Conversation

Notices

1. Embed this notice
   Matt Hamilton (eriner@noauthority.social)'s status on Friday, 14-Feb-2025 12:53:01 JST Matt Hamilton

   The codegen LLMs I've been using for the last quarter are quite good, given that they fit on my single GPU.

   Beyond boilerplate, given a description of a program, it can reliably produce functional code to do one-off things. Useful for both writing tests and writing implementation based on tests I've written manually.

   It's also quite good at creating a styled webpage based on a description, as well as fixing CSS issues when given commentary about the visual bug.

   I ❤️ Artificial Indians.

   • Embed this notice
     Robert R. (iloveamericanews@noauthority.social)'s status on Friday, 14-Feb-2025 12:53:36 JST Robert R.

     in reply to

     @eriner

     What are you using? I would love to play around with AI code generation.

   • Embed this notice
     Matt Hamilton (eriner@noauthority.social)'s status on Friday, 14-Feb-2025 13:24:19 JST Matt Hamilton

     in reply to

     • Robert R.

     @ILoveAmericaNews Here are the two models I use most often: https://noauthority.social/@eriner/114000086682956219

     I wrote my own wrapper around llama.cpp to handle dynamic model switching based on my task, but LM Studio is probably the best place to get started using LLMs locally on your own hardware if you're just starting out.

   • Embed this notice
     Robert R. (iloveamericanews@noauthority.social)'s status on Friday, 14-Feb-2025 15:35:02 JST Robert R.

     in reply to

     @eriner

     Thanks. AI coding is super exciting.

   • Embed this notice
     Matt Hamilton (eriner@noauthority.social)'s status on Friday, 14-Feb-2025 15:35:22 JST Matt Hamilton

     in reply to

     • vectrex / 56001 DSP lover

     @petaluma Presently I use qwen2.5-coder-14b-instruct for dev work, primarily Go and web stuff (Vue, Tailwind, etc). I use mistral-nemo-instruct as my general purpose rubber duck.

   • Embed this notice
     vectrex / 56001 DSP lover (petaluma@c.im)'s status on Friday, 14-Feb-2025 15:35:25 JST vectrex / 56001 DSP lover

     in reply to

     @eriner Deepseek was too noisy for me so I went back to llama3. Also DS couldn't generate DSP assembler code very well while llama3 did great! I mean it can't do signal gen code but it can generate enough code for an interpreter / monitor.

     What's your fav model for code gen?

     Robert R. repeated this.
   • Embed this notice
     Matt Hamilton (eriner@noauthority.social)'s status on Friday, 14-Feb-2025 15:35:26 JST Matt Hamilton

     in reply to

     My productivity is probably 2-3x what it was prior to integrating LLMs into my dev work, if not greater.

     Others may have a different experience, but being able to write *some* of the code and instruct the LLM to finish the monotonous parts is great. It frees my mind to think about what I need to do next while I wait for it to churn, as opposed to having to be fully engaged typing out loops and conditionals for 2-5 minutes.