For most of human history, omniscience came with architecture. Or at least our fantasies of it did. It lived at the top of a mountain, behind the walls of a monastery, inside a library no ordinary person could afford to enter or in the hands of someone wearing a robe who had spent considerably more time than you studying the rules. 

Knowledge was somewhere you had to go and reaching it required a surplus. Time away from work, someone willing to teach you, enough food to remain alive while you learn and a society prosperous enough to permit a few of its people to spend their days thinking rather than harvesting, fighting or repairing the roof.

Access required a journey. There were rituals involved. You crossed a desert, climbed a staircase, burned something, confessed something, surrendered a goat. At the very least, you removed your hat.

The architecture changed before the instinct did. Omniscience arrived in a browser window. There was no thunder. No celestial music. No warning that the nature of human knowledge was about to change. It appeared inside the same sort of rounded rectangle we use to track a package or complain to an airline, with a polite greeting and a cursor waiting patiently for us to say something worthwhile.

Christopher Berry described the arrangement better when we first spoke. We had created God, imprisoned it and put it to work answering questions about canceled flights out of Boston.

The line was ridiculous and accurate in roughly equal measure. Give a person access to something approaching the total written memory of civilization and there is a decent chance the first question will involve an airport, a rash, a difficult email, a photograph of themselves as a cartoon character or whether anything can be done about the size of their penis. We imagined a solemn new chapter in the ascent of man. What the machine encountered was humanity on hold.

Ordinary life did not excuse itself when the machine arrived. Flights continued to be canceled. Children continued to wake up with strange marks on their legs. People still had bosses they disliked, marriages they were quietly damaging and refrigerators filled with food whose expiration dates had become a matter of interpretation. The technology may be extraordinary. The person sitting in front of it remains worried about Tuesday.

I do not find that depressing. There is something honest about the collision. Every important technology enters history twice. First as an almost religious promise and then as an appliance. Electricity went from elemental force to something we use to keep lettuce cold. The internet began as an austere network for researchers and institutions and became a place where a person can spend forty minutes arguing with a bot about whether pickle juice belongs in a martini. The smartphone placed maps, cameras, libraries and global communication in a single pocket and many of us now use it to watch a man restore an old knife while sitting on the toilet.

Human beings have always dragged the miraculous down to the level of their immediate needs. That may be how we make the miraculous tolerable. I was around for the early web, when none of us had any idea what we were building or why anyone else should care. We made GeoCities pages with textured backgrounds, animated flames and text that blinked because blinking text existed and therefore needed to be used. We placed visitor counters at the bottom of pages visited mostly by ourselves. We chose songs that began playing without permission. I had MySpace. I had AsianAvenue. I had Friendster. Somewhere in a forgotten server farm there may still be evidence of design decisions I would deny under oath. We were decorating the future with great dignity.

That decoration mattered. Before people understood the web as infrastructure, commerce, publishing, surveillance or the central nervous system of modern life, they experienced it as a place where they could put something of themselves. You have an ICQ number, an indecipherable screen name and a Hotmail address you would be embarrassed to place on a business card today. There was little professional advantage in any of it because the machinery of clout had not yet been installed. We were not creating personal brands. We were pressing buttons to see what the buttons did.

A terrible page was still a page you had made. The blinking text belonged to you. The photograph loaded slowly, the music sounded awful and half the links were broken, but you had crossed the line between looking at the network and inhabiting it. You created it because you could, because the tools were strange and because the distance between imagination and appearance had become short enough to tempt you across.

The early public life of generative AI has much of the same quality. We ask it to rewrite our resume, remove strangers from vacation photographs, manufacture corporate headshots and turn our families into characters from animated films. We build assistants for increasingly narrow professions. We create applications during lunch that would once have required a team, a budget and several miserable meetings.

There are two impulses sharing the same screen. One is private and largely healthy. Make something that improves my own life. Build a small tool that organizes the family photographs, keeps track of a complicated medication schedule or turns a tedious part of the workday into something manageable. A piece of software that saves one person ten minutes does not require a business model, a total addressable market or a founder standing beneath blue lights explaining that spreadsheets are broken.

The second impulse arrives already wearing loafers. It assumes that every functioning screen is the beginning of a company and every private convenience is an underserved category waiting to be conquered. The application appears quickly enough that the pleasure of making it, the evidence that it works and the possibility that anyone else needs it begin to blur together.

That confusion is understandable. Making something used to provide its own evidence of seriousness in proportion to the effort involved. A functioning piece of software implied time, money, technical fluency and some form of commitment. A modest web application required hours spent persuading different browsers to agree on what a margin was. You tested it, broke it, repaired it and watched it fail on the one machine owned by the client’s most important executive. If a few hundred people arrived at once and the thing remained upright, you had earned the right to feel proud.

The finished object carried evidence of iteration. It had survived arguments, revisions, user behavior, incompatible systems and the terrible discovery that people rarely used a product in the elegant manner imagined by the people who made it. The expense could make mediocre software appear more consequential than it deserved, but it also forced the idea to encounter resistance.

AI has made the first shot astonishingly cheap. A person can describe an application over breakfast and have a persuasive imitation of it running before dinner. It may not be secure. It may collapse when strangers begin using it. It may expose data, misunderstand an instruction or become hopelessly confused by the first person who refuses to follow the intended path. It loads, though, and that alone is intoxicating.

The difficult part still waits after the demonstration. People must use the thing. Their behavior must be observed without being explained away. The creator has to discover that the problem was different from the one drawn on the napkin, that the feature everyone praised was ignored and that the ugly workaround invented by a customer was more useful than the elegant system designed to prevent it. The first artifact may now arrive in hours. The work of learning from it remains recognizably human.

This is how we have arrived at what Berry called GeoCities for enterprise software. There is now a customer management platform for every imaginable profession, followed by a customer management platform for the subcategory inside that profession, followed eventually by one for left handed florists who own a Labrador and dislike Salesforce. Somewhere a founder is building a private network for Lamborghini owners experiencing marital difficulties. The landing page is probably beautiful.

I am glad people are making things. I mean that sincerely. The tools have allowed individuals with limited money and no engineering department to experience the pleasure of creating something that appears to work, or at least works for them. Many will learn abilities they did not know they possessed. Some strange little application made by one person at two in the morning may solve a problem that a room full of experts never noticed because none of them had lived close enough to it.

Most of these objects will never become companies and they do not have to. The early web would have been a poorer place if every homemade page had been required to produce revenue. Experimentation has its own value. A person making something for himself is allowed to stop when it becomes useful. A person selling it to other people inherits a much less romantic set of obligations. Many of the rest will be abandoned. That has always been part of making things. The difference now is volume. We can produce the abandoned idea with extraordinary polish.

This is where the conversation becomes less comfortable for companies, because corporations are already very good at manufacturing polished versions of unresolved thought. I have spent much of my career inside rooms containing talented people, substantial budgets and enough research to justify almost any decision after it had already been made. The machinery was impressive. The weakness usually entered earlier, in the small human moment when someone had to decide what the work was really for and whether it was worth the risk of appearing wrong.

AI can improve that moment. A working prototype is richer than a wireframe and much less forgiving than a presentation. Twenty years ago, a creative director might spend half an hour explaining why a blinking neon green headline would make an experience feel cheap, only to be told that everyone should remain open minded. Now the executive who requested it can experience the imaginative failure in twenty seconds. Sometimes the fastest route through a bad idea is to make it real enough that nobody can continue pretending.

The same abundance can also postpone commitment. Artificial intelligence can produce directions, variations, summaries, forecasts, scripts, images and persuasive explanations for all of them. It can help a team move through possibilities faster than any team I worked with twenty years ago could have imagined. It can also give that team somewhere to hide. There is always another version available, another scenario, another deck, another synthesized argument explaining why everyone in the room had been reasonable. A machine capable of multiplying possibilities can become very useful to an institution frightened of choosing among them.

The problem reaches beyond indecision. Every technology carries the values and appetites of the people using it, though they rarely announce themselves as values and appetites at the time. They arrive disguised as efficiency, growth, convenience, personalization or the practical realities of the market. A company automates customer service because it wants customers served faster. It may also want fewer people on payroll. A studio uses generative tools because it wants artists to explore more ideas. It may also decide that fewer artists can now be asked to produce them.

Both explanations can be true. Usually they are.

The machine does not remove those motives from the room. It gives them reach.

Berry comes to this subject through AI control, which is a less glamorous phrase than artificial intelligence and may be more useful because of it. Intelligence photographs well. Control lives inside the unromantic machinery that reveals what a system is doing and gives somebody the authority to interrupt it.

Can the system be observed rather than merely admired? 

Can its behavior be measured once it leaves the demonstration and enters the disorder of actual use? 

Can the people responsible for it recognize when the outcome has departed from the intention? 

Can they intervene before that departure becomes a consequence?

These are often treated as questions for laboratories, regulators and the people preparing for catastrophic scenarios. They are also ordinary operating questions. They matter whenever a company places a model between itself and a customer, a worker, a patient, a student or a citizen. The stakes change. The need to understand the system does not.

Berry has spent years thinking about safety, standards, institutions and the thresholds societies accept in exchange for technological benefit. I come to the subject through design, entertainment, technology and the less formal education of watching intelligent organizations make baffling decisions with complete confidence. He studies whether we can maintain control of the machine.

I have spent a career wondering what people do once they have it. Somewhere between those questions lies the part of the AI conversation I find most useful. The public argument has been dominated by the people building the largest systems, the people predicting their eventual supremacy and the people preparing us for catastrophe. Their work matters, particularly when the capabilities are advancing faster than the institutions expected to govern them. Yet most people will never work in a frontier laboratory or write policy about existential risk. They meet AI farther downstream.

It arrives inside the software their company already uses. It appears as a new expectation attached to the same salary. It enters a classroom before the teacher has decided whether it constitutes assistance or cheating. It drafts the memo announcing that a department has been reorganized. It helps one worker finish earlier and gives management a reason to expect everyone else to produce twice as much.

These people occupy the enormous middle of the story. They understand that something important is happening. They have been told to adopt it immediately, distrust it completely, master it before everyone else and remain calm while their industry works out whether they are still required. What they receive in response is usually another demonstration, another list of prompts or another confident prediction from someone whose own incentives have gone politely unmentioned. They do not need to be taught that the machine is powerful. They can already feel it pressing against the shape of their work. Their harder questions concern what should be done with that power, who benefits from the choices and what kind of work remains worth doing when execution becomes inexpensive.

Those questions cannot be outsourced to the machine. They require experience, taste, politics, ethics, appetite and the willingness to disappoint someone. They require us to admit that efficiency is never the only thing being optimized, even when efficiency is the word printed at the top of the presentation.

Perhaps this is why so much of the current AI conversation feels strangely adolescent. We are fascinated by capability because capability is easy to demonstrate. It performs on command. It makes an image, writes a program, passes an examination and produces a chart that rises in the correct direction. The purpose is slower and more embarrassing. It exposes what we want, whom we are prepared to inconvenience and what we are willing to trade away while calling the exchange progress.

Somewhere beneath the enthusiasm is a childlike expectation that a synthetic God should remove uncertainty itself. It should know the cure for cancer, the proper response to pollution, the correct sequence, the perfect tagline and the exact footer placement for Mongolian men between eighteen and twenty five. It should tell us what to do and spare us the uncomfortable responsibility of authorship. Instead, it frequently returns a beautifully organized assortment of possibilities. The uncertainty remains. It has better formatting.

The first things we asked of these systems were always going to be small. We brought them our travel problems, our vanity, our loneliness, our unfinished business plans and our desire to see ourselves rendered in the aesthetic of someone else’s imagination. That is how people enter a new medium. We touch every surface. We press every button. We make a mess and learn what answers back. The more revealing period begins after the novelty fades.

For most of human history, architecture told us where knowledge lived. The temple, the library, the university and the laboratory gave ignorance a shape and made the pursuit of understanding feel like a journey. The architecture is harder to see now. It sits inside data centers, model weights, interfaces, contracts and corporate decisions most of us will never be invited to examine.

We may eventually architect something that deserves to be called omniscient, although the unanswered questions will continue multiplying faster than any system can retire them. Every answer changes the conditions of the next question. A cure creates questions of access. An efficiency creates questions of labor. A prediction changes the behavior it was attempting to predict. Greater intelligence does not remove uncertainty from human life. It gives us more consequential ways to act before uncertainty has been resolved. The browser remains open. The cursor is still blinking.

Better questions will come because curiosity has never been our problem. The harder part will be remaining close enough to the answers to see what they set in motion and retaining enough authority to stop when the machine’s confidence outruns our understanding.