Does any other model besides GPT4 not work at all?

Hello!

I’ve recently tried Cursor, initially using the invaluable GPT4 model, and I found that it really works: it suggests code changes to my js files and integrates them when I click “Accept”.

As I understand it, even though GPT3.5 doesn’t understand the code as well as GPT4, the ability to “accept” and integrate the code would still be the same, because I believe that this should be part of the Cursor, and has nothing to do with which model is used.

However, the fact is that with any model other than GPT3.5 or 4, it will just give me some code snippets (as an example) and, when I select “Accept”, it will either insert them into my original file at random or simply overwrite the code in the original file!

I’ve been using GPT3.5 on web pages for a long time. It understands my intentions very well. When I provide it with my code and ask it to modify the code to my specifications, it is able to return to me the complete modified code.

However, frustratingly, it seems that for me personally, GPT3.5 is useless on Cursor, and all other models are useless.

No offense here, I really think Cursor is a great product, but I would still like to venture a guess:

Cursor has been neutered in order to let us use GPT4, while using another model, the model will not know which file I am currently asking it to look at, and will not make changes to the current code, in short, using another model means that the normal functionality of Cursor will not be available.

I’m hoping I’m wrong, and that I may not have mastered some of the tricks. Please tell me what I’m missing? Much appreciated!

Despite the assurances of manufacturers of models prior to GPT-4, when addressing real-world tasks, they significantly or somewhat fall short. For instance, models from the Corporation of Light and Goodness—in my system—are incapable of even working with plugins. (GPT-3.5, on the other hand, handles them quite well.) When attempting to code anything beyond a calculator, it spouts utter nonsense (GPT-4 also generates semi-nonsense 50% of the time, but if you make corrections, there’s a chance to achieve something adequate.) Something adequate attached to GPT-4 (outperforming other models to a lesser extent) as of today can only be done by Mistral Large (but it’s nowhere near worth 80% of the token cost of GPT-4).
So, as of today, of course, you can add it, but working with them will be unpleasant, inaccurate, and completely pointless—in terms of coding. Just so you know, I’m just an end user, this is my personal opinion.

Hi there, would love some more information about how Cursor is falling short.

Are you talking about the cmd-K feature or applying through chat? If it is cmd K, it would be helpful if you flagged the suggestions so we can continue to improve the feature (Cmd-shift-P → Developer: Report Bad Command K Suggestion).