Epistemic status: I am confident in the coordination-problem diagnosis; considerably less confident in the specific enforcement architecture, which I expect to be substantially wrong in its current form (see sections 6 and 7).
The bottleneck is not the architecture, it's whether the actors who'd need to negotiate are willing to sit down at all. If that ever happens, the technical work could be the easy part. Section 8 has what I'm actually asking for.
The A.L.E. Project explores one possible way out of this problem. Rather than presenting a finished regulatory system, it proposes a concrete governance architecture built around physically verifiable constraints — including energy consumption, semiconductor supply chains, compute infrastructure, and orbital thermal dissipation — together with complementary verification and enforcement mechanisms.
The architecture is deliberately open to revision — a starting structure for the actors who would have to negotiate, not a finished system.
The full framework, The A.L.E. Project, is available on SSRN:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7187658
What follows is a summary of the central problem, the proposed architecture, and the questions I believe need to be answered.
Methodological Note
This post and the underlying paper are my own work. I used AI systems as tools — for research, stress-testing arguments, and editing — under my direction, and I'm responsible for everything here, including where section 7 describes AI-generated critique changing my mind.
1. The Infrastructure Coordination Challenge
Most governance proposals begin by asking how to govern increasingly capable models. That is an important question.
The A.L.E. Project does not attempt to replace capability-risk or alignment discussions. Rather, it explores an additional layer of analysis that may become increasingly relevant as frontier systems scale: how do competing actors coordinate around a technology whose development depends on rapidly expanding physical infrastructure?
Every major actor in the frontier ecosystem faces essentially the same dilemma. A laboratory that slows development risks losing ground to competitors. A government that acts unilaterally risks strengthening rival jurisdictions. An investor who becomes cautious risks watching competitors capture the upside. This is a multipolar trap in the fullest sense of Scott Alexander's "Moloch" essay — coordination collapsing under competitive pressure, applied here to compute. Many actors may recognize the same risk while remaining individually incentivized to accelerate.
This isn't unique to AI — telecom, aviation, and semiconductor manufacturing show similar patterns, where competitive pressure pushed collective investment past what coordination would have produced. Frontier AI now shows the same shape: a single frontier training cluster can draw power comparable to a major industrial facility.
Historical experience also suggests that efficiency gains alone may not resolve this pressure. Under conditions consistent with Jevons's Paradox, reductions in the cost of computation can increase, rather than decrease, aggregate demand for compute.
None of this by itself proves a governance intervention is necessary. But it raises a coordination question that most existing governance discussions treat as secondary — and it's worth being explicit that this is closer to the compute-governance literature (Sastry et al., 2024, "Computing Power and the Governance of Artificial Intelligence" https://arxiv.org/abs/2402.08797 ; Shavit, 2023, "What Does it Take to Catch a Chinchilla?" https://arxiv.org/abs/2303.11341 ) than to alignment or capability-risk work. The A.L.E. Project sits in that tradition and tries to extend it into a full sequencing and enforcement architecture.
Readers already steeped in this forum's own compute-governance tradition — Anderljung and Carlier (2021, https://forum.effectivealtruism.org/posts/g6cwjcKMZba4RimJk ) on compute as a natural governance node, and the June 2026 post "Beyond the Threshold" - Taha Iqbal (https://forum.effectivealtruism.org/posts/LLR4G4EXDcao4WYFB/beyond-the-threshold-designing-compute-governance-that ) will recognize the starting premise. What follows tries to go further: a full sequencing, verification, and enforcement architecture built on it, not just an argument for why compute is governable in principle.
Why add another proposal here? Existing compute-governance work mostly does one of two things: establishes that compute is governable in principle (Anderljung and Carlier; Sastry et al.), or advocates near-term levers inside institutions built around US and allied interests. What's neglected is the sequencing and enforcement architecture connecting the two, treating non-aligned participation — China above all — as a design constraint from the start. Whether it fills that gap well is a separate question (section 8).
Even readers unconvinced by long-term AI risk arguments may find reason for concern here. Uncontrolled escalation creates not only governance challenges but capital-allocation risk, infrastructure exposure, and the risk of large-scale value destruction through competitive overbuild.
This isn't an abstract problem waiting for a solution to appear before it matters. The one Track 1 government-to-government dialogue on AI safety, held in Geneva in May 2024, produced no formal follow-up; as CSET's Helen Toner put it in late 2025, the US and Chinese governments are "barely talking at all" on this. Track Two channels aren't filling that gap by default either — they only work under conditions this paper takes seriously in section 6.
2. The Hobbesian Bootstrap Paradox
This observation led me to what I call the Hobbesian Bootstrap Paradox.
No sovereign government can credibly design or impose enforceable operational limits on frontier technology without the cooperation of the very actors it seeks to regulate. Yet those actors cannot safely offer that cooperation or voluntarily slow down unless a credible governance structure already exists to protect them from defection by competitors. The authority required to enforce coordination depends on coordination existing first.
This is a specific instance of a general commitment problem in international relations — actors would benefit from a binding agreement but no one can credibly commit first without a third party to enforce it. Many proposals assume regulation alone breaks it. I'm not convinced.
3. Why Physical Constraints?
Most governance proposals focus on digital artifacts: model weights, training runs, deployment protocols, organizational commitments. These may all be useful. But they share a common weakness — software can be copied and moved across jurisdictions in ways that physical infrastructure cannot.
The A.L.E. Project explores a different anchor: can governance be tied to constraints that remain observable even when software does not?
The framework anchors on physically verifiable infrastructure — energy consumption, chip supply chains, compute infrastructure, thermal signatures — because physical systems leave footprints harder to conceal than code. A 500 MW terrestrial data center shows up in aggregate grid draw; any orbital compute system can't dissipate heat without a radiator bounded by the Stefan-Boltzmann law — a hard physical constraint, not a policy choice, and not specific to any one operator — Kepler Communications, Axiom Space, and Starcloud are already testing smaller-scale orbital compute under the same physics (current publicly-claimed specs for SpaceX's Starmind/AI1 are vendor-declared, not verified — treated here as provisional, a distinction I tried to hold throughout; more on where that slipped, in section 6).
Physical constraints are not proposed as sufficient by themselves. They're the foundation that additional verification layers build on top of. Ordinary chip telemetry (power, thermal, utilization) is mature and already deployed; the cryptographic, capability-detecting layer this needs — Shavit (2023, https://arxiv.org/abs/2303.11341 ); FlexHEG (Petrie and Aarne, https://arxiv.org/abs/2506.03409 ) — is an active hardware-governance research frontier, not yet a deployed product for this purpose. The objective isn't to replace every other governance tool; it's to establish one anchor that's difficult to evade even when everything else fails.
4. The Montreal Protocol Analogy — And Where the Precedent Actually Points
The framework draws on the 1987 Montreal Protocol — not for the chemistry, for the sequence: a framework convention (Vienna, 1985) before a binding protocol with numbers (Montreal, 1987). I borrow the sequence, not the substance.
The A.L.E. Project mirrors that sequence: a private-sector technical agreement among labs first — a Vienna-like framework — then intergovernmental ratification filling in binding numbers. DuPont only backed limits once it had commercially viable substitutes. Do frontier AI labs have an equivalent to "just race harder"? I don't know. The paper poses that question; it doesn't answer it.
5. What the A.L.E. Project Actually Proposes
The full paper develops this into an operational architecture: tiered enforcement based on physical metrics (installed power, radiating surface, verified FLOP), a technical oversight body with anti-capture composition rules, a computational freeze paired with an "Asymmetric Convergence Corridor" so it doesn't just lock in the current leader's advantage, an incentive structure aimed at making Chinese participation rationally preferable to staying outside, and orbital infrastructure as a jurisdiction-evading pathway a terrestrial-only framework would miss. That's deliberate: the goal is to remove as much technical work as possible from the path of whoever eventually sits down, so agreement doesn't have to start from a blank page.
The specific thresholds, committee structure, and implementation details should not be treated as fixed points. Some are almost certainly wrong. The purpose isn't a finished treaty text — it's a starting architecture built to be criticized, refined, or replaced.
Any serious change to the current trajectory will necessarily affect existing interests, incentives, and competitive positions. That is not a feature of the problem I have overlooked; it is part of the problem itself. The objective is therefore not to pretend that coordination has no distributional consequences, but to design mechanisms through which the interests and advantages affected by coordination can be rebalanced elsewhere in the architecture.
6. What I Don't Know
The paper names its own weak points explicitly; here are the ones most worth surfacing.
Who convenes the first conversation. Not the first treaty — the first informal conversation among people with real access to the labs that matter. As an example, Pugwash (Cold War scientists meeting before governments signed nuclear treaties) needed a convener with technical credibility in both blocs and no commercial stake. I don't know who plays that role for an AI ecosystem larger and more commercially entangled than Pugwash-era physics. This isn't hypothetical: a case study on this forum by rani_martin (https://forum.effectivealtruism.org/posts/ggiCDnYcSKLxwFbBv/the-pugwash-conferences-and-the-anti-ballistic-missile) found Track II dialogues with infrequent contact and low trust tend to underperform.
Whether evasion is fully closed. The paper's Aggregation Rule targets fragmented training across many small clusters — network logs, synchronization APIs, cloud audits — and Heuristic Telemetry targets the efficiency gains that decouple FLOP from capability, the same gap between reported training compute and delivered capability that DeepSeek's release made public. The Aggregation Rule is close to deployable; Heuristic Telemetry's cryptographic layer (Shavit 2023; FlexHEG, https://arxiv.org/abs/2506.03409 ) is real active research, not yet fielded, and neither has been tested against an actor actively trying to evade it. Neither has any comparable regime: IAEA safeguards, the NPT, and the Chemical Weapons Convention were all deployed against real adversaries without first being pre-tested against a fully motivated evader. This doesn't clear a bar no comparable regime has cleared.
Whether the industrial-privacy mechanism actually scales. The paper proposes a way for labs to prove compliance (via energy consumption and declared hardware efficiency class) without exposing weights or architecture to a committee that could include competitors. It's designed for a small number of catch-up actors; whether it holds at the scale of every lab above threshold is untested.
Whether orbital verifiability is real yet. Infrared satellite monitoring of orbital compute is not uniform or universally available today, and not every method of concealing a radiating surface is known or knowable at this point.
Uneven rigor across sections. I tried to consistently separate established fact from vendor-declared projection (I think I mostly succeeded on the orbital-thermodynamics section), but some claims — particularly about how fast China's chip industry is closing the gap with Western foundries — are stated with more confidence than the underlying data currently supports.
If the underlying coordination failure is real, identifying a credible convening mechanism may matter more than refining any single governance principle further. Without a forum capable of attracting the actors who matter, even the most carefully worked-out proposal stays theoretical.
7. Two Things That Happened Since I Wrote This
Since finishing this paper, two things happened that bear directly on its central design choice — how compliance actually gets enforced — and I'd rather put them in front of you than quietly patch the paper before anyone notices.
First, a tool-assisted check, not a comparable event: an adversarial critique run with DeepSeek — a model's output, not a diplomatic signal or anyone's official position, and explicitly told the architecture is meant to be rewritten — returned a structural critique: the Technical Committee has one seat per major bloc, major changes need an 8-of-11 supermajority no single bloc can block, and automatic enforcement (compliance within 48 hours, no case-by-case deliberation) means any signatory pre-commits to a body it doesn't control.
Second: on July 16, China and 28 other countries signed the agreement founding WAICO in Shanghai, with the UN Secretary-General attending; at the WAIC opening the next day, Mr. Xi Jinping called for states to stop stretching national security to place one country's above another's. Against my own Level 4 sanctions (full exclusion from certified chips, cloud, markets), the critique lands: automatic sanctions through a Committee without a great-power veto read as unilateral security logic in multilateral clothing, whatever I intended.
The obvious fix is to route serious enforcement (Levels 3-4 — partial and full exclusion, respectively) through something analogous to a UN Security Council model, combining broad support with a veto for designated major powers, instead of automatic Committee-triggered sanctions. But that runs straight into the problem Section 2 of this post is about: the entire reason for physical, verifiable constraints instead of political ones is that no government can credibly commit to limits it can unilaterally lift. A veto is precisely the tool a government would use to lift them.
So: does adding a veto for legitimacy quietly re-create the coordination failure this framework exists to solve? Or is a framework nobody trusts enough to join, because it looks like unilateral enforcement, worse than one with a controlled escape hatch? Those may not be the only two options — weighted voting, qualified majorities short of an absolute veto, or graduated response schemes might thread the needle better than either extreme. I don't think I currently have the right answer, and I'd rather find out here than after this becomes a formal proposal.
Other suggestions are welcome too.
8. An Invitation to Falsify the Thesis
I'm not posting this to seek agreement. I'm posting it because I want the central thesis stress-tested.
In particular:
1. Is the Hobbesian Bootstrap Paradox a useful description of the coordination problem facing frontier AI, or is it fully subsumed by existing commitment-problem literature with nothing added?
2. The veto question raised in section 7 — absolute veto, automatic enforcement, or something in between (weighted voting, qualified majority, graduated response)? Pick a point in that space and defend it; I don't think the pure binary captures the real design problem.
These aren't the only failure modes a 27-page governance architecture has — they're just the ones I see most clearly from where I'm sitting.
The premise I started from is simple, in the effective-altruist sense of the word (the way I understood it): there is a coordination problem in frontier AI development, and that problem is what produces a Moloch dynamic — not necessarily any single actor's bad intent. If that premise is right, the fix probably isn't the specific architecture in this paper, or any particular set of rules. It's coordination itself. Get two, three, or more of the relevant actors willing to sit down, and the tooling — verification mechanisms, thresholds, committee design, whatever the negotiation actually needs — is a far more tractable problem than getting people to the table in the first place. I'm glad to keep revising this architecture, build a different one, or replace any single mechanism inside it, toward that end. The hard part was never going to be the paperwork.
If you work on these questions professionally — at GovAI, CSET, RAND, or inside a frontier lab's governance team — I'd particularly value:
1. A pointer to whether the industrial-privacy compliance mechanism is fundamentally broken at scale, not just untested.
2. Even if you can't make an introduction yourself: a pointer to how conversations like this typically get convened in practice — precedents, forums, intermediaries — would help more than another opinion on the architecture. I'm aware this ask is disproportionate to my current standing; I'd rather hear that said directly than have it go unsaid.
3. A better design for Level 3-4 enforcement than either an automatic trigger or a great-power veto — both of which this post admits are flawed.
This post gives the architecture. The full paper develops the enforcement regime, committee design, convergence mechanism, and incentive structure in detail.
The complete 27-page framework, The A.L.E. Project, is available on SSRN:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7187658
I'm looking for failure modes and falsification attempts — and, more than anything, for someone willing to help get that table started; if the physical-constraints thesis is fundamentally wrong, I'd rather find out now, because the more urgent question may not be what rules to write, but who needs to be at the table before anyone writes them.
