Haven’t used Cursor yesterday but since today I also encounter really bad quality answers (using Claude 3.5). Does anyone know the cause? Trying gpt-4o now…
I was very skeptical of this theory until it happened to me the 3 days ago with a seemingly easy enough request for sonnet 3.5. I tried 3 times in the composer (restarting the session, new reformulated prompt) and they all failed. It was changing code that it wasn’t supposed to.
I fired up VS Code and gave the task to copilot agent with sonnet 3.5, 2 prompts later it was done and working.
I heard that Microsoft host their own instance of Sonnet, so they don’t depend on the API from Anthropic. So now when Sonnet 3.5 underperform, I no longer fight it and just switch to copilot.
Yep, idk what’s happening, but definetly, Cursor is becoming useless… Went from using it the whole time, to… being tired of rolling back and starting from scratch.
My theory is:
user asks for newer models
Cursor adds them
but they use same prompting and params (temperature, etc.) for all models
OMG Yes! I switched to Cursor after using CodeBuddy for a while but getting annoyed that CodeBuddy just kept making the same stupid mistakes over and over again. Cursor seemed like magic.
But now, I don’t know why, but it really seems to be getting dumber. I have wasted like $10 today just fixing a few simple things that, admittedly, I should probably just be able to fix myself. But I’m stubborn. I even use ChatGPT to help me explain to Cursor what I want to do to because my original version gets more and more riddled with typos and obscenities as I try for the 20th time to get it to so something. It’s getting really expensive.
It’s failing to find the files it should use and is constantly making new ones. It’s ignoring the cursorrules file. It’s ignoring my naming conventions.
I should just take the day off. It’s been off to a really frustrating start today trying to get Cursor to help me fix some fairly simple and straightforward things in my app. I even through the hassle of writing out a bunch of different queries to help it understand how the data is structured and it just didn’t get it.