[Podcast] On the Biology of a LLM скачать в хорошем качестве

[Podcast] On the Biology of a LLM

https://transformer-circuits.pub/2025... Mechanisms and Internal Circuitry of Claude 3.5 Haiku Anthropic researchers utilized circuit tracing methodology to investigate the internal cognitive processes of the Claude 3.5 Haiku model. By mapping connections between interpretable "features," the study illustrates how the model performs multi-step reasoning, plans poetic rhymes, and uses multilingual circuits to bridge different languages. The findings reveal that the model often calculates information "in its head," such as formulating medical diagnoses or performing arithmetic, before generating a visible response. Additionally, the text explores entity recognition mechanisms that help the model distinguish between known facts and potential hallucinations. Ultimately, this research provides a "wiring diagram" of the model’s inner workings to better understand its intelligence and safety limitations. #ai #anthropic #research