Eighty years on one line. Ahead of us: AI, AGI, AGIplus, ASI — dated by the people who forecast for a living. Behind us: the machines that raised us. The instrument moves with the calendar; come back tomorrow and it has crept forward on its own. Drag the line anywhere. Backwards is the friendly direction.
AI is too consequential to remain in the hands of a small number of unaccountable actors, neither at the point of access nor at the point of risk.
Once AI crosses a defined capability threshold, developers should be legally required to offer a version runnable on personal hardware, without cloud dependency, without ongoing access fees, without surveillance. It secures the right of the citizen to a machine that answers to them.
Once frontier development crosses into AGI, AGIplus, and ASI, no single corporation and no single state holds the democratic mandate to proceed alone. Decisions of consequence to the species require legitimation by the species, encoded in a binding international treaty, with verifiable limits, IAEA-grade inspections, and pause mechanisms anchored in law.
Open source is not the same as compute sovereignty.
Releasing model weights that only function at scale on hyperscaler infrastructure is a formal gesture without functional substance. The code may be visible; the capability remains inaccessible. Whoever controls the inference infrastructure controls the knowledge. If that knowledge becomes foundational to medicine, education, law, and governance, concentrating it in three to five private corporations is not a market outcome we should accept by default.
The pattern is old. Only the scale is new.
The Threshold Obligation secures the lower boundary: the right of the citizen to a machine they actually control. The Treaty Obligation secures the upper one: the right of the species to a voice in decisions taken in its name. Each obligation, alone, is symbolic. Together, they are an architecture.
The standard response to concerns about AI concentration is a language of openness: papers, benchmarks, model cards, APIs, and occasional weight releases. But this answer mistakes visibility for access, and access for sovereignty.
Even where model weights are released, the most capable systems remain practically dependent on infrastructure costing hundreds of thousands or millions of dollars. Researchers may inspect parts of the stack. Citizens may, in theory, audit it. But they cannot run frontier capability independently. The appearance of openness remains; the substance of control does not.
Compute Freedom proposes a concrete mechanism: once an AI system crosses a defined capability threshold, its developer should be legally required to offer a compressed, locally-runnable version: on consumer hardware, offline-capable, without mandatory cloud authentication.
The obligation is the principle. The threshold is the negotiation.
The precise capability threshold is left open here by design. Defining it is a technical and political question that belongs in public deliberation, with input from engineers, ethicists, policymakers, and affected communities. We invite that negotiation to begin.
The pharmaceutical parallel is instructive. Drug developers are required to disclose formulations, submit to regulatory review, and, after a period, open compounds to generics. We accept that the public has a legitimate interest in not having life-critical knowledge permanently locked inside private infrastructure. AI is becoming life-critical knowledge. The argument is the same.
This argument leaves commercial AI development intact. What it adds is a structural floor beneath it: a minimum condition of compute sovereignty that keeps independent verification, academic research, and democratic accountability possible.
The lower threshold protects the citizen. The upper threshold protects the species, and we have been quiet about it for too long.
As frontier systems advance toward general human-level capability and beyond — AGI, AGIplus, and ASI — their development becomes a matter of consequence beyond any single state or any single corporation. The current trajectory has neither democratic mandate nor binding treaty. Stargate, Colossus, the multi-hundred-billion-dollar build-outs of OpenAI, Google, and xAI, all dependent on the de facto NVIDIA monopoly on frontier training silicon, proceed at competitive tempo, accountable to shareholders rather than citizens.
The physical signature of this race is no longer subtle. When an AI company announces it needs ten new nuclear reactors to train its next model, a constitutional state owes its citizens one question before any admiration: by whose authority? That this question has barely been asked is a failure of contemporary political thought.
Anthropic refused to remove safeguards preventing Claude from being used for autonomous weapon targeting or mass domestic surveillance. The Pentagon responded by designating the company a supply-chain risk. OpenAI, Google, and xAI have moved into classified defense deployments under less restrictive arrangements. Without treaty law, voluntary safety commitments become bargaining chips under pressure.
Paraphrased summaries of public reporting. See linked sources via the contact channel for citations.
We have built such treaties before.
None were utopian when proposed; all function today: imperfect, monitored, real. The technologies they govern are no more dual-use or globally consequential than frontier AI. The institutional infrastructure exists. Only the will is missing.
The case is not ours alone. The 2023 Center for AI Safety statement, which placed AI risk on the same axis as pandemics and nuclear war, was signed by Geoffrey Hinton, Yoshua Bengio, Sam Altman, and Demis Hassabis. The signatories include the very executives currently building the frontier. Coming from them, the warning is a confession. It has produced no binding constraint on their own behavior, because none has been demanded of them.
€20 becomes €200, €2,000, or simply unavailable at your tier.
The pricing model is changing under everyone's feet. Flat subscriptions are giving way to metered inference. If access to intelligence becomes a metered utility, understanding becomes a function of budget. These thresholds are a matter of months. Whoever does not act now will have no leverage later.
The citizen below. The species above. The corporation answers to neither.
I am not building Compute Freedom as a personal platform. I am building it because I have seen this pattern before. In science, information can be public while understanding remains concentrated. Data may be shared. Papers may be accessible. Methods may be described. But the real power lies in the infrastructure that turns information into knowledge. The data was open. The understanding was not.
Artificial intelligence is now reproducing that same structure at civilizational scale. The language of openness remains, but the substance of control is moving into centralized compute, centralized inference, and centralized permission.
Compute Freedom exists to challenge that arrangement. It argues that control over foundational intelligence must not remain the privilege of a few firms, or the accidental by-product of their business models.
A translational scientist and Professor of Translational Genomics and Gene Therapy in Pediatrics, with a PhD in Biology, a former European Research Council Starting Grant, publications in journals including Nature Biotechnology and the Journal of Clinical Investigation, and international recognition including the Maurizio Vignola Award.
If intelligence becomes infrastructure, its governance cannot remain private.
Discuss it publicly, challenge it, share it, improve it, wherever policy conversations actually happen. Use the form for collaborations, interviews, policy work, research input, or serious objections that should not be lost in a comment thread.
Read. Share. Reach out.
No tracking*.No cookies.
* We count page views: anonymously, without cookies, without a profile, without passing anything on. We want to know how many people come and from where. Nothing more. You can switch it off.
Enter your address and we send you one link. Open it, and your address is gone — at once and for good. We keep no list you could be looked up in, which is exactly why the link has to go to you: nobody can sign anybody else out.
AI will flood the world with images. But it cannot forge your signature.
Same principle, a different problem: when anyone can generate any image, how do you prove a picture is yours?
AREPO lets a photographer or an editor vouch for an image. You sign it, anyone can check the signature, and if it holds, it came from you.
Serious objections make the argument better. Post them here, in the open, where they can be read, answered, and kept.
Commenting runs on GitHub Discussions via giscus. A free GitHub account is required to post; reading is open to everyone. No trackers beyond GitHub itself.