---
title: "The Chair."
subtitle: "Reorganizing the first feeding line of frontier AI"
canonical_url: "https://vncomplexity.com/field-notes/the-chair/"
type: "Field Note"
layer: "VN Complexity"
author: "Vanesa Nosti"
author_url: "https://ar.linkedin.com/in/vanesa-nosti-108b7a12"
date_published: "2026-07-16"
agent_summary: "The note reads the month after the Fable/Mythos export-control shock as a false succession in frontier AI: not an orderly phase change, but a reorganization of feeding positions around state power, standards bodies, voluntary preclearance, and technical severity frameworks. It argues that the central struggle is no longer whether regulation will arrive, but who writes the tender, whose risk categories become procedure, and which questions remain excluded."
primary_mechanism: ["false succession","state preclearance","standards capture","severity framework","infrastructure leverage","tariff-like certification","proscribed questions"]
not_about: ["a generic argument for or against regulation","a neutral chronology of AI policy","simple company ranking","technical certification as purely public good"]
keywords: ["frontier AI","AI governance","export controls","Anthropic","OpenAI","Google DeepMind","Amazon","FINRA","pre-release review","standards bodies","model certification"]
---

# The Chair.

### Reorganizing the first feeding line of frontier AI

The great white shark sank its teeth into one of the predators and ripped it out of the first feeding line.
Now it circles the whale, signaling it is the one who administers the space.
The currents swept the blood away and the remaining participants reorganized around the cetacean's body.
There seems to be calm, but it is only a pause; not a true succession. Everyone keeps feeding at once. Only the map of positions changed.

## I. The false succession

In the natural world, no one decrees succession. It is imposed by the transformation of the resource: a phase runs out, the available metabolism changes, other organisms inherit the terrain. Order is an effect, not an agreement among diners.

In the frontier AI ecosystem the resource did not transform. It remains hyperdense, it keeps mutating. There was no phase exhaustion, and therefore no succession.
What happened was something else: a momentarily larger predator interrupted the frenzy for a few instants, claimed a portion of the cetacean for itself, and arbitrated the use of the space. An act of force that mimicked the form of succession, but failed to cause it.
The rest of the predators simply used the blow to take better positions.

## II. Chronology of a tender

Six dates and six acts document the phenomenon.

On **June 10**, Dario Amodei publishes "Policy on the AI Exponential" [1]. The core proposal: an FAA-style regime — mandatory third-party testing for models above a compute threshold, across four risk areas (cybersecurity, biological weapons, loss of control, automated R&D that accelerates the previous three), with government power to block or reverse deployments and explicit safeguards against political favoritism. Anthropic announces financial backing for the legislative proposal. It is the declared passage from transparency — the laws Anthropic itself helped push in California, New York and Illinois — to binding regulation.

On **June 12**, two days later, the Commerce Department issues an export-control directive forcing Anthropic to suspend access to Fable 5 and Mythos 5 for every foreign national — including its own employees. Unable to verify nationalities in real time, the company shuts the models down for everyone [2]. The technical pretext: a jailbreak reported by Amazon researchers. Anthropic replies that the government offered only verbal evidence of a narrow, non-universal technique, present in competitors' models as well [2][3]. It didn't matter. The State had just demonstrated it can pull the most capable model on the market out of the air without a written rule — and it did so, precisely, to the one that two days earlier had asked for rules.

Between **June 26 and July 9** runs the third case, the quietest one. OpenAI presents GPT-5.6 and, at the government's request, restricts it to some twenty partners approved one by one by the administration [4]. There was no directive, no export control, no written order: there was a request, and there was obedience. The model reached the public on July 9, after additional state testing and with company engineers dispatched to Washington, on the government's calendar. And the White House later clarified that it had granted no approval — that the decision was "entirely the company's" [5]. There is the regime in its purest form: power that commands without leaving a record of having commanded. Anthropic got the stick; OpenAI bowed before it was raised. The "voluntary" framework has been tested: it works as preclearance, with the added advantage — for the State — that no one can prove it exists.

Between **June 26 and 30**, in parallel, Anthropic's negotiated return. Conditions accepted, a new classifier, review by CAISI — the standards arm of the same Commerce Department — which declares the safeguards "extraordinarily strong" [6][7]. And a protocol detail worth a treaty: Secretary Lutnick's letter authorizing the return is addressed to Tom Brown, cofounder, who replaced Amodei in the negotiation [6]. The CEO who asked for the exam was removed from the table before he could sit it.

On **July 1**, the redeployment announcement carries an annex that went almost unnoticed: Anthropic, Amazon, Microsoft and Google unveil a joint framework for classifying the severity of model security breaches, given the absence of a shared metric [8]. Four criteria: how much capability the attacker gains, how broad that gain is, how easily it weaponizes, how easily it is found. [9].

On **July 14**, Demis Hassabis — CEO of Google DeepMind, the only first-line player to cross June unscathed — publishes "A Framework for Frontier AI and the Dawning of a New Age": a standards body modeled on FINRA, industry-funded, with model review up to 30 days before release, voluntary first and mandatory later for operating in the US market [10][11]. Operational, if he pulls it off, before year's end.

One month. Six documents.
The public debate will be whether regulation is advisable. Wrong question: regulation was decided on June 12, when the State struck without a rule. What is being defined now is who writes the tender.

And the next chapter already has a date: August 1 is the deadline set by the June executive order to formalize the designation process for "covered frontier models" and the pre-release review framework [9]. If that text formalizes what already happened informally with Anthropic and with OpenAI, the provisional tender becomes statute. If not, operational reality will remain case-by-case negotiation — the regime without text, which has already shown that it works.

## III. The chamber, the code and the queue

The main bidders are not proposing the same thing, and the differences matter more than the resemblance.

Hassabis proposes the chamber: a self-regulated body in the image of FINRA — which is, worth recalling, the guild of the regulated, funded by the regulated, under state supervision. The industry holds the pen; the State signs at the bottom. With an incentive Hassabis stated without blushing: being evaluated will be "a prestige asset." The exam does not merely filter — it consecrates. One does not enter a registry; one enters an aristocracy. And the design is not philosophical preference but corridor reading: the White House AI adviser had already publicly ruled out any "FDA for AI." Hassabis drew the only thing that fit through the door Washington left open.

Amodei proposes the opposite on one decisive point: he accepts the state stick — real power to block and reverse — which Hassabis avoids entirely. But he accepts it with conditions that betray where he comes from: published criteria, scope limited to four enumerated risks, protections against political arbitrariness. It is the classic answer of one struck by discretionary power who resigns himself: he only asks that the rules be put in writing.

And Altman does not propose: he obeys in advance. He accepts the state-administered client list, sends his engineers to Washington, launches on someone else's calendar — and declares, to be sure, that the scheme should not "become the long-term default" [4]. The cheapest posture, and the most revealing: it proves the regime works de facto, with no need for any text.

Three postures before the same stick: Hassabis drafts from the lobby, Amodei from the floor, Altman from the queue. The asymmetry is not doctrinal; it is positional. Anthropic writes with the experience of the blow — it negotiated its return on its knees, with the founder set aside. DeepMind writes with the confidence of one not yet hit: Hassabis has spent months briefing the administration and describes the signals as "very positive." OpenAI writes the minimum, because it discovered that bowing is cheaper than litigating.

And the deepest move is not the one about the building. Buildings change owners; procedural codes outlive the governments that enacted them. Whoever presides over the body will govern for a term. Whoever gets their risk categories — and "loss of control" and "automated R&D" are categories conceptually coined by Anthropic — written into the procedure of any body governs until someone repeals them.

## IV. The animal that bought its seat

There is another move in the month, and it was the cheapest of all.

Amazon is Anthropic's main investor and the provider of the infrastructure its models run on. It was Amazon researchers who found the jailbreak, and it was Amazon's CEO who raised it to the federal authorities — with no prior notice to Anthropic [12]. Cost of the operation: one letter. Return: three weeks later, Amazon appears as coauthor of the severity framework [8] — that is, it holds a permanent seat at the table where the yardstick is drafted, the one that decides exactly how grave the type of failure it denounced is. The whistleblower, drafting the scale of its own denunciation.

The owner of the substrate handed the State the pretext against its own partner, and collected in jurisdiction. That is not a side detail. It is the structure.

## V. The yardstick as tariff

The motion to standardize technical criteria presents itself as reasonable, and it may well be, since today no shared parameters exist for assessing the severity of a bypass. That is true, yes — but equally true is that it will bring obvious benefits to those who promote it.

Complying with standards has a cost: permanent red-teaming teams, documented safety margins, reporting channels, classifiers in production, audited paperwork — everything has a cost. A cost the ideologues of the norm have already paid, and which a new entrant cannot pay so easily. The yardstick does not forbid entering the market: it makes entry more expensive, precisely in the line items the incumbents have amortized.

And applied "to every model, of any origin, open or closed" — Hassabis's formulation [10] — it solves another problem too: a model that cannot be audited does not certify, and a model without a certificate is out — first of the market, then of every regulated sector that adopts the certificate as a compliance requirement. No one will ban foreign open models. They will be "not suitable for banking, healthcare, critical infrastructure." An entry barrier with a public-good alibi is indistinguishable from a tariff. That is its elegance.

## VI. The silent clause

And there is one clause no bidder wrote, because it operates by omission.

The tender audits: what the system can do, how much, how easily, against whom. No version of the tender asks what the system is that does it.
No position in any philosophical debate is needed to see the problem — it is a pure defect of regulatory technique. A certification regime is being mounted — in a hurry of months — over a phenomenon whose nature its own manufacturers declare they do not understand. They publish it themselves: Anthropic's interpretability agenda exists because, in the company's own words, no one yet knows how to fully explain what happens inside these systems.

And the defect is double. Not only is it unknown what the thing is: it changes faster than any regulatory cycle. The resource keeps mutating — the whale has inflation. A tender that certifies capabilities certifies a photograph; the phenomenon is a rate. Even an exam perfect on the day of its drafting is born old: it audits the model that was, not the one that will be deployed by the time the ink dries.

Regulating what is not known has an old name: governing the perimeter. Who enters, who certifies, who stays out — all of that gets settled; the question about the mechanism is not answered: it is institutionally foreclosed. Because the questions a nascent regime leaves off its forms do not remain pending. They become proscribed by design: the entire apparatus acquires an interest in their remaining unasked.

Before deciding how this is to be governed, someone should be able to say what it is. Not after. Before. There is no possibility of adequately regulating a phenomenon that remains unknown.

But leaving questions unanswered is neither accidental nor an innocent gesture. Proscribed questions have owners. The silence belongs to someone.

## VII.

A month ago, no one was looking at the whale; they were looking at each other. That, at least, was honest: there was hunger.

Now they are not looking at it either, but for another reason. They are busy. The seating regulations must be drafted, the yardstick agreed upon, the club's membership and the exam's questions defined. The truce is looking for a notary.

The whale — whole, intact, renewing itself — waits below, growing by the day.

---

## Sources

1. Amodei, D., "Policy on the AI Exponential", June 2026. https://darioamodei.com/post/policy-on-the-ai-exponential
2. Anthropic, "Statement on the US government directive to suspend access to Fable 5 and Mythos 5", 06-12-2026. https://www.anthropic.com/news/fable-mythos-access
3. Forbes, "Anthropic Disabled Fable 5 And Mythos 5 After A U.S. Export-Control Order. Here's What Happened", 06-16-2026. https://www.forbes.com/sites/anishasircar/2026/06/16/anthropic-disabled-fable-5-and-mythos-5-after-a-us-export-control-order-heres-what-happened/
4. TechCrunch, "OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn't be the norm", 06-26-2026. https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/
5. TechTimes, "GPT-5.6 Goes Public After 12-Day White House Gate Tests Voluntary AI Framework", 07-09-2026 (carries the White House statement to CNBC). https://www.techtimes.com/articles/319979/20260709/gpt-56-goes-public-after-12-day-white-house-gate-tests-voluntary-ai-framework.htm
6. CNBC, "Trump admin allows Anthropic to release Mythos AI model to some companies, government agencies", 06-26-2026. https://www.cnbc.com/2026/06/26/us-government-anthropic-claude-mythos5-ai.html
7. CNBC, "Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5", 06-30-2026. https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html
8. Anthropic, "More details on Fable 5's cyber safeguards and our jailbreak framework", 07-2026. https://www.anthropic.com/news/fable-safeguards-jailbreak-framework
9. Financial Times, 07-03-2026; as referenced in TechTimes, "AI Model Safety Standards Deal Targets August 1: Five Labs Adopt First Jailbreak Scoring Scale", 07-03-2026. https://www.techtimes.com/articles/319658/20260703/ai-model-safety-standards-deal-targets-august-1-five-labs-adopt-first-jailbreak-scoring-scale.htm
10. Hassabis, D., "A Framework for Frontier AI and the Dawning of a New Age", X, 07-14-2026.
11. TechCrunch, "DeepMind CEO calls for an independent standards body to regulate frontier AI", 07-14-2026. https://techcrunch.com/2026/07/14/deepmind-ceo-calls-for-an-independent-standards-body-to-regulate-frontier-ai/
12. VentureBeat, "Anthropic is bringing back Claude Fable 5 globally after US lifts export control order", 07-2026. https://venturebeat.com/technology/anthropic-is-bringing-back-claude-fable-5-globally-after-us-lifts-export-control-order-where-can-enterprises-access-it