Introduction [[2022-09-24-AI_Copilot|2022 Copilot]] [[2024-08-24-Claude_Sonnet_Copilot|2024 Copilot]]
I used VS Code with Copilot in 2022 and changed to different plugins such as ContinueDev and ClaudeDev in 2023. In 2024/08, Karpathy posted in X that he switched to use “Cursor”, a branch from VS Code but adding the LLM capability. It is a startup company where founders come from MIT. Since I believe Karpathy’s judgement by watching his YouTube videos, I decided to try Cursor to use Cursor.
Cursor can be seen as Method 2, i.e. non-agentic approach.
There are two method to setup the VS Code:
- Easy setup but expensive: ClaudeDev
- Slightly complicated but flexible and much cheaper: ContinueDev + Sonnet + Qwen
A table compares method 1 and method 2:
| VS Code | Method 1 | Method 2 | Comment |
|---|---|---|---|
| LLM | Claude-3.5 Sonet | Claude-3.5 Sonnet | Input $3/M-tokens Output $15/M-tokens |
| Plug-in | ClaudeDev | ContinueDev, Cursor | |
| Function | Agent for task | Code and Auto-fill | |
| Agent | Claude | – | No agent for Method 2 |
| Code | Claude | Claude | |
| Autofill | – | Owen2 | Owen2 is local, free! |
| Comment | Agent is inmature Cost is expensive |
Example
I used transformer model as an example.
Prompt
兩年內 LLM (大型語言模型) 帶來的變化令人驚訝。我在2022年寫了一篇關於使用 VS Code + Copilot blog。GitHub Copilot是由OpenAI的Codex模型驅動，而Codex模型是GPT-3模型的後裔。每月支付10美元使用這麼舊的產品實在有點離譜。新的小型語言模型（例如，Microsoft Phi3）可能表現得比Codex更好。
Reference
Sonnet-3.5 ClaudeDev Extension: https://www.youtube.com/watch?v=5FbZ8ALfSTs&t=678s Sonnet-3.5 + Qwen2 + ContinueDev: https://www.youtube.com/watch?v=Vxsx7Il-KMA