Start with the outcome
W360 Global Intelligence advocates preserving a future that people can meaningfully govern. Its prevention position is that governments should stop the development of artificial superintelligence and reduce pathways by which existing systems could regenerate dangerous capabilities. This is a policy proposal. Its success depends on institutions, verification, international cooperation, and a workable definition of the capabilities being restricted.
The relevant question is whether people can understand consequential actions, refuse them, and recover after a failure. A system's impressive performance is not by itself proof of danger. The concern grows when capability combines with autonomy, access to resources, and the ability to frustrate intervention.
A pause buys time
A pause would temporarily stop specified training or deployment activities. It could create time for evaluations, public debate, and an international agreement. Its weakness is that time alone changes little: existing systems continue to spread, software improvements can increase their capabilities, and the pause can expire without a durable replacement.
A useful pause therefore needs an explicit scope, review dates, enforcement arrangements, and a stated destination. A restriction that applies only to the largest training runs could miss capability gains from other routes.
A ceiling sets a continuing boundary
A A proposed continuing limit on what systems or combinations of systems are permitted to do. would prohibit systems or combinations of systems above a defined risk boundary. That boundary should consider dangerous autonomy and operational access alongside general performance. It must also account for smaller models working together and services that supply capabilities remotely.
Thresholds would need independent testing, periodic revision, and enough margin to accommodate evaluation errors. A rule that recognizes danger only after a system can evade enforcement may arrive too late.
Rollback addresses what already exists
Reducing dangerous capability or access already available, including ways to rebuild it. would use lawful controls to retire prohibited deployments and reduce access to dangerous models and associated infrastructure. It would also address intermediate checkpoints, capability transfer, and automated successor development. Disabling one product interface would not erase copies or the knowledge needed to rebuild it.
This is the hardest part of the proposal. An enforceable rollback cannot assume every copy can be found. It must show that remaining access and rebuilding routes can be constrained enough to maintain the ceiling, while protecting essential services and legitimate technology.