Hacker News Reader

I'm not sure if that's already exposed as an option, but if you're happy to run this app locally, then LiteLlm exposes local models using the OpenAI API. That means you only need to change the endpoint, tell LiteLlm to use for example Mistral and it should "just work". Even if not implemented yet, it would be a trivial change.

View on HN

> It would be a trivial change.

My favorite line in the tech industry.

Trivial change, except for also having to change all the prompts to work well with Mistral. LLMs are not equal, you have to ask each one "nicely" in a different way.

You can change the prompts if you want better results. But you don't have to do it. I've replaced GPT endpoint with LiteLlm a few times already and had to adjust the prompt only once. If I can't see the difference, why bother?

The project uses the ai npm package from Vercel which does interface with a lot of different LLMs. But I imagine there is a little work for it to just work with any provider.

I think it would be really cool to get it running with something like Mistral in the browser: https://webllm.mlc.ai/

Reply on news.ycombinator.com