The signal was buried in a single paragraph. A spokesperson for OpenAI, relaying what an unnamed insider claimed Sam Altman told staff, sketched three possible roads to superintelligence: an IAEA-style international audit regime, a CERN-style consolidation of compute holders, and outright compute monopoly. No date. No named source. No operational definition of the phrase "super AI." And yet within that sparse transmission sits the most revealing thing a frontier lab has said out loud in years โ not about superintelligence, but about what they believe the only enforceable lever over it is.
I've spent nine years watching this industry narrate itself, and I've learned that the most honest documents are the ones nobody meant to publish. This is one of those. Let me walk you through what's actually in it, because the interesting content isn't the three paths. It's the convergence.
The Three Paths All Point at the Same Physical Object
Strip the framing away and look at the mechanics. Path one, the IAEA model, would task an international body with inspecting advanced models and auditing safety measures โ which in practice means inspecting training clusters, because you cannot meaningfully inspect a model without inspecting the machines that made it. Path two, the CERN model, explicitly merges compute holders into a shared international facility. Path three is compute control, stated plainly.
Three options. Zero of them govern model weights. Zero of them govern algorithms. Zero of them govern research culture, publication norms, or open-source release. Every single one reaches for the same object: silicon, cooling, and square footage.
That is not a coincidence, and it is not a failure of imagination. It is the direct consequence of verifiability. Compute is the only part of this stack that can be counted, tracked, and physically seized. A chip has a serial number and a customs declaration. A datacenter has a power draw you can measure from orbit with thermal imaging. Model weights are information โ losslessly copyable, transmissible across borders, compressible into a few gigabytes that fit on a phone.
This is where the IAEA analogy collapses, and it collapses hard. The IAEA's safeguards work because nuclear material is physically irreproducible in the ways that matter. You can fingerprint isotopes. You can do mass-balance accounting โ every gram of fissile material in, every gram accounted for, discrepancies flagged. Fissile material cannot be emailed. A 70-billion-parameter model can be, and routinely is.
So when someone says "an IAEA for AI," they are either describing something that will inevitably degrade into compute governance the moment you operationalize it, or they are describing a press release. There is no third option. The physics doesn't allow one.
The Label Swap That Deserves a Second Look
The reporting attaches the name "the Eliezer Plan" to the compute-monopoly path. This is worth slowing down for, because it is a directional inversion.
Eliezer Yudkowsky's public position โ laid out in the 2022 "AGI Ruin" thread and his 2023 Time op-ed โ is a global, multilateral moratorium enforced by international agreement, with the explicit logic that no single actor should be permitted to win unilaterally. The premise of his argument is that unilateral advantage-taking is precisely the failure mode that ends the game.
What Altman reportedly described is the opposite: one company achieves recursive self-improvement first and monopolizes compute to hold the position. That is not Yudkowsky's plan. That is the worst-case outcome Yudkowsky's plan exists to prevent, restated as an option on a whiteboard.
Now, there is a charitable reading. Altman may have been describing what happens if someone adopts Yudkowsky's logic and executes it unilaterally โ a scenario sketch rather than a citation. From a single unattributed paragraph, I cannot rule that out, and I won't pretend otherwise.
But here's why the charitable reading doesn't fully rescue it. Attaching a critic's name to the implementation of their own nightmare does two things at once. It borrows the critic's moral authority โ "we're taking this seriously, we named the serious option." And it softens the concept by making it discussable. Once compute monopoly has a label and a spot in the taxonomy, it stops being the thing you don't say and becomes a reasonable option that has been considered and set aside. Naming is not neutral. Naming is how an unacceptable idea enters the room.
And note the phrasing: the plan is terrible, mentioned only to cover all possibilities. Name it, then disown it. That is the standard softening maneuver for high-risk proposals, and it works because the disavowal is what people remember while the concept quietly takes up residence in the discourse.
The Voluntary Standards Body Is the Tell
The paragraph mentions a joint safety standards organization forming among the three parties by year-end or early next year. In the same breath, it concedes there is disagreement over who bears responsibility for model safety testing.
That concession is the whole story. Testing responsibility is not one clause among many. It is the only clause that determines whether the organization is a brake or a rubber stamp. Who defines the threshold? Who runs the evaluation? Are results published, and in what form? Is there a veto, or only a communiquรฉ?
Disagreement on that question means the likely output is a statement of shared principles. Shared principles cost nothing. A shared veto costs everything, which is exactly why it won't be in the first draft.
A standards body without a veto is not governance. It is branding with a letterhead.
The Contradiction Nobody Has Resolved
Here is the sentence that should be underlined twice: Altman reportedly favors the IAEA-style path, which requires ceding a degree of sovereignty to an international regulator, while simultaneously stating that OpenAI cannot afford to hand competitive advantage to rivals.
Both cannot be true under a neutral regulator. They can only both be true under one specific design: a regulator whose rules are written by, or heavily shaped by, the current leader. That is not a conspiracy charge โ it is a structural observation with a long historical track record. Regulated entities that participate in designing their regulators tend to get regulators that favor incumbents. Licensing regimes favor those already licensed. Compliance costs are trivial for the leader and existential for the challenger.
The same ambiguity infects the phrase about setting a speed limit on AI development. Measured against the current frontier, a speed limit costs the leader least and the laggard most. Under a safety reading, that sentence is responsible. Under a competitive reading, it is a moat. The two readings are indistinguishable from the outside โ and that indistinguishability is itself advantageous to whoever holds the lead. That is the sharpest ethical ambiguity in the entire transmission, and it deserves to be recorded as such rather than resolved by charitable assumption.
The Chapter That Wasn't Written
Three paths were offered. All three are centralizing. That is the finding.
The missing options are not obscure. Open weights. Capability diffusion. Defensive democratization โ the argument that a world with many capable, widely distributed systems is safer than a world with three controlled ones, because no single point of failure can be captured or coerced. None of these appear in the reported taxonomy.
The option set was trimmed before the discussion began. The power to set the agenda beats the power to choose from it, and this document shows the agenda being set.
There is a second omission, this one geographic. By the 2023โ2025 window this conversation sits in, the regulatory environment moved repeatedly โ US executive orders, the EU AI Act entering its enforcement phases, export controls rewritten more than once. Against that backdrop, any "global" body is realistically a Western three-party arrangement with voluntary accession. China, the EU, the open-source community, Meta, xAI โ absent from the frame entirely.
A governance architecture covering three labs facing ten to fifteen frontier developers worldwide is not a regime. It is a leaky bucket with good intentions. The leakage rate determines how fast the bucket empties, and nobody has published the leakage rate.
The Answer to the Riddle Is Whether There's a Pause Button
One question cuts through all of this. Does OpenAI's own Preparedness Framework contain a hard pause provision that can actually be triggered?
If yes โ has it ever been triggered? If never, what is its real-world constraint value? A framework that has never stopped anything is a document, not a control. If no, then the entire three-path discussion is upstream of the only mechanism that would matter, and the paths are theater.
The second question is what the meeting was for. Decision solicitation, risk education, or expectation management โ preparing staff for an unpopular decision by pre-framing the menu? The reported text cannot tell us, and the reaction of employees is not recorded. That absence is informative in itself.
And then the question that makes everything else provisional: what does "super AI" mean here? A system that surpasses humans at nearly all cognitive tasks, or something narrower? The definition sets the urgency by an order of magnitude. It determines whether the relevant window is two years or twenty. Without that number, no timeline attached to any of these paths can be graded, and every estimate floating around the discourse is a guess wearing a suit.
What I'm Watching Next
The compute frame is not a strategy. It's a constraint discovered under pressure โ the last measurable object in a stack that has become unmeasurable everywhere else. That's why all three doors lead to the same room.
What matters now is narrower and more concrete. Watch the charter of that joint standards body when it appears. Read the testing-responsibility clause first, before anything else, and check for the word "veto." Watch whether any lab in the arrangement publishes an evaluation it failed. Watch whether the definition of the threshold capability is written down in numbers a third party could measure, or in adjectives a lawyer could defend.
In the meantime, the honest position is this: the most revealing part of the leak was never the three paths. It was the shape of the room they were drawn in โ a room with three doors, all of them locked from the same side, and no window labeled open weights.
That's the signal in this particular patch of static. The rest is still noise.