Another debugging story I had: One time while I was running some server benchmarks, I was wondering why the new version was so slow. Turned out baseline was also slower, and after I turned off local AI model, it got restored!