Given Enough Time, Everything Seems Inevitable

There is a tendency to describe artificial intelligence as though it appeared suddenly, the consequence of a handful of recent breakthroughs followed by an even smaller handful of companies discovering that the breakthroughs could be commoditized. This is a useful account if one wishes to begin the story somewhere around 2017. A longer view is less reassuring. Artificial intelligence sits at the end of a history in which intelligence has repeatedly encountered limits imposed by its biological substrate and, with no apparent understanding of where the process was leading, found ways around them.

The first problem was isolation. A thought occurring in one brain was of little use to another unless the second brain happened to have much the same thought independently. Language provided a workaround. A thought could be compressed, transmitted and then decompressed, albeit imperfectly, inside another brain, combined there with other thoughts and returned changed. Ideas no longer had to be invented separately by every mind that needed them. Intelligence had acquired a means of networking. Whatever intelligence had been before language, it could now operate across multiple brains. It had become collective.

This solved one problem, but collective intelligence was still inconveniently dependent on the continued presence of the people doing the thinking. Memory was unreliable, distance was troublesome and death remained, as ever, a fairly severe form of data loss. Writing separated a thought from the organism that had produced it. What one mind knew could persist after that mind disappeared, travel farther than a voice could carry and be encountered by another person centuries later. Memory was no longer confined to the thing doing the remembering. Intelligence had acquired persistence.

Persistence created a problem of abundance. Once ideas could survive, there were progressively more of them worth preserving and distributing, while copying remained a slow manual process performed by people whose accuracy and enthusiasm varied. Printing changed the scale. The same external memory could be reproduced cheaply, with reasonable fidelity, and placed before thousands of minds rather than a fortunate handful. Libraries became larger, books became common. Intelligence had acquired distribution.

The conditions under which intelligence operated had now changed substantially without much changing the biological machinery itself. Each new mind arrived in a world already dense with preserved thought, much of it produced by people long dead and much of it dependent on earlier thought whose authors were deader still. A medieval scholar and a modern physicist possess essentially the same kind of brain. The difference between what they can know lies largely in the enormous external structure of preserved thought available.

But the brain was still doing the work. Books could preserve an equation indefinitely; they could not solve it. A proof could survive its author by centuries, but another mind still had to reconstruct the reasoning. The next step was not simply to preserve what intelligence had produced, but to describe some of the operations by which it produced it.

Mathematics increasingly made that possible. A procedure could be stated precisely enough that different people following the same steps would obtain the same result. Even uncertainty proved less resistant than it appeared. Much of ordinary reasoning consists of making judgments from incomplete evidence, estimating what is likely and revising those estimates as new information arrives. Probability gave that process mathematical structure, and Bayesian inference showed that even apparently intuitive judgments could, at least in part, be reduced to repeatable operations. Intelligence had discovered that even its guesses contained structure. Parts of reasoning could now exist as procedures independent of the particular mind performing them.

Once parts of reasoning could be expressed as procedure, the next leap was stranger: imagining that the procedure might not require a mind at all. Babbage’s Analytical Engine proposed a clockwork machine capable of carrying out operations that had previously existed only as instructions for people. If an operation could be reduced to a sufficiently explicit sequence of instructions, there was no obvious reason those instructions had to be executed by the same kind of thing that had devised them. The Analytical Engine was never completed as Babbage imagined it, but that mattered less than the distinction it exposed. Procedure and performer could be separated. The answer did not become less correct because no mind had experienced the steps along the way.

Mechanical calculation presented a relatively simple case because the standard of success was external. Either the answer was correct or it was not. But if machinery could perform increasingly complicated operations once thought to require a mind, when did the result become something we were prepared to call intelligent? Answering seemed to require knowing what was happening inside the machine and comparing it with something we understood only imperfectly inside ourselves, at which point we are drifting uncomfortably into metaphysics.

Turing’s solution was to decline the question. Rather than define thought, consciousness or intelligence before proceeding, he proposed judging the machine by observable performance. If its responses were sufficiently convincing that a human interlocutor could not reliably distinguish them from those of another person, the unresolved nature of whatever was happening inside the machine no longer prevented the comparison from being useful. Biology had not been disproved as the natural home of intelligence. It had simply ceased to be a requirement of the test.

The machinery now existed in principle, but early computers were expensive, scarce and astonishingly limited by modern standards. That changed with remarkable speed. Moore’s Law became shorthand for the relentless growth in computing capacity that the semiconductor industry then spent decades behaving as though it were obliged to fulfill. Computation became relentlessly faster, smaller and cheaper, but some of the hardware that would later matter most had been developed for entirely different reasons; graphics processors designed to render video games would eventually prove surprisingly useful for another kind of calculation.

More computation did not immediately produce more intelligence. Early artificial intelligence largely assumed that if the relevant facts and rules could be stated precisely enough, a machine could apply them. This worked impressively in narrow problems and considerably less well when confronted with the real world. Language, image recognition and other complex problems defied this rules-based approach. A human child could recognize a dog more reliably than a programmer could specify everything that made something a dog.

Machine learning reversed the approach. Instead of specifying every rule, we could give a machine examples and allow it to discover useful regularities for itself. Traditional programming began with rules and produced answers; machine learning began with examples and learned how to produce more answers like them. Neural networks pushed this further because what they learned did not need to take the form of rules intelligible to their creators. The idea was old, but increasing computational power and the availability of vast quantities of data allowed it to be attempted at practical scale.

Large language models were a natural extension of this approach. They were shown enormous quantities of language and learned the patterns governing what was likely to follow what. Predicting the next fragment well enough began producing paragraphs, translations, computer programs, explanations, arguments and answers to questions the systems had never encountered in precisely that form.

Intelligence produced language; language accumulated; intelligence produced machinery; machinery learned from the accumulation of knowledge; and something recognizably intelligent appeared on the other side.

Science fiction imagined this moment repeatedly. The machines awaken. They rebel, enslave us, exterminate us, merge with us or assume the management of civilization after we prove unequal to the task. These stories share the comforting assumption that the decisive event will be an act: artificial intelligence becomes an actor, humanity remains another actor, and one eventually does something consequential to the other. The more unsettling possibility is that no such event is necessary. The important transition may already have occurred simply because intelligence no longer requires a biological mind for every operation we once assumed only a biological mind could perform.

Natural selection produced intelligence in biological organisms; intelligence then spent thousands of years working around the constraints of its biological origins. If biological and machine intelligence continue developing under different conditions, there is no particular reason to expect them to remain alike. Speciation does not require one branch to exterminate the other. It requires only that different conditions begin carrying them in different directions.

AI may therefore be something stranger than the latest product of human ingenuity. It may be an emergent property of intelligence itself. We may simply be the generation to come to terms with the fact that biology no longer has an exclusive claim on it.

Bridgetown, Barbados

Homework:

• Stanislaw Lem, Golem XIV (1981)
• Charles Stross, Accelerando (2005)
• Greg Egan, Diaspora (1997)

A Note on the Writing

This essay argues that intelligence has spent a very long time finding ways to operate beyond the individual biological mind. It therefore seemed slightly perverse to exclude an LLM from the process.

The argument, structure and voice are mine. GPT-5.6 Sol helped with research, challenged weak claims, proposed alternatives and did a considerable amount of redrafting and compression. The useful part was iterative: trying a formulation, discovering what it had accidentally changed, cutting the explanation, restoring the distinction and doing it again. The less useful tendency was equally consistent. Left unsupervised, the model would cheerfully smooth a difficult idea until the difficulty and the idea had disappeared.

The final irony is difficult to avoid. An essay about language becoming an external record of intelligence, and that record becoming material from which another form of intelligence could emerge, was written by passing language back and forth between a biological mind and one of the resulting machines. At some point the process itself became an example of the thesis.