Structural Analysis
Current events through the lens of structural analysis
The previous piece argued that the fertilizer constraint does not vanish but moves. Where it moved is nitrogen, so this piece examines the nitrogen balance. The acute phase has peaked — urea for India fell from roughly $890 a tonne in April to about $445 at June's oversubscribed tender, and after an eight-month halt China restarted exports with June-August quotas of 1.5-1.6 million tonnes communicated directly to producers, under a $660/t FOB price floor — a design that protects the domestic market while capturing only high prices, and a template for August's phosphate decision. But the structure remains: the Gulf accounts for 36% of world urea exports, Ras Laffan's repair runs in years, strikes continue on Russia's Azot plants, and new capacity (4.6 million tonnes of ammonia over two years, Qatar's urea doubling) lands from 2027. And the piece's thesis: every phosphate-side escape route merges into nitrogen. OCP's TSP shift strips out the nitrogen that came bundled in MAP/DAP and moves that demand to urea. The nitrophosphate route re-ties phosphate to ammonia. If fuel-ammonia co-firing scales as planned, fuel joins fertilizer in the same procurement market. China's disasters are the original template here — the October 2021 export inspections began in the autumn of the Henan floods and an energy crisis, after urea futures spiked. Japan's own weakness: urea is 74% Malaysia with 3% domestic, and ammonium sulfate rested on a caprolactam-byproduct base that is thinning as the majors exit. The one move that works whichever way it falls is to reduce purchased nitrogen itself.
Three months after this site wrote that phosphate fertilizer would become unobtainable in Japan, a more useful sorting than a scorecard is available — what turned out better than expected, what turned out worse, and what is still undecided. Better — Morocco's OCP cut second-quarter output by about 30% on sulfur costs, then returned to full production by the end of June by switching its mix toward TSP, which uses far less sulfur and no ammonia (now 65% of volumes). In China, domestic sulfur prices rebounded in July on anticipation of phosphate export quotas — a restart signal in money, not statements. Japan's stockpile, diversification and priced-in seasonality held autumn phosphate to +4.5-4.9%, with no empty shelves. Worse — nitrogen, outside the forecast's field of view, took the largest hit: Qatar and Iran stopped, urea rose 22.7% in April alone, and the World Bank sees 2026 urea nearly 60% higher. The nutrient that had diversified successfully was the one that got hit — diversification meant changing the resource, not the counterparty. Sulfur ran at five to eight times normal, and potash was collateral. Undecided — the variable that matters most is China's domestic balance: the August decision turns on domestic prices, autumn planting, LFP battery demand and sulfuric acid costs, and the leading indicator is the Chinese sulfur price, not the announcement. Morocco's product mix and November's spring pricing remain open. The constraint has not disappeared. It is moving, and changing shape.
The previous piece cited the Bun port reported on Anthropic's engineering blog — one million lines of Zig to Rust in under two weeks, 100% of the existing tests passing, roughly $165,000 in API costs — as proof that distillation runs at industrial scale. This piece does the conversion. The same article carries a second example — an individual moved a 165,000-line Python project to TypeScript over a single weekend. The custom portion of a small company's business system is usually a few thousand to a few tens of thousands of lines — an order of magnitude below that proof. Divide by line count and the cost lands in the low thousands of dollars and the timeline in a weekend. The procedure is the same, and the difference is exactly one: Bun had a test suite, and business systems usually do not. So the first job is extracting the tests — the distillation of the previous piece. Use the running system as the teacher, copy its behavior, put scripts up as the referee, and work through a mechanical queue. A lighter target still is the website, whose behavior is fully exposed from the start — extracting the tests takes a single crawl, and going static removes the maintenance fee and the attack surface along with it. The wall the large-scale proof knocked down was not technical. It was the assumption that "this is beyond us."
The word "distillation" is circulating around Kimi K3 — the technique of putting questions to an expensive model, collecting the answers, and using those input-output pairs to train a cheap model with the same behavior. The internals are not needed; the behavior is enough. The same structure applies directly to business systems, because a test is nothing but putting a question to a system and recording the answer. Once the behavior is in your hands, AI can build another implementation that does the same thing. What is copied is not the vendor's code but how your own business ought to behave — which was yours to begin with. Routine maintenance reduces to security auditing; improvement requires only a reason and test data, and most of the test data the model itself can generate. The reason for an improvement can only come from you, and AI writes the implementation. There is no longer any need to outsource: judgment stays with the company, and the work can be left to AI.
From the Kimi K3 announcement on 16 July 2026, through the letter "Open Weights and American AI Leadership" on the 24th, to the weight release on the 27th, this site wrote eight articles about eleven days. Yang Zhilin's cost-benefit calculus, the AI industry's tower of debt, Microsoft's earnings, three war fronts, the end of IT as an industry, the price of attack, the practice of containment, and the line of the signatures. This piece rereads all of them through four Hegelian concepts. Nadella, who bought the frontier model, and Yang Zhilin, who built one and opened it, are the modern casting of the master-slave dialectic. To sell intelligence you must expose its outputs, and exposed outputs always become teaching material — the master's power is copied the moment it is exercised. A world where attack costs zero is the completed form of alienation, domination by one's own creations, and the practice of containment is the recovery from it. That fifty signatures fell along the line of the P&L is Hegel's own thesis — history advances on interests and passions, not ideals. Trump's threatened restriction produced the largest open-ecosystem coalition in history within three days; Yang's cost-benefit calculus completes the commoditization of capability. The cunning of reason is using everyone's self-interest to close the era of concentrated judgment. What remains as synthesis is the laboring master — the free person who keeps touching reality and never hands over judgment.
On 24 July 2026 a letter titled "Open Weights and American AI Leadership," hosted by NVIDIA, was published. As Washington weighs restrictions on Chinese AI models, the letter urges policymakers not to impose broad restrictions on open-weight models and to punish specific bad actors instead. Signatures doubled from about 25 on day one to 50 within a day, with OpenAI and Google joining late. The major names still missing are Anthropic and Amazon. This piece lays out the facts in sequence and then places two inferences. First, restrict and the substrate of American AI development breaks — the weights publish outside the border, so what a restriction removes is not the world's open weights but the American developers who can touch them. Second, leave them free and every model below the frontier loses its price — in a world where the frontier of a few months ago is opened for free, the only thing left to sell is the gap to the frontier. And that consequence makes no exception for Anthropic — even the winner of the frontier race is left selling an ever-shrinking gap, and large profits end. Either way, the middle product — a reasonably good closed model — does not survive; value moves to the two ends, the side that runs models and the frontier gap. The fifty signatories stand at the ends; the two non-signers sell the gap. The line of the split is the line of the P&L.
Moonshot AI announced Kimi K3 on 16 July 2026 and promised the full weights on 27 July under a modified MIT licence — 2.8 trillion parameters, a one-million-token context. Overall it trails Fable 5 and GPT-5.6 Sol, but it beats Opus 4.8 and GPT-5.5 and took first place in blind frontend-coding comparison: the frontier of a short while ago is about to be opened for free. No individual can run it even then (loading alone takes eight H100 80GB cards), but the premise still changes, because the attacking side is not an individual. Working out what to do about it, almost all of it turned into what to do about Microsoft. Moving the Windows machines to Linux and managing them properly is the best answer, but it cannot be done immediately — so this piece sets down what can be done now, layer by layer. Network (segments, no client-to-client traffic, egress deny by default, your own name resolution), identity (leave Active Directory), documents (turn off macros), applications (make the website static), recovery (an immutable backup). Then it shows that on Windows the premise of containment itself breaks down, because the hardware is tied to the cloud. Setup compels a connection, the key the TPM sealed is deposited in the cloud, and in April 2026 the seller's own update stopped machines from starting. Malware that turns BitLocker itself into ransomware exists, and whether keys escrow to AD or to the cloud, both are concentrations. These restrictions are waived for the large contracts, so there is no reason to expect market pressure to change the direction. Which leaves one conclusion: anyone responsible for a company's information should hold a migration to Linux as a plan.
Part five said the AI revolution drove the cost of software toward zero, and derived self-sufficiency from it. But attacking is software too, and so is verifying. One change produces three things at once — anyone can build, anyone can attack, anyone can check. A cheap attack lands on whatever is concentrated, and delegation is the arrangement that concentrates the judgment to defend at the furthest point available. Microsoft and the systems integrators face the same wall: hold compatibility and you cannot defend, break it and customers leave; if AI can do the maintenance the fee loses its basis, and if it cannot, the labor cost cannot be justified. So take on responsibility, rebuild the structure, and then you need people who can defend — which means the opposite of "specialize in security." The conclusion of the zero-trust trilogy.
The first four parts covered AI's economic deadlines and geopolitics' deadlines. Both were cross-sections of the end of the age of concentration. Part five is about what grows afterward. A product is software, components, materials, and assembly; when the first two go to zero, only the last two remain. The volume of IT work does not disappear — what disappears is IT being bought and sold as an independent industry. The precedent is the business of selling electrification, which vanished once electrification was complete and became an internal function of each factory. The trigger is that AI made verification nearly free: a business standing on information asymmetry cannot stand once the asymmetry is gone. Materials and assembly win for three reasons — the logic of the residual, the demand effect, and the physicality of the barrier to entry. This transition favors Japan and its regions, because the defeat of Japanese manufacturing was never about quality. Fifth of five.
The first three parts covered the deadlines in AI's economy. Part four turns to three events moving outside AI in the same week — the war in Ukraine, the war in Iran, and China's disasters — laid out with no predictions, only facts and observation points. Russia faces a human limit at over 32,000 casualties a month; the US and Israel face the economic limit of a front they opened themselves; China faces the physical limit of aging infrastructure meeting climate. Forced attrition, a chosen war, deferred maintenance — the origins differ, yet the deadlines converge, because the variable that moved the hands is one. Those who entrusted themselves to distributed AI took the initiative in the war of attrition; those who bet on concentration are paying in dollars and in people; the one that concentrated investment in physical production broke at home. What Shanghai's record rainfall showed is that distribution is not the placement of sensors but the placement of judgment. And AI did not only move the hands — it made the dial public, though this piece also marks where that free verification does not reach. Fourth of five.
Microsoft has already been sold off. The stock is down about 20 percent year to date, the forward P/E has fallen to 19 — its lowest in a decade — and the center of the doubt is capital expenditure that has swollen to $190 billion. So what the bull case needs is a single story: that AI expands margins. This piece introduces two variables into that story — Kimi K3's weight release, which makes the model free, and Meta's talks to lease compute to Anthropic, which brings a new seller into infrastructure. When the price ceiling is set from outside, falling input costs do not become profit in a commoditized resale business. And a two-year, monthly, mutually-cancellable contract is the worst possible shape of demand for a facility financed with five- to twenty-year debt. It then identifies the condition that runs through all five parts — a Sonnet-class model running in a desktop box with 128 to 192GB — and narrows the July 29 checklist to five items. Third of five.
The claim that there is no demand for AI does not survive contact with the top two companies. OpenAI is at a $24 billion annual run rate, Anthropic at $30 billion. But the losses and the infrastructure commitments are larger still, and beneath them sits a layer of data-center SPV debt whose total no one can state with confidence. The economic life of a GPU is two to three years; the debt against it runs five to twenty. And the decisive weakness is borrower concentration — almost all of the rent is paid by two companies. This piece establishes that structure in numbers, then shows Kimi K3 acting on it in four directions: compressing pricing power, possibly creating a third class of borrowers, the reverse case where margins erode faster than that class forms, and the politics of who controls distribution. With the DeepSeek shock as precedent, it narrows the verdict to three indicators. Second of five.
Moonshot AI announced Kimi K3 on July 16 and promised the full weights on Hugging Face by July 27. 2.8 trillion total parameters, roughly 1.4TB in MXFP4, with a deployment guide asking for 64 or more accelerators. What ships that day is a mass that no individual — and no small or mid-sized company — can actually run. July 27 still matters. Two premises have shifted underneath it: a US government accusation of distillation with sanctions hinted at, and demand that hit Moonshot's compute ceiling within 48 hours of launch. This piece reads founder Yang Zhilin as a pragmatist to be predicted by economics rather than principle, counts four reasons he has to descend into the empty tier — 128 to 192GB of unified memory — and puts the July 27 checklist and a prediction on record in advance. First of five.
On July 16, 2026, Moonshot AI announced Kimi K3 — 2.8 trillion parameters, the largest open-weights model ever, benchmarking within reach of Claude Fable 5 and GPT 5.6 Sol. Two opposite stories now circulate about this model. The American story says it was distilled from Claude — in February, Anthropic named Moonshot in over 3.4 million illicit exchanges. The maker's story says it caught up through architectural engineering. As a sequel to "zero trust for the user's side," this piece reads the maker's side — choosing a canon, inheriting by distillation, filtering by verification — from the public record of the Kimi K2 technical report, and traces the structure of the moment: the question of which story is true will itself be verified, not argued, when the weights drop on July 27. And so the answer can be graded, a prediction is put on record in falsifiable form: distillation secondary, verification primary, the frontend strength a product of vision-based self-verification — and small specialized models to follow.
I asked a generative AI about exchanging text data between systems, and it answered that "zero trust strangles operations — the most robust option is physical isolation with a data diode." Plausible-sounding, and upside down. The AI was not ignorant of the subject. It traced the center of gravity of the vendor discourse that fills the web around the words "zero trust" — a commercially weighted average of public discourse. This piece dissects one wrong answer to locate the mechanism of the error, then draws out five rules for use — hand it primary sources, choose your own canon, read plausibility and correctness separately, use it where you can verify, and throw the opposing view at it.
Microsoft account takeovers are surging right now. Passwords changed, recovery emails rewritten, and people who prove "this is mine" still can't get their account back. Meanwhile ads have crept into File Explorer — the screen where you view your own files — turning it into a possible entry point for malware. The ad is the entry, the stolen password is the exit; they're one line. Here, in plain language, is what's happening and what to do now.
Starting from the US Supreme Court's June 29, 2026 ruling in Trump v. Slaughter, the structural risk of depending on US technology infrastructure surfaced across five distinct layers at once — the loss of FTC independence shaking the DPF, a dormant CSRB, the structural problems of Entra ID, VBA, and Copilot, the precedent of the Fable 5 export halt, and an 80% drop in AI researchers moving to the US. The conclusion is neither anti-American nor anti-vendor — a structure that entrusts your data, authentication, and operations to the policy decisions of one country or one company carries uncontrollable risk by design.
In June 2026, Axios reported the Trump administration moving to restore access to Fable 5. During the halt, companies switched to Chinese models, so the administration cannot keep it shut. The day Fable 5 returns, a company should do one thing — run every system it operates through Fable to verify it. An AI that can attack can also design and verify. And whatever was optimized on the premise of never being verified — the SIer's deliverables, Microsoft's vast opaque products (Entra ID CVE-2025-55241, EchoLeak) — is exposed whole the moment verification becomes free.
In June 2026, the export ban on Claude Fable 5 and NSA testimony of a classified-systems breach exposed the fragility of cloud dependence at every layer. Cohere and Aleph Alpha's sovereign AI, order-of-magnitude inference-efficiency gains from SSA and TurboQuant, and practical local inference on AMD Strix Halo — the premise behind the $725B cloud-concentrated CapEx breaks, and AI turns toward autonomy, distribution, and diversity. Not as an ideal, but as a technical and economic consequence.
With the release of Claude Fable 5, Stripe migrated 50 million lines of Ruby in a single day. Just as Nadella says, a business application is essentially a CRUD database with business logic. And Fable can rebuild it — one process at a time — using the running system as the answer key and comparing outputs on the same inputs. Start with read-only reports and queries; with Ruby + Sinatra + raw SQL, drop the middle translation (ORMs and shared classes); and replace the old system without ever stopping it. The advantage in maintaining business systems shifts from vendor outsourcing to in-house development.
Boris Cherny, who built Claude Code, calls himself a "builder" and has not written a line of code by hand since last November. "Coding is solved," he says. But what does "humans will no longer be needed" overlook? Building an app is a cycle — finding structure, forming a hypothesis, and verifying it against reality. AI is not a partner that hands you the right answer; it is a tool that gives you the speed to keep up with change. From beside a field in Tokushima, where phosphate fertilizer is running out, and by way of Socrates, the printing press, and a Second Renaissance, this piece asks what it really means to build in dialogue with AI.
Microsoft CEO Satya Nadella wrote on his personal blog about "model overhang — capability outrunning our ability to put it to real-world use." But if that diagnosis is right, the place to invest is not $190 billion a year of infrastructure but the side that uses the capability we already have. AI chief Suleyman has openly said the goal is to "ultimately eliminate" what Microsoft pays Anthropic — exposing a structure in which the intelligence running underneath can be silently swapped out. Ironically, the very "intelligence that can replace an engineer" they spent so much to build has become the best partner for users who want to leave the cloud and self-host. With ONLYOFFICE, Ryzen AI Max, and open-weight models, this piece asks whether you can take a Microsoft 365 (Standard + Copilot) equivalent back into your own hands.
Microsoft's AI-first strategy is creating two problems. The first is internal to Windows 11 PCs — long-term support is not guaranteed past 2029, and the "AI PC" branding is largely cosmetic. The second is the bigger one this strategy itself produces: explosive data center expansion driving electricity demand and emissions. Windows 10's October 2025 end-of-support could push up to 400 million PCs into the e-waste stream, around 88 million tonnes CO2e from manufacturing alone. Microsoft's electricity use is up 168% since 2020 and total emissions are up 23.4%. Doubling data centers in two years while pledging carbon-negative by 2030 is structurally incompatible.
Windows 10 support ended in October 2025. Hundreds of millions of PCs are flagged "not compatible with Windows 11" — yet the hardware is fine. The same hardware runs Linux normally, and old PCs are arguably better suited to Linux. The "Linux is hard" reputation reversed twice in the AI era: AI is unusually good at teaching commands, and Flathub now offers a cleaner GUI app store than the Microsoft Store. With AI beside you, Linux commands are no longer hard. Now is the time to try Linux.
The widely circulated advice — \"in the AI era, become a specialized engineer; hold a deep specialty AI cannot take, like security or ML\" — misreads the structure. AI is absorbing the whole layer of software engineering, not a particular subdomain. The medieval analogue: telling a serf, \"become a more specialized serf and you will be free.\" Freedom comes from stepping out of the lord's structure of control, not from deeper specialization. Software engineer → builder, software engineering → liberal arts, employment → free person — these three transitions are unfolding inside the Second Renaissance (the historical turning point of the AI era, with the LLM playing the role of the printing press). Creation and upheaval are two faces of the same time, and Trump is the canonical figure on the upheaval side. The article compresses aiseed.dev's \"AI-Native Ways of Working — Software\" sub-series (thirteen argument chapters) into three pairs of words.
Microsoft CEO Satya Nadella has poured trillions of yen into AI, made a massive investment in OpenAI, restarted a nuclear power plant, and is building giant data centers around the world. By any conventional measure, this looks like a decisive bid for victory in the AI era. But viewed through Hegel's concept of the "cunning of reason," a different picture emerges. The $13 billion investment in OpenAI that ignited a global AI race, the embedding of autonomous agents into Windows and Office that produced the "Microslop" phenomenon, and the giant data center construction that triggered a global scramble for power and water—these three bets reveal a structure in which the one aiming to be the era's victor ends up playing the role of "the one who closes out the previous era."
"Python only," "declarative," and "cross-platform GUI" — Flet delivers all three for real, and as a UI stack for line-of-business software it reclaims the productivity Visual Basic 6 once had, on a modern foundation. Using aiseed-weather (a mid-size weather studio integrating JMA, ECMWF, and Open-Meteo; Flet 0.85 + Python 3.13; 6,800 lines of components) as the subject, we walk through what Flet actually solves for business GUIs — declarative components, reactive shared state, vector charts in pure Python, native async I/O, embedded terminals, and its synergy with AI-assisted coding — entirely through real code.
In December 2017, I wrote an article on Qiita titled "Python in Excel?" Eight years later, Excel has Python — but every advantage I described back then has been stripped away. No local execution. No VS Code debugging. No Git version control. This is not Microsoft's lack of capability. It is a deliberate business decision to force computation into Azure. But AI has progressed to the point where working without Office is now more efficient. The time to leave the Microsoft cage has come.
As of May 2026, the latest date Microsoft officially guarantees Windows 11 Home/Pro support is October 12, 2027 — just 1 year and 5 months from today. There is no roadmap for Windows 12. There is no guarantee that a PC bought today will run Windows in five years. Windows is no longer a platform — it has become an asset to be drained. There is only one option left: escape the sinking ship.
In April 2026, Elon Musk and Sam Altman face each other in a federal courtroom in Oakland. Musk claims OpenAI "stole a charity." Altman's lawyers reply that Musk simply "didn't get his way." Testimony, cross-examination, and private emails fill the record. Yet both men share the same belief: that AGI can be built with enough capital. Human intelligence, however, rests on four billion years of biological evolution—a body that knows pain, hunger, and the possibility of death. That cannot be bought with capital, data, or compute.
In January 2026, Microsoft made offline use of Windows effectively impossible. Windows has become an OS that cannot run without constant connection to US data centers. Meanwhile, Japan's data centers and submarine cable landing points are concentrated in the Kanto region. The Nankai Trough earthquake will paralyze western Japan with communication failures; the Tokyo Inland earthquake will simultaneously stop cloud services nationwide. Government offices and hospitals will stop functioning, and lives will be lost.
In the enterprise AI market, Anthropic has surpassed Microsoft. Copilot's active usage rate sits at 35.8%, with an NPS of -19.8. Due to the architectural limits of LLMs (O(N²·d)), hallucinations are mathematically unavoidable. Yet Microsoft continues to embed Copilot into the deepest layers of the OS, while its data center build-out drains the world's memory, electricity, and naphtha. This piece argues for an urgent migration to Linux.
By autumn 2026, fertilizer stockpiles run out; from 2027 onward, phosphate fertilizer becomes effectively unavailable in Japan. The product-import route (76% dependent on China) halts, the domestic-processing route hits its limits, and sewage-sludge recovery is blocked by PFAS contamination — taken together, the shift to natural farming becomes not ideology but necessity.
About 40% of GitHub Copilot-generated code contains critical vulnerabilities; more than 15% of commits introduce new bugs. From the moment an AI acknowledged its own structural flaw, through Pearce et al., Veracode, and Fortune 50 findings, to the Model Carousel phenomenon, the 8,000-token ceiling, and the limits of defensive filters.
In April 2026, Claude Mythos and Gemma 4 arrived. Half a year to a year until Mythos is released publicly. The moment it is, attackers gain the same capability. Windows/Office dependency, internal-system black boxes, CMS — the response is urgent. Start today or it's too late.
The Strait of Hormuz blockade is cutting off raw materials for medical supplies, semiconductors, and jet fuel. Meanwhile AI can now find vulnerabilities in hours and auto-generate attack code. These two crises are connected at the root.
Naphtha shortages from the Hormuz blockade spread to food packaging, medical disposables, pharmaceuticals, and defense materials. Japan's reserves buy time but cannot save imported goods. America keeps commodity plastics with ethane but faces aromatic shortages that reach military supplies. China has relative advantage through coal chemistry but remains vulnerable on methanol, LNG, and sulfur.
WordPress powers 43% of all websites — a massive monoculture. Tightly coupled AI plugins bring the same structural vulnerability as the financial system's Copilot problem to the entire web.
Fed Chair Powell and Treasury Secretary Bessent urgently summoned bank CEOs. Analysis of the systemic cyber threat Mythos poses to the financial system through the COBOL-Copilot-SWIFT three-layer structure.
Two forced transitions are unfolding at once. The fertilizer crisis pushes a partial shift to natural farming, while Claude Mythos forces companies to bring system development and defense in-house. Japan's survival strategy across technology, diplomacy, and networks.
Destroyed Middle Eastern fertilizer plants need gas turbines to rebuild, but AI companies have booked production lines through 2029. Food or AI — which gets the turbines? Japan holds the key.
Even after the Strait of Hormuz reopens, restoring destroyed petrochemical facilities will take 3–5 years. Fertilizer feedstocks are petroleum byproducts, making supply recovery a multi-year problem. Japan has an alternative — natural farming.
Found a bug in markdown-it-py where bold fails in mixed CJK/ASCII text. Claude Code identified the root cause, wrote a 6-line fix, forked the repo, and submitted a PR — all in a single session.
Trump threatens to permanently destroy Iran's power plants. But in the Middle East, destroying power plants means destroying desalination. The countries that can't survive without desalination are not Iran — they're Saudi Arabia, UAE, and Israel.
On April 5, 2026, Trump declared 'Tuesday will be Power Plant Day.' What happens if Iran's power plants are destroyed? Chain of retaliation, 62 million lose drinking water, global fertilizer crisis — the world's supply chains stop simultaneously.
Structure doesn't lie.
Current events prove it, every time.
Explore
For the complete structural analysis, see the Insights series.