Hi folks, On 11/9/24 01:26, DebGPT wrote:
This is an experiment, by letting LLM go through all 369 emails from debian-devel on Oct.
I received lots of feedbacks from the experiments, from positive ones to negative ones. It wasn't discouraging to see negative feedbacks since that is usually what would happen when people see something that has not appeared in the past. However, I see the core value of this experiment as the opportunity that the community and I can learn from it. Speaking of AI summary in the context of Debian community, I think it brings more harm than benefit if broadcasted to general audience -- unless the industry can realiably address the hallucination issues. While AI summary might be helpful to some extent, it in fact requires a certain level of expertise from the user, in order to properly tell its good parts and bad parts, and make use of its good parts. However, assuming a public general audience with such expertise (or at least tolerance) is not practical -- there might be people who consume the bad parts as well along with the good. We simply cannot bear the cost of forcing people to learn to tell the truth and the hallucinations. Whenever the community needs to face more interaction with AIs, this experiment can serve as an example for precaution. Apart from that, I'll continue trying to explore the ways to make such new technology useful since I'm interested in it. I can leak some of my plans in this regard: 1. Let LLM answer the NM templates (maybe with debian policy or debian developer reference in context) and see the percentage of questions that can be answered correctly. Even if I don't do it, maybe new DD applicants will. 2. Continue adding features to DebGPT and make a major release. Since the first time when DebGPT was announced, I wrote many new stuff to this tool -- and it gradually became my daily terminal LLM tool (I cannot find a better one on pypi). One of the interesting features is to edit file inplace, automatically git add, generate git message and commit. This has been very reliable and useful for simple tasks like adding documentation and type annotations in DebGPT's python code. Really saves lots of time. An real example for the described fully automated pipeline: https://salsa.debian.org/deeplearning-team/debgpt/-/commit/4735b38141eafd6aa9b0863fc73296aa41562aed What I did is just type the instruction in natural language, then the implementation is automatically committed in git. For me that's a fun part.
