What
changed

In 1965, a computer lived behind glass.

You could see it from the corridor: a room of cabinets with tape reels turning, kept cold, and between you and it a counter staffed by people in white coats.

If you wanted the machine to do something, you definitely didn’t touch it. You wrote your program on paper, punched onto cards, handed the deck across the counter, and came back the next day. If the deck had a mistake in it, you came back the day after that. The people behind the counter were called operators, and for about fifteen years they were the only human beings who were allowed to touch a computer.

Nobody planned this. The machines were expensive, fragile and difficult to use, and it made sense to have specialists tend them. But something else happened at the same time, which is that the operators became a class.

They had their own vocabulary, their own certifications, their own salaries, and a great deal of informal power over who got their work run and when. Ted Nelson, writing in 1974, gave this class the name that stuck: the computer priesthood. His pamphlet Computer Lib was a manifesto against it, and its first line was that you can and must understand computers now.[1]

Nelson won, in the sense that the priesthood he described is gone. What is less often noticed is that it has come back four times since, in different clothes, and that we are living inside the fifth instance right now. This edition is about the pattern, and about what it says regarding the people who currently stand between you and an AIA.I. model and charge for the privilege.

§ The pattern, four times

The operators were dissolved by time-sharing, and then by the personal computer. By 1981, the machine sat on your desk and answered to you. Nobody in a white coat.

The second priesthood was the database administrators. Through the 1970s and 1980s, the data was a thing that lived somewhere, tended by people who knew where and how, and if you wanted a number out of it you filed a request and waited.

Edgar Codd’s relational model, in 1970, was on its face a paper about data structures.[2] It was also, quite deliberately, an attack on the priesthood: Codd wanted ordinary people to ask questions of the data in something like English, without an intermediary. SQL was the result, and then the spreadsheet, and then the dashboard. The database admin still exists, but the gatekeeping role is gone. Anyone with a login can query anything they are allowed to see.

The third was the webmaster, which is a word people under forty may not have encountered as a job title. From about 1994 to 2003 it was one — with a salary band, a conference circuit, the whole nine yards. The webmaster’s power rested on the fact that putting words on the web required knowing HTML, where the server was, and how to get files onto it. Blogger arrived in 1999, WordPress in 2003, and within a few years the title was useless.

The skill had not become worthless; the interface had become easy, and the premium for standing in front of a hard interface fell to zero.

The fourth is still visible, because it’s dissolving as we watch. From roughly 2009, when the first DevOpsDays was held in Ghent, the person who could operate cloud infrastructure became the most expensive engineer in the building.[3] The skill was real; the systems were hard. But managed platforms absorbed the difficulty piece by piece, and the premium has been falling for five years. The people who held it have mostly moved up into architecture or sideways into the next thing.

Four cycles. A genuinely hard interface creates a genuine bottleneck. A class forms around the bottleneck and, being human, acquires the trappings of a class: titles, credentials, a conference, a salary premium, a faint contempt for the laity. Then the interface gets easier — because someone builds the thing that makes it easier, and someone always does — and the class dissolves in under a decade, usually while its members are still describing the skill as permanent.

A priesthood is not a skill. It's a temporary interface gap with a salary attached.

§ What the priesthood actually is

It’s worth being precise about this, because the priests themselves are never precise about it, and cannot afford to be.

The value of an intermediary is exactly equal to the difficulty of the interface they stand in front of. Not their intelligence, not their training, not their years of experience: the difficulty of the interface. When the interface is hard, that value is high and feels earned. When the interface becomes easy, the value goes to zero regardless of how good the intermediary was, and it goes there suddenly, because interfaces do not get easier gradually. They get easier in steps, each of which is a product launch.

This means the priesthood is always, structurally, working against itself. Every good operator wanted a faster turnaround; every good DBA wrote tools so people would stop asking her for numbers, and every good webmaster built templates. The competent priest automates the priesthood. The only priests who last are the incompetent ones, and not for long.

It also means the priesthood’s own signals of maturity — the certifications, the salary surveys, et al — are lagging indicators. They appear when the class has become visible enough to be sold to, which is usually a few years after the interface has started to close. When the seminaries open, the church is already emptying.

§ The fifth priesthood

You know its names. Prompt engineer. AIA.I. strategist. Head of AIA.I. transformation. The consultancy’s AIA.I. practice. The two-day certification. The founder whose product is a thin layer between you and a model that you could, with mild effort, talk to yourself. In 2023 a frontier lab advertised a prompt-engineering role at a salary that made headlines, and the headlines were the point: they told a great many people that a new priesthood was hiring.[4]

Every marker from the previous four cycles is present. A genuinely hard interface: in 2022 getting good output from a model did require craft, and the craft was real. A class forming around it, with titles and credentials and a contempt for the laity (“most people don’t know how to prompt”). A salary premium. A conference circuit. And now, right on schedule, the seminaries: courses, certificates, bootcamps, arriving in numbers that already exceed the roles that ask for them.

And the counter-signal, also on schedule: the interface is closing. Models now rewrite the user’s prompt before answering it. Agents chain their own tools without being told how. The gap between what a skilled operator can get out of the system and what a distracted person typing a sentence can get out of it — the gap that is the priesthood — has been narrowing every quarter for two years, and the thing narrowing it is the product itself. Previous priesthoods were dissolved by someone else’s product. This one is being dissolved by the very tool it claims to interpret, which is a new thing under the sun and is why this cycle will be the shortest.

§ What is different this time, and what is not

Two things are different. The first is speed: the operators lasted fifteen years, the webmasters nine, the cloud priesthood is on its way out after about twelve, and the AIA.I. priesthood is unlikely to see its seventh birthday, because the interface is closing at the pace of model releases rather than at the pace of an industry. The second is that the tool is eroding its own intermediaries, which means there is no external product to wait for and no lag to hide in.

But one thing is not different, and it’s the thing that matters most for anyone deciding what to do about all this.

Every previous priesthood had two parts fused together, and everyone mistook them for one. There was operation — the skill of getting the machine to do the thing. And there was accountability — being the person who answered for what the machine did. The operator signed the log. The DBA vouched for the number. The webmaster was who got called when the site went down. When the interface closed, operation went to zero. Accountability did not. It moved: to the analyst who now ran her own query and put her name on the result, to the marketing manager who now published the page himself and owned the typo. The work of answering for the output never disappeared; it was redistributed to whoever now pressed the button.

That is the part of the fifth priesthood that survives, and it is the part the priests are least interested in, because it is not glamorous and cannot be certified in two days. The machine will write the memo, the brief, the diagnosis, the code, the contract. Someone still has to be the human whose name is on it — who can be asked why, who can be sued, who can be struck off. The value that used to attach to “knows how to use the model” is migrating, right now, to “will stand behind what the model produced.” The insurers have already noticed. The professional bodies are noticing. The labour market has not yet, which is what makes it a slow variable and this an edition rather than a news item.

§ Second-order effects

If the pattern holds, four things follow.

The intermediary premium goes to the interface builders, not the intermediaries. The money made by webmasters went, in the end, to WordPress and Squarespace. The money made by the cloud priesthood went to the managed platforms. The money currently made by AIA.I. intermediaries will go to whoever makes the intermediary unnecessary — which, unusually, is the model companies themselves. Expect “AIA.I. transformation” revenue at consultancies to peak and reverse well before anyone announces it has.

The training ecosystem is the first casualty. Courses and certificates are sold to people entering a priesthood, and the flow of entrants dries up before the flow of leavers begins. Anyone whose business is teaching people to operate the model should assume they are selling into a closing window.

Accountability becomes the scarce, priced thing. Expect liability insurance for AIA.I.-produced professional work, expect sign-off requirements in the licensed professions, and expect the salary premium to reappear — attached not to “AIA.I. skills” but to whoever is willing to be named. This is a return to an older pattern, not a new one: the engineer stamps the drawing, whether or not she drew it.

The people who did best in every previous cycle were the laity who picked up the tool early and kept their own jobs. Not the priests. The accountant who learned VisiCalc in 1980 did better than the person who became a “spreadsheet consultant.” The general rule is: use the thing, put your name on the result, and do not build a career on standing between other people and the machine. The gap you are standing in is closing.

§ What this edition is not saying

It is not saying that skill with these systems is worthless; it is saying the skill is being absorbed into the interface, as it always is, and that the premium for it is a wasting asset. It is not saying the machines are good, or bad, or that the jobs of the laity are safe; those are fast variables and other people’s essays. It is saying that a specific class of role is forming exactly the way four previous ones formed, and that we have four complete records of what happens next.

The claims that follow from this are on the ledger below, with confidence attached. If the fifth priesthood is still standing in 2029, this edition will say so, on this page, under Revisions.

[1] Theodor H. Nelson, Computer Lib / Dream Machines (self-published, 1974). The pamphlet is two books bound back to back; Computer Lib is the polemic, and “priesthood” is its recurring word for the people who controlled access to machines. Nelson was not the first to use it — the phrase was in the air in the early 1970s — but he is the reason it stuck.

[2] E. F. Codd, “A Relational Model of Data for Large Shared Data Banks,” Communications of the ACM 13, no. 6 (1970). The paper’s opening sentence says users should be protected from having to know how data is organisedorganized in the machine. That is an interface argument before it is a mathematical one.

[3] The first DevOpsDays conference was organisedorganized by Patrick Debois in Ghent in October 2009; the term itself dates from that event. The salary premium for the roles it named is well documented in the annual developer surveys from about 2014 onward, and so is its recent flattening.

[4] The posting, from Anthropic in early 2023, listed a range that topped out above US$300,000 and was reported by most of the technology press. It is worth noting that, two years later, the same industry ships models that rewrite the user’s prompt before acting on it. The job advertisement and its own obsolescence arrived from the same handful of buildings, about two years apart.