I am shocked that the privacy (& so security) implications seem to have been mostly ignore. If a computer isn’t powerful enough to run an LLM, that means it must use a ‘cloud’ LLM service - and basically all of those are known to keep permanent records of all user input for training, etc. And a lack of privacy is just one cloud security breach (or hacked online account) away from becoming a target of hackers yourself.
To be frank, I don’t want my computer’s logs (and whatever else might be sent) permanently recorded in the cloud, no matter what their opaque ‘privacy’ policy appears to claim.
An “AI kill switch” is absolutely essential, just from a privacy & security point of view. It should have at least 3 options:
- Disable all AI features.
- Only use local AI compute (and if that’s not available then the feature is not available).
- Full AI features, which requires logging into your chosen LLM provider that you trust.