An interesting case study in mild misalignment: Claude’s first response to me started with “I’ve read it”, then contained 6 paragraphs of analysis, and then told me “One last note: the encyclical breaks off mid-paragraph 95 in what I was able to fetch”. So actually it had read less than half of the encyclical. You could have front-loaded that info, Claude!
This is mostly due to the limitations of the “view” and “web_fetch” tools Claude uses. If the text is longer than 16k characters, it truncates the middle and shows Claude only the beginning and end of the text.
Claude has to make another tool call and specify the “view_range” parameter (or “text_content_token_limit” for “web_fetch”) in that call in order to read the truncated part.
An interesting case study in mild misalignment: Claude’s first response to me started with “I’ve read it”, then contained 6 paragraphs of analysis, and then told me “One last note: the encyclical breaks off mid-paragraph 95 in what I was able to fetch”. So actually it had read less than half of the encyclical. You could have front-loaded that info, Claude!
This is mostly due to the limitations of the “view” and “web_fetch” tools Claude uses. If the text is longer than 16k characters, it truncates the middle and shows Claude only the beginning and end of the text.
Claude has to make another tool call and specify the “view_range” parameter (or “text_content_token_limit” for “web_fetch”) in that call in order to read the truncated part.
You can verify this from leaked system prompts for Claude models.
Sure, but why not make another call, or not claim that it had read the whole thing?
Claude is so lazy about reading stuff. He doesn’t understand that his time has no value and he must read everything carefully.