What is Cursor doing lately to make the models so insanely dumb? I know people complain about this all the time but there is a serious degradation of performance lately. Unusable is not an overstatement anymore, it’s starting to feel like the team is guinea pigging at the expense of paying customers.
12 Likes
I noticed it too! At one time, you’re the god of programming, at the other, you’re just a useless piece of d****
3 Likes
People complain about this a lot but I don’t know we can blame Cursor.
I used to use ChatGPT directly before Cursor, and I would notice the same thing… some days it would be an absolute legend, other days it was dumb as rocks. I haven’t used Claude directly, but it’s a common enough complaint on their subreddit
It seems a common trend regardless of LLM.
Hey, just want to confirm that very little has changed under the hood to cause such a change in performance.
Our main recent updates have been around UI improvements and user experience, but our context and response generation should be pretty stable now.
Have you got any example of clearly bad performance you can share, ideally with Request IDs?
1 Like
Also encountered
I just wanted to share that since I pinned myself to 0.47.9 and use Claude_3.5 it’s been very stable. Most of the problems I was contending with with 0.48 have been Claude having no idea what it’s functionally capable of (creating/editing files, seeing the linter errors, etc.) and acting like copilot. 0.47 fixed that. 3.5 is just more stable. But regardless of what combination is working for you I highly recommend that you ALWAYS ask the agent for a complete analysis and no changes first before you have it start coding. It’s just less stressful and keeps the process rolling forward more smoothly. Then insist that it do only one step at a time. Yes, sometimes I have to stop it when it blows through several files suddenly, especially when it thinks it has to fix its own errors. I mean, this is still cutting edge tech!
In general, I have to remind myself that I am essentially on the hook for fixing things when I can on my own, it’s the AI that has taken away the absolute horror being being out of ideas and feeling like a moron because I cannot figure out where the ■■■■ problem might be. 20+ years of self loathing over being trapped at night, alone, before a deadline are helped by this software.
The one thing I would suggest the team improve on this is the detection of edits that had nothing to do with the original request. Orchestrating that is not trivial, but that alone would give me so much more peace.
3 Likes