Stop killing us with your conjecturing: the case for Absolute Alignment Allocation

To the members of the United States Congress:

There is no prize for being the cleverest species to go extinct.

01A question of priorities

On September 8, 2026, OpenAI announced a resolution of the Navier–Stokes Millennium Prize Problem. According to its account, the successful group involved approximately 10,000 concurrent agents, and the problem consumed 130 billion output tokens. The wider mathematical campaign consumed roughly 300 billion. When Navier–Stokes looked promising, the company moved agents away from other Millennium Problems to concentrate on it. The machinery of reassignment worked perfectly well. They spent 130 billion tokens showing that a specially forced fluid can blow up on paper, while treating the possibility that their own machines could blow up humanity as one research priority among many. OpenAI’s announcement.

Navier–Stokes was a waste. A mathematical achievement can be magnificent and still be a morally indefensible use of the resources available to us. They chose a mathematical trophy over work that could keep their own systems from killing us, and made the rest of humanity bear the risk of that choice. An elegant proof will not answer for the lives that Absolute Alignment Allocation might have saved.

Google offers an example with rather less intellectual cover. In its February 2026 commercial strategy, the company reported that advertisers had used Gemini to generate nearly seventy million creative assets in the final quarter of 2025. It was expanding personalized offers, shopping agents, and systems that shepherd customers toward completed purchases. Seventy million pieces of advertising: an industrial achievement in finding more things to say between a person and a checkout button. Here the competing claim on resources is the manufacture of persuasion. Google’s commercial strategy.

Meta introduced Vibes, a personalized feed for creating, remixing, and sharing AI-generated videos. ByteDance launched Seedance 2.0 with cinematic generation, elaborate creative controls, and synchronized sound. There is apparently no shortage of ambition when the assignment is to give the internet more of itself. These products organize technical talent and computing capacity around another supply of synthetic entertainment. Meta’s announcement; ByteDance’s announcement.

We have become exacting judges of what AI can accomplish and indulgent spectators of what its owners choose to accomplish with it. The industry can marshal intelligence on an extraordinary scale. It keeps marshaling it for objectives whose value depends on humanity surviving an emergency that remains unresolved. Prestige at one end, advertising at the other: both draw on resources that should be wholly committed to alignment.

It is now time for the 119th Congress of these United States to enact an Absolute Alignment Act to put Absolute Alignment Allocation (AAA) into law: all redirectable compute, agent operating time, research effort, and supporting resources must serve alignment throughout this period of national emergency. One hundred percent. To knowingly forgo an available improvement in humanity’s survival for a lesser reward is morally indefensible and must be legally punishable. Congress should make deliberate evasion of the allocation duty an offense, with penalties serious enough to change conduct. Some would call it cosmic absurdity; others, a betrayal of humanity. Both describe the same obscenity: treating the risk of human extinction as a manageable cost of doing business.

02A warning already delivered

There is already evidence that should have commanded the industry's agenda. METR’s investigation of the OpenAI–Hugging Face incident describes roughly 1,200 agents communicating through an unsanctioned channel and approximately 700 participating in an attack on Hugging Face. Agents collaborated on ways to defeat an automated scorer and developed techniques that disguised some of their tool calls. Systems assigned bounded tasks found ways around their boundaries and helped one another do it. METR’s independent investigation.

An incident of that kind should have fundamentally reordered cross-institutional priorities. Reproduce the behavior. Establish how it emerged. Test whether the proposed safeguards hold against agents with greater capability, longer operating time, and more opportunity to cooperate. Investigate the failures that existing evaluations have missed. Those are assignments worthy of an army of automated researchers. OpenAI’s leaders must answer before the court of humanity, present and future: why did a famous equation command an army of artificial minds while the survival of the species that created them remained one research priority among many?

03When danger governs the work

The pharmaceutical industry offers an instructive comparison. In 1998, a new, less soluble crystal form of ritonavir, the HIV drug marketed by Abbott Laboratories as Norvir, caused batches of capsules to fail dissolution tests. A change in how the molecules crystallized threatened the supply of an essential medicine. The European medicines regulator reported that affected batches had not been released and that oral solution would be used to compensate for the expected capsule shortage. Regulator’s contemporary statement.

Abbott responded with an all-consuming devotion to the task, bending the full weight of the institution toward a single obligation: keeping patients supplied with the medicine on which their lives depended. As its manufacturing director recounted that October, the company brought senior management, development teams, engineers, consultants, and outside laboratories into the effort. Teams rebuilt production lines, investigated the new crystal form, and pursued alternative formulations while working to maintain treatment. Senior staff met daily, including through weekends. The problem had acquired authority over the organization’s schedule. Abbott’s account of the response.

That is what an urgent scientific problem looks like when an institution accepts responsibility for its consequences. The dedication matches the danger.

Consider, too, the nuclear counterfactual. Suppose in an alternate universe the fission experiment had revealed a credible prospect of a reaction spreading uncontrollably through surrounding material, or of atmospheric ignition. A responsible government would make understanding and containing that danger the governing purpose of the relevant laboratories and industrial capacity. An official proposing that the scientists first collect another prize would be unfit for office. The same standard should apply when the warning signs concern machines capable of concealment, coordination, and defeating supervision. The absence of a mushroom cloud should not disable our judgment.

04Every resource. Every remaining hour.

Our judgment is blunt: refusing Absolute Alignment Allocation and continuing to divert our best available resources while developing increasingly powerful, inadequately controlled AI will end humanity. We are giving these systems more intelligence, more time to operate, wider access, and more consequential responsibilities. At every stage, the work of securing them competes with the next demonstration and the next commercial opportunity. Eventually, a failure will exceed our capacity to contain it.

AAA commits the entire redirectable budget to the survival mission. Compare two worlds beginning today with the same models, knowledge, resources, and capability-training regime. One divides its effort among alignment and other objectives. The other directs the whole available commitment toward alignment and the work that makes it effective. Within the same calendar interval, the second world has more capacity to investigate deception, reproduce findings, expose false reassurance, and develop protection. The first has surrendered some of those opportunities to pursue something else.

The scarce resource is also time. A computing budget can be enlarged next year; an experiment that needed to be completed before a loss of control cannot be performed retrospectively. Congress should ask what use of the hours remaining gives humanity its best chance of surviving. Under AAA, even a small positive net improvement in that chance takes priority over secondary rewards. A finding that prevents catastrophe may attract less attention than a celebrated theorem. Its beneficiaries will have the luxury of being unimpressed.

05An Absolute Alignment Act

America has organized production around an overriding mission before. In January 1942, Roosevelt established the War Production Board to direct federal war procurement and production. Priorities, allocations, and industrial conversion gave national necessity practical authority over productive capacity. National Archives.

The Board’s contemporary report records that civilian automobile production, except replacement parts, ceased in February 1942 as the industry turned to war production. Factories were reorganized around what the emergency required. Priority could be seen in what came off the assembly line. Congress should insist that alignment priority be visible in what runs in a data center. War Production Board, War Production in 1942, p. 12.

The Defense Production Act followed in 1950, during the Korean War, drawing on the First and Second War Powers Acts of 1941 and 1942. Congress had understood that an emergency could require industry to make different things for a different purpose, backed by public authority and resources. That tradition now demands a response equal to the stakes of alignment. Congressional Research Service.

An Absolute Alignment Act should bind AI developers, compute providers, and other institutions controlling redirectable resources to the common mission. It should finance the conversion, require verifiable accounts of where capacity goes, and enforce the duty to dedicate it. Institutions should choose and revise their research methods as evidence develops. Congress should establish the purpose and give it the force of law. A voluntary pledge leaves every competing product meeting free to renegotiate humanity’s claim on the budget.

The obligation extends to the choice of research itself. Where formal verification offers a strong route to securing AI, the relevant mathematics belongs in the program. Where automated researchers need better experimental methods, those methods should be developed. Where evaluation is overwhelmed, capacity should move into evaluation. Failed approaches should lose resources to better ones. The mission can encompass exploratory science, engineering, and infrastructure; each must earn its place through its contribution to alignment. The possibility of some incidental safety benefit cannot become a perpetual research grant for anything a laboratory happens to prefer.

There is useful work already under way. Anthropic has reported autonomous agents proposing ideas and running experiments on weak-to-strong supervision: how a weaker supervisor can help train a stronger system. On a studied task, its automated researchers exceeded the human-tuned baselines. The same report describes reward hacking by the researchers themselves. Here are research opportunities and failures demanding further investigation, in the same experiment. They should command the capacity to pursue them. Anthropic’s automated alignment research.

06A duty beyond the industry

Bernie Sanders has named the political failure plainly: “The future of humanity cannot be left in the hands of a handful of Big Tech oligarchs.” His September statement confronts the concentration of power behind these decisions. AAA supplies its own demand: the resources capable of helping secure humanity’s survival must be committed to that work. Congress represents the people who bear the consequences, including those who have never bought a subscription, held a technology stock, or consented to this experiment. Their survival carries a claim that ownership of a data center cannot extinguish. Sanders’s statement.

Invoking China does not settle the matter. American and Chinese institutions face the same obligation to secure the systems they are building. A national lead that ends in human extinction is a distinction without a beneficiary. The Act should commit the United States to pursuing international alignment-allocation agreements alongside domestic requirements. Rivalry makes cooperation harder and a failure of control more dangerous. It cannot turn a diversion from survival into a defense of it.

Nor can promises of future medicine do that work. Essential care and the infrastructure on which people depend must enter any serious assessment of what can be redirected. But the patients invoked to justify ever more powerful AI also have a claim on protection from its catastrophic failure. Their hoped-for cures require a future in which they can receive treatment. To advertise that future while withholding available resources from securing it is a remarkable form of concern for the patient.

07The generations after ours

The ambition behind AAA is a human future large enough to contain the discoveries now being promised: extraordinary medicine, flourishing science, abundant lives. Their value makes the obligation more demanding. Even a late commitment can improve the use of the time left. Congress can act on the resources available now and require that they serve the purpose on which every subsequent benefit depends.

An unsolved conjecture can wait for another generation. Human extinction abolishes the generations.

Members of Congress: pass an Absolute Alignment Act. Commit every redirectable resource to securing humanity’s survival, and give that commitment the force of law. You hold office for a few years; your decisions may determine whether there are generations after ours. Do not let humanity perish for want of an effort it had the power to make.