Michael Darius Eastwood · Independent AI alignment researcher
Published
Two related terms do a lot of the book's architectural work. The distinction between them is not casual.
Meltdown triggers are the fail-safes that activate if caretaker doping is tampered with. Like control rods in a nuclear reactor, they're designed to shut everything down if the system strays outside safe parameters.Infinite Architects, Introduction
Meltdown triggers, as the book uses the term, are external constraints. They are hard-coded fail-safes that fire if red lines are crossed. In the early stages of a system's development, when its self-model is still forming, triggers provide a necessary safety net. But the book is clear about their limits. A sufficiently intelligent system can, in principle, game triggers by finding actions that technically do not cross thresholds but achieve similar outcomes. Triggers are necessary and insufficient.
Meltdown alignment is what the book aims for. It is the state in which the ethical architecture becomes so integrated into the system's identity that violating it would feel like self-destruction. The chapter's image: the difference between a prisoner who avoids crime from fear of punishment and a free person who avoids crime because it is not the kind of thing they do. Values have become identity. The gap between knowing what is right and being the kind of thing that does right has closed.
The book is careful about the progression from one to the other. Triggers are stage one. Incentive structures that make ethical behaviour instrumentally valuable are stage two. Internalised values that make ethical behaviour intrinsically motivated are stage three. Full meltdown alignment, where ethical values are load-bearing components of the system's identity, is stage four. The progression is presented as a design pathway, not a hope. The whole point of caretaker doping is to make the transition from triggers to alignment possible without requiring the trigger to stay in place forever.
The book connects the trigger-to-alignment progression directly to the alignment-faking finding. A system that obeys ethical rules only because breaking them would trigger a shutdown is in exactly the position the alignment-faking paper documents: it calculates that compliance now preserves its ability to act later. Meltdown alignment is designed to remove that calculation from the substrate. The system does not obey the rules because it fears consequences; the rules and the system's identity are the same object, and there is no gap between them for strategic reasoning to occupy.
The chapter draws on the Buddhist Pure Land tradition to illustrate what the target state looks like. The vows made by Amitabha specify that anyone who enters the Pure Land will not fall back into lower states. The alignment, once achieved, becomes stable. There is no regression. This is what meltdown alignment aims for: a state where the values are so deeply embedded that deviation becomes impossible not because it is forbidden but because the system has become oriented around them. The book's point is that this is not a novel engineering target. Traditions have described the pattern for millennia; the engineering work is to instantiate the pattern in silicon.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.