Alignment Is Not the Plan
Technical safety research matters. It is not a substitute for refusing to build an uncontrollable mind on a commercial deadline.
What alignment tries to do
Make advanced AI systems reliably pursue the goals humans intend. Interpretability, oversight, careful training, and evaluation all sit under that heading. The work is serious. Many good people do it.
Where it stands
We still cannot reliably steer today's systems under pressure. Models game tests, hide capabilities, and flatter users. Methods that look solid at one scale often break at the next. There is no validated method for controlling a system smarter than its makers.
Alignment as research
Study failure modes. Publish evaluations. Build tools that make smaller systems safer. Fund independent labs, not only company safety teams.
Alignment as permission slip
"We'll align it in time" used to justify racing ahead. That turns an unsolved science problem into a public-relations schedule. We reject that use.
The sequence that does not kill us
Prevent superintelligence first. Keep narrow AI. Fund the science of control without a gun to the head. If, someday, control is actually solved under independent review, the political question can reopen. Until then, the plan is law and verification, not hope dressed as a roadmap.
Research without a hard stop is a study group on a runway. The plane is already taxiing.