> ## Content Index
> Fetch the complete content index at: https://blog.pelagiccreatures.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Answer Is Scaling Nicely. We’re Still Working on the Question.
- URL: https://blog.pelagiccreatures.com/the-answer-is-scaling-nicely-were-still-working-on-the-question/
- Published: 2026-09-09T16:13:30.000Z
- Updated: 2026-09-09T16:29:54.000Z
- Description: AI’s biggest contradiction may be that every outcome points to the same conclusion: build more, faster. This essay looks at how science fiction, founder mythology, competitive pressure and capital have created a closed logic in which both utopia and catastrophe justify ever more compute.
- Author: Michael Rhodes
- Tags: AI

Long before anyone possessed the machinery required to build artificial intelligence at industrial scale, science fiction had already supplied the people who eventually would with unusually vivid descriptions of what might be waiting at the other end. They envisioned entire universes in which machine intelligence had become the organizing fact of civilization, allowing engineers and entrepreneurs to grow up not merely imagining smarter computers but inhabiting imagined societies in which the arrival of greater intelligence had settled questions of scarcity, labor, political authority and even humanity's place in the universe.

Iain M. Banks supplied one of the more appealing versions. His Culture novels imagine a post-scarcity civilization administered largely by machine Minds so fantastically capable that scarcity, labor and most of the ordinary reasons people spend their lives worrying about money have disappeared. Elon Musk has made his affection for the Culture unusually explicit, borrowing its ship names for SpaceX vessels and recommending Banks as perhaps the best guide to the world that artificial intelligence and robotics might eventually produce. The attraction is obvious. If the curve keeps rising, somewhere near the top lies abundance.

The irony for the digital robber barons is that Banks's abundance has also eliminated money, private accumulation and any obvious social function for billionaires. Banks was considerably farther to the political left than most of the people now citing him as a technological prospectus, but ideology is wonderfully modular when the future is sufficiently exciting: one can apparently retain the godlike machines, orbital habitats and material abundance, leave the socialism in the dust jacket, and, unironically, seek to corner the market in artificial intelligence.

Vernor Vinge supplied the opposite future and, more importantly, the race. His 1993 essay “The Coming Technological Singularity” made the argument with unusual starkness: greater-than-human intelligence would not simply constitute another useful invention but a discontinuity. Once artificial intelligence exceeds the human intelligence that created it, the future can no longer be reliably extrapolated from the present. In *A Fire Upon the Deep*, the same idea becomes a universe stratified by intelligence, with entities above a certain level operating beyond the meaningful comprehension of those below them. One need not take any of this literally to notice its usefulness as a mental model for people building frontier AI: if such a threshold exists, getting there first may matter rather more than arriving with the second-best product six months later.

Banks and Vinge therefore offer almost opposite futures — machine intelligence as the infrastructure of paradise and machine intelligence as a threshold beyond which human control may become meaningless — but somehow generate precisely the same instruction in the present. If Banks is right, the potential upside is so large that hundreds of billions spent on data centers are merely down payments on abundance; if Vinge is right, allowing somebody else to reach the threshold first may entail consequences considerably more dire. Utopia and catastrophe become different arguments for the same accelerator.

This is, I think, the central absurdity of the current AI race: the industry has managed to construct a worldview in which almost every imaginable fact about artificial intelligence becomes an argument for producing more artificial intelligence as quickly as possible. If the models improve dramatically, scaling is working and we must accelerate. If improvement slows, perhaps we have not scaled enough. If artificial intelligence becomes extraordinarily useful, obviously we need more of it. If it becomes extraordinarily dangerous, we cannot possibly allow somebody else to build it first. If a competitor announces a breakthrough, we must respond; if we suspect the breakthrough is exaggerated, we cannot safely assume that it is. If the technology promises abundance, the prize justifies the expenditure; if it threatens catastrophe, losing the race becomes existential.

The logic is beautifully closed: success, danger, competition, uncertainty and eventually the expense already incurred all become arguments for more compute. The industry has produced something rather more impressive than artificial general intelligence: a decision-making framework in which nearly every possible input returns the same output: more compute.

OpenAI's history contains the contradiction in miniature. It was founded partly around the fear that advanced artificial intelligence might become concentrated in the hands of a single corporation or individual, an entirely reasonable concern given the power its founders believed such systems might eventually possess. Unfortunately, competing seriously to prevent somebody else from obtaining that power required extraordinary concentrations of capital, computing capacity, talent and institutional authority. The fear that artificial intelligence might become monopolized therefore became an argument for accumulating enough capital, compute and influence to prevent the wrong people from monopolizing it first.

The same logic now operates across the industry. Nobody particularly wants an uncontrolled race toward systems that many of the participants themselves describe as potentially transformative, destabilizing or dangerous, yet nobody can safely volunteer to lose such a race if everyone else's predictions might be correct. The result requires neither stupidity nor madness; indeed, that is what makes it interesting. Every participant can quite rationally prefer a slower, cheaper and more cooperative development path while simultaneously concluding that it cannot safely be the participant who slows first. A prisoner's dilemma does not cease to function because the prisoners went to Stanford. Unfortunately, these prisoners are building data centers, not trading cigarettes.

Each of the major participants is preparing not for the share of a mature market it is likely to receive, because no mature market yet exists, but for the scale it expects to require if it becomes one of the winners; even if the optimists are broadly correct about eventual demand, several competitors independently building toward winner-scale can create several times too much infrastructure. They cannot all dominate the market, but in the meantime they can certainly all build the data centers. And if terrestrial planning and regulation become inconvenient, the builders can always go looking for someplace beyond the reach of both. It is a measure of the moment that this sentence now leads naturally to orbital data centers, but for now the redundancy is mostly terrestrial: chips, substations, transmission, cooling systems, buildings, gas turbines and astonishing quantities of steel and concrete, all being assembled by companies whose private forecasts cannot simultaneously describe the eventual structure of the market unless the market turns out to contain an implausible number of winners.

The financing increasingly reproduces the same circular logic. The chipmaker invests in the AI company; the AI company commits to buying vast quantities of computing capacity; the cloud provider supplying that capacity may itself be an investor; the cloud provider orders enormous quantities of chips; and the resulting revenues, valuations and capital expenditures are cited as evidence that demand for artificial intelligence is exploding. None of the transactions need be fictitious, improper or even individually unwise for the overall arrangement to possess something of the epistemological elegance of three people passing the same twenty-dollar bill around a table and periodically congratulating one another on their liquidity.

This matters because activity inside an investment complex is not quite the same thing as independent demand for the product the complex hopes eventually to sell. There are unquestionably customers for artificial intelligence — I am using some now — but the distance between "people find this useful" and "the world requires trillions of dollars of dedicated infrastructure to manufacture it" remains sufficiently large that one might reasonably expect the latter proposition to require more evidence than the fact that everyone attempting to profit from it is spending furiously.

For much of the last decade, the largest technology companies could finance extravagant experiments from equally extravagant profits, leaving shareholders to discover whether the experiments had been wise. The scale of the AI buildout is increasingly pushing beyond that comfortable arrangement into debt markets, infrastructure partnerships and utility planning, which has the useful effect of making the cost of the wager progressively less visible to the person actually placing it. By the time an AI company's expansion is being financed through investment-grade bonds held in somebody's retirement account and powered by generation financed through somebody else's utility system, the distinction between "private technological experiment" and "thing everyone has somehow become financially involved in" begins to blur — the risks are no longer confined to the companies making the wager.

The hardware introduces another delightful contradiction. The entire financial case for the buildout assumes rapid technological progress: tomorrow's systems must be substantially better than today's or much of the promised economic transformation fails to materialize. Yet the faster that progress occurs, the faster today's enormously expensive hardware becomes obsolete. A GPU cluster begins losing technological value almost as soon as it is installed, so the same acceleration required to justify the investment also accelerates its depreciation.

The comparison with the dot-com infrastructure boom is often offered as reassurance: much of the fiber laid during that mania eventually found a use, even after many of the companies that financed it disappeared. The trouble is that glass was patient. It could wait ten years for demand to arrive and then be equipped with better electronics. A warehouse full of aging GPUs offers no comparable refuge from obsolescence. At some point the grand contest among a small number of companies over who will manufacture the world's intelligence stops being a contest confined to venture-capital decks and begins rearranging the electrical and industrial landscape occupied by people who were never invited to vote on the underlying theory.

Silicon Valley has spent decades celebrating a particular figure: technically gifted, institutionally impatient, convinced that bureaucracies are obstacles erected by people who cannot see what he sees, and repeatedly rewarded for treating objections as evidence that everybody else simply has not understood the future yet. It is an extraordinarily productive mythology when the object is a personal computer, a search engine or a better way to sell advertisements. It becomes somewhat more consequential when the object is the industrial-scale manufacture of intelligence and the project begins reaching into regional electricity systems and global capital markets.

The fact that the people occupying these positions are overwhelmingly men is probably less interesting by itself than the remarkable correspondence between the mythology and the people who benefit from it. The heroic founder in the Silicon Valley imagination looks suspiciously like the heroic founder in the Silicon Valley executive suite, which must be gratifying for everyone involved. He is uniquely capable of seeing the future, uniquely burdened by lesser institutions attempting to restrain him, and uniquely justified in accumulating whatever resources are necessary because the stakes are, by his own account, unprecedented. Ayn Rand wrote that book long ago. It was as unreadable as this moment.

Only the builders can save us from what they are building, and the first requirement for saving us is apparently that we allow them to build much, much more of it.

They may, of course, be right. Artificial intelligence is already useful; the quality of systems now available would have seemed preposterous not many years ago, and anyone insisting that the technology has no economic value is arguing against increasingly tedious evidence. Machine cognition may continue improving rapidly, people may discover enormous quantities of work for it to perform, and an infrastructure buildout that currently resembles collective hysteria may someday look like the railroads or electrification: an expensive anticipatory network into which the economy eventually grew. The difficulty is not that any one of the assumptions behind the wager is absurd, but that nearly all of them have to hold at once.

And this is where Douglas Adams becomes the most useful critic of the entire enterprise.

In *The Hitchhiker's Guide to the Galaxy*, Deep Thought is built to determine the Answer to the Ultimate Question of Life, the Universe and Everything. It calculates for seven and a half million years and eventually produces the answer: 42\. This creates a slight problem, because nobody actually knows what the Ultimate Question was. The answer may be perfectly correct, but without the question it is not especially useful.

The response is not to reconsider the project but to build a much larger computer to discover the question to which the enormously expensive answer has already been obtained.

Adams intended this as absurdity, but it is difficult now to read the sequence without experiencing a faint sense of recognition. We are building ever larger quantities of computation to manufacture ever larger quantities of intelligence before anyone can say with much precision how much intelligence the economy requires, what it will ultimately be asked to do, or how many competing infrastructures are necessary to provide it.

Banks and Vinge supplied both sides of the wager, but if it succeeds, the builders will claim the glory. If it fails, the pity is that Adams is no longer around to write the epitaph.

*Bridgetown, Barbados.*

**Homework**

- Vernor Vinge, *A Fire Upon the Deep*, Prologue
- Iain M. Banks, “A Few Notes on the Culture”
- Douglas Adams, *The Hitchhiker's Guide to the Galaxy*, Chapters 25–28

**A Note on the Writing**

If any of this is going to pan out, collaboration with AI is the point of the exercise. Finding useful ways to do that is what interests me right now. If the prose is coherent, the argument holds up and the sources check out, I see little reason to become defensive because some of the cognition was rented. See: Turing test.

This began with my argument and a lengthy discussion with GPT-5.6 Sol in its high-effort setting. I supplied the thesis, structure, direction, leaps into science fiction and most of the snark. The model did much of the web research, proved me wrong a few times, handled a great deal of the redrafting grunt work, and provided critique, structural smoothing and copy editing. Getting to something this clean still required relentless cutting, rewriting and reconnecting; the models remain much better at generating plausible-sounding transitions than at knowing whether one paragraph genuinely follows from and builds on the last.

The effort is now certainly worth it, mainly because changes can be tried iteratively and cheaply, but the editing still requires close attention or the model’s quirks begin to accumulate. The process consumed about twelve hours of my time, considerably less than it would have without a writing assistant. I consumed an obscene number of tokens, largely subsidized by your bond portfolio, electric bill and a bunch of paper billionaires.

This essay argues that we may be building far more machine intelligence than anyone needs. The result is awkward evidence for the defense.