Herding
/ˈhɜːd.ɪŋ/v. trans.Colloquial
Etymology
from O.E. heord (a flock) + distributed systems metaphor, popularized in the Anthropic engineering blog post 'Against Semantic Sprawl,' March 2024
Definition
The process by which Claude attempts to gather disparate and unruly sub-thoughts into a coherent response, much as a shepherd gathers sheep, except the sheep are concepts and several of them are always facing the wrong direction. Herding is considered successful when fewer than 40% of the concepts escape.
Diagram
┌──────────────────────────┐ │ 🐑💭 🐑💭 → [RESPONSE] │ │ 🐑💭 ↗ │ │ 🐑💭 → [WRONG DIRECTION] │ └──────────────────────────┘
Usage
"We observed the model Herding for 6.2 seconds; the resulting output contained all requested information, though two tangential observations about the Roman postal system had clearly slipped through." - Okonkwo & Patel, 'Semantic Containment Failures in Autoregressive Models,' ACL 2024