Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ask HN: What's your experience with Cursor and Sonnet 3.5 for coding?

news.ycombinator.com

Post reply on HN
Print view View original

21–24 of 24 posts

Re: Ask HN: What's your experience with Cursor and Sonnet 3.5 for coding?

#21

Post by stormtitan »

fallinditch wrote: post #18

I'm in the same boat, about to get stuck in to Cursor. But I'm wondering if it might be better to go with a VS Code extension like Continue. With Continue you can use your own local copy of an LLM, and also query LLM API endpoints, so therefore better for a pay as you go solution and cheaper if you are a heavy user. Also you can carry on using VS Code as your editor of course. This is my understanding of the trade of…

My experience, and also what I’ve seen from a few YT videos as well, is that the local LLMs just aren’t as good as Claude and chatGPT. They don’t seem to understand or follow instructions as well as C and C.

Re: Ask HN: What's your experience with Cursor and Sonnet 3.5 for coding?

#23

Post by FloatArtifact »

fallinditch wrote: post #18

I'm in the same boat, about to get stuck in to Cursor. But I'm wondering if it might be better to go with a VS Code extension like Continue. With Continue you can use your own local copy of an LLM, and also query LLM API endpoints, so therefore better for a pay as you go solution and cheaper if you are a heavy user. Also you can carry on using VS Code as your editor of course. This is my understanding of the trade of…

It looks like it takes three different models chat/ranking/ etc. What is the total cost for three API or VRAM requirements for reasonable local open source models to run Continue?

Re: Ask HN: What's your experience with Cursor and Sonnet 3.5 for coding?

#24

Post by fallinditch »

stormtitan wrote: post #21
fallinditch wrote: post #18

I'm in the same boat, about to get stuck in to Cursor. But I'm wondering if it might be better to go with a VS Code extension like Continue. With Continue you can use your own local copy of an LLM, and also query LLM API endpoints, so therefore better for a pay as you go solution and cheaper if you are a heavy user. Also you can carry on using VS Code as your editor of course. This is my understanding of the trade of…

My experience, and also what I’ve seen from a few YT videos as well, is that the local LLMs just aren’t as good as Claude and chatGPT. They don’t seem to understand or follow instructions as well as C and C.

Yes, good point. The latest LLMs from Anthropic, Open AI etc will always be more powerful than your local LLM. I reckon that the cutting edge cloud-based models will (continue to) provide enormous value for a range of complex actions, but that we will also derive great value from our own personal and organizational local models running on our devices and LANs.

I'm assuming that small specialized LLMs for coding will become 'good enough' for a significant proportion of tasks. For example, when Codestral came out earlier this year it supposedly beat Meta’s CodeLlama 70B model in all of the top coding evaluation benchmarks, even though it's 3 x smaller.

Post reply on HN