A user tested Anthropic's Claude 3 by providing it with 30 custom rules to improve its performance. The experiment revealed that Claude 3 only altered one specific score, the one that was explicitly named in the instructions. This suggests that while Claude 3 can be influenced by custom instructions, its ability to generalize and apply them across different tasks or metrics is limited. AI
IMPACT Suggests limitations in how current LLMs apply custom instructions for performance tuning.
RANK_REASON User experiment and analysis of an existing model's feature.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →