The technological singularity: an anecdote, a conditional and a forecast

The “technological singularity” often arrives wrapped in borrowed authority: it is presented as John von Neumann's theory about the day machines surpass people. The archive tells another story. Von Neumann published no such theory. Stanislaw Ulam recalled a conversation; I. J. Good framed a conditional argument; Vernor Vinge assembled several routes to a discontinuity and added a deadline.

Separating the three texts is not a minor attribution dispute. An anecdote is assessed by provenance, a conditional by the validity of its premises, and a forecast by its outcome and date. Blending them lets a failed prediction retain the prestige of a mathematician who never wrote it.

The transferable skill is to decompose any AGI or singularity claim into five boxes: definition, mechanism, necessary conditions, observable indicator and time horizon. When the last two are missing, the statement is commonly a possibility or narrative rather than a testable forecast.

1958: Ulam left a recollection, not a von Neumann theory

In May 1958, Ulam published John von Neumann 1903–1957 in the Bulletin of the American Mathematical Society. It was a 49-page posthumous tribute covering the mathematician's life, personality and work. Within a remembered conversation, Ulam mentioned accelerating technology and changes in human life that appeared to approach a historical singularity.

The documentary unit matters. Ulam is writing in 1958 about an earlier conversation; the passage is not a paper by von Neumann, a definition of machine intelligence or a model of self-improvement. It supplies no variable, mechanism or date. Calling it “von Neumann's theory of the singularity” converts a biographical memory into a technical thesis that does not exist.

The method for a famous quotation is simple: find the document, identify author and genre, read the surrounding paragraph, and ask whether the quoted person wrote those words or somebody else remembered them. An illustrious surname cannot repair a broken attribution chain.

1965: Good stated a conditional consequence

I. J. Good published Speculations Concerning the First Ultraintelligent Machine in volume 6 of Advances in Computers. The commercial original is paywalled, while Virginia Tech preserves an institutional reprint identified as such. Good defined an ultraintelligent machine as one able to surpass humans greatly in every intellectual activity.

His core was an implication: if designing machines is one of those intellectual activities, a machine already superior to humans could design a better one. An intelligence explosion might follow. The conclusion depends on the first machine existing, on its superiority including research and design, and on subsequent improvements being repeatable.

Good did not date that first machine. Nor did he establish that every cycle would be fast, cheap, verifiable or cumulative. The chapter discussed possible architectures, learning, the brain and control; it was a broad technical speculation rather than an empirical law of growth.

The conditional remains intellectually powerful because it locates a real question: can a system improve the process that creates its successor? But “if A, then B” does not establish that A will happen or when. Confusing logical possibility with probability is one of the debate's common shortcuts.

1993: Vinge added routes and a clock

In March 1993, Vernor Vinge presented The Coming Technological Singularity at VISION-21, a symposium sponsored by NASA's Lewis Research Center and the Ohio Aerospace Institute. The essay placed the technological means to create greater-than-human intelligence within 30 years. That horizon led to 2023.

His singularity was broader than “an AI rewrites its own code”. He listed four routes: awake computers with superhuman intelligence, large networks waking as an entity, interfaces amplifying human intelligence and biological improvement of human intellect. The thesis concerned a rupture beyond which ordinary models for predicting the future would fail.

The date makes the text answerable to evidence. By 2023, there was no accepted public demonstration of an entity possessing the general superhuman intelligence Vinge described. This does not make the essay useless: it identifies which part was a forecast and allows its literal deadline to be recorded as unmet.

A serious forecast preserves the edition and date rather than moving the milestone whenever it nears. It also states what observation would count as success. Without a criterion, every benchmark improvement can be relabelled retrospectively as “the beginning”, and the thesis never risks failure.

AGI, superintelligence and singularity are three concepts

AGI generally concerns breadth: learning or performing many kinds of cognitive task and transferring knowledge to new problems. Superintelligence adds a level: performance beyond humans across a broad set of intellectual activities. Singularity describes a historical consequence or prediction boundary, not merely a capability score.

Legg and Hutter proposed a formalisation of intelligence as the capacity to achieve goals across a wide range of environments. Their measure is general and mathematically explicit, but a practical approximation is not straightforward. The difficulty of defining intelligence itself prevents “AGI” from being treated as a directly observed label.

Levels of AGI proposes separating performance from generality. A system can be superhuman at a narrow task and remain far from general competence. The framework also separates autonomy and risk from capability itself: doing something well does not mean doing it safely or without supervision.

AlphaFold illustrates the denominator. Its 2021 paper documented high accuracy in protein-structure prediction. That is a specific scientific achievement. It does not demonstrate general task learning, autonomous improvement of AI research or a historical discontinuity.

An explosion needs a chain, not one leap

First there would need to be broad competence, or at least exceptional ability in AI research. Second, the system would have to propose useful improvements to models, data, algorithms or hardware. Third, those proposals would need experiments, resources, manufacturing, energy and validation. Fourth, results would have to feed another cycle quickly enough.

Gains would also need to compound. A process may improve early and later hit limits in data, compute, physics, coordination or knowledge. Designing a hypothesis is not testing it; generating code is not deploying it; increasing capability does not eliminate mistakes or misspecified goals.

Speed belongs to the mechanism. If each cycle depends on years of chip fabrication, biological trials or human proof review, feedback may matter without being explosive. If improvements interfere with one another, adding their separate benefits exaggerates the whole.

Evidence for a singularity would therefore have to measure more than final performance. It would need new research outcomes, improvement per cycle, cycle time and cost, the share of valid proposals, observed limits, and the external work supplied by people and infrastructure.

How to read growth charts

A curve for compute, investment or benchmark score describes one variable rather than the whole conclusion. Extrapolation must explain why the regime will continue, whether the measure has a ceiling, whether the test set changed and whether the measured resource is the bottleneck. A rising line does not select an AGI date by itself.

Measured and promised also belong on separate lines. “The system scored X under this protocol” is a result; “this trend will automate research” adds an inference; “therefore singularity arrives in year Y” adds a mechanism and deadline. Every arrow needs its own evidence.

Benchmarks can saturate because models improve, because contamination exists or because the test no longer represents the frontier. A broader battery reduces risk, but generality requires transfer, adaptation, robustness and retention when the problem changes, not many examples from one family.

A template for the next headline

Definition: does AGI mean average-human, expert, superhuman or autonomous performance, and on which tasks? Mechanism: is the proposal scale, new algorithms, tools, research improvement or human interfaces? Conditions: which resources, permissions and outside validation does it require?

Indicator: what observable result would refute or support the thesis? Preserve the dataset, version, denominator and human contribution. Horizon: is there a date or range, and does it remain fixed? Without one, a claim may guide research but cannot demonstrate forecasting accuracy.

The genealogy leaves three distinct lessons. Ulam teaches not to turn a memory into authorship. Good teaches the difference between a conditional consequence and a probability. Vinge shows the value of accepting a deadline that can fail. Together they offer a more rigorous way to discuss the future: text in view, assumptions visible and a clock that does not move.