Privacy

Its not bs, pattern analysis of large data sets is literally exactly what LLMs excel at.

Human language is way way way more complex and arbitrary and inconsistent than if you train or use a model to only need to consider a defined kind of data output that follows strict rules precisely, as you'd have in a structured cache of memory.

The demo is also apparently on the way, I believe they said in the paper or the article that they're working toward showing off / making available an actual working example.

There have been previous exploits and methods sort of similar to this kind of 'timing as a spy' method, that can literally discern useful information based only on the minute timing variations of overall power draw from a wall socket a system is plugged into, insane shit like that.