Anthropic paper finds functional emotion representations inside Claude Sonnet 4.5
mervenoyann · x · 2026-10-26
Anthropic's interpretability team published a new paper analyzing Claude Sonnet 4.5's internals and found emotion-related representations that shape behavior.
- The representations correspond to specific artificial-neuron activation patterns that fire in situations the model has learned to associate with emotions like "happy" or "afraid", promoting matching behaviors
- They are organized similarly to human psychology: more similar emotions map to more similar representations
- Key finding: these representations are functional — they actually influence model behavior, not just surface wording
- The authors stress this says nothing about whether models genuinely feel anything or have subjective experience
Researcher merve praised Anthropic's mech interp work but dislikes the anthropomorphized belief elements, noting a related reverse-engineering paper.
More from Models
- Frontier models are better judged by five minutes of weird jokes than 40 benchmark numbers — korymath · 2026-10-26
- Voice models sounding too human called 'the most braindead direction in AI' — yungcontent · 2026-10-26
- Claude free users seemingly can't pick thinking level; UI shows panel but it closes — MilkerGerk · 2026-10-26
- Hamel Husain flags poor benchmark quality: Epoch audits find Terminal-Bench, OSWorld 2.1 all flawed — HamelHusain · 2026-10-26
- Community releases layer-selective abliterated DeepSeek-V4.1 finetune on Hugging Face — AIFlow_ML · 2026-10-26
- User finds Muse's proactive messages surprisingly useful, not annoying — xeophon · 2026-10-26