News for hackers. Stuff that matters.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by gaze on Thursday October 31, 2024 @03:23PM (#42010494)
I guess this is as good a place as any to ask -- what's everyone's favorite AI code assist tool?
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by nicce on Thursday October 31, 2024 @04:17PM (#42011054)
Zed's integrated tools have been more than enough for me.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by mattnewton on Thursday October 31, 2024 @03:29PM (#42010558)
cursor. First one I've tried that seems like it's more than a neat demo.

- but I'm weird and I usually disable tab completion, I find having generations popping up while I'm typing slows me down, I gotta read them and think about it, feels like it's giving me ADD. So I've always kinda been a copilot hater. Lots of people find this more productive, and a fancy version of it is on by default in Cursor. However, Cursor implemented a bunch of different interfaces well not just the copilot one, and I find the chat window in your editor for churning out boilerplate or refactors is a huge productivity win personally. There are a lot of one-off refactors that are annoying enough that I wouldn't want to dedicate an afternoon to them but now they are taking me just a few minutes of reviewing AI changes.

Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by written-beyond on Thursday October 31, 2024 @03:56PM (#42010838)
Exactly why I never went with getting copilot, I instead got a chatgpt subscription and prompt for stuff I need.

I do sort of regret it too, sometimes you just want to give more context and it's a hassle at that time. Figuring out what is it you need to paste to ensure the model has adequate context to generate something valid. Also, Claude is magnitudes superior to ChatGPT anything. Both are terrible at implementing abstract completely unique code blocks, however ChatGPT is significantly more "markov-y" when it comes to generating any code. When Claude gets things wrong it feels like a more human mistake.

Anyway, with 50% of HN obsessing over Cursor, is it worth it? I couldn't get it to open projects I have in WSL2 and I kind of gave up at that point. I've gotten far with Claude's free tier and $20 for just cursor seems steep for something that's not as stable.

Have you assessed Zed auto completion or read about others experience of it? Zed seems like something with a more stable foundation than any of these VSCode forks.

Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by infecto on Thursday October 31, 2024 @05:09PM (#42011538)
Cursor has been my favorite so far but I also have never tried Codium. Copilot was the winner prior but honestly its just tab completion. I tried Jetbrains but it felt janky and slow. Cursor tab completion feels nicer, its super fast and will do updates based code changes. I like being able to quickly get it to write some code updates and it returns in a green/red line like a github PR. The flow is really nice for me and I am looking forward to the future.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by mattnewton on Thursday October 31, 2024 @07:13PM (#42012500)
My trajectory was sublime -> vscode -> cursor. I tried Zed, but didn't need any of the collaboration features, didn't notice any speed increases vs normal vscode usage, and was generally just less productive than vscode with all the extensions I had configured. Cursor imported all of those, and my keyboard shortcut muscle memory still worked right after install. For me at least, moving over to it was completely seamless.

Like you, I started using LLMs in a chat window copy pasting code back and forth, (but I think Claude 3.5 sonnet was the first model I felt it worth the hassle). The cursor workflow is basically indexing all your files in your codebase to figure out what to copy over into the LLM, and then scraping the LLM output to figure out where to paste it in your code. It works like 95% of the time, which is way more than I was expecting, and falls back to the manual copy paste workflow pretty easily. Also it comes at a time where models are good enough to be handed gobs of context and trusted with more than a handful of lines at a time (Claude Sonnet 3.5 really shines here).

The whole experience just feels very polished and well thought out. One great (not really well documented?) feature is a .cursorrules file that gets invisibly pasted into all the context windows where you can say things like "use double quotes in javascript" or "prefer functional paradigms and always include type annotations in python" to avoid having to make those edits for consistency and style if the LLM fails to pick them up from the surrounding code. You can commit this file and have teammates get the same part of the prompt. Now everyone's autocomplete consistency and style is improved.

It's difficult to state how nice having it automatically do this copy paste back and forth just a keyboard shortcut away feels, it really is much better than I expected. So yes, I would try Cursor.

Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by shadowmnifold on Friday November 01, 2024 @05:47AM (#42015372)
I am about 70% through my first paid month with Cursor.

I should say I have budgeted that I am going to pay $20 for something AI right now no matter what.

Cursor is worth it, but I should have had an exact project in mind when I subscribed. I don't think I am going to get my money's worth completely this month because of how generous Claude's free tier is. I do like just asking the web ui random things sometimes.

At one point a few weeks back I did have overlap with o1, Cursor paid and Claude free tier all at the same time. I know I am not subscribing again to chatGPT until the non-preview o1 is out.

I do like the Cursor UI but there is something about copy/paste with the chatbot that I also like in conjunction with non-AI VScode or jupyter. Like the difference between oil paints and pastels. Both probably have their place in the tool belt.

Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by jonathaneunice on Thursday October 31, 2024 @03:54PM (#42010815)
Cursor.

All in on tab completion and its other UI/UX advances (generate, chat, composer, ...)

Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by Alifatisk on Thursday October 31, 2024 @03:32PM (#42010582)
Cursor.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by aberoham on Thursday October 31, 2024 @04:26PM (#42011139)
aider-chat
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by victorbjorklund on Thursday October 31, 2024 @04:45PM (#42011330)
Aider AI.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by emmanueloga_ on Thursday October 31, 2024 @05:04PM (#42011492)
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by ghawkescs on Thursday October 31, 2024 @05:53PM (#42011941)
Same question, but for VSCode plugins. Besides copilot what is everyone using? Claude support is a huge plus.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by Y_Y on Thursday October 31, 2024 @06:41PM (#42012300)
Emacs.
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by sunaookami on Thursday October 31, 2024 @05:59PM (#42011997)
Cody
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by edm0nd on Thursday October 31, 2024 @06:21PM (#42012169)
I've been loving Claude Sonnet for python
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by acheong08 on Thursday October 31, 2024 @09:18PM (#42013166)
I don't use AI much but I do have CopilotChat.nvim for when I do need it. o1 and Claude support was just added yesterday. (Disclaimer: I am one of the maintainers)
Re:Support for Claude Sonnet 3.5, OpenAI O1 and Gemini 1.5 Pro by d4rkp4ttern on Friday November 01, 2024 @07:53AM (#42016041)
Zed - open source, written in rust and hence extremely fast. I’ve always been a jetbrains user and every time I tried VSCode it never stuck. For such people zed is great since you can set up the config to use jetbrains key maps. Also you can get it to work with open LLMs via ollama or groq or cerebras; the latter two require an unpublished hack; here’s my config, thank me later :)

https://gist.github.com/pchalasani/9e71c58d2f846412b253ae0ec...