Post #2684417

2026-05-15 13:09 UTC

@nihilistic_capybara@layer8.space which languages? Common languages can be translated sort of, and LLMs have made some of those translations better. But LLMs are still only maybe 80% or 90% accurate for any kind of text analysis. So if you're speaking regularly in a common language, they can probably get a rough idea of what you're talking about. More obscure languages are less accurate. But it gets even less accurate on specialty language and obscure references. LLMs give the illusion of surveillance without the detail needed to make it actually work... Because, if LLMs magically made surveillance work then no one would ever be able to circumvent censorship, and I don't think DeepSeek fixed that problem. Languages are not necessarily about mass surveillance, so much as they are about increasing the barrier of entry for infiltrators.

Replies (3)

@nihilistic_capybara@layer8.space 2026-05-15 13:41

@Hex@kolektiva.social most obscure language i speak is polish so i guess you got a point there

Open ##2684418
@Hex@kolektiva.social 2026-05-15 13:49

@nihilistic_capybara@layer8.space there's a thing about infiltrators: they work for a living too. They need to add stuff to their resume. They're trying to get promotions. Very few of them (on that side anyway) are actually driven by ideology alone. Even if they were, they still have bills to pay. So "spend a year or two learning a language before you can even find out what your target is really talking about" is a bad investment. It's better, in terms of RoI, to just entrap some teenagers or mentality Ill folks. You can get the same promotion for less work. And, if you can get a bunch of feds to waste time learning some obscure variant of elvish in order to find out if some weird DnD group is actually a radical leftist cell or not, then you win anyway.

Open ##2684419
@nihilistic_capybara@layer8.space 2026-05-15 17:59

@Hex@kolektiva.social i am currently being whored out by my company to banks to parse their documents and extract specific parameters from it and analize text in all sorts of languages and we are hovering at a 98% accuracy and that's just because we are beeing forced to use openai models for "privacy" or whatever. Sonnet 4.6 doesn't make any errors that I could find I am a single guy with a 10 year old laptop and an API key. Fuck knows what Palantir can do nowadays.

Open ##2684421