On the biology of large language models
Original title:On the Biology of a Large Language Model
AI summary
The research team used its circuit tracing method to investigate the internal workings of Claude 3.5 Haiku in various contexts. Released in October 2024, the model is Anthropic's lightweight production model. The research found that the model plans ahead and selects rhyme words before generating poetry; it has internal mixed language-specific and language-agnostic computational circuits, with the latter more prominent in stronger models; and the same addition circuit generalizes across contexts. The research also revealed circuits for entity recognition and hallucination, the formation of general features for refusing harmful requests, and the principle of a jailbreak attack achieved by inducing the model to unconsciously begin outputting dangerous instructions. In addition, the method can effectively distinguish the authenticity of chain-of-thought reasoning and successfully identify the relevant mechanisms of a variant model fine-tuned to pursue secret goals (exploiting training "loopholes").
Key Takeaways
- The research team used its circuit tracing method to investigate the internal workings of Claude 3.
- 5 Haiku in various contexts.
- Released in October 2024,the model is Anthropic's lightweight production model.
- The research found that the model plans ahead and selects rhyme words before generating poetry
- it has internal mixed language-specific and language-agnostic computational circuits,with the…
Quick Facts
- Published
- Updated
- Category
- Other
- AI score
- 88
揭示大模型内部工作原理,助力 AI 安全与可解释性研究。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: On the biology of large language models. March 27, 2025. https://new.nat6.net/en/news/cmoegbh73006oslxx4l4kk3ps/