47 minutes ago · 15 min read3079 words · Tech · hide

In my last post I shared how a model running on the box under my desk was able to rebuild most of my life from the sessions I had sent to AI labs. This scared me so much that I decided to run a follow-up experiment: if a local model was enough to build that profile, maybe I could come up with a way to leverage a local model to limit the amount of personal information and sensitive data leaked to inference providers. My first idea was to build a privacy filter, so I spent the last two weeks building a privacy filter, and I have to admit that I am not proud of how this experiment went (this is what you get when, on top of all that I do in my day-to-day, I let a background agent autonomously work on an unspecified problem without active supervision).For a few days I had a number that looked great, a 69% drop in what a model could correctly say about me after the filter. I was so excited that I had even started writing this post around it to publish it last week. Then I realised the…

Commenting on Bubbles has been removed. Why? If you want to talk about this post, write to the author, use the comments on their blog if it has them, or reply with a "RE:" post on your own blog. Suggestions? feedback@bubbles.town