Greetings. I agree in general about the concerns being raised here.
Mathias Gibbens - 26.10.26, 15:12:35 CET: > I also want to point out there appears to be a penalty for anyone > uploading to NEW who chooses to ignore LLM-generated complaints[2] and > a bonus for uploads which make the LLM happy[3]. There should not be any penalties for not reacting to LLM reviews. These can provide hints to DFSG team and ease their work. But for a maintainer of a package it should be entirely optional to react to it, unless various points have been deemed legit by members of the DFSG team. I have seen some of these reviews that did make sense to me from a first glimpse. However that does not take away the general concerns being raise here. > Picking at random a successful[0] and unsuccessful[1] LLM-generated > review from the dashboard, the successful review consists of 734 words > (~2.5 printed pages) and the unsuccessful review weighs in at a > whopping 1,535 words (~5 printed pages). […] Regarding length of the output, there might be ways to guide the involved LLM to produce shorter output. Examples of prompts that apparently have been proven to be successful for various purposes may be found within the Fabric project by Daniel Miessler¹: https://github.com/danielmiessler/Fabric/tree/main/data/patterns One of the greater challenges with LLMs seems to be to guide them to produce concise, accurate and to the point output. Recently in an experiment to develop a script I told a LLM just to send me the one line that needs to be changed to fix up an error message I posted to it, cause I simply got annoyed it was repeating all the other stuff all over again. It did then just send that one line. [1] I tested some of these, but certainly did not evaluate all of these. So your mileage may vary. Best Regards, -- Martin
