had accepted as necessary become choices we have to make.