HACKER Q&A
📣 raincole

What's the state of art of local/open "Copilot"?


I've seen a lot news about local LLM and people's attempts to fine tuned llama 2 on HN. But surprisingly I haven't seen people to build a local version, or at least open API version of Copilot.

What's stopping people from eating GitHub's cake?


  👤 version_five Accepted Answer ✓
Maybe I'm misunderstanding your question. Off the top of my head, there is codellama and starcoder, I'm sure I've seen various other ones. These are downloadable coding models that you can run on your compute. Did you mean something else?

👤 dtagames
What's stopping them is the millions of dollars in compute costs that it takes to train the model, as well as the expense of developing the interface with the IDE to make Copilot perform as well as it does.

👤 smoldesu
> What's stopping people from eating GitHub's cake?

Iunno, convenience? Most devs are schlepping along on an M1 Macbook Air that's lucky to hit 5 tokens-per-second on a local model. They could be running a local model and heating up their lap, but most would rather pay Microsoft $10.

Coincidentally, the local programming models are quite nice nowadays. If you're interested, I recommend trying it out.


👤 muzani
There's sruff like https://codeium.com/

In fact, there's a lot of coding pair program products. Some open source fueled, some OpenAI fueled, some are doing their own weird thing years before Copilot was released.