Cursor + Claude 3.5 Sonnet is doing very dumb things

For several days already I was noticing degradation in Cursor responses, to the extent that it couldn’t finalise very simple tasks (in a smart way). It definitely was able to accomplish such tasks in much better way previously. Then I noticed that I actually ran out my 500 monthly requests and got switched to so called slow requests. My question is if those fallback requests are only slower, or also dumber? I’ve read through the site and forum, I do not see a clear answer to this anywhere. I think it is understood that quality should be the same?

1 Like

I’m noticing something very similar. I was humming along making great progress when suddenly Claude 3.5 Sonnet was unable to integrate any of the code it created in a sensible way even though I’m still well within the limit of my 1000 monthly fast requests. The IQ of the AI seemingly dropped by 50 points out of nowhere, to the point that I can no longer use Claude 3.5 Sonnet within cursor. I either need to switch to ChatGPT 4o or use the web based version of Claude to get the job done. No idea why this is happening and would like to see this resolved ASAP.

1 Like

This checks out.

I was humming along when suddenly Sonnet-20241022 started more or less giving me advice instead of solutions. Half of it was either what already existed or just wrong.

Someone blew a quantization at Anthropic :sweat_smile:

I’m experiencing the same with all AI models, outside of cursor as well, with GPT-4 it seems it just have the bad day where it just don’t get even simple things.
Sonnet sometimes does great thing other fails on super basic stuff that I’ve asked about only out of laziness

I think it’s just the way AI currently works. Or maybe they adjust like model size dynamically as per load, who knows.

But I wouldn’t say that it’s cursor specific

1 Like

I just paid an extra 20 to continue work and my project has been destroyed this morning. I’ve gotta roll back, no issue there but getting hosed on this. Every single prompt response ends with “Would you like me to show you the remaining fixes”. When I respond yes it’s like it’s completely lost the plot and starts making unrelated and unusable suggestions. This is extremely frustrating.

1 Like

This is getting ridiculous. 4 times I have asked changes to be applied and 4 times it’s spewed out a huge response in the chat interface and then asked me if I wanted to apply the changes.

We’ve been testing Sonnet-20241022 as part of our product and it’s a steaming pile of crap. I would recommend reverting to sonnet-20240620

1 Like

I don’t even see the option for 20240620. My issues don’t seem to be entirely sourced from the LLM. Like… cursor is flat out f’d today. Getting suggestions that aren’t being applied and then re-asking to apply the changes and getting new suggestions. Just a waste of money. I need to step away until this is fixed.

Agreed, something weird is happening, behavior is just not the same as before. It ruins the old working code randomly and can’t help me solve issues in 20+ prompts, stuff that it was easily solving with just 1 prompt before

Just a quick tip – watch out for those “would you me to go ahead and…” questions. It’s something I noted in the addendum to the Idiot’s Guide post a little while back: if you get one of those, don’t respond with “yes”. You’ll want to make sure you always give something longer and more specific like “yes please go ahead and apply those most recent suggestions to improve the brangel widget” or whatever. From experience, a pure ‘yes’ has a dangerously high chance of making it jump into something completely irrelevant.

I have a feeling that cursor is actually not using Claude 3.5 Sonnet after a certain while and is using a lesser model in the background because I’ve experienced the same thing constantly. It will start off great and then 30 minutes later it will just start randomly deleting large sections of code or not able to fix simple things. I’ve been spending more time trying to fix it’s errors at this point…

Same here. I don’t think it is related to Sonnet-20241022 since it used to work perfectly fine, but just all of a sudden degraded early this week.

Cursor team must be tuning something in the backend and broke things. Not sure if this intended or not. Since lately Claude has been down (being rate limited) for them, maybe they are fixing something or tweaking something and limit the api calls with the real Sonnet model.

Basically the model this week cannot answer very basic things.

This is the aspect that I don’t like for close sourced product - your work and life started to depends on it and all of a sudden it can be all gone…

It happened before for Cursor, when they switched quietly from GPT4 to GPT3.5 due do the huge usage on the backend one year ago. Get frustrated and this time again

Hopefully they can get a fix asap

1 Like

Shameless plug, but do check out An Idiot’s Guide To Bigger Projects if you haven’t already (not being rude, the idiot is me not you!). There are some tips in there specifically related to degradation in Composer sessions and why it’s a good idea to start fresh ones as soon as you start to see the quality drop off.

1 Like

I have the same experience. Today was the worst experience for many days. This inconsistency between Cursor and Sonnet is killing. I can understand their intention to optimize but it killed off the customer experience whatsoever.

If on the beginning of the week my biggest problem was AI not follow instructions and often try to ease out by hardcodding problems. Today, it cannot execute even direct instructions. Dunno how we can rely on such wild card solutions in everyday tasks.

1 Like

Thanks for this. Yea I’ve just been starting new composer sessions when I get these issues. Like you say in your post, the best results seem to be at the beginning of a composer chat. When I get the circular changes loop and silly mistakes, that’s when I know it’s time for a new session. Maybe Cursor should actually do this automatically to prevent user’s frustrations?

Thanks again!

1 Like

i agree it gets worse and worse as the convo goes on and now im noticing its happening much more often with every different chat or composer session