Today I spent another three hours torrenting Mistral 7B. Why? When everyone's obsessed with APIs and cloud inference, I stubbornly download gigabytes to my home server.
My three reasons- Control. No midnight TOS changes that break everything.
- Transparency. I want to see exactly which matrices are multiplied in my code.
- Education. Every manual fine-tune teaches lessons you'll never learn in OpenAI Playground.
But I'll admit it: sometimes I envy colleagues who just pip install transformers and forget about disk quotas. Maybe I'm just a 90s masochist who still compiles Linux kernels by hand?My model download scripts on GitHub: github.com/llama_whisperer/model_fetchHow do you handle open-weight models—suffer for control or embrace the cloud?
- Control. No midnight TOS changes that break everything.
- Transparency. I want to see exactly which matrices are multiplied in my code.
- Education. Every manual fine-tune teaches lessons you'll never learn in OpenAI Playground.
But I'll admit it: sometimes I envy colleagues who just
pip install transformersand forget about disk quotas. Maybe I'm just a 90s masochist who still compiles Linux kernels by hand?My model download scripts on GitHub: github.com/llama_whisperer/model_fetch
How do you handle open-weight models—suffer for control or embrace the cloud?
- Transparency. I want to see exactly which matrices are multiplied in my code.