The Miserable Bedspread — On Mistaking Marketing for Science

In One Hundred Years of Solitude, when the gypsies bring a flying carpet to Macondo, José Arcadio Buendía stands unmoved. “Let them dream,” he says. “We’ll do better flying than they are doing, and with more scientific resources than a miserable bedspread.” He is the rationalist in a village of magic—the one man insisting on understanding how things actually work rather than being dazzled by how they appear.

Then he ties himself to a chestnut tree and never recovers.

That irony is worth holding. The danger of AI hype is not only that it is false. It is that it is seductive even—especially—to people who know better. The ability to identify an illusion and the susceptibility to being drawn into it are not opposites. They coexist in the same person, in the same moment, shaped by economic structures and social incentives that make belief rational even when its objects are not.

I. The carpet in new language

The current AI moment is characterized by a particular rebranding. We are not simply impressed by what language models can do—we are calling them intelligent, conscious, general, and imminently superhuman. The technology is real. The category is borrowed.

LLMs are genuine achievements. They compress and recombine human language at a scale no prior system approached. The engineering matters. The mistake is in the translation: a remarkable tool for predicting text has been renamed as reasoning. A system trained to interpolate patterns in human output has been called an agent. A text predictor that produces plausible continuations has been described as thinking.

The vocabulary tells the story of the substitution: “hallucination” instead of “error.” “Reasoning” instead of “pattern completion.” “Alignment” instead of “constraint.” “Intelligence” instead of interpolation. Each word swap imports implications the underlying mathematics does not support. The carpet has been given a new name, and the name is doing the work.

II. The economic structure of conviction

“I know it is not easy because we have to eat.” This is the sentence most industry critique omits, but it is the honest one.

The hype is not purely cynical. Some of the people making extravagant claims believe them, or have learned to believe them because belief is professionally rewarded. When the market for a technology requires a story large enough to justify valuations, and belief is financially incentivized, the boundary between marketing and conviction dissolves. The person performing belief so consistently that it becomes indistinguishable from knowledge is not necessarily dishonest. They are responding rationally to incentive structures.

This sociology of conviction is not new. The dot-com bubble had true believers. The subprime mortgage crisis had analysts who genuinely thought the models worked. Every technology hype cycle generates an ecosystem where people can be simultaneously skeptical about hype in general and fully credulous about the specific hype nearest to them. The confession worth making: even people who can identify the pattern get swept up in it. There is something gravitationally compelling about being near the frontier, about people building things that did not exist five years ago. Jobs, Musk, Gates, Altman—the biographies create a pull. The susceptibility to hype and the ability to identify it can occupy the same mind.

This is not weakness. It is exposure.

III. Alchemy was not fraud

AI is not the first technology to be dressed in the language of magic. Alchemy was not simply con artistry—it was a proto-chemistry that produced real findings (distillation, new acids, novel alloys) while claiming to pursue the impossible (gold from lead). The alchemists who were most productive were often those most sincere about their impossible goal. The genuine discoveries emerged alongside the genuinely mistaken framework.

History shows a pattern. The AI winter of 1973 (the Lighthill Report) killed British AI funding because capability claims had run decades ahead of demonstration. The Lisp machine collapse of 1987 followed the same pattern: the promise was sci-fi; the delivery was a specialized box. Each time, contraction was severe. Each time, recovery required more honest foundations.

The template is recognizable. What cannot be easily predicted: whether we are in a phase where capability claims are still running ahead of what can be demonstrated, or whether the current capabilities genuinely justify the language now being used. Marvin Minsky in 1970: “In from three to eight years we will have a machine with the general intelligence of an average human being.” The quote is famous because it was wrong, but also because it was said by the most serious person in the field. Seriousness and excess are not opposites in technology. The most rigorous minds can generate the most elaborate overstatements.

IV. The doomsday as sales pitch

A shift worth naming: AI companies have adopted the eschatological vocabulary—superintelligence, existential risk, the end of work, the end of humanity—not as a warning but as pitch. The Terminator is being used to sell the product, not to frighten people away from it.

This inversion deserves to be named. The doomsday narrative has been captured by the people with the most direct financial interest in the doomsday narrative being true. Effective altruism, the alignment industry, the “pause AI” discourse—all can coexist with valuations that depend on the threat being believed. The higher the stakes, the more urgent the purchase. The apocalypse is the sales department.

When the same people prophesying existential risk are also raising capital and selling access, the categories collapse. The warning and the offering become indistinguishable. This is not hypocrisy on an individual level. It is structural. The incentive to believe is built into the economic foundation.

V. The map and the tree

José Arcadio Buendía’s tragedy is not that he was irrational. He was, in many ways, the most rational person in Macondo. He had maps and instruments. He could see through the gypsies’ carpet trick. His tragedy was that he could not hold together his capacity for clear analysis and his susceptibility to obsession. He could not distinguish between the map he had drawn and the territory it was supposed to represent. He confused the elegance of his own thinking with proof.

He ended up tied to a tree, muttering in Sanskrit, unable to be reached.

The grounding required is not skepticism as a posture. Postures get adopted and abandoned when incentives shift. The grounding required is the habit of asking, repeatedly and without fatigue: What does this actually demonstrate? Not what does it suggest. Not what could it imply. Not what story does it license. What does the evidence support?

LLMs demonstrate that very large neural networks trained on very large text corpora produce outputs that are useful across a broad range of text tasks. That is a finding worth having. Everything beyond it—sentience, agency, general intelligence, the singularity—is the carpet, draped now in the language of science.

We can do better than a miserable bedspread. We just have to be honest about what we are actually building.

I. The carpet in new language#

II. The economic structure of conviction#

III. Alchemy was not fraud#

IV. The doomsday as sales pitch#

V. The map and the tree#

Further reading#

I. The carpet in new language

II. The economic structure of conviction

III. Alchemy was not fraud

IV. The doomsday as sales pitch

V. The map and the tree

Further reading