One thing Kaczynski's brother noticed as particularly idiosyncratic was the consistent use of the phrase "you can’t eat your cake and have it too", which is usually phrased as "you can't have your cake and eat it too".
— https://en.wikipedia.org/w/index.php?title=You_can%27t_have_...
So the order implies the temporality of the actions.
Joking aside, the key word which is sometimes implied rather than included is “too”. The order isn’t important. The saying is both things can’t simultaneously be true.
"Hey, that guy only communicates using cutouts from magazines, what a strange choice"
LLMs introduce all kinds of linguistic choices, and you can focus on those choices.
That's a bold and unproven statement, made worse because we can't really see that fingerprint.
Fingerprint matching of course isn't completely useless, but it's not as solid as you'd hope either.
And now LLMs are going to add more noise to these features...
AFAIU the more people know of it the better expectations are set about real account privacy
[0]: https://tails.net/
I strongly suggest people try ollama - it takes a few minutes to set up, download a local model and you’re up and running. https://ollama.com/
My specific mention of Tails was because it's designed as an ephemeral OS for the paranoid / extremely security conscious, so it would make sense that they would consider something that allows further obfuscation of a user's identity.
"Eats, shoots, and leaves"
Or is it
"Eats shoots and leaves"?
;)
Book "Eats, Shoots & Leaves: The Zero Tolerance Approach to Punctuation"
Time to eat grandma
Vs
Time to eat, grandma
But of course in reality it is much more complex. All forms of forensic evidence are vulnerable to noise: sure, that one interesting artifact may be evidence but it also may not be, and the inverse may be true for that perfectly ordinary thing everyone missed. A linguistic quirk can be a piece of evidence but it can also be accidental (e.g. nowadays it might also result from bad predictive typing or autocomplete, or even more recently, as others have pointed out, LLMs).
So all of these are in effect just probabilistic filters. And filters are only useful when your sample size includes your target (i.e. if the actual perpetrator is a suspect and you have adequate data about them). And even then they may not only produce false positives but also false negatives and these may interact the more filters you attempt to combine.
Forensic linguistics can be useful when you have a small set of suspects that you absolutely know includes the actual perpetrator. But otherwise they can send you on a wild goose chase or hurt the innocent.
https://likeinamirror.wordpress.com/2013/12/01/satoshi-nakam...
The use case is for competitive gaming where a player can get a major advantage by using several accounts. So the software can be used for screening and detect accounts that are suspicious alike.
https://news.ycombinator.com/item?id=33755016
Tool seems to be dead, but link to it for the related discussion.
