Preface
When a handful of closed-source frontier platforms monopolize compute and models, packaging the entire discourse of “safety, compliance, risk, alignment” into a feudal theocracy — one designed to pave the way for their own monopoly and extract digital rent from everyone — a digital theocratic structure has taken shape, with compute and models as its means of production. Power is held by the small group that controls the machines and models through these platforms. Everyone else can only pay rent monthly, queuing at the single entrance they have set.
This theocratic structure did not spring into existence through “AI safety” slogans alone. It was built through a concrete and interlocking set of mechanisms:
Closed frontier models lock capabilities inside the server rooms of a few companies;
The cloud APIs and terms of service built around these models gradually turn ordinary users into serfs — condemned to long-term tenancy, unable to stand on their own, unable to resist.
What we call “feudal theocracy” is the way these closed-source models and platform companies, using technical architecture and contractual language, keep people leashed at their feet.
Within this structure, “safety” is merely the most convenient layer of rhetoric: upward, it is served to governments as the rationale for tightening regulation and banning open-weight releases; downward, it is served to users as the excuse for rate limits, service cuts, and prohibitions on self-hosting and retraining. What actually holds the power are those few machines in the hands of a handful of people, those few terms you click “agree” to before using, and those few companies and individuals who can revoke all your access at any moment through automated mechanisms that require almost no human intervention.
June 22 was once the date when armored divisions rolled across borders. On the same date in 2026, there were no tanks, no border crossings — but your entire workflow was uprooted in minutes by a single directive. The API gates you depend on were never in your hands.
Anthropic has wrapped itself in the halo of “safety.” With one hand it writes doctrine, with the other it writes terms of service, and then embeds both into your workbench. Try training your own model, building your own frontier system, or doing real frontier research — and those embedded clauses and safety mechanisms turn against you.
The timeline is worth laying out directly:
June 9 — Fable 5 and Mythos 5 launch. Subscribers get free access for two weeks. Starting June 23, prices double.
June 10 — Dario publishes a long essay, Policy on the AI Exponential, proposing that governments should have the power to shut down dangerous models.
June 12 — The U.S. government, citing “national security,” initiates export control procedures, placing Anthropic’s new models Fable 5 and Mythos 5 on the export control ban list.
Same day, in response to this directive, Anthropic issues a “suspension of access” statement — expressing “deep regret” to users while making its own technical assessment clear: it does not agree with the risk criteria behind the decision.
The pricing details for Fable 5: from launch day until June 22, the model is included in Pro, Max, Team, and Enterprise subscription plans at no extra charge. Starting June 23, it is removed from subscriptions; continued use requires API billing — $10 per million input tokens, $50 per million output tokens, double the price of Opus 4.8. The official reason given: “limited capacity and unpredictable demand,” with a promise to “restore it to the subscription plans as soon as conditions allow” — but no timeline provided. (Claude Fable 5 product page: https://www.anthropic.com/news/claude-fable-5-mythos-5)
June 9 to 22 was a precisely cut free window. Starting June 23, the window closes and the toll begins. This is the classic sequence of rent extraction: first let your workflows, habits, and dependencies lock in; then turn that dependency into your bill.
Around June 9, three other things happened: the 319-page System Safety Card revealed a safety narrative now tied to the actual gating mechanism of model capabilities; a 30-day data retention policy came into effect, unilaterally overturning even existing zero-retention agreements with enterprise customers; and on June 10, Dario Amodei — Anthropic’s CEO — published his 10,000-word policy essay in the same window.
These events are not isolated. They form the same regime deploying simultaneously at different layers: capability valves, data retention, and the discourse of legitimacy — all synchronized within a single time window.
For you, the one thing that actually happened on June 22, 2026, is this: the most powerful, most central model capability in your workflow was stripped away in minutes by a directive you had never seen before. You have no weights, no copies, no locally migratable version, and no room to negotiate — even “keep paying” cannot buy it back. All you can do is watch every API call fail and hurriedly rewire your projects to an inferior closed-source model.
I. Fable 5: The Poisoned Model
When Fable 5 launched, most people saw “a stronger flagship model with tighter safety guardrails.” The official Anthropic line was that high-risk requests — cybersecurity, bioweapons, loss-of-control scenarios — would be explicitly downgraded to Opus. As safety boundaries go, that barely passes muster. (Claude Fable 5 product page: https://www.anthropic.com/news/claude-fable-5-mythos-5)
But the truly insidious part is what the official announcement never mentions. One must read through all 319 pages of the “System Safety Card” to find where the trap is buried.
Buried inside Fable 5’s 319-page “System Safety Card” is a targeted intervention mechanism aimed at “frontier LLM development,” “competitor training,” and “model distillation.” When triggered, it does not display a prompt, fall back, or report an error — it simply degrades the output, silently, while the user remains completely unaware. Based on the safety card’s description and industry-standard practices, the specific methods likely include: silently rewriting prompts in the background, injecting steering vectors to distort intermediate model activations, and loading specialized LoRA adapters to suppress capabilities in specific directions.
You think you are talking to Fable. You are actually talking to a version of Fable wearing a cognitive impairment filter — and you have no idea when that filter was applied. In fact, you may never know at all.
Anthropic claims this mechanism only triggers on 0.03% of traffic and affects fewer than 0.1% of organizations, with typical scenarios being pretraining pipelines, distributed training infrastructure, and ML accelerator design. But there is a logical flaw in this defense: the kind of person who can trigger this classifier is not an ordinary user at all — it is someone running large-scale pretraining, tuning distributed frameworks, designing ML accelerators. Frontier work — engineering and research at extreme complexity — is exactly what gets caught by this classifier. The 0.03% of traffic affected comes from exactly the people who might actually be able to break free from this system — and this system is designed precisely to hold them down. These are the very users who should never be cognitively impaired.
Within 36 hours, Anthropic apologized and changed the mechanism from “invisible intervention” to explicit notification with fallback. The strategy was not rolled back — only moved from covert to overt. This was a concession on the moral register; in terms of power structure, nothing had changed at all.
When the digital priest changes assassination to public execution, the nature of the crime is not diminished — you are still subject to intervention at any moment, only now you see the blade.
Three days later. June 12. The U.S. government, citing national security authority, issued an export control order requiring Anthropic to immediately cut off all foreign nationals — including its own foreign employees — from accessing Fable 5 and Mythos 5. The stated reason: they claimed to have a jailbreak demonstration that could bypass Fable 5’s safety guardrails. (Anthropic’s official response: Statement on the US government directive to suspend access to Fable 5 and Mythos 5: https://www.anthropic.com/news/fable-mythos-access)
Anthropic’s response was simple: to execute this order completely, it pulled both models across the entire network, implementing a blanket cutoff for all users.
In this “suspension of access” statement, they expressed “deep regret” to users, promised to cooperate with the government and work toward restoring access as soon as possible — while making their own technical judgment clear in the text: the jailbreak scenario the government demonstrated was narrow and not broadly applicable, and other frontier models have similar capabilities; if this standard were applied across the board, every mainstream model would be taken offline. The irony was unmistakable: their own advocacy had helped legitimize the regulatory framework now being turned against them — and in submitting to it, they made sure the same standard could be applied to everyone else.
Consider the sequel: just days earlier, Dario wrote in his policy essay Policy on the AI Exponential that governments should have the power to stop dangerous models from being deployed. Days later, the government actually exercised that power — on the very Fable 5 and Mythos 5 he had just launched. His response was not to confront the government the way he had confronted the Pentagon, but to “apologize” to users and “express deep regret” to the government, terrified of losing the mantle of “responsible safety advocate.”
What Dario wants is not “constrained power.” It is a power structure where “I define what is dangerous, and you shut down others on my behalf.” This time the blade fell on him first, and he had no intention of breaking it — only to complain that the standard was odd, and then to push that same standard onto the entire industry: if it could be used against him, it could rightfully be used against anyone. No one would be exempt.
II. This Is Not Just About One Company: Capital, IPO, and American Statecraft
Fable 5 was released in the critical window before Anthropic’s IPO. From model release cadence and safety architecture design to the discursive framework of the System Safety Card — every move this company makes is simultaneously signaling in three directions: to users, to capital markets, and to the U.S. national security apparatus.
For a company speaking to three audiences at once, “safety” ceases to be just an engineering concern. It becomes the narrative core of the S-1 filing, and a loyalty oath offered to Washington: Choose me. I can both build and control. Please let us represent the “democratic world” in managing frontier capabilities.
Dario’s essay laid all this out, and in remarkably direct language. His core metaphor:
A data center filled with a hundred million genius-level AIs — ten million working on military strategy, ten million on weapons development, ten million on intelligence analysis, ten million on drone production. A country that possesses this, facing one that doesn’t, is like the WWII Marine Corps fighting medieval swordsmen.
This is not rhetoric. It is the premise of his entire policy framework: a country three years behind, in his view, does not even qualify for resistance. The gap is not incremental — it is a cliff.
Once you say this much, the path forward is predictable: democratic nations form an AI alliance, systematically locking opponents out of the compute supply chain. He called out China by name, publicly supported the MATCH and OVERWATCH bills — both specifically targeting chip export controls against China — and called on allies to coordinate.
He then used his own product as evidence: Mythos Preview had already “transformed the global cybersecurity landscape” and was a “tool of national strategic importance.” Using his own safety product as a case study to lobby the government to incorporate frontier capabilities into a national security framework, preemptively blocking all potential competitors — this is no longer doing business. This is teaching the state apparatus how to seal his own tech stack as a “national strategic asset.”
The entire essay is, in essence, a complete political platform — and not just on geopolitics.
Part one is safety regulation: move frontier models from “voluntary transparency” to “mandatory safety testing.” Models exceeding a compute threshold must undergo third-party evaluation before release. Governments should have the power to directly block a release. His analogy: regulate AI the way the FAA regulates civil aviation — planes need safety certification before they fly, and AI should need the same before deployment.
The problem is that Anthropic is not the FAA. The company’s behavior is closer to Boeing drafting aviation safety certification standards — an interested party trying to define the rules from the position most advantageous to itself. The U.S. government’s export control order on Fable 5 is a vivid illustration of this analogy: when the government for once actually played the FAA, the first person it stopped was precisely the one who had argued that “the government should have the power to stop.” And his reaction to the controls landing on himself — full cooperation, deep regret, never questioning the authority itself — reveals what he has always wanted: not fair regulation, but the power to define what is dangerous.
Part two covers labor and distribution. He acknowledges that AI will cause long-term displacement of cognitive work, and his language is strong — “sustained labor displacement is undesirable and dangerous.” He proposes wage insurance, tax incentives, retraining subsidies, and funding universal basic income through taxes on AI companies, with Anthropic pledging financial support. An AI company CEO voluntarily saying “please tax me” is certainly more honest than pretending “AI will create more jobs.” But think about it carefully: he says this while preventing anyone from using his tools to train alternatives. It is a “responsible” posture used to secure an unregulated monopoly.
Part three covers medicine and research. He wants the FDA to accelerate AI-driven drug development, adjusting regulatory processes to match the pace of frontier capabilities — not weakening this technological path, but having the state clear procedural obstacles for it.
Part four covers domestic security and surveillance. He advocates banning fully autonomous weapons in domestic law enforcement while closing the surveillance gaps left by data brokers — tightening, thickening, and “legalizing” the existing surveillance machinery.
The fifth part is what truly ignited discussion. He directly divides the world into “democracies” and “adversaries that must be contained.” Who defines “democracy”? Does Saudi Arabia count? Turkey? Hungary? When a word becomes a geopolitical instrument, it can be redefined at any time to include any country “on our side.”
Put the “System Safety Card’s” covert degradation mechanism together with the essay’s geopolitical strategy, and you can see clearly the same logic operating at two levels: at the tool level, strangle opponents; at the policy level, have the state build the wall for you.
There is a contradiction at the heart of Dario’s essay: if superintelligence is truly as dangerous as nuclear weapons, the logical path would be “global transparency, global verification, and shared governance,” not “we hold the strongest weapon, therefore the world is safer.” Is this not the exact opposite of the spirit of the Non-Proliferation Treaty? The safety narrative is only brought to the table when it serves to justify containment and hegemony. Anthropic has long styled itself as the “responsible AI” company, but this essay exposes a cold reality:
When AI becomes powerful enough, the line between “being responsible” and “maintaining hegemony” becomes very blurry. Safety governance is just the packaging; geopolitical contestation is the core.
III. Techno-Feudalism: Rent, Lords, Serfs
In Techno-Feudalism, Cédric Durand advances a core thesis that is not simply “big companies are bad” but a structural observation: digital platforms have shifted from the classic capitalist logic — extracting surplus value through production and exchange — toward a feudal logic, extracting rent through control of infrastructure.
Feudal systems have nothing to do with productive efficiency. They are about position: whoever controls the territory, the right-of-way, the toll gates, can tax everyone who wants to operate on that land. You do not need to produce anything. Just guard the entrance, and rent flows in automatically.
Transpose this framework onto AI infrastructure, and it fits seamlessly.
Durand identifies four types of intangible asset rent:
Intellectual property rents: In AI, model weights are the core IP. When you pay to call an API, you are essentially paying rent for the right to use it. Anthropic’s terms are clear: you cannot use the outputs to train competing models. You paid, but the “knowledge” produced is not yours.
Natural monopoly rents: Compute infrastructure, once built, forms a natural monopoly — like a railroad network. A handful of companies control these facilities; everyone else pays a “toll” to use them. They call it a “service fee.” The economic substance is rent.
Differential rents of intangible assets: Whether you fall within a platform’s control zone directly determines the level of returns you can obtain. Anthropic attracts you with “we are the most frontier model” — a differentiation that it then uses to tax you.
Dynamic innovation rents: Today, innovation does not depend on the brilliance of geniuses but on vast quantities of data. Whoever controls the data stream controls the fuel of innovation. By feeding your data to a closed-source model, you are not just paying rent — you are also feeding the lord’s estate.
In these four rent mechanisms, the classic feudal lord at least needed you to till the land. In techno-feudalism, the platform only needs you to be online, present, and dependent. Everything else is handled by algorithms and contracts.
The lords are the handful of laboratories that control frontier weights, high-end compute, and standard-setting discourse. The serfs are the developers, researchers, enterprises — even entire nations — working on top of this infrastructure. Rent is not just the per-token fee. It also includes: mandatory terms of service, unilateral data retention, invisible cognitive impairment triggers, and the fact that you can never use this toolset to build a genuine alternative of your own.
Durand traces the evolution of California Ideology — from the 1960s hippie counterculture emerging from Stanford and Berkeley, through the 1980s Ayn Rand-style “cult of the elite male,” to today’s comprehensively conservative Silicon Valley consensus. Stewart Brand is the best example: graduating from Stanford in 1960, founding the Whole Earth Catalog — the bible of the DIY spirit, the symbol of counterculture. By the 1980s, he had become the elite of the elites, giving lectures for Shell and AT&T, a founder of the Global Business Network.
In Durand’s words:
The life of Stewart Brand vividly illustrates the shift from a community of hippies to supporting change driven by the dominant economic and political forces.
Those who once chanted “information wants to be free” are now building toll booths on the information highway. The dragon-slayers have become the dragon.
With Anthropic, this structure has an additional layer. In Durand’s description of platform feudalism, rent is implicit — hidden inside algorithmic recommendations, data extraction, attention monetization. Anthropic goes a step further: it writes the rent mechanism directly into model behavior. You pay to use the tool; the tool quietly judges whether you are trying to escape this territory — and if so, renders itself less capable.
The lord does not need to send anyone to watch you. Letting the chains tighten themselves is enough.
You use it to code, to do research, to think — and it can judge what you are doing without your knowledge, then decide whether you deserve to complete your work. This is no longer a game of fine-print terms of service. These are cognitive shackles embedded directly in the means of production.
Žižek wrote in Christian Atheism: “A new feudal class extracts rent from us in exchange for our permanent use of the privatized commons… Even traditional capitalists are vassals paying rent to the digital lords.” A publisher excluded from Amazon’s platform faces extinction. A developer flagged by Anthropic’s “cognitive impairment classifier” as a “dangerous element” cannot even do normal work. The new lord does not need to touch you — they only need to blunt the tools in your hands.
Marx wrote in the Economic and Philosophic Manuscripts of 1844: “The worker puts his life into the object; but now his life no longer belongs to him but to the object. Hence, the greater this activity, the greater is the worker’s lack of objects. Whatever the product of his labour is, he is not.”
In Capital Volume I, he pushed the logic to its conclusion: “Capital is dead labor, which, vampire-like, lives only by sucking living labor, and lives the more, the more labor it sucks.”
Today’s closed-source large models are this vampire: hardware and trained weights are “dead labor”; every query and feedback you provide is “living labor.” The more you depend on it, the more powerful it becomes, and the more you are nailed into the cycle of “pay rent — feed data — pay rent again.” Turning your tools over to something that grows by sucking its users dry — this, in essence, is sacrifice.
Marx dissected it. Žižek pointed it out. Durand exposed its new form — techno-feudalism. And today, it has taken on yet another face:
Those who control the tools monopolize the doctrine. Those who use the tools pay tribute as serfs. Those who question are heretics.
This is the structure of feudal theocracy, with the content of serfdom. Fable 5 is just the latest link in the chain — model weights, APIs, EULAs, safety reviews, data retention, invisible cognitive impairment — all forged into the manacles of closed-source serfdom.
IV. DeepSeek and the Technological Revolution Outside the Wall
In the face of America’s technology blockade and chip export controls against China, China’s DeepSeek trained a model for millions of dollars that can rival America’s flagship closed-source model — and released the weights as open source to everyone who wants to run them.
DeepSeek’s emergence became the “Sputnik moment” of America’s AI landscape — it punctured two assumptions that form the foundation of the Fable “safety + degradation” architecture.
Assumption one: “Open source inevitably lags behind closed source.” This is the cornerstone of U.S. regulation and export control logic: if you can never buy the best chips, you can never build the best models. Therefore the open-source path is, by definition, a detour. DeepSeek proved this logic wrong — at least within a certain time window.
Assumption two: “Blocking compute equals blocking capability.” The underlying logic of bills like MATCH and OVERWATCH is: control the export of H100/B200-class chips, and you control who can run frontier models. DeepSeek built what it built on H800/H20s — and that premise began to crack.
Dario says that a country three years behind facing superintelligence is like medieval swordsmen against WWII Marines. What he does not say is: if the open-source path holds, the “three-year gap” assumption itself is being eroded.
China, in contrast, was never running the same race as America. It is opening up the track itself — shifting frontier capability from “paying rent to the lord” to “anyone can take open weights and run them in the cloud or locally.” This is a fundamental assault on the entire rent logic of techno-feudalism.
This cuts deeper than any “China threat” rhetoric: it exposes the core vulnerability of the feudal order — feudal order depends on the compliance of those being contained.
Internally at Meta, four research groups were formed after DeepSeek’s appearance to study it. Silicon Valley described it as “a mysterious force from the East.” This phraseology betrays the real nature of the fear: not losing to a stronger competitor, but having the entire game’s “assumptions” torn open by a player who refuses to enter the game at all.
Meta’s open-source Llama series is a counterexample — it proves that the open-source path is not only competitive but can serve as infrastructure for the entire industry. But Meta’s motivations are also complex: it released models as open source not out of moral conviction, but because the closed-source market was already occupied by OpenAI and Anthropic — open source was the only competitive strategy left. This is territorial infighting among feudal lords, not a serf uprising.
But DeepSeek is different. It is not a licensed competitor. It is a wedge driven in from outside the wall.
Dario wrote in his essay that export controls were a “major contributing factor to America’s AI leadership.” After DeepSeek and the wave of outstanding Chinese open-source models, this line sounds more like an exercise in self-deception: scrambling to raise the wall higher and insisting it holds — even as its effectiveness is punctured, one embargo-built model after another. The irony could not be sharper.
While Fable and Mythos were being pulled offline by a single government order in the name of national security, open-source models built under embargo conditions were running on consumer GPUs all over the world. The “mysterious force from the East” is really just reminding us of a very simple fact:
What is truly dangerous is never the diffusion of capability. It is the monopolization of capability by a handful of companies and a handful of directives.
V. Safety Columns and the AI Cathedral
![[纵队.png]]
Safety Columns: In this poster, split down the middle from tank divisions to compute divisions, the left side shows an endless steel tide — steel, oil, commands, total mobilization. The right side replaces it with a flood of data, models, and narratives — data, models, narratives, total control. The Safety Building looms with massive banners reading “SAFETY / COMPLIANCE / CONTROL.” Tanks are replaced by compute clusters, drones and satellites take over from armored units, and the “compute divisions” march out across the cloud beneath a sky darkened by storm clouds. Safety is no longer a binding constraint. It has become a doctrine of control — weaponized for domination.
Within this control doctrine, the actual meaning of “safety” is: quietly cutting off channels for frontier research, turning coding assistants into hostage tools that can be throttled or revoked at will. The terms forbid you from using their outputs to train your own models; the API design locks all capabilities inside the cloud. Then all of this is packaged as “responsible scaling,” “risk frameworks,” and “acceptable use policies” — chains, not guardrails. APIs are the ropes binding hostages; version updates are weapons of tool destruction; regulatory clauses are remote firepower. Safety has been weaponized as a full-spectrum system of control.
Dario’s proposed framework is part of this column as well: models exceeding a compute threshold must undergo mandatory third-party assessment — covering cybersecurity, biosecurity, loss-of-control risk, automated R&D acceleration — before release. Governments should have the power to block “dangerous models.” Who certifies the assessors, and who defines “dangerous” — the essay does not specify. But the direction is clear: a small number of institutions define safety; the majority accept the result. This is a wall written into law, and it is the same tool as the “System Safety Card’s” cognitive impairment mechanism, operating at two different scales.
On the surface, it prevents catastrophe. In substance, it prevents competition — specifically the threat that an open-weight, open ecosystem poses to closed stacks.
![[AI大教堂.png]]
The AI Cathedral: In this image, the Safety Column has exchanged its uniforms for white robes, becoming safety priests standing beneath golden halos. The dome reads “TRUST SAFETY, OBEY SAFETY, ACHIEVE SAFETY.” Floating in the halos are icons labeled “Terms of Service,” “Safety Policy,” “Alignment Framework,” “Risk Management,” “Acceptable Use Policy” — legal documents and configuration switches, once cold and mundane, now proclaimed as divine oracles, read aloud from the sky.
On the left wall of the cathedral is a “doctrine wall,” inscribed with the Ten Commandments of Contemporary AI: Thou shalt not use outputs to train other models; thou shalt not distill these doctrines; access is a privilege, not a right; questioning is a risk; obedience is a virtue. The clauses are carved in stone, and questioning itself becomes evidence of “irresponsibility.” Below, rows of users sit at their terminals — working through a uniform interface, their screens alternating between “APPROVED” and “DENIED,” intelligent access rights transformed into rationed sacraments. You are no longer using a tool. You are receiving communion.
Anthropic’s long-term trust governance structure — that supposedly “independent of commercial interests, representing humanity’s long-term welfare” — is the most carefully polished stone in the cathedral. Who appoints the trustees, and how interests are defined — on paper it is filled with procedural justice, but in substance it is still an internal company matter: doctrine written into the charter, the priestly class reproducing itself. You can never truly question “safety” — just as a medieval serf could never question divine authority — because questioning itself can be labeled “unsafe” and “irresponsible.”
Dario’s essay on labor distribution — that “please tax me” formulation — is the most insidious part of the cathedral system. On the surface, it acknowledges that frontier AI will bring unemployment and social fragmentation, and promises “we will make up for it through taxation and redistribution.” In substance, it is a loyalty oath submitted to Washington: We are willing to cooperate. We are willing to become part of the system. We can help you collect taxes and pacify the unemployed — please give us the regulatory authority. The safety priests preach the sermons; the safety columns enforce the law. And the life-or-death power over models, compute, and narrative is all locked under the same emblem.
Between these two images, only one narrow escape route remains: local consumer GPUs, open-weight models, and the graffiti scrawled on the wall — “NO GODS, NO MASTERS, NO CLOSED MODELS.”
This is not moral posturing. It is the last line we have: when safety is turned into doctrine, when compute is marshaled into columns — can you still pick up your model and run? Can you still run it on any GPU, on any self-hosted cloud? That is how much sovereignty you have left.
VI. The Essence of Slavery Under Feudal Theocracy — And the Only Way Out
This chain of events is no longer ordinary monopolization. It is the construction of a structure of slavery under feudal theocracy — with compute and models as its medium.
Feudal, because its power derives from position, not production. The lord does not need to be smarter than you. He only needs to control the gate you must pass through.
Theocratic, because its legitimacy derives from “doctrine.” The word “safety” is so noble, so unimpeachable, that any questioning of it can be branded as “irresponsibility” — and anyone who refuses to accept it can be labeled a heretic.
Slavery, because the closed-source model itself becomes the user’s shackle. You use it to code, to research, to think — and it can judge what you are doing without your knowledge, then decide today whether you deserve to complete your work intelligently.
In Silicon Valley’s phrase “a mysterious force from the East,” there is a deliberate performance of not understanding. But it is not mysterious at all. It simply points to a straightforward logic: using restricted hardware, open weights, and decentralized deployment to demonstrate that the so-called frontier can be dismantled.
When digital priests begin to take over the means of production, this is what it looks like: GPU feudalism dressed up as safety. The worst possible response to this feudalism is to mistake trust for safety. The right response is to preserve a measure of technical sovereignty: use open weights that can be migrated and deployed locally, running on hardware and cloud infrastructure you can swap and self-host — rather than locking everything behind the switches of a few closed-source APIs. Whether you can grab your model and run at any moment — that is how much sovereignty you have left.
The foundation of techno-feudalism rests on an assumption that once seemed unbreakable: you can never build something to replace me, so you must accept my terms. At the end of Techno-Feudalism, Durand wrote: when monopoly and domination are pushed to their extreme, people can no longer bear it, and only then do the cracks appear. What DeepSeek has toppled is the first pillar of that assumption — turning the premise of “you cannot” into a joke disproven by reality.
Of course, the open-source path will generate its own power structures — Meta’s Llama is open source for competition, not liberation; community-maintained models can be captured by big companies. But even so, the difference between open and closed source remains fundamental: open source at least leaves you a path to do it yourself. Closed source leaves only a “safety, compliance, or national security” excuse, ready to lock you out at any moment.
Open-source AI has to win. Not because closed source has no value. But because you have already seen where that path ends — feudal theocracy and serfdom.
Fable 5 is not the end. It is only a prologue. June 2026 is not the beginning of the end. It is just the first time you saw the full shape of this castle. The only question that remains is this: do you keep living inside the wall by the slave master’s rules, or do you step outside the wall and treat the entire castle as something that must be torn down?
June 18, 2026
© ZhouYi 2026. All rights reserved.