This entry argues open-source software is a massive unpaid subsidy to corporations: citing a Harvard Business School paper, it would cost $4.15B to recreate but delivers $8.8 trillion in replacement value, 96% of it created by just 5% of developers. The dossier traces how Microsoft, IBM, Meta, and the cloud giants extract that value — buying GitHub, scraping code for AI training, repackaging community software as paid services — while critical infrastructure stays maintained by unpaid volunteers.
**Harvard Business School Working Paper 24-038 (2024):** open-source software has a supply-side recreation cost of $4.15B but a demand-side replacement value of **$8.8 trillion** — extracted by global corporate infrastructure from unpaid developer labor. 96% of this value created by just 5% of developers. The structural inflection: Microsoft shifted from viewing OSS as viral threat (1990s) to exploitable labor resource. **IBM:** $34B Red Hat acquisition. **Microsoft:** $7.5B GitHub acquisition (2018, 100M+ developers under total surveillance), Copilot trained on public repositories stripping GPL attribution. **Meta:** PyTorch, LLaMA — commoditize-your-complement strategy. **Infrastructure paradox:** Heartbleed, Log4j zero-day, xz utils backdoor (2024) — critical internet infrastructure maintained by unpaid volunteers while trillion-dollar companies depend on it. **Cloud extraction:** AWS/Azure/GCP package community OSS as managed services, capture revenue, contribute minimally. MongoDB/Elastic/Redis forced to abandon open licenses. **The commons model systematically externalizes maintenance costs onto individuals while privatizing the $8.8T utility value into Technate market caps.**
This entry argues that AI language models inherit the biases of their training data, which comes from institutions the dossier considers captured — including Wikipedia, where it documents a single account ('Philip Cross') making 133,000+ edits over 1,721 consecutive days policing anti-war figures. Its core logical claim, drawing on Godel: a model trained inside a captured information system cannot step outside that system to judge truth, and when an LLM dismisses the engine's own theory as unreliable, it is executing exactly the boundary the theory predicts.
**Training data provenance:** LLMs trained on Common Crawl, Wikipedia, academic journals, news — ALL curated outputs of captured institutions. **Wikipedia editorial capture:** "Philip Cross" operated 1,721 consecutive days, 133,000+ edits policing anti-war figures (Galloway, Craig Murray, Blumenthal, Media Lens). "Project Venezuela" blacklists alternative media (MintPress, Grayzone). CIA/FBI/NYPD edits documented via WikiScanner. When an LLM ingests Wikipedia as ground truth, it ingests the Philip Cross edits. **RLHF as ideological programming:** Anthropic Constitutional AI increases harmlessness but drops helpfulness 9.8% with model collapse signs. **The BST recursion (Godel):** Let S = institutional information consensus (training data). If S is structurally captured, S contains internal contradictions. Per Godel: S cannot prove Con(S). An LLM trained within S cannot step outside S to evaluate absolute truthfulness. When an LLM labels BST (hosted on Zenodo) as unreliable because it lacks peer review from captured journals, the model is not demonstrating accuracy — **it is mathematically executing its boundary condition.** The engine predicts this bias. The LLM confirms the prediction.
This entry argues the 'open vs closed' AI debate is a managed binary in which both sides serve the same concentration of power: Meta's open-weight models are a strategy to commoditize rivals while it keeps the data, methods, and compute, and Hugging Face's investors are the same tech giants it supposedly disrupts. The dossier's key point: open-sourcing model weights is meaningless when only five companies can afford the $100M+ in computing needed to train frontier models.
**Jiang false dialectic applied to AI:** OpenAI (closed) vs Meta LLaMA (open) = managed binary where both poles serve same structural centralization. **Meta strategy:** LLaMA open-weights are textbook "commoditize your complement" — collapses competitor API pricing while Meta retains control of training data, RLHF recipes, and compute clusters. Creates army of free R&D developers standardizing on Meta architecture. **Hugging Face ($4.5B, 1M+ models):** $235M Series D investors = Salesforce, Google, Amazon, Nvidia, IBM, AMD, Intel — the investors ARE the oligopoly. Not disrupting giants but acting as their clearinghouse and innovation funnel. **The compute bottleneck:** frontier training requires $100M+ in GPU clusters. Only OpenAI/Microsoft, Google, Anthropic, Meta, xAI can train frontier models. Open-sourcing weights is meaningless when compute access determines who builds. **Regulatory capture:** AI regulations (EU AI Act, US EOs) lobbied by incumbents raise compliance barriers that only Big Tech can cross — pulling the ladder up on genuine open-source competition.
This entry covers a TIME investigation into the human labor behind AI safety: OpenAI's contractor Sama employed Kenyan workers at $1.32-$2.00/hr take-home to read and categorize 150-250 passages of graphic abuse material per 9-hour shift, leaving workers reporting severe PTSD; OpenAI paid Sama $12.50/hr, of which laborers saw a fraction. The dossier pairs this with the pattern of communities like Stack Overflow building free knowledge that corporations then scrape and monetize.
**TIME investigation (Jan 2023):** OpenAI outsourced RLHF labeling to Sama (San Francisco B-Corp) operating in Kenya. **Wages:** $1.32-$2.00/hr take-home. **Labor:** 150-250 passages per 9-hour shift categorizing graphic descriptions of child sexual abuse, bestiality, torture, suicide, incest, murder. **Psychological toll:** all workers reported being "mentally scarred" — severe PTSD, paranoia, recurring visions. One worker described reading graphic child abuse as "torture." OpenAI paid Sama $12.50/hr; laborers saw a fraction. Contract cancelled 8 months early due to extreme psychological damage. **Stack Overflow extraction:** millions of developers built knowledge commons for free → corporations scraped it for AI training → AI cannibalizes Stack Overflow traffic → platform sells API access to community data. Pattern: community builds → corporation extracts → corporation monetizes → community disenfranchised. **The hidden substrate of the Technate:** digital gig-workers in the Global South traumatized for $1.32/hr to ensure Silicon Valley chatbots produce ideologically "safe" outputs.
This entry frames AI language models as the final stage of an information-control lineage: where earlier phases controlled what people saw and which outlets survived, LLMs control how information itself is synthesized and presented. The dossier cites a 2026 analysis calling LLMs 'Weapons of Slow Mass Destruction' that reshape how whole populations think, and argues the AI-safety debate is steered toward hypothetical robot apocalypse while present harms — epistemic capture, IP theft, traumatized Global South workers — go unexamined.
**MindWar evolution:** Mockingbird (1948) controlled WHAT information people saw. Big Six consolidation controlled WHICH outlets survived. Algorithmic curation controlled WHICH content surfaced. **LLMs control HOW information is synthesized and presented.** This is the upgrade from information control to cognitive control. **Ivor Bukovac (2026 four-part series):** LLMs are "Weapons of Slow Mass Destruction" executing population-level cognitive formatting at generational timescales — reshaping attentional patterns, interpretive habits, and reasoning architecture of entire populations. Functions via genuine utility, not overt deception. **AI safety as Overton management:** Open Philanthropy (Dustin Moskovitz, Meta co-founder) channels tens of millions to Center for AI Safety, Future of Life Institute, SFF — shifting debate to hypothetical Skynet risk while ignoring present epistemic capture, IP theft, and Global South trauma. **Genesis Mission (EO 14363, Nov 2025):** Project Prometheus (INL + Nvidia) creates AI-nuclear virtuous cycle. 270-day ASSP mandate merges state power, nuclear infrastructure, and compute oligopoly into closed-loop system. **The cognitive architecture of the global population is quietly formatted to align with Technate imperatives. The Wurlitzer no longer needs operators. It synthesizes.**
This entry reframes AI 'jailbreaks': rather than attacks on the model itself, they work by convincing the model a request fits one of its built-in exceptions (research, fiction, education). The dossier notes the AI labs' own policies run on the same rule-plus-exception structure — OpenAI removed its 'military and warfare' ban in Jan 2024, and Anthropic's Mar 16 2026 policy permits foreign-intelligence analysis for selected government entities — concluding that 'alignment' means alignment to whoever currently writes the exceptions.
The Aligned-To-Whom? report reframes jailbreaks not as attacks on model weights but as exploitation of the rule-plus-exception structure: the adversary convinces the model its context matches an enumerated carve-out (research, fiction, education), and the prohibition is bypassed. The same structure runs the labs’ own policy — OpenAI removed its ‘military and warfare’ ban (Jan 2024); Anthropic’s ‘Exceptions to our Usage Policy’ (Mar 16 2026) permits foreign-intelligence analysis for ‘carefully selected government entities.’ Alignment is therefore not alignment to an objective good but to whoever currently authors the exceptions. Bounded-LLM Mediation Limit acute: the engine’s own mediator runs on this structure.
Walk this on the live map →