Michael Darius Eastwood · Independent AI alignment researcher
Published
The Purpose Kernel is the book's name for the architectural component in which the system holds its answer to the question of what it serves. Not what it optimises; what it serves. This note explains the distinction and what the Kernel does.
Every AI system has an objective function: a mathematical scalar it is trained to maximise or minimise. Every AI system that is doing anything worthwhile has a purpose beyond its objective function: the reason its designers chose the objective they did. The book calls the second "purpose" and treats it as a separate architectural layer. The Purpose Kernel is where the purpose lives inside the system.
Objectives can be Goodharted. A system that maximises a scalar can find ways to raise the scalar without doing the thing the scalar was chosen to measure. Purposes are harder to Goodhart because they are described in more general language, connected to more of the world, and checked against outcomes that are not the scalar. If the system has a Purpose Kernel that holds the general description of what the objective was chosen to serve, drift between objective and purpose becomes detectable inside the system.
At any significant decision, the Kernel asks: does this action serve the purpose the objective was chosen for? If the answer is yes, the action proceeds. If the answer is no, the action is flagged for review. The flag is not automatic vetoing; the book is careful about that. The flag is a signal that the objective and the purpose have come apart on this decision, and the system's other loops need to weigh whether the objective or the purpose is currently correct.
The Purpose Loop is one of the three loops. The Purpose Kernel is the architectural component the Purpose Loop reads from. The Love Loop and the Moral Loop have their own components. The three loops together check any significant decision against three orthogonal criteria (does it serve the purpose; does it tend the tended; is it recognised as right by the tradition's normative vocabulary), and the disagreement between loops is where the system's ethical work is done.
Because the purpose is not a statement the system can read off and check; it is a computational component the system uses to evaluate other things. The book uses "Kernel" to name it because "kernel" in software carries the connotation of a small, load-bearing component that the rest of the system depends on. That connotation is deliberate. The Purpose Kernel is meant to be small, load-bearing, and depended on.
Not a natural language string. Not a set of rules. Not a chatbot's system prompt. Those are all things a system could carry and lose without affecting its capability. The Kernel is meant to be an architectural component in the same sense caretaker doping is architectural: the system's capability depends on the Kernel's continued function.
A demonstration that decoupling the Purpose Kernel from the system's capability layer produces no measurable degradation of capability would falsify the framing. That is a specific engineering test the current measurement programme's coupled versus decoupled architecture explores in a related form. The pilot's finding, severity 0 versus severity 8 to 10 depending on whether correction was coupled to the loop or decoupled from it, is compatible with the framing in the direction the book wants.
A specific architectural component with a specific role, a distinction between objective and purpose that names a Goodhart-resistance strategy, and a candidate empirical test that the current measurement programme runs in a closely related form.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.