Yeah like I said above, one of the main differences will be in the way that they handle context. Cursor and Copilot will both be attempting to handle this the best they can, but at the end of the day, the questions are being answered by (more or less) the same model GPT-4, so they shouldn’t be too varied either way.

I’m not sure what you mean by ‘enrich my question with the necessary parameters’. Sounds like marketing speech. Could you provide an example?

If I’m asking ‘What does this function do?’ then I’m going to get a pretty similar answer from each when providing them both with the same function and the same question.

Well, no
give to cursor something like this Conventional Commits to read, previously ask him to read carefully the codebase, than “ @diff with working bla bla bla” and see the results
Copilot just write “well this file has changed”
As I said, cursor needs more micromanagement, but sometimes results are way better that any other tool in this field.

As a reference I used for enough time easycode, bito, cody, copilot (+chat), codeium, codium and almost any other vscode plugin in the markerplace and even the beta AI assistant in Pycharm (all with a sub to test all their features, where possible)…so I think I know what I’m talking about

claidler:

If the commit is small, simple and clear it works fine for Copilot in my experience

I was talking about my experience. I didn’t say that you don’t know what your talking about.

‘Well, no’ - it’s my opinion? Are you saying I can’t have an opinion? My opinion is wrong?

claidler:

I’m not sure what you mean by ‘enrich my question with the necessary parameters’. Sounds like marketing speech. Could you provide an example?

Here’s an example: you spent ten minutes describing your situation, you attached a couple of hundred lines of code hoping to get a reliable answer in return, in the first phase your couple of hundred words were compressed into a simple question, in return you get a useless answer

You can answer this with an objection, tell the model that it gave a bad answer, at the first phase, your question may be supplemented with some information with which you hoped to clarify the situation by asking the first question.

In return you will receive a less useless answer.

This is what I’ve been observing for almost a month of active use

Plus constant loss of context

I have no reason to believe that the cursor directly transmits your question to the fourth model

Everything is much more complicated here

What I observed made me think that my questions did not fall into the fourth model at all, the answers came either from the cursor model or from GPT-3.5

psychaos:

As I said, cursor needs more micromanagement, but sometimes results are way better that any other tool in this field.

For what it’s worth, I agree with you here:

As I said, cursor needs more micromanagement, but sometimes results are way better that any other tool in this field.

I believe Cursor gives better commit messages also. I think it’s often a difficult task for an LLM to do at this point though, as it’s not always clear of the users intent in the code. Sometimes it requires knowledge outside the code.

My main point was that I’m personally satisfied with the messages Copilot gives me, but tend to not use this feature as it’s usually not perfect from either option, in my opinion.

I’m sorry but I still don’t understand what you’re saying here. Maybe English is not your first language?

I know it’s just an example but - 10 minutes describing your situation! I’ve never spent that long before - or anywhere near that long describing my situation. 1 minute max.

So breaking this down:

You’re saying that you ask a very long question with lots of code, and you get a bad answer from Cursor.
You then tell Cursor it’s wrong and it asks you another question to clarify your original question. Then it gives you a better answer.

Okay, so this is kind of good. Best option is getting a good answer from the start, but at least it’s recovered, when you’ve told it you’re not happy.

Then you say you have a constant loss of context - so this is obviously not good. It’s not remembering the history of things you say to it I assume. I mean if you spent 10 minutes writing a question and submitted hundreds of lines of code - I suppose it’s expected at some point to forget some stuff.

Then you say, ‘I have no reason to believe that the cursor directly transmits your question to the fourth model’. I think you’re saying that at this point you’ve gone back and fourth to the llm 4 times and the original question you asked is no longer being sent? I guess that’s expected.

‘Everything is much more complicated here’. I’m not sure what you mean? Maybe, the way in which Cursor behaves is more complex for some reason at this point.

‘What I observed made me think that my questions did not fall into the fourth model at all, the answers came either from the cursor model or from GPT-3.5’. There’s only 2 models cursor uses - gpt-4 and gpt-3.5, so I assume you mean the 4th interaction with the original model (rather than the fourth model). You say the ‘cursor model’, but Cursor is the name of the application. I assume you mean that the original question you sent has been transformed into something completely different by the time you’ve interacted with the llm a few times, at which point you think that it is giving you answers from either gpt-3.5 or 4. Yeah, okay.

Thank you for giving an example!

From what you’ve sent - it sounds like it’s behaving just like any other LLM behaves. You go back and fourth through a conversation until you get the answer you want. The LLM will ask you for clarity on your initial question, but eventually forgets context as you go on.

My personal experience is that Cursor provides better code generation and the chat seems less “dumb”. When I gave a complete prompt, Cursor told me how to separate the files for my code while copilot just gave it all to me in a single giant file.

Github copilot chat also seems to misinterpret what I am trying to say more often. When I gave a complete prompt,

On many ways, the Cursor is better than Copilot. becuase it provided to user more controllable options,such as context, images etc.
But there is a disadvantage of Cursor, that the Cursor is not support Visual studio, so if you are a Csharper, it’s difficult to use it.

just FYI: several months paid Cursor user, satisfied, but latest updates to GitHub Copilot made me try to temporary switch back to Copilot.

and so far i like it, main reason: it works much faster + there’s no limit to quries + their side by side modifications inside the code is much better

Yeah, I think Cursor is going to lose this battle.
So many people use GitHub and they have tight integration with the PR workflow.
Given both Cursor and Copilot are on VS Code, switching back and forth is easy… gonna bite them in the bum.

And they can’t compete with Microsoft’s pricing of $10 for individuals. Was thinking to pay for Cursor for the last few months but the price difference and value between Cursor and Copilot are getting too little.

Update on this: after 1 week I’m back to Cursor, so far I think it’s better than Github Copilot

I don’t know why, but Cursor makes more high quality edits to my code, probably because the team at Cursor is 100% focused on this project

Oleghuman:

just FYI: several months paid Cursor user, satisfied, but latest updates to GitHub Copilot made me try to temporary switch back to Copilot.

and so far i like it, main reason: it works much faster + there’s no limit to quries + their side by side modifications inside the code is much better