Thought discovers it has a desk job
In the spring and summer of 1957, J. C. R. Licklider began following himself around with the intellectual equivalent of a stopwatch. He wanted to know what one moderately technical person actually did during the hours he described as work. The least expensive available research subject was Licklider, and at first he appeared cooperative.
The study immediately developed an administrative problem. Keeping a record of his activities was itself an activity, which required a record, which threatened to require another record. Licklider declined the invitation to spend the remainder of his career documenting the documentation.
What he did record gave him pause. By his informal estimate, about eighty-five per cent of his “thinking” time went into getting into a position to think: finding information, calculating, plotting, translating incompatible measurements and performing the small clerical ceremonies that precede an insight. Several hours might be spent making six experiments comparable. Once they were comparable, the answer could take seconds.
Licklider set out to measure thought and found paperwork wearing its clothes.
This was not a universal productivity law. It was one person’s inadequately sampled self-study, and Licklider said so. But it helped sharpen the proposal he published in 1960 as “Man-Computer Symbiosis.” Computers would handle the routinizable preparation; people would formulate problems, set goals, choose criteria and evaluate results. The ideal was not a mechanical oracle. It was a partnership in which each participant did the work the other did badly.
Licklider wanted to shorten the interval before insight. He did not ask the question that concerns us now: what happens if the interval between insight and consequence becomes too short for anyone else to enter?
The future develops an edge
A year after Licklider’s self-surveillance, the mathematician Stanislaw Ulam published a memorial essay about his friend John von Neumann. Ulam recalled a conversation about accelerating technology and an approaching “essential singularity in the history of the race,” beyond which ordinary human affairs could not continue as before.
The provenance matters. This was Ulam remembering a conversation after von Neumann’s death, not a transcript; the remark concerned technological change in general, not modern artificial intelligence. It does not give us a date, a mechanism or permission to put a red circle around a Tuesday.
It does give us an image. A singularity, in this usage, is not merely a very large event. It is a boundary beyond which the coordinates that guided us to the boundary stop being reliable. The road does not necessarily explode. The map simply declines further responsibility.
The singularity may be less something that happens in the future than something that happens to the future.
That distinction was central to Vernor Vinge’s influential 1993 essay. Vinge is often remembered for placing the transition somewhere between roughly 2005 and 2030. The later edge of that window has not yet arrived, and no agreed operational test tells us how the forecast should be scored. A conjecture should not become a countdown merely because its subject is dramatic enough to sell commemorative clocks.
Vinge’s more durable point was epistemic. If minds substantially more capable than ours began producing knowledge and redesigning the systems that produce it, the world beyond that threshold might become opaque to extrapolation from ordinary human history. The difficulty would not be that nothing happens after the wall. It would be that we cannot responsibly draw the rest of the tour from this side.
There is no accepted date or agreed scientific definition for a technological singularity. Von Neumann’s reported remark, I. J. Good’s intelligence explosion and Vinge’s epistemic horizon are related ideas, not interchangeable experimental results.
The last invention, conditions apply
In 1965, the statistician and former Bletchley Park cryptanalyst I. J. Good supplied the mechanism that now powers much of the singularity story. Imagine, he proposed, an “ultraintelligent machine” capable of surpassing every human intellectual activity. Designing intelligent machines is one of those activities. The machine could therefore design a better machine, which could design a still better one. Each improvement might shorten the path to the next.
The argument is recursive and almost indecently elegant. Place intelligence in a loop with the ability to improve the loop, and the interval between generations contracts. Good called the result an intelligence explosion and suggested that the first ultraintelligent machine might be the last invention human beings need make.
Then came the condition history tends to leave in the packaging: the machine would have to remain manageable enough to help us keep it under control. The last invention arrived with terms and conditions. We remembered the product name.
The loop also lives in a material world. Better designs still require chips, electricity, cooling, networks, factories, data, tests and permission to deploy. An improved system must be able to tell a real improvement from a confident new error. Software may move quickly; power grids and fabrication plants remain discouragingly three-dimensional.
None of this disproves recursive improvement. It prevents a logical possibility from dressing as a demonstrated timetable. Present AI systems can accelerate coding, search and parts of scientific work. That is consequential. It is not, by itself, evidence that an autonomous intelligence explosion has begun.
If a system could reliably improve the capacities needed to improve itself—and could obtain the physical resources and authority to implement those changes—positive feedback could make progress difficult to forecast. Whether every link in that chain will hold is the open question, not an inconvenient detail.
The important unit may not be intelligence alone. It may be the time between a proposed change and the world’s ability to absorb it. A machine can shorten the first interval while leaving the second stubbornly human.
Four clocks stop agreeing
Put four clocks on the museum wall. They are not empirical averages; they are an explanatory exhibit. Real laboratories can take decades, governments can move overnight and companies can spend three quarters selecting a meeting platform. The clocks describe a growing structural mismatch.
The laboratory clock may generate a candidate answer in hours or days. The company clock translates answers into products, jobs and incentives over weeks or quarters. The legislative clock defines a harm, hears objections and constructs a rule over months or years. The human clock needs childhoods to learn, relationships to acquire trust and a considerable portion of a life to discover which ambitions were borrowed.
Now give all four clocks one decision. A model recommends a new way to diagnose illness, educate a child or screen an applicant. The system updates. The company deploys. The hospital, school or employer revises policy. A regulator begins an inquiry. The person affected must decide now, with a body, job or child already inside the experiment.
I call the mismatch institutional jet lag: systems occupy the same present while updating their assumptions at incompatible rates. The phrase is an interpretive metaphor, not a recognized scientific diagnosis. It names the queasy interval in which one institution has arrived in a future that another has not yet learned to regulate—and a person has been asked to live there without luggage.
This horizon will not reach everyone at once. Those who control the fast clock may retain a detailed internal map. Those receiving its decisions encounter a black box and are told that opacity is simply what progress looks like from economy seating. Unpredictability can be genuine and still be distributed unequally.
If an answer arrives before the people affected have learned how to ask the question, is that speed—or a new distribution of authority?
Government is not merely a slow computer. Some of its delay is failure, capture or ritual. Some is the time required for evidence, representation, appeal and consent. A person is not a defective institution because trust cannot be installed during a maintenance window.
The singularity, viewed this way, need not appear as a vertical line on a graph. It may arrive as widening disagreement among clocks: not the instant the machine becomes a god, but the interval in which the systems changing the world outrun the systems that make change legitimate.
Which pauses belong to civilization?
Return to Licklider’s eighty-five per cent. Much of the delay he found was genuine drudgery: searching, transcription, incompatible formats, graphs drawn by hand. Nobody needs to preserve a badly labeled filing cabinet as a sanctuary for the human spirit.
But some intervals that resemble friction contain something else: checking a source, replicating a result, explaining a choice, hearing an objection, obtaining informed consent, allowing an appeal and discovering that the confident first answer was wrong.
Slowness is not automatically wisdom. Bureaucracy can hide power, exhaust vulnerable people and keep obsolete systems alive long after the farewell luncheon. Speed is not automatically intelligence. A process can produce an answer faster than it can produce a reason anyone is entitled to accept.
The moral object is not delay. It is answerability: preserving enough shared time for the people affected by a decision to understand it, challenge it and participate in determining what it is for. Some pauses are empty corridors. Others are the room in which responsibility catches up.
Licklider imagined the machine finding the paper, plotting the graph and proposing the next experiment before he finished his coffee. That partnership can be liberating. What the machine has not automatically done is choose a legitimate goal, decide who bears the risk, explain the result to everyone affected or make those people share one future.
The corridor sign—Please do not make plans beyond this point—does not mean that nothing lies ahead. It means that the public calendar has stopped binding all participants equally. Some are building beyond the wall. Some are regulating what crossed it yesterday. Some have just received revised terms for a life they planned under the previous model.
The future may not disappear. It may simply stop waiting for us to agree on it.
The first machine to end the future may not predict tomorrow. It may make tomorrow arrive before we have finished naming today.
Exhibit record
- J. C. R. Licklider, “Man-Computer Symbiosis” (1960), including his 1957 informal time-and-motion study
- Stanislaw Ulam, “John von Neumann, 1903–1957,” Bulletin of the American Mathematical Society 64 (1958)
- I. J. Good, “Speculations Concerning the First Ultraintelligent Machine” (1965)
- Vernor Vinge, “The Coming Technological Singularity” (1993)
- Karl Popper, The Poverty of Historicism (1957), on limits to predicting the growth of knowledge