Found on Rachel Woods on Twitter: “Has anyone experienced or studied prompt drift? When prompts were performing well at one point in time, but due to model changes, degrade in performance? Seeing this on one of our projects.” Twitter on 2023-05-22 12:45:30 +02:00.

Here’s something interesting that shows that human language isn’t the best way to interact with AI and expect consistent results.

Has anyone experienced or studied prompt drift? When prompts were performing well at one point in time, but due to model changes, degrade in performance? Seeing this on one of our projects.