Unverified70% confidenceFactExact time
In a paper published by Anthropic in 2023, sparse autoencoders were used to find tens of thousands of interpretable features in Claude, including emotional states and programming language types.
2
Sources
70%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
VerifiedAnthropic published mechanistic interpretability research on the internal feature interpretation of Claude in 2024, discovering millions of interpretable feature units.75% similarUnverifiedAnthropic published a paper in 2024 demonstrating how to locate neuron groups related to concepts such as 'honesty' and 'harm recognition' in the Claude model, and how intervening in these neurons can alter model behavior.73% similarUnverifiedAnthropic于2023年发表论文《Towards Monosemanticity》，首次在百万参数级别模型中识别出数万个单语义特征70% similarPartially VerifiedIn 2024, Anthropic's mechanistic interpretability team used sparse autoencoders to extract millions of interpretable concept units from the Claude model.69% similarUnverifiedIn multiple documents published between 2023 and 2024—including the Model Spec—Anthropic acknowledged it cannot rule out the possibility that Claude has functional emotions, and listed model welfare as a formal research topic68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/514832API
curl https://kongchang.com/api/v1/knowledge/claims/514832MCP
get_claim(id=514832)