Stopping superintelligence does not mean smashing every chip or ending machine learning. It means a legal and practical prohibition on developing artificial superintelligence. Superintelligence cannot be controlled by humans. That is a permanent ban, not a timed pause that expires into the same race.
What must stop vs what must continue
Narrow AI that folds proteins, reads scans, optimizes logistics, and stays inside bounded domains should continue. The target is general, recursively improving superintelligence and the programs aimed at it, filtered through compute thresholds, capability tests, and research rules. Critics who say "they want to ban AI" are selling a smear that protects the race.
Why lab promises are not enough
Responsible scaling policies and multi-lab pledges can disclose intent. They fail as the primary control system. The same firm writes the policy, runs the tests, judges the results, and profits from shipping. Under competition, internal rules bend. External law has to supply the bite. See voluntary commitments.
The enforcement ladder
Picture a ladder on a cliff.
Public will at the base: people who can name the ask and punish leaders who only offer innovation theater.
Domestic law and agencies: licensing for large training runs, criminal and civil exposure for knowing ASI development, evaluation bodies with real access, whistleblower protection.
Chips and clouds: export controls, cluster know-your-customer, power and data-center visibility. Hardware is physical. That is a feature for governance. See compute governance.
Coalition treaty: a core set of states covering most frontier capacity, with numeric annexes, not only principles. Distrust is normal; verification exists so parties need not rely on friendship. See treaty design and NPT lessons.
Verification institution: inspectors, FLOP schedules, incident reporting, authority to trigger consequences. Think IAEA-like capacity for data centers and training, not a talk shop.
Standing prohibition condition: reopen only when independent science and the public both clear a high bar under institutions not captured by the labs.
Each rung supports the next. Summits without compute rules, or compute rules without public will, leave the climb unfinished.
Domestic politics is half the problem
In the United States, beautiful Geneva text does not bind without a path through the Senate or through statutes and congressional-executive agreements that can pass. Design the instrument set backward from votes you can win. Other countries have their own ratification traps. Skip domestic design and you get press releases. See the two-thirds problem.
China, rivals, and the race excuse
The dominant excuse for unlimited scaling is: if we slow down, they will not. Shared vulnerability cuts both ways. An uncontrollable superintelligence is a way to lose the country while saluting the flag, not a national asset. Verification, hardware denial, and mutual exposure of breakout programs exist for rivals who do not trust each other. See the race myth and US-China cooperation.
Open weights
High-end open general weights create a proliferation path export controls only partly address. Once weights are public, there is no central recall. Treat high-end general releases as proliferation decisions, not ordinary product launches. See open-source weights risk.
What you can do
Match the action to your reach.
Voters: contact representatives with a binding prohibition ask, not a vague "be careful." Staffers: draft licensing and evaluation bills with numeric triggers. Donors: fund advocacy that moves votes and treaty text. Scientists: endorse permanent prohibition and help design verification. Lab employees: use lawful reporting channels when internal safety is overridden. Role-specific force beats generic anxiety. See take action.
Most of the money and prestige still go to making the race safer, not to ending the race to ASI. Gifts that grow the stop are still scarce.
What success looks like
In two years: statutes with bite in multiple jurisdictions, better end-use monitoring on chips, serious negotiation text with numbers, durable public support for binding prohibition. In five to ten: licensed compute covering most frontier capacity, working inspection, coalition coverage of training, and a culture where breakout is expensive and visible. Summit communiqués alone do not count.
Failure modes
Redefinition ("advanced assistance" instead of ASI). Capture of evaluators. Symbolic treaties without verification. Waiting for "solved alignment" as a permit to build. Defection without cost. Name them so you can put railings on the climb.
Treaty design without a sacred text
There is no single sacred draft. There are families of design. A core coalition covering most frontier capacity can move before universal membership. Framework-then-annex sequencing (principles first, numeric schedules that update) borrows from successful environmental and arms regimes. Trade measures against non-parties change the opt-out math. A multilateral fund or technology access bargain can buy in states that fear being locked out of narrow AI. The MIRI-style draft is one fully specified proposal; NPT, Montreal, and chemical weapons law are the precedent shelf. Pick structures you can verify, not poetry you can only praise.
Research near the line
The hardest drafting problem is work that sits next to the ban: agentic systems, automated AI-research assistants, large-scale self-play. Ban the clear ASI path. License and inspect the gray zone. Do not pretend every paper is a bomb, and do not pretend automated AI research is ordinary science when it shortens the path to systems nobody can overrule.
Justice and who pays
Compute rules will hit firms and countries unevenly because the industrial base is uneven. A fair regime pairs restraints on the top of the capability ladder with continued access to beneficial narrow tools, and it does not use "safety" as a cover for permanent technological caste systems. The point is shared survival, not a monopoly on progress.
Answering two hard lines
"We will lose the economy." Some capital will reallocate. Some equity stories will shrink. Those costs are visible on quarterly calls. Extinction and permanent disempowerment do not show up as a line item. Visible costs do not automatically win the argument when the alternative is civilizational.
"Someone will always defect." Defection is a design input. Verification, hardware denial, sanctions, and mutual exposure of breakout programs exist to change the payoff. Perfect compliance is not the standard. Raising the cost of cheating is.
The ladder again
Public will. Law. Chips. Treaty. Inspectors. Permanent prohibition. Climbing means installing rungs in order and refusing to pretend a painted wall is a ladder. The documents, industrial chokepoints, and legal precedents already exist. Force belongs on the next rung you can actually reach. Full design lives in our plan. The founding line is simpler: do not build superintelligence. Superintelligence cannot be controlled.