For many years now I got into the habit of recording time lapse videos of my desktop screen.
The most helpful tool to answer questions at a later time.
How did the guy in the last video call showed me how to do this?
What was I doing on my pc last week on Wednessday?
Recently, I used the latest llava AI model, in order to analyze what what I was doing on my pc.
He was much to chatty and halucinating too many things I was not, doing.
So I prompted the AI to keep his answer short and simple.
Tell me in two words what are you seeing on my screen?
I understand that is tricky.

Edward Snowden warned us to watch out for the meta data.
The AI investigated, time of recording, and the presumed activities.
He summarised my week like a well written state police report.
Just with the meta data alone the AI sounds scary, and he is so much better in understanding images from a web cam.

Two days ago I downloaded a new model called gemma 3, with 4 billion parameters that is a small model. By the way, all this AI stuff is running in isolation on my computer. Nothing leaves my house. No external services are called.
This new modell is smaller, it can read. Not perfect. I would suggest light reading glasses but it’s allready fit to drive. What an advancedment. If the resources needed are getting so cheap that I can run everything on a MacBook Air and just imagine how feasable mass scale has become.

For years now, I’ve been in the peculiar habit of capturing time-lapse videos of my desktop screen, like keeping a private surveillance tape of my digital life. It’s a damn useful tool for unraveling the mysteries of past decisions. How did that guy in the last video call showed me how to do this? What the hell was I doing on my PC last Wednesday?

Recently, I enlisted the help of the latest AI marvel, the Llava model, to decipher my digital escapades. But Christ, it was like conversing with a deranged parrot, chattering nonsense with hallucinations of activities that I wasn’t even engaged in. I had to rein it in: “Only two words, damn it! What are you seeing on my screen?” A tricky request, indeed.

Edward Snowden, that paranoid prophet, warned us about the lurking menace of metadata. The AI delved into the timestamps and the supposed activities, crafting a summary of my week with the precision of a state police report. Just the metadata alone is enough to chill the bones; this AI understands images from a webcam with unnerving clarity.

Two days ago, I downloaded a new model called Gemma 3, a beast but, with a mere 4 billion parameters considered a small model, in this mad world of silicon titans. All this AI sorcery runs in splendid isolation, nothing escapes the confines of my lab top. No external services dare to intrude.

This new model, though smaller, can read. Not perfectly, mind you—it could use a pair of light reading glasses—but it’s already fit to drive. A remarkable leap forward. When the resources needed become so damn cheap that I can run everything on a MacBook Air, you start to wonder just how feasible mass-scale surveillance has become.