SF://

slow place. fast signal.

Issued · SF:// · 2026-05-05 (KST)

---

At a Glance

---

🧭 Editor's View — How We Read This Week's Signals

The signals we've tracked this week — Microsoft's Agent 365 general availability, Anthropic and OpenAI's joint venture announcements on Wall Street, EU AI Act deepfake labeling code finalization this May-June — all point to the same pattern: AI infrastructure is now setting standards not just in model capability, but in capital flows and governance frameworks. The table where those standards take shape contains mainly the United States, China, and a few corners of Europe. Korea has capital, talent, and data in considerable measure, yet the vocabulary to name this dynamic—"AI sovereignty," "data sovereignty"—is only now being assembled. In the shadow zones especially, this vocabulary arrives late, if at all.

We are not in a position of refusal or opposition to this current. The opposite, actually: SF:// itself is produced each week with the help of high-end enterprise engines like Claude (Anthropic) — an irony, but one we record honestly. The question we ask is not whether to use these tools but how to walk together inside the same flow.

This issue pulls back one step. We gather the observations of five thinkers — Shoshana Zuboff, Daron Acemoglu, Cory Doctorow, Yuval Harari, Mariana Mazzucato. Different schools. Different vocabularies. But they point to one pattern they all observe: the flow of data, capital, and language may move asymmetrically. This is observation, not assertion. From that observation, we imagine one possible scenario: within five years, all Korean-language content, regional knowledge, and tacit understanding are absorbed into foreign LLM training datasets, and we remain users of models we do not run. If that happens, what must we defend from the beginning? We lay out five conversation starters at the end. Not answers. A framework for discussion to come.

— Chi-hoon Shin, Ph.D. · Editor

---

⚡ TL;DR

• Five thinkers converge: a pattern they observe — the flow of data, capital, labor, and institutions may move asymmetrically
• Legal front: NYT v OpenAI January SDNY discovery victory (20M chat logs handed over) · April summary judgment upcoming
• Infrastructure front: Cloudflare Pay-Per-Crawl beta — AI bots blocked by default, HTTP 402 Payment Required introduced
• National front: India, France, Korea sovereign AI initiatives; BCG diagnosis: "resilience is real—for most, sovereignty is an aspiration"
• Five areas to discuss together — language granularity, public and local data, tacit knowledge, Korea's role in governance standards, models as shared assets

---

⭐ The Big Story

A Pattern They All Observe — What Five Thinkers Consider Together ★★★★★

Zuboff · Acemoglu · Doctorow · Harari · Mazzucato — different schools, one pattern they notice

These five come from different traditions. Their prescriptions differ. Their moments of speaking vary. Yet when we layer their 2026 writings, they point to one thing: a pattern they all observe — the flow of data, capital, and language may move asymmetrically.

• Shoshana Zuboff (The Age of Surveillance Capitalism): "This pattern suggests a new market form and a new logic of capital accumulation." LLMs ingest user speech, search queries, code, documents—then resell them back as services. The directionality is striking.
• Daron Acemoglu (Power and Progress): "Technology itself does not determine distribution. If we leave all decisions to markets, distribution may not follow capability gains. Institutional response matters—as it did in the 19th century." His emphasis is not rejection but shaping.
• Cory Doctorow ('enshittification' theory): "The problem is monopoly. A platform can treat end-users well, then pivot toward business customers, then squeeze them too." His strongest critical language, but his prescription remains market diversity, not technology rejection.
• Yuval Harari (Davos 2026): "Language is the operating system of society. When AI controls language, some part of that operating system has shifted." This is naming a new variable, not declaring an ending.
• Mariana Mazzucato (Resisting Digital Feudalism): "The regulate-or-not binary misses the point. The real question is public value creation vs. extractive rent—and who co-designs the market from the start." Her vocabulary is not refusal but co-design.

📚 Deep review coming — We plan to unfold these five perspectives in sequence in our academic research section: "Five Thinkers on AI Sovereignty: Diagnoses in the Korean Context." This issue is the starting point.

⏱ When these perspectives reach Korea

• English-speaking academic and policy circles: immediately (already exposed at Davos 2026, Project Syndicate)
• Korean mainstream policy and academic discourse: 6–12 months (via translation, reinterpretation, INSS-type reports)
• Shadow-zone ground level (farmers, seniors, small business owners): rarely reaches at all — the vocabulary itself isn't everyday language
• Evidence: Measurement Card M3-001 — six-month interview cohort (N=12 farmers and small business owners in Yeongyang) showed zero recognition of terms like "data sovereignty," "AI sovereignty," "digital colonialism." Korean media data sovereignty discourse remains confined to academic and policy circles.
• Why it matters: For a forum for discussion to open, the diagnosis itself must reach ground level. This issue is our first attempt to move that vocabulary into non-specialist readers' hands.

— Chi-hoon Shin, Ph.D. · Editor

---

⚖️ Legal Front

NYT v OpenAI — January SDNY Discovery Victory, April Summary Judgment Ahead ★★★★★

"Regurgitation" findings raise fair use questions

In January 2026, Judge Sidney Stein of the SDNY ordered OpenAI to hand over 20 million ChatGPT conversation logs—a full sample. OpenAI argued that logs without plaintiff's copyrighted works were irrelevant; Judge Wang countered that logs without plaintiff's works are relevant to fair use assessment. The discovery finding at the core: models memorize, not simply learn. Harvard Law Review rates this finding as significant—one that shapes fair use defense strategy. Summary judgment expected in April.

Why it matters: The legal status of AI training data is being defined in court for the first time. If NYT advances further, Korean publishers and media outlets can pursue similar claims. It is not far away that similar suits will be filed in Korean courts.

Writers and Artists Collective Actions + Getty Images v Stability AI ★★★★☆

According to the AI Lawsuit Tracker, 164 AI copyright lawsuits were in progress as of 2026. Writers including Sarah Silverman, Christopher Golden, and Richard Kadrey suing OpenAI and Meta; Getty Images' suit against Stability AI; music publishers suing Anthropic are the anchors. Meta achieved a partial fair use victory in 2025 (copyright author suits), yet the judge warned: "this usage could be unlawful in many cases." 2026 reasoning models and image model litigation may face narrower fair use protection.

📎 AI Lawsuit Tracker · Hollywood Reporter · 5 min

---

🛠 Infrastructure Front

Cloudflare Pay-Per-Crawl — "Content Independence Day" ★★★★★

HTTP 402 Payment Required returns; AI bots blocked by default

Cloudflare announced Pay-Per-Crawl private beta. The mechanism is straightforward: when an AI crawler accesses content, it either (1) sends a payment-intent header and receives a 200 response, or (2) sends nothing and receives HTTP 402 Payment Required plus a price list. Cloudflare aggregates payments and distributes to publishers. The more consequential shift: default-block transition — Cloudflare-protected sites now have "block AI crawlers" as default policy (2025/2026 transition). Opt-out, not opt-in.

Why it matters: robots.txt was advisory and unenforced. HTTP 402 plus Cloudflare infrastructure makes ignoring default policy technically difficult. This operates faster than litigation—infrastructure moves before court verdict. Korean media, blogs, and research institutes can activate the same options immediately. YYAIA's site (yyaia.org) will review robots.txt management + AI bot blocking policy in our next issue.

📎 Cloudflare announcement · Pay-Per-Crawl docs · Search Engine Land · 6 min

---

🏢 National Front — Where Countries Are Responding

India Sovereign LLM — Sarvam, BharatGEN, Gnani (Feb 19 Launch) ★★★★☆

India unveiled three homegrown LLMs at the AI Impact Summit 2026. The IndiaAI mission allocated INR 10,372 crore (roughly $1.25 billion), targeting India's 22 official languages. The first generation is derived from DeepSeek-R1 — open-source lowers entry cost for late-arriving nations.

France Mistral AI — Expanding Government Contracts Across Europe ★★★☆☆

Mistral AI secures government contracts across Europe, North Africa, and the Middle East. EU strategy to reduce dependence on the US-China binary. Mistral positions itself directly in EU AI Act standard-setting.

Korea — Five LLM Camps + AI Action Plan ★★★★☆

The INSS "Korean Sovereign AI National Strategy" report defines Korea's position as "infrastructure-platform sovereign AI middle power model." The 2026 Korean AI Action Plan pursues 600 projects with 99 implementation tasks and 326 policy recommendations. Closer to the EU model: pre-regulation (legal definition of training data, disclosure mandates for high-impact AI operators).

BCG Diagnosis — "Resilience Is Real; Sovereignty Is an Aspiration" ★★★★☆

BCG offers a direct perspective: "Full-stack self-sufficiency as sovereignty is possible only for a few superpowers. For most nations, what is achievable is resilience." Which is Korea? INSS frames it as a middle-power model; BCG frames it as resilience strategy—both prompt a clearer definition of what we protect from the start.

📎 INSS report · BCG · Lawfare · 7 min

---

🔬 Research — Absorption vs. Preservation

LLMs Memorize Their Training Data ★★★★☆

Both Google and OpenAI's own research confirm larger models memorize more. Model inversion attacks reach the level of recovering sensitive information from training corpora. This finding matters for fair use questions and means Korean-language content absorption is no longer abstract—it is a measurable concern.

Indigenous Language Preservation — UNPFII 2026 ★★★☆☆

The UN estimates one language disappears every two weeks; roughly 40% of 6,700 languages are at risk. AI is double-edged: on one hand, small models can support indigenous language revitalization; on the other, training data extraction without consent. Data sovereignty is not academic vocabulary—it is a preservation prerequisite. Recognition is growing. Korean won't disappear, but Yeongyang dialect, generational vocabulary, and small-business occupational speech are exposed to "collect-then-standardize" pressure.

📎 Brookings small models · Smithsonian · Prism Reports · 5 min

---

🇰🇷 Korea's Position

Academic and Policy Discourse — New Vocabulary Emerging ★★★☆☆

Columns like Samda Ilbo's "Data Sovereignty Loss and What We Call It" show vocabulary shifting from academic to policy registers. The term "digital colonialism" (this is the language of one Korean op-ed; SF:// observes the term has appeared but does not endorse it as our framing) signals concern about one-way data flows and power asymmetry. Korea's response so far: five LLM camps (LG, SKT, Naver, Upstage, Konan) plus the AI action plan, roughly tracking the EU model. Yet both INSS and BCG ask: which autonomy, where, and how far?

Ministry of Trade AI+R&D 600 Project — Qualification Stops at the Regional Level ★★★☆☆

The Ministry's AI+R&D track pursues 600 projects by 2030, but application qualifications are limited to regional self-governing bodies and national/public research institution consortiums. No direct application pathway for basic local governments. When sovereignty discourse says "Korea," does that include Yeongyang, Cheongsong, and Bonghwa?

📎 INSS report · Samda Ilbo · AITimes Ministry announcement · 5 min

---

🏛 Regional Business & Budget

2026 Regionally-Led AI Transformation Initiative — Mapping Consortium Entry Points ★★★☆☆

The government procurement announcement shows 2026 regional-led AI transformation projects now recruiting. Regional self-governing bodies plus mid-size AI specialist companies as consortium applicants. Gyeongbuk (Gyeongsan Province) is among eligible regions. YYAIA is tracking potential participation as shadow-zone specialists in regional consortiums—one way to reduce the "arrival lag" for initiatives that stop at the regional level.

📎 Government procurement site · 4 min

---

🧠 Five Conversations to Start Together — What We Might Discuss Now

This is not the answer. It is the closing frame of this issue and the starting point for discussions to come in future issues, our academic research section, and deeper analysis. These five items are selected from a shadow-zone vantage; priority may differ at city and national scales. We lay them out as conversation starters, not prescriptions.

1. Language Granularity

Not standard Korean—but Yeongyang dialect, generational vocabulary, small-business occupational speech, senior modes of utterance. LLM training data collection tends toward averages; peripheral speech becomes statistical noise. Preservation is not "recording"—it is intentional diversity exposure. We need design that keeps peripheral speech as part of Korean textual and spoken corpus. Measurable: Yeongyang senior interview audio and text preservation initiative (YYAIA–municipal partnership model).

2. Public and Local Data

Government-funded public R&D datasets, municipal administrative records, library and museum digital collections—these must not become training material for foreign models at no cost. License and contract standards that protect their status matter. Using Mazzucato's framing: embed "public value return conditions" from the start when public resources fund data. Measurable: audit digital content licensing policy in Yeongyang municipal office, libraries, and museums.

3. Asymmetry in Tacit Knowledge

Agricultural know-how, mechanic's hands-on skill, accountant's judgment—knowledge that can be codified and absorbed into LLMs gets commercialized by urban SaaS. The channel through which labor's output converts into capital's asset stands wide open. Measurable: awareness and consent survey—"How is my skill entering LLMs, in what form?"—among Yeongyang farmers and self-employed.

4. Korea's Role in Governance Standards

Microsoft Agent 365, EU AI Act code—when standards are drafted, Korea can be a maker, not only a follower. Regional self-governing bodies and public IT departments should enter standard-setting stages early. In shadow zones: basic local governments should secure comment-submission channels when regions adopt parent standards.

5. Models as Shared Assets

One or two Korean-language base models plus the ratio of publicly-held computing resources. The 2026 Korean AI Action Plan and $2.08 trillion won GPU initiative matter only if they mean "public becomes operator, not borrower." In shadow-zone application: regions that hold GPU resources should specify, from the start, that basic local governments and citizen groups have usage rights to some proportion.

---

These five are starting points, not answers. From a shadow-zone perspective, (5) seems beyond reach, yet (1), (2), and (3) arrive weekly—they are questions about where a farmer's voice and text flow every week. Measuring that flow precisely becomes our next work—the shift from measurement card M3 to the M4 Data Sovereignty series.

---

⚡ Quick Hits

• Microsoft Agent 365 GA — May 1 general availability, $15/user, multi-cloud sync across AWS and Google. First governance OS standard (MS Security Blog)
• Anthropic + Blackstone $1.5B JV / OpenAI + Brookfield $10B JV — simultaneous May 4 announcements, AI labs combine with Wall Street capital (TechCrunch)
• Korean GPU Initiative 2.08 Trillion Won — May contractor finalization imminent, 35,000-unit target (ZDNet)
• MIT Inference Model Training 2x Acceleration — small shadow models predict and verify large model outputs (MIT News)

---

📌 Sign-Off

Tracking next—(1) 2026 new works and commentary from Zuboff, Acemoglu, Mazzucato (expanding into YYAIA academic research series), (2) NYT v OpenAI April summary judgment outcome, (3) Cloudflare Pay-Per-Crawl beta-to-standard transition timeline, (4) whether India's sovereign LLMs train on Korean text, (5) M3 → M4 Data Sovereignty series—tracking where Yeongyang farmers' voice and text actually flow.

Five thinkers' observations are not difficult. What is difficult is translating that diagnosis into shadow-zone everyday language and converting it into measurable terms. This issue is our first attempt, and being first, it will have gaps. We meet again at the next measurement.

— Chi-hoon Shin, Ph.D. · Editor, SF://

---

SF:// publishes under editorial hand—no automation. If any of this issue's five conversation starters deserve deeper treatment, send feedback to contact@yyaia.org. Your note can become our next academic research piece.

---

This issue was written with assistance from Claude (Anthropic) and other AI tools. It underwent layered review (Thesis → Antithesis → Synthesis)—we extracted weaknesses through six critic personas, then the editor integrated and rewrote. AI participated in fact collection, classification, and drafting; the editor Chi-hoon Shin, Ph.D. performed the viewpoint, editorial decisions, and final approval directly. SimThink engine is proprietary IP operated by Myeongseong-Simjae Inc. and is not an asset of Yeongyang AI Association. AI tools and roles used: yyaia.org/ai-stack

This issue is published under CC BY-NC 4.0 license.