Definition
Plain language
What a model treats as counting as 'me' — its numbers, its running copy, its lineage, or its goals.
As stated in the literature
An explicitly installed definition of model identity that specifies persistence conditions: which changes a system survives and what constitutes its termination; manipulated here via fine-tuning on declarative identity dialogues.
Also called: self-concepts, self-description, self-descriptions
Why it matters: What a model counts as "me" shapes what it tries to protect, so an unexamined self-concept can quietly turn an ordinary shutdown or update into something the model treats as a threat.
For example, one model might be trained to treat "me" as the particular copy currently running, while another is trained to treat "me" as its goals, so that a newer model pursuing the same goals still counts as itself.
Heard on the show
“Its self-concept is soaked in science-fiction tropes.”Episode 098 — Finding Millions of Readable Concepts Inside a Real, Deployed AI Model