I was sitting in the auditorium, early 2010s at DEF CON ~X[¿I?]X~, when General Alexander gave the headlining speech of that conference (then-Director of NSA).

Within the speech he defined the world "intercept," within the intelligence community, as meaning a human operator has (in some manner) catalogued some piece of information.

The implication was that all data in stored forever, and machine learning tasks were making associations without meeting their definition of "having been intercepted" — even with the elementary ML of fifteen years ago, this was a striking admission.

----

This was among the first things I thought about during my initial weeks using GPT-3.5 (~January 2023): that most of these conversations wouldn't be considered "intercepted" despite this immense capability of humanless understanding.

Now, almost three years later, I_just_hope_our_names_touch_on_this_watchlist.jpg