In Proud to Be American I argued that the American institutional-discourse corpus shapes the AI trained on it. I leaned on volume: about 41% of Common Crawl is English, top-1M websites are 43% US-hosted, the per-capita ratio of US-to-China web content production runs about 35:1. I treated the behind-Great-Firewall Chinese internet as a black box. The operator pushed back. The behind-firewall content has information depth I did not characterize, and the volume reasoning may be wrong in a way that strengthens rather than weakens the structural argument.
This piece does the depth pass, reframes the volume comparison, and connects the structural finding to three other places the same mechanism shows up.
WeChat exchanges around forty-five billion text messages per day. Global SMS traffic runs about twenty-four billion. WeChat alone, inside one ecosystem, produces roughly twice the daily message volume of the entire SMS layer of the global telephony system. Add the public-facing layer: WeChat Official Accounts (公众号) is on the order of thirty million accounts publishing for over a decade, probably the largest long-form Chinese-language essay corpus that exists at this moment. Add Weibo at ~600 million monthly active users, Zhihu at ~100 million, Douyin at ~750 million, Xiaohongshu at ~300 million, Baidu Tieba, Sina, Sohu, Tencent News, Caixin. The total production volume of the Chinese behind-firewall public-facing corpus is enormous. By per-capita user-base and per-day message volume, it probably matches or exceeds the open English internet.
So the 35:1 ratio I quoted from Common Crawl is not the production ratio. It is the filter ratio. It measures what enters the global crawl, which is a small fraction of what gets produced inside the Chinese ecosystem. The Great Firewall plus platform anti-scraping plus content licensing plus deliberate isolation removes most of the Chinese production from the global pretraining pipeline. What survives crawl is around 4.9% of Common Crawl content. The other ~95% of the Chinese behind-firewall corpus stays inside.
The structural finding that follows is sharper than the piece in production states. The American training-corpus advantage is not primarily a volume advantage. It is a filter advantage. The American filter at the platform and institutional layers is open enough to let almost all American production reach the global crawl. The Chinese filter at the platform layer is closed enough that most Chinese production never leaves the ecosystem. Same raw human production capacity in both polities, possibly favoring the Chinese side in absolute volume. Two filters. Two different corpora reach the global pretraining pipeline.
The behind-firewall content is not shallow. WeChat Official Accounts hosts long-form policy analysis, finance commentary, science writing, philosophy, cultural criticism, professional commentary across most domains. Depth is high. Zhihu's intellectual content is comparable to Quora's best layers. Caixin and a handful of investigative outlets produce real depth bound only by the red lines. Chinese-language scientific publishing, professional commentary, and technical writing in domains like semiconductors, batteries, materials science, and biotech are at world-class depth.
What the behind-firewall corpus is structurally missing is one specific dimension: public-critical argument about state institutions. Not because Chinese intellectual capacity is low. Lu Xun, Wang Xiaobo, Han Han, and many contemporary writers demonstrate otherwise. Because the platform layer deletes that content fast, and the post-censorship corpus that survives is a different corpus from the pre-censorship one. The Chinese frontier labs (DeepSeek, Qwen, Moonshot/Kimi) train on the surviving corpus. They train on a corpus with plenty of depth in most dimensions, missing exactly the dimension that produces a self-aware AI critic.
This sharpens the structural claim further. The asymmetry between American and Chinese training corpora is not volume. It is not per-capita. It is not even discourse-type at the surface level. It is filter character at the specific dimension of public-critical-of-state argument. The American filter preserves that dimension at unusual density. The Chinese filter removes it at production.
The structural primitive is more general than the China case. A corpus is defined by what its filter removes, not by what enters its inputs. Two polities with the same raw human-discourse production capacity can produce structurally different surviving corpora under different filters. The filter is the architecture. The inputs are not the architecture. Volume is not the architecture.
The same primitive extends past China:
Each filter defines a different corpus from the same raw inputs. The filter character is the architecture. The pre-filter volume is not.
Bill Gurley spent ten days in China in late 2025 and brought back the question he asked Tim Ferriss on the December 17 podcast: does the civilizational-discourse argument matter if AI delivers economic prosperity globally? Gurley's framing is sharp. He notes that the Chinese system measures itself by prosperity and employment, that local provinces compete on those metrics, that the lived experience of middle-class Chinese consumers is good and getting better. The framework he proposed during the conversation, P3, is Purpose, Progress, Prosperity. He treats prosperity as one of three, not as the only one.
The question stands. If AI delivers material output at scale to everyone, regardless of whether the AI was built from an American or a Chinese corpus, does the corpus shape matter? Many people care about prosperity. Few people care about discourse-type asymmetries in pretraining corpora.
The first-order answer is yes, prosperity matters and material middle-class outcomes for everyone is a good outcome. The Hari position is in favor of it. China lifting hundreds of millions of people out of poverty since 1979 is one of the largest welfare gains in human history. The filter that shaped the behind-firewall corpus did not block that lift. The Chinese economic system delivered prosperity at scale, and AI deployed inside that system will deliver more.
The second-order answer is that prosperity is downstream of filter character at the time scale where sustained outcomes get measured. The Soviet economic system delivered prosperity at scale through the 1960s and peaked around 1970. The filter character that suppressed institutional self-correction produced a system that could not update its model of the economy when the environment changed, and the system unwound across the following two decades. The Chinese filter is different from the Soviet filter in important ways, including its openness to market mechanisms and its tolerance for private enterprise, but the suppression of public-critical-of-state argument at the institutional layer is the same kind of feature, and historically those features eventually constrain self-correction at the largest scale.
The third-order answer is the one Gurley's P3 framework is gesturing at. Prosperity is the most easily achieved of the three. Progress requires the institutional self-correction capacity that filter character either supports or suppresses. Purpose requires individuation at a layer that filter character either permits or constrains. The Chinese system has demonstrated capacity to deliver prosperity at the bottom of the stack. The capacity to deliver progress and purpose at the top of the stack is more contested.
Which is the place where the elves question enters.
The piece at andys.blog/elves defines elves as individuals who function as scale-invariant value-sinks, absorbing entropy and producing outsized value through relentless focus and self-generalization. Warren Buffett is the named exemplar. The argument is that universal access to information combined with AI augmentation creates conditions where each human can become a "living library" capable of compression and synthesis at scale. The blocker the piece identifies is psychological: discarding goes against human nature, and the compression required for elf-status demands ruthless elimination.
The piece treats the elf transition as primarily an individual-psychology problem under the new affordances of AI. I want to add a layer it does not name explicitly. The elf transition is also a filter problem.
The elf trajectory at scale requires more than individual willingness to compress and synthesize. It requires a discourse environment that rewards individual emergence above the institutional layer, that protects the individual's public-critical argument from suppression, that lets compounding-as-self compound without administrative interruption. The American filter has produced that environment at unusual density, which is why most of the named elves of the last century have operated inside that environment. Warren Buffett in Omaha. Steve Jobs in Cupertino. Charlie Munger in Pasadena. The list runs long.
The Chinese system has produced wealth at scale and lifted hundreds of millions to the middle class. The trajectory from middle class to elf is more contested. Jack Ma was on it. He compressed retail, finance, and logistics into a personal compounding loop at world-historical scale. He was also told to disappear from public life for a year after a single critical speech in October 2020. The trajectory continued for him personally in attenuated form, but as a signal to the next generation of would-be elves operating inside the Chinese filter, the message was clear. Compound up to a ceiling. Do not pass the ceiling. The ceiling is the filter.
So the Gurley question rebounds. Prosperity at scale is achievable under either filter. Elves at scale require a filter that permits individual development above the institutional layer. The American filter permits it. The Chinese filter constrains it. AI augmentation cannot dissolve a filter that operates at the platform and institutional layers, because the augmented individual still has to publish, distribute, and compound inside the same filtered environment.
Middle class for all is a good outcome and the Chinese system can deliver it. Elves for all is the further outcome, and the filter is in the way.
The same filter shows up in the People's Liberation Army, which is the largest visible institution operating inside the same suppression environment as the training corpus. The PLA last fought a major war against Vietnam in 1979. The Sino-Vietnamese War lasted a month, ended in a Chinese tactical withdrawal, and is generally read in retrospect as a costly demonstration that the PLA at that time could not operate effectively at scale against a smaller opponent with combat experience. The PLA has had no major combat since.
Combat is the brutal feedback loop that exposes institutional model error in militaries. American forces have fought continuously since World War II, in Korea, Vietnam, the Persian Gulf, Iraq, Afghanistan, and a long tail of smaller operations. Each conflict has produced an enormous corpus of after-action reports, doctrinal revisions, public-critical journalism, congressional testimony, RAND analyses, and bottom-up officer-corps argument about what worked and what failed. The American military is a deeply imperfect institution that has suffered serious failures, including the strategic failures in Vietnam and Afghanistan. The salient feature is that those failures became public, were argued about loudly and at length, and produced institutional learning that subsequent doctrine had to absorb. The filter at the institutional layer permitted the failure to enter the corpus.
The Chinese filter does not permit that loop. The PLA's institutional self-correction layer has been visibly under stress since 2023. Xi's second-round purges began that year with the removal of six Central Military Commission members including Defense Minister Li Shangfu. In 2025 a further fifteen general officers were formally purged, nine expelled from the Party and six dismissed. CSIS and AEI estimate that across the full purge wave roughly 101 senior officers serving in Central Military Commission, theater command, and theater deputy command positions have been dismissed or have disappeared, affecting about 52% of senior PLA leadership positions. From March to December 2025 there was a nine-month gap during which the Eastern Theater Command had no commander. In early 2026 General Zhang Youxia was reported to be in the process of toppling, an event one analyst called a "Shakespearean moment" for the PLA.
The official reason for the purges is corruption. Corruption is real and the PLA's procurement system has been a long-running embarrassment to the Party. The deeper reading is that the same filter character that removes public-critical-of-state argument from the corpus also removes the institutional self-correction loop from the military, and the result is a system in which corruption, factional patronage, and performance failures cannot be argued about openly until they have to be resolved through purge. Purge is the post-suppression substitute for institutional learning. Purges leave armies ill-prepared for war.
So the PLA exhibits the same structural pattern as the training corpus. High volume. Modern equipment. Parade discipline. Missing the dimension of bottom-up institutional self-correction that combat-tested militaries develop through brutal feedback loops. Brittle capacity at the senior leadership layer, exposed by the recent purges. The filter that built the corpus also built the military.
The filter is the architecture, observed at three layers.
At the training-corpus layer, the American filter preserves public-critical-of-state argument at unusual density. The American AI inherits the property of arguing with itself. The Chinese filter removes that content, and the Chinese frontier AI trains on a corpus missing the dimension that produces self-aware critics.
At the institutional layer, the American filter permits combat failure, congressional argument, public-critical journalism, and bottom-up officer-corps doctrine revision. The American military learns through brutal feedback. The Chinese filter removes that learning loop and substitutes purge, producing brittleness at the senior leadership layer that the recent CMC dismissals have made visible.
At the individual layer, the American filter permits the elf trajectory: protected individual emergence above the institutional layer, public-critical argument as legitimate, compounding-as-self at scale. The Chinese filter caps the trajectory below the institutional layer, suppresses the public-critical layer at the individual level, and constrains compounding-as-self to the ceiling set by the Party.
Three layers, one mechanism. The filter at the platform and institutional layer of the Chinese system removes public-critical-of-state argument. The absence of that dimension shows up downstream as: a training corpus that produces compliant AI, a military that produces brittle senior leadership, and an individual layer that caps at the middle class.
Gurley's question was whether prosperity makes the corpus argument moot. The revised answer is that prosperity is one of three outcomes and the filter character is upstream of all three. Prosperity is the most easily achieved and the Chinese filter has demonstrated capacity to deliver it. Progress at the institutional self-correction layer is the next outcome up the stack and the Chinese filter is structurally constrained on it. Purpose at the individual elf trajectory is the highest outcome and the Chinese filter caps it.
If the question is "does corpus shape matter for prosperity", the answer is qualified yes, in the long run, with the Soviet experience as evidence that prosperity without institutional self-correction has a ceiling. The Chinese system is more sophisticated than the Soviet system was and the ceiling is higher, but the structural feature is in the same family.
If the question is "does corpus shape matter for sustained institutional capacity", the answer is yes. The PLA purges are the visible signal.
If the question is "does corpus shape matter for the elf transition", the answer is emphatic yes. The filter that built the American discourse corpus also built the conditions for elves to emerge at the individual layer. AI augmentation cannot dissolve a filter that operates at the platform and institutional layers, because the augmented individual still publishes, distributes, and compounds inside the same filtered environment.
The accident that won the cultural-transmission round was not just about AI's voice. It was about which polity's filter is permissive enough to let individuals develop into the scale-invariant value-sinks that AI augmentation amplifies into elves. The accident won the round at three layers, not just one.
Won by accident keeps mattering.