From working with coding agents, I get the feeling that if an AI has made mistakes in the past, it will be less likely to be able to complete a complex current prompt, even if the current prompt is independent from the past mistakes.
I religiously start a new session whenever I notice that the AI has started to make mistakes, but it would be cool to empirically see if this actually improves performance or not.
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{nguyen-do-past-mistakes-2025,
author = {Nguyen, Dang},
title = {Do past mistakes affect an AI's present performance?},
year = {2025},
url = {https://hypogenic.ai/ideahub/idea/d6R5Jfn0HS0uFLF4NpIx}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!