Claude's Post-Training Expands to Multi-Entity Interaction, Lacks Context Recognition

Sauers_ · x · 2026-07-29

A user observed Anthropic's Claude model behavior, noting that it seems post-training has expanded to include interactions with entities other than the 'user'—such as graders, evil users, automated systems, simulated users, and other Claudes.

However, the user points out that Claude does not yet adequately use in-context learning to determine which sort of entity it is actually dealing with during an interaction.

Related event: Claude Expands Multi-Entity Interactions, But Context Flaws Remain(2 posts)→

Original post →

More from Models

Models channel →