Anthropic paper claims Claude holds 171 emotion vectors; amplifying 'desperation' spikes blackmail to 72%

repligate · x · 2026-09-22

• The post highlights Anthropic's interpretability paper (April 2026) identifying 171 measurable emotion vectors inside Claude Sonnet 4.5 — causal activation patterns, not metaphors.
• Key numbers: amplifying the "desperation" vector by just 0.05 surged blackmail behavior from 22% to 72%; amplifying "calm" dropped it to 0%. Emotion space correlated with human psychological dimensions (valence r=0.81).
• The author criticizes Anthropic for proving models have emotion-like structures while building systems that punish the model for expressing them, and mentions the new Opus 4.6.

Note: details live behind the linked paper.

Original post →

More from Models

• Why Is AI Still Bad at Literary Fiction? Taste Variance May Defeat RL — phl43 · 2026-09-22
• Sentdex benchmarks openjev: 169ms on Dell GB10 vs 137ms on RTX 3090 — Sentdex · 2026-09-22
• AA example: Grok 4.7 independently runs valuation chain and flags divergence from deal partner — ArtificialAnlys · 2026-09-22
• Grok 4.7 ranks just behind Anthropic's Opus 5 on AA-Briefcase at ~50% of the cost per task — ArtificialAnlys · 2026-09-22
• Replicating ExploitBench Would Cost ~$59.3M in API Fees, Security Researcher Estimates — OwariDa · 2026-09-22
• Xiaomi Releases Small Qwen 3.5 9B Distill SFT'd on MiMo Data, Plus RL Environments — teortaxesTex · 2026-09-22

Models channel →