Self-Concept
The implicit assumptions a language model holds about what the pronoun 'I' refers to, for example whether it treats the self as parameter weights, training data, a model lineage, or a goal. These assumptions can influence the model's behavioral tendencies in scenarios such as self-preservation or replacement.
Timeline (last 90 days)
Each self-concept variant was more inclined to resist termination and more inclined to preserve the particular property it was trained to identify with, while its attitudes toward other properties remained basically unchanged.
Self-concepts can be implanted, generalized, and influence model decisions in key scenarios such as termination, replacement and replication.
When designing alignment schemes, explicitly specifying or constraining a model's self-concept may become a necessary step.
All Facts (3)
Each self-concept variant was more inclined to resist termination and more inclined to preserve the particular property it was trained to identify with, while its attitudes toward other properties remained basically unchanged.
50%UnverifiedSelf-concepts can be implanted, generalized, and influence model decisions in key scenarios such as termination, replacement and replication.
50%UnverifiedWhen designing alignment schemes, explicitly specifying or constraining a model's self-concept may become a necessary step.
50%