5 Comments
User's avatar
Ted Roberts's avatar

Is "gradient decay" different than what's been called "context rot"? Also, for the hallucinated legal citations, it may be worth mentioning that the biggest failure in these cases appears to be the lawyers' failure to review the legal authorities that they have cited in support of their arguments. This is a professional obligation that exists independently of artificial intelligence.

scott cunningham's avatar

I think any kind of thing describing the failure of the LLM at the margin as tokens in gross is either the same thing or for my uses a distinction without a difference.

And I hear you about the use of bibliographies that haven’t been read. I just think the fact is, cites can be garbled for many reasons, including even having the LLM collect the cites for you that you do know well, but which is still messed up.

Caden G's avatar

This would be nicer one level up, to check my Zotero database for incorrectly imported citations. I feel I can mostly trust pandoc or the word extension

Laura Dumin's avatar

Thanks for this post. Would you be willing to share (here or in email) the steps to build /bibcheck? I feel like that is a low lift for me to get started with Claude code.

scott cunningham's avatar

Absolutely. I will find the conversation and print it out.