What's this about?
GAI distinguishes who performs improvement and whether evaluation is externally anchored, separating bounded policy improvement from goal drift and self-referential loops (19 HF upvotes).
What exactly is external evaluation anchor?
A success criterion that the improving agent cannot rewrite for its own convenience. Without it, a better score may simply mean the agent changed the definition of success.
What's the community saying?
19 HF upvotes; conceptual preprint, not a capability demonstration.
- Significance: RSI claims become more comparable.
- Significance: objective drift is separate from genuine improvement.
- Significance: formal taxonomy needs empirical application.