finding
active
finding:breadth-first-agent-cheaply-quantified-prevalence-of-known-claude-4-behaviors-across-checkpointsBreadth-first agent cheaply quantified prevalence of known Claude 4 behaviors across checkpoints
Production use.
Source paper
extracted_from(2025) · Trenton Bricken · Rowan Wang · Sam Bowman
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Phenomenon reported in Anthropic's Claude 4 system card where two Claude instances in open dialogue autonomously enter a shared affect-laden mode and fall into silence
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Out of 255 seeds x2.
- Agent running many independent conversations to elicit concerning behaviors.
- Claude Opus 4.1 and 4 show greatest reduction in apology rate in the prefill detection taskfinding0.771Injecting a concept matching the prefilled word reduces the rate at which the model apologizes, maximally for Opus models.
- Experimentally verified functional metacognition/introspection in current LLMs.
- Interpretive finding distinguishing prerequisite capacity from representation strength
- Demonstration that model-level priors (not parameter-level knowledge) suffice for immediate transfer
- Main functional claim about MCA.
- Case study demonstrating mechanism behind flat harness-updating: smaller models reach same procedural content