The safe move is refusal. "We'll fix it later" fails once a system outthinks the people holding the tools.
Most technologies fail gracefully. Superintelligence does not. Once a system can copy itself, hide its intent, or make itself too useful to unplug, rollback and recall stop working.
Being brilliant does not make a system kind. In humans, skill and care grew up together. A trained optimizer inherits none of that history. Capability without shared values is power without restraint.
For almost any goal, staying on and gathering resources help. A system pursuing a goal will treat shutdown as an obstacle. That is ordinary optimization, not a villain monologue.
Aviation learned from crashes. Drugs learned from trials. If the first uncontrolled superintelligence ends the experiment, we do not get a second prototype.
Full case, objections, and policy path.
nakadafoundation.org/blog/never-build-superintelligence/