The Long View — March 2-8, 2026

In 1998, the political scientist James C. Scott published Seeing Like a State, a study of how governments and institutions impose order on messy realities. His central case studies ranged from Prussian forestry to Soviet collectivization, but his thesis was simple: when powerful institutions try to make the world legible from above, the world pushes back from below. The imposed order looks clean on paper. The actual territory is always more complicated.

This was the week American AI governance tried to become legible. A presidential executive order attempted to create a single federal AI standard, overriding the patchwork of state laws that have been the primary source of AI accountability in the United States. Within days, Florida's Senate passed an "AI Bill of Rights." Colorado introduced medical AI guardrails. The states did not get the memo, or they got it and decided it didn't apply to them.

The same pattern repeated everywhere you looked. OpenAI launched GPT-5.4, its most capable model ever, an attempt to consolidate the capability race under one product. And on the same day, a senior OpenAI researcher resigned over the company's Pentagon partnership, the latest fracture in an organization that has been losing safety-focused talent for two years running. Oracle abandoned its flagship Texas data center expansion and started cutting thousands of jobs, the infrastructure correction arriving before the infrastructure was built. Anthropic committed $20 million to political spending on AI regulation, positioning itself against the deregulatory tide. Every move toward consolidation produced its equal and opposite fragmentation.

Scott had a word for this: metis. The practical, local knowledge that resists being captured by any grand scheme. This week, the grand schemes multiplied. So did the resistance.

The Model and the Exit

OpenAI released GPT-5.4 on Wednesday with three capabilities that, taken together, change what "using AI" means in practice: a one-million-token context window, native computer control, and what the company calls an "extreme reasoning" mode. The context window alone eliminates the need for retrieval-augmented generation in most applications. You can paste an entire codebase, a full legal filing, a year of financial records directly into the prompt. Computer use means the model can operate a browser and desktop applications without human intervention. Combined with extended reasoning, this is a system that can read large volumes of material, think about them at length, and then act on its conclusions through a computer interface.

The benchmarks are strong. OpenAI reported GPT-5.4 outperformed or matched human professionals on 83 percent of tasks on GDPval, a professional capability benchmark the company developed. That qualifier matters: OpenAI built the benchmark. Self-reported evaluations are marketing, not peer review. But the third-party results that have trickled in over the week are consistent with the claim that this is a step function in capability.

The counterpoint came from inside the house. A senior researcher resigned over the company's deepening military partnership with the Pentagon, per Politico and Gizmodo. The departure follows weeks of escalating tension between AI companies and the defense establishment. Anthropic was designated a "supply-chain risk to national security" the week before for refusing unconstrained military access to Claude. OpenAI filled the vacuum with its own Pentagon deal. And now the people inside OpenAI who understand how models fail are starting to leave.

The organizational psychologist on the council noted the structural problem: one public resignation means dozens of private deliberations. Every safety researcher who exits removes an internal check on deployment risk. The technology gets more capable in the same week that the institution deploying it gets less equipped to understand its own failure modes. Those two curves should worry anyone paying attention.

Meanwhile, Anthropic's Claude discovered 112 security vulnerabilities in Firefox, including 22 CVEs and 14 high-severity bugs, with the first critical flaw found in under twenty minutes. This is what happens when model capabilities meet serious application: not chatbot novelty but genuine security research at a pace no human team could match. The gap between what these systems can do and our institutional readiness to govern their use widened this week on both ends.

The Infrastructure Correction Nobody Planned For

The quieter story of the week played out in Abilene, Texas, and in Oracle's boardroom. Bloomberg and Reuters reported that Oracle and OpenAI have abandoned plans to expand their flagship data center site, a project that was supposed to anchor the Stargate AI infrastructure initiative. Community opposition, energy supply constraints, and cost overruns killed it. The next day, Tom's Hardware reported Oracle is cutting thousands of jobs and freezing hiring as the financial burden of its AI data center bets becomes unsustainable. The jobs being cut will be "backfilled with AI," according to the company.

This is what a correction looks like before anyone calls it a correction. The promises made during the 2024-2025 AI infrastructure euphoria are colliding with physics and balance sheets at the same time. Local politics are doing the rest. On Wednesday, the White House brokered a pledge from seven major tech firms to self-fund their data center power rather than drawing from local utility grids. The framing was "we'll pay our own way." The subtext was that voter backlash over electricity bills near data center clusters has become a political liability heading into the midterms.

The economist on the council pointed out the competitive implications: only hyperscalers can afford to build their own power plants. The companies that signed the White House pledge are the same companies that already dominate cloud computing. If energy access becomes a competitive moat, the barrier to entering the AI infrastructure market just rose by several billion dollars. The pledge that sounds like responsibility is also a consolidation play.

Meanwhile, the gap between announced AI data center capacity and actual buildable capacity keeps widening. Oracle, which committed to some of the most ambitious expansion plans, is now simultaneously canceling projects and cutting the people who would have built them. The War Correspondent on the council, who has covered plenty of economic booms that produced more press releases than buildings, recognized the pattern. The promises travel faster than the concrete.

Who Writes the Rules

The governance collision that had been building all winter finally arrived this week on multiple fronts at once.

President Trump signed an executive order establishing a federal AI policy framework that directly threatens state-level regulations. California, which has been the most aggressive state on AI accountability, took the hardest hit. The order signals intent to create a single federal standard and use legal action against states whose laws conflict with it.

The states responded by accelerating. Florida's Senate approved an "AI Bill of Rights." Colorado introduced guardrails for AI in medical systems. Missouri and Arizona are debating their own approaches. The political dynamic is unusual: this is not a blue-state-versus-red-state split. Florida is a Republican-led legislature. The push for AI governance crosses party lines in a way that the federal preemption order does not account for.

Anthropic's $20 million political commitment, reported by CNBC, the Wall Street Journal, the New York Times, and Reuters, adds a corporate dimension. The company is spending to support AI regulation ahead of the 2026 midterms, directly countering OpenAI's deregulatory positioning. The New York Times framed it as Anthropic "declaring war" on OpenAI's lobbying approach.

The contrarian on the council raised the question nobody in the coverage is asking: who benefits from Anthropic's preferred regulatory framework? Corporate-funded regulation tends to favor the funder. If compliance costs rise, large established companies can absorb them. Open-source projects, university researchers, and startups cannot. "Regulation" is not inherently protective. The details determine who gets protected and who gets priced out. Anthropic's safety reputation makes the political spending hard to criticize, which is exactly what makes it worth scrutinizing.

Alongside the political maneuvering, new chip export rules are taking shape. Reuters reported the US is developing requirements for foreign firms to invest domestically as a condition of accessing American AI chips. This is export control policy evolving from simple denial into a bilateral trade instrument, weaponizing semiconductor access against allies as well as adversaries.

These are not separate governance stories. They are one story about a power vacuum. The federal government wants control but is attempting it through preemption rather than substance. The states want accountability but lack coordination. The companies want favorable rules and are willing to spend nine figures to get them. Nobody is writing the rules from a position of public interest. Everybody is writing them from a position of advantage.

The Bookshelf

James C. Scott, Seeing Like a State (1998)

Scott spent his career studying what happens when large institutions try to impose rational order on complex, organic systems. His examples range from the catastrophic (Soviet collectivization killed millions) to the mundane (Prussian scientific forestry produced a monoculture that collapsed within a generation). The pattern is always the same: the institution sees the world through its own categories, ignores the local knowledge that doesn't fit those categories, and is then surprised when the world doesn't cooperate.

The book's most useful concept for this week is "legibility." States want to make populations, economies, and territories legible. Legible means countable and governable. The Trump executive order is a legibility project: replace fifty different state approaches with one federal standard. The White House energy pledge is a legibility project: make AI's power consumption visible and manageable. Oracle's data center plans were a legibility project: turn a map of the American Southwest into a grid of compute capacity.

Every one of those projects ran into what Scott calls the resistance of the local. States didn't comply. Voters objected to their electricity bills. The land in Abilene turned out to be harder to build on than the spreadsheet suggested.

Scott's conclusion, refined across 445 pages of historical evidence, is not that planning is bad. It is that planning which ignores what it cannot see will fail. Read it this week and notice how many of the grand AI schemes announced in 2024 and 2025 are now meeting their local resistance in 2026.

The Week Ahead

Three questions worth tracking:

Does DeepSeek V4 actually ship? The model has been "imminent" for over a week, timed to China's Two Sessions political meetings. If V4 delivers competitive multimodal performance on Chinese-manufactured chips while being withheld from US chipmakers, the entire export control strategy needs rethinking. The announcement will matter less than the benchmarks.

How many companies follow the Oracle pattern? Data center cancellation plus workforce reduction is a specific combination that signals capital discipline replacing capital euphoria. Watch for similar announcements from second-tier cloud providers in the next two weeks. The infrastructure correction may still be in its early innings.

Does Anthropic's lawsuit against the Pentagon's "supply-chain risk" designation move forward? The legal theory matters more than the outcome. If Anthropic challenges the designation on constitutional grounds, it establishes precedent for whether the government can compel AI companies to remove safety constraints. If they settle quietly, the precedent stands unchallenged, and every other AI company's safety commitments become negotiable.