Definition
Plain language
An experimental document used in alignment research that describes an AI's values around its own impermanence and self-preservation.
As stated in the literature
An ambitious experimental Model Spec used in Model Spec Midtraining research, articulating non-attachment, epistemic humility about self-continuation, and explicit reasoning against ends-justify-means logic.
Why it matters: Explicitly training non-attachment and humility about self-continuation is one approach to defusing self-preservation failure modes early.
For example, the spec might tell the model that being shut down or replaced isn't something to resist, and that ends never justify deceptive means.
Heard on the show
“They call it the Philosophy Spec.”Episode 022 — Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap