4 ms·
These papers usually have poor stability to prompting and rerunning. It would be nice if we had some kind of meta-evaluation metric where rewriting the prompt c
by arjie 5mo ago
These papers usually have poor stability to prompting and rerunning. It would be nice if we had some kind of meta-evaluation metric where rewriting the prompt conditions or varying the input params could be used to determine how stable a result is.
Regardless, it's definitely true that AI agents have different priorities from us. That's what alignment is about anyway.
- Chu4eeno 5mo agoIt's probably because they care more about the headline than figuring anything out: https://github.com/kennethpayne01/project_kahn_public/issues/1 https://github.com/kennethpayne01/project_kahn_public/issues... So you create leading prompts like that, and re-run until you get a publishable session.