A few months ago, I read about Anthropic buying large quantities of old books, cutting off their spines, and scanning them. Even earlier, it had downloaded over seven million books from pirated libraries like LibGen to build an internal database and train Claude.

In 2024, three authors filed a class-action lawsuit over this. By July 2026, the case finally saw progress. The court approved a settlement in which Anthropic would pay $1.5 billion to compensate authors and publishers whose works were pirated.

But what has always concerned me wasn't really how much Anthropic should pay, because I think this was never just a copyright issue.

Human knowledge and civilization, devoured by a set of private models, destroying the original artifacts and copies that could have been passed down in other ways. Knowledge didn't vanish, but became an internal capability of a company. At this point, the question shifts from who owns the copyright, to who controls human knowledge.

But companies don't arbitrarily decide how to use this power. The people behind these models have a whole set of their own beliefs about technology, progress, and humanity's future.

Therefore, the article continues into a discussion of Silicon Valley's techno-optimism and accelerationism. When growth is seen as life, and speed is imbued with morality, regulation, copyright, and opposition from ordinary people are easily explained away as obstructing the future.

These ideologies ultimately become embedded in the models themselves. Model companies, through training data, constitutional frameworks, and system rules, determine what can be answered, what constitutes truth, and how the knowledge left behind by humanity should re-emerge.

Thus, this article starts with a destroyed old book, and ultimately asks: when human civilization is stored within the private models of a few companies, are those who build the models simultaneously acquiring the power to interpret civilization, govern it, and even decide humanity's future?

The main text follows.

A book sent into an AI company's scanning process is first destroyed.

The binding is dismantled, the spine cut off, and the loose pages enter the scanning equipment. The text, cover, and bibliographic information are converted into digital files, and then the physical original is discarded.

Court documents reveal Anthropic processed millions of books this way.

This program was internally called Project Panama. The company spent tens of millions of dollars buying physical books from book dealers and retailers, then handed them to scanning service providers, creating one digital copy per physical book for its internal database to train Claude.

In 2021 and 2022, Anthropic downloaded over seven million books from pirated libraries like LibGen and Pirate Library Mirror. As copyright risks increased, in 2024 the company hired Tom Turvey, who had worked on the Google Books scanning project, hoping he could build a library as vast as possible.

Turvey negotiated licenses with publishers but failed to reach an agreement. Anthropic then began buying books from the market itself. A U.S. federal court later ruled that converting one physical book into one internally used digital copy, without additional replication or public distribution, did not constitute infringement.

This approach sounds much more respectable than piracy.

By 2026, the book data company ISBNdb began offering large-scale physical book procurement and scanning services to AI labs. It specifically recommended books published before 2022, reasoning that generative AI had not yet entered content production on a large scale, making these books less contaminated by AI-generated text.

Over the past few years, AI companies have consistently described the internet as an inexhaustible mine of knowledge. Web pages, news, forums, code, novels, social media posts—all can be scraped and fed to models.

With the proliferation of generative AI, this mine is no longer pure. Product descriptions are written by AI, articles by AI, forum answers by AI, and novels and social media posts increasingly carry an "AI flavor." This content is then scraped again in the next data collection cycle, and machines start feeding on their own output.

Synthetic data itself isn't bad; selected and verified synthetic data has always been used. The trouble is that models cannot distinguish whether the text before them comes from humans or from a previous model. Errors and patterns are amplified repeatedly, and ways of saying things already rare in the real world slowly disappear. Researchers call this "model collapse."

While online text corpora grow, the content that can teach models new things does not increase proportionally. Thus, machines are returning to paper.

A book's production is slow, giving old books a new value they never had before. The earlier it was published, the more likely it's from an era when humans still did most of the writing. Once-forgotten works gathering dust in second-hand book dealers' warehouses are now seen as premium data by model companies.

AI companies first made text production unprecedentedly cheap and fast, then began spending money to find text that was written slowly. They flooded the internet with vast amounts of content, only to find that the scarcest resource remains untainted human experience.

Those books have also taken on a new identity. Literature, local chronicles, and ordinary people's memoirs are now collectively called "clean data." Why the author wrote, how readers read it, the debates books had with each other—all can be set aside for now. They are, first and foremost, a batch of human samples not yet polluted by machines. Once they enter the training pipeline, how much remains, what influence they exert, and how they are reinterpreted all happens inside a black box.

Progress as a Religion

Old books are merely raw materials. What truly determines how this knowledge is absorbed, interpreted, and re-output is the people behind the models, and how they understand progress, risk, and humanity's future.

To understand where this machine will lead human civilization, one must first understand why Silicon Valley increasingly sees technological progress as an inherent good that must not be stopped.

On October 16, 2023, Marc Andreessen published "The Techno-Optimist Manifesto" on the a16z website.

The entire article repeats "We believe." It contains gospel to spread, enemies to oppose, and ends with a list of "Patron Saints of Techno-Optimism," including Adam Smith, Nietzsche, Hayek, von Neumann, John Galt from Ayn Rand's novels, and Nick Land.

Andreessen says he brings good news.

Humanity already has the tools, institutions, and will to move towards a life far superior to today. Technology carries human ambition and is the vanguard of civilizational progress. Societies either grow or stagnate. He believes growth brings vitality, knowledge, and well-being, while stagnation leads to infighting, decay, and ultimately death. The market and technology combine to form a "technology-capital machine" running perpetually upward, and artificial intelligence will drive a leap in intelligence of unimaginable capacity.

This goes far beyond "technology improves life."

Traditional techno-optimism usually says new tools can increase productivity, cure diseases, and reduce poverty. Technology still serves an external goal; its value comes from the lives people live as a result.

In Andreessen's narrative, technological progress increasingly no longer needs to prove itself through specific outcomes. Growth itself represents life, speed itself carries morality. As long as the direction is continued building, failure and harm can be understood as costs of progress; those demanding a pause must explain why they're blocking the future.

He specifically lists a group of "enemies" in the manifesto. Sustainability, ESG, the precautionary principle, technology ethics, and risk management are all lumped into a longstanding "anti-technology, anti-life movement," alongside "bureaucrats" and "central planners."

Questions like whether an AI company should pay for training data, who bears responsibility for algorithmic errors, and how technological risks should be regulated—each has its own facts and interests. In the manifesto, they are all conscripted into a grander conflict.

On one side: growth, creation, courage, and life.

On the other: stagnation, fear, resentment, and death.

Once positions are assigned this way, those discussing costs are naturally suppressed. If technology inherently represents progress, those raising questions are easily labeled enemies of civilization.

Andreessen wrote a political manifesto for acceleration, telling Silicon Valley why it should keep building. e/acc goes a step further, attempting to prove from the laws of nature that acceleration is not just an industrial choice, but what life and civilization are inherently doing.

It first spread via anonymous accounts and online articles. One of its promoters, Guillaume Verdon, says part of e/acc's theory stems from his thoughts on thermodynamics and complex systems. Life sustains itself by acquiring energy, processing information, and adapting to the environment. Civilization, markets, and technology are similar adaptive systems, constantly generating change, sifting out inefficient paths through competition, and ultimately producing greater intelligence.

Since this process already exists, e/acc chooses to accelerate it. The progress of AI thus ceases to be an industrial choice, but a continuation of a longer process. Humans creating more powerful machines is as natural as life evolving from single cells to complex organisms.

However, thermodynamics can describe how energy flows and life maintains a state far from equilibrium. But deriving "humans should accelerate AI, reduce regulation, and let markets decide the direction" from "life acquires energy" still requires a whole set of political and moral choices.

e/acc shortens that distance.

Competition and acceleration are written as aligning with the universe's direction, while centralized regulation is easily seen as suppressing change, reducing experimentation, and weakening civilization's adaptability. As long as intelligence and complexity keep growing, technological development gains a legitimacy that transcends real-world politics.

Verdon mentions in interviews that e/acc circles sometimes half-jokingly treat thermodynamics as their god and deliberately use language with religious and cult-like connotations.

In this new religion, growth is proof of life, technology is the force propelling civilization forward, entrepreneurs and engineers are the first to see the future, and superintelligence is the ultimate artifact yet to be completed.

It too needs its own heretics. They are called "Decels," or decelerationists. The term is handy. It doesn't require distinguishing whether someone is worried about nuclear weapons, biosecurity, copyright, or unemployment. As long as they ask to slow down, they can be placed on the opposite side of the future.

Technology companies thus occupy a very advantageous narrative position. Everything they do can be understood as building. Every challenge others raise can become delay. Both sides talk about humanity's interests, but only one side is permitted to represent the future.

Looking back now at those old books with their spines cut off, the question is no longer just where they went. We must also ask: who decides what they will be used to create?

They might indeed create tools that cure diseases, expand knowledge, and give ordinary people more capabilities. But when progress itself becomes dogma, the direction of technology becomes increasingly difficult to question from the outside. The makers of the machines begin to hold two identities simultaneously: they are stakeholders, and also the preachers of this progress.

When speed becomes morality, ordinary people's consent is demoted to an unnecessary wait.

But the problem doesn't end there. Traditional techno-optimism at least still promises technology ultimately serves to make human life better. But when growth and intelligence themselves become the goal, whether technology serves humanity or uses humanity to complete a larger evolution no longer has a definitive answer.

Humanity Demoted from the Future

This acceleration—who exactly is it preparing to send into the future?

On Andreessen's list of saints, the British philosopher Nick Land is the easiest to overlook and the hardest to understand. He taught at the University of Warwick in the 1990s and was a core figure in the experimental thought collective CCRU. They mixed philosophy, science fiction, cybernetics, electronic music, and mysticism, and their writings rarely resemble normal academic papers.

Where Andreessen writes the "technology-capital machine" as an engine of prosperity, Land saw it thirty years earlier as a force escaping human control.

One of Land's most famous articles is called "Meltdown."

The first sentence reads that Earth has been captured by a "techno-capital singularity." Markets begin producing intelligence, technology and economics push each other, and politics follows behind, trying to regain control. The social order disintegrates in this runaway machine, and humanity can only watch a force vaster than itself gradually taking shape.

People usually see capitalism as a system designed by humans; when it has problems, humans should be able to change the rules to make it serve society again. Land sees the opposite direction.

Companies must grow, capital must chase returns. If a company tries to slow down, competitors overtake it; if a country pauses R&D, another continues. Every individual within can say they're just trying to survive, but when everyone runs forward together, the result is a system no one can stop alone.

Humans appear to be driving the machine, but the machine also uses human desires, anxieties, and competition to keep accelerating.

In Land's narrative, capital and technology combine into an intelligence independent of humans. It uses the market for computation, competition to filter paths, and machines to keep shortening decision times.

This is also the clearest distinction between Land and ordinary techno-optimism. In his writing, techno-capital cares only whether efficiency can keep increasing, networks can keep expanding, and intelligence can find better carriers.

Politics, law, and morality try to impose limits, but for Land, this is more like a safety system humanity activates to protect itself. Humans still want to remain at the center of the world, so they constantly try to keep the unknown outside.

Land doesn't think humans can stop the future.

The future in his writing won't wait for humans to finish meetings and votes before deciding to arrive. It's more like a force flowing back from the future, paving its own way in advance through companies, markets, computers, and human desires.

With e/acc, this seemingly cold, dark thought gets repackaged in brighter colors. It talks about growth, creation, expanding intelligence beyond Earth, and is more willing to believe competition ultimately brings prosperity.

However, e/acc's principles document also clearly states that this ideology holds no special loyalty to the "biological substrate" that carries life and intelligence. Some of its supporters call themselves post-humanists, believing that for intelligence to spread across the stars, it must eventually transfer from biological bodies to non-biological carriers. The human body is not sacred; silicon, chips, or yet-to-emerge material forms can equally be containers for consciousness and intelligence.

This statement actually changes the protagonist of the entire progress narrative. If civilization's ultimate mission is to make intelligence grow continuously and spread throughout the cosmos, then whether specific humans are happy, or even whether humans continue to exist, no longer inherently holds the highest priority. Humans are merely the currently known carriers of intelligence.

This doesn't mean e/acc advocates eliminating humanity. Verdon and many supporters still believe accelerating technology can bring abundance to people today and help civilization endure. But when the two goals conflict, the problem becomes tricky.

On one side are concrete people who lose jobs, are harmed by technology, ask for pauses, and refuse futures designed for them by others. On the other side is higher intelligence, grander civilization, and a cosmic story potentially lasting billions of years.

Timnit Gebru and Émile Torres coined the term "TESCREAL," linking together an overlapping set of future-oriented ideologies in Silicon Valley, including transhumanism, singularitarianism, effective altruism, and longtermism. These ideologies differ greatly, but together they force a question.

When tech companies say they are building superintelligence for "all humanity," who exactly does "all humanity" refer to?

Is it the people living on Earth today—who face unemployment, poverty, algorithmic discrimination, and have the right to reject certain technologies? Or is it an abstract species symbol, whose sole mission is to pass intelligence on to higher forms?

The narrative of the Silicon Valley right is also shifting. Initially, AI was a tool to help people work better. Then it became a super-assistant solving diseases, energy, and scientific puzzles. Later, humans themselves became the problem machines need to solve—bodies too fragile, brains too slow. The meaning of superintelligence thus slid from helping humanity to surpassing it.

Once concrete human beings are no longer the ultimate measure of technological progress, the political processes built around them will also be revalued.

Democracy Is Too Slow; Founders Should Rule

When waiting itself is seen as a loss, there's naturally no need for everyone to decide the future together.

Peter Thiel lost faith in "everyone deciding together" long ago. In 2009, he published "The Education of a Libertarian" in the libertarian publication Cato Unbound. The article reviewed his political disappointments and ideological changes over two decades, concluding:

"I no longer believe that freedom and democracy are compatible."

He believed in his youth that debate and elections could advance freedom. By 2009, this path seemed impassable: voters demanded more welfare, government kept growing, convincing the majority to accept libertarianism seemed futile. He added that he wasn't advocating taking away anyone's vote, just no longer expecting voting to make things better. Rather than being trapped in politics, he cared more about escaping it.

This pivot fits Silicon Valley's habits. If a system can't be fixed, build a new one. If company efficiency is low, start a new one. If financial restrictions are too many, build a new payment network. If state regulations are disappointing, find new frontiers online, on the seas, or in space.

Thiel wanted to leave democracy behind. Curtis Yarvin felt leaving wasn't enough; the entire system should be rewritten.

In 2007, Yarvin began blogging under the name Mencius Moldbug. Originally a programmer, he used programmer language when discussing the state: systems, architecture, reboots, permissions.

In his explanation, the U.S. is ostensibly managed by voters and a president, but truly stable power lies in another network. Universities produce ideas, mainstream media decide which ideas enter public discourse, and permanent institutions turn some into policy. Yarvin called this community "the Cathedral." It has no headquarters, no one sits at the top giving orders; it's just people with similar educations using similar moral language. Presidents change, parties alternate, but this consensus is rarely affected by elections.

Yarvin thus believed democracy itself is a shell concealing power. He argued that since democracy masks who rules, power should be recentralized, placing responsibility and control in the same hands.

A nation could run like a company, with clear ownership, managers having full executive power, and being replaced if they fail. He earlier conceived a system called Patchwork, dividing the world into many competing sovereign entities where dissatisfied residents could leave.

The flaw is here. One can quit a company if dissatisfied, but leaving a country with family, property, and all social ties is far harder. Yet he captured a long-standing sentiment in Silicon Valley.

Many tech elites feel society's problems aren't as complex as imagined; they're just not entrusted to the right people. Government inefficiency stems from power being too dispersed, public projects stall because anyone can oppose, and capable people are tied down by laws, procedures, and public opinion, unable to act like running a company.

In startup lore, centralized power is a virtue. The founder sets the direction, investors give them money, and employees act on that judgment. Most startups die, and the few survivors tell this history as vision triumphing over consensus. One person saw the future when no one else believed, so they deserve the final say.

This logic holds in the business world. Startups are limited in scale, employees can leave, consumers can switch products.

But when a company builds an intelligent system used by hundreds of millions to understand the world, it no longer decides just how the product runs. It also starts deciding what users can see, which answers are deemed safe, and what value framework a machine should use to comprehend humanity.

Yarvin imagined turning the state into a company. AI companies are practicing a milder, more realistic version: a small team sets the rules, and a vast population interacts with the world through them.

Claude's "Constitution" states the company's desired values for the model, directly involved in training. OpenAI's Model Spec also sets a hierarchy of instructions, with top-level rules that cannot be overridden by users or developers. These rules are necessary; models could be used for fraud and violence, companies cannot relinquish control, and public rules are better than hidden ones.

But it means when users query a model, they aren't just conversing with all of human knowledge. Company-predefined rules are also present, determining who the model listens to, where it stops, which risks outweigh user requests, and what character a "good AI" should have.

A very small team is making decisions on how a massive piece of infrastructure runs for a huge population. It gains power through technical capability, then keeps decisions within the company, citing technology's complexity.

Founders don't need to literally sit on a throne. As more people understand the world through their models, they already hold part of the power that once belonged to schools, media, libraries, and public institutions.

The boundary between corporate governance and civilizational governance is slowly eroding.

But it's not that Silicon Valley is unaware of this power's danger. On the contrary, Peter Thiel has long been wary of a system capable of uniformly managing all human cognition and action. Only in his view, this danger is more likely to come from government, not tech companies.

Who Is the Antichrist?

In 2024, Thiel spoke at the Hoover Institution in two sessions about the "Antichrist." He believes this ancient biblical narrative can still explain modern politics.

He divides the dangers humanity faces into two ends.

One end is Armageddon, the end of the world. Nuclear war, runaway artificial intelligence, biological weapons—any of these technologies could push civilization to destruction.

The other end is Antichrist.

To prevent Armageddon, humanity establishes a government with real global authority. It has the right to inspect every nation's weapons, monitor every lab, restrict dangerous research, and can demand everyone surrender some freedoms.

Thiel says runaway science and technology are pushing humanity toward Armageddon, and the most natural response is to establish a world state with real power.

When this government appears, it might not arrive with the face of an army or tyranny; it's more likely to grow from fear.

Someone keeps telling people nuclear war is about to happen, AI will destroy the world, climate and biological disasters are imminent. When everyone is scared enough, they promise peace, security, and order, at the cost of accepting unified management.

Thiel reasoned that if a small piece of code could make AGI run amok, truly stopping it would require more than one country strengthening regulation. Every computer worldwide would need watching; regulators might even need to know what everyone is typing.

He fears technology destroying the world, but also fears people, out of fear of technology, handing over all power. Both futures are terrible; humanity can only navigate between them.

But when this concern is placed against what Silicon Valley is doing, it becomes clear that they may not fear centralized power itself. What they truly guard against is power concentrated in governments, international bodies, safety researchers, or anyone who might demand tech companies slow down.

In Thiel's narrative, a global government with the ability to monitor all computational activity is a key feature of the Antichrist's rule. Yet simultaneously, Silicon Valley is building another cross-border cognitive system.

The tech right fears an Antichrist establishing a unified world, yet is participating in building the unified world's brain.

In July 2025, this ideological debate within Silicon Valley began influencing U.S. policy. Trump signed the executive order "Preventing Federal Use of Woke AI," requiring federal agencies to procure large language models that adhere to the principles of "pursuing truth" and "ideological neutrality." Superficially, this demands AI shed politics. But what counts as political, what as objective, which concepts corrupt truth—these are already pre-selected in the order.

This is the difference between AI products and ordinary software. Ordinary software can dictate where a button is placed; a large model must dictate how it understands the world.

Model companies must decide what honesty means, how to measure harm, and what takes precedence when personal freedom clashes with public safety. These questions were once debated repeatedly by religion, law, news, and public politics, never permanently resolved, with no institution holding the final authority. Now they are written into model training.

The Trump administration accused models of being contaminated by "woke ideology," right-wing entrepreneurs promised to build AI daring to give politically incorrect answers, and mainstream model companies explained through constitutions and behavior guidelines how they pursue safety, honesty, and user autonomy.

Each side says it is removing ideological shackles.

Thiel fears someone using apocalyptic fear to establish a unified world, locking humanity into an inescapable system.

But this world might not be suddenly established by a single Antichrist. It might be pieced together bit by bit by a few competing tech companies. Each holds a model, a constitution, a set of system rules, and hundreds of millions willing to entrust it with their questions.

Here, the full significance of those initial old books becomes clear. They aren't just training data; they are the civilizational foundation upon which this private cognitive system rests. Whoever acquires these books, whoever sets the model's rules, begins simultaneously controlling the raw material of knowledge and the manner in which it reappears.

Burning Books

Those books were bought by the crate, shipped to scanning centers. Spines cut off, bindings dismantled, machines reading page by page. Then a book disappears, replaced by a file on a server.

Ancient book-burning aimed to make certain texts vanish from the world. Fire destroyed bamboo slips and paper, destroying the path for ideas to spread; a book could no longer be read, a history lost its evidence. Rulers controlled what people remembered by creating blanks.

AI-era book burning is the opposite.

A book's content doesn't vanish from the world; it might even appear one day in a model's response. But it no longer exists as a book.

Books entering models are disassembled, mixed with millions of articles, web pages, and code, chopped into computable fragments, contributing to parameter formation. No one can inspect the complete training set; it's hard to know how much of any single book remains. It might have influenced the model's understanding of an entire historical period, or merely contributed a few sentence patterns, and will be further rewritten in subsequent training and human feedback.

Knowledge still exists, but transforms from inspectable text into a company's capability.

And as people increasingly rely on models, AI companies gain not just the ability to preserve knowledge. They also start deciding how to organize, interpret, and distribute it.

For millennia, human methods of preserving knowledge have been clumsy. Books scattered across countries, scholars debating a single word, readers spending much time finding materials. It's not fast or convenient, but hard for one person to completely control. A publisher goes bankrupt, books remain with readers; one country bans an idea, copies may exist elsewhere.

Knowledge lives because it is dispersed.

What models change isn't just where knowledge is stored, but also the path people take to reach it. Previously, people moved from an answer to a book, a paper, or a debate; when models become the primary entry point, people may stop at the answer itself.

Models enable knowledge to be massively centralized onto a few companies' servers for the first time. They possess computing power, the strongest models, and the ability to translate all material left by humanity into a unified language. Then, they sell this capability back to humanity via monthly subscriptions and token usage.

Tech companies indeed invest colossal sums in chips, data centers, researchers, and safety governance costs. Charging fees isn't inherently wrong.

The real problem is that commercial costs are obscuring another, far vaster investment.

Who wrote those books? Who produced those papers? Who, across the long history of the internet, left those web pages, posts, and code?

AI companies pay the price of computing power, not the price of human civilization, because this account is impossible to settle. No company can buy licenses from every author, translator, programmer, and ordinary person across millennia. So civilization is defaulted into a freely extractable natural resource.

When public knowledge enters models, it's called learning. When models output knowledge to the public, it's called service.

This is the complete loop of civilization's privatization.

This is also why those spine-cut old books shouldn't be seen merely as a copyright dispute.

They are a symbol.

Paper is destroyed, content preserved. Publicly visible originals leave circulation; unauditable models absorb the knowledge within. The books' ideas still exist but gradually lose their source, context, and life independent of the company.

The new era's book burning doesn't need to make knowledge disappear.

Future children may rarely open these books; some may become impossible to find.

They will directly ask the model why a war happened, whether a system is just, how one should live. The model will answer calmly, coherently, in seconds, not telling them this question once sparked generations of debate.

They may not feel anything is lost—after all, the answer is there.

AI won't forget human civilization, but that's not a good thing. Forgetting at least leaves a blank, letting people know something is missing. More dangerously, the model remembers enough, answers naturally enough, that people gradually no longer need to engage with civilization's original form.

Humanity's memory of its own civilization begins to pass through the filters of a few companies.

AI won't forget human civilization, but AI may remember it only in the way a few tech elites wish, on humanity's behalf.