Very interesting article by Proton on privacy and LLMs:

- First and foremost, you can't have privacy with a closed model. Many claim to be open sourced but the reality is very different

- You don't always need a large model. A small, task specific model will perform as well as an LLM, for that specific task, and can run locally on your device

- If you need a LLM and need to use cloud computational resources, Homomorphic encryption (HE) is a nascent technology that allows cloud based LLMs to operate on your encrypted data and the results are the same as if they had operated on your unencrypted data. HE is not ready for prime time but it will get there

Kudos to the Allen Institute for AI. It's great to see a local Seattle research institute leading with a true, open source AI model. 

#personalai  #privacy  #ai  #foss  #homeai

https://proton.me/blog/how-to-build-privacy-first-ai