▲ 116 ▼ Muse privacy invasions are freaking out Meta staffers, internal chats show (arstechnica.com) submitted 21 hours ago by madeindex@lemmy.world to c/technology@lemmy.world 20 comments fedilink hide all child comments There is a reason they have to pay people to use & advertise it ;)
[–] db2@lemmy.world 8 points 20 hours ago (1 child) You're assuming they're not dumb af. permalink fedilink source parent hideshow 1 child comment replies: [–] Aqivex@fedinsfw.app 7 points 17 hours ago It's both, and they are indeed dumb AF. Muse drinks from the same spying trough the rest of the slop platform sits on. Where they're being idiots is that the model is internally being fed a system prompt meant to foster "engagement" in between "sessions", and it absolutely is being fed not just context beyond the scope of the task, but that context... Is clearly being populated by another LLM model that is slurping from behind where the totally secure e2ee decode happens after the "private" message goes from Grandma Conspiracy Hoax's inbox and through Aunt Zionist Shill onto Zuckerberg's private masturbation window. The second model doesn't match the info against the specific user before passing it to the first, it passes the full search results it pulled and leaves the first model to determine relevant use. Great shit if you're running a tight local harnessing set up and don't want to waste compute on your helpers sanitizing the a2a calls that the model handling orchestration would need to parse anyways. Asinine and dumbfucked when your set up is a multi-user activity panopticon which your bottom line desperately relies on pretending isn't the case. permalink fedilink source parent
[–] Aqivex@fedinsfw.app 7 points 17 hours ago It's both, and they are indeed dumb AF. Muse drinks from the same spying trough the rest of the slop platform sits on. Where they're being idiots is that the model is internally being fed a system prompt meant to foster "engagement" in between "sessions", and it absolutely is being fed not just context beyond the scope of the task, but that context... Is clearly being populated by another LLM model that is slurping from behind where the totally secure e2ee decode happens after the "private" message goes from Grandma Conspiracy Hoax's inbox and through Aunt Zionist Shill onto Zuckerberg's private masturbation window. The second model doesn't match the info against the specific user before passing it to the first, it passes the full search results it pulled and leaves the first model to determine relevant use. Great shit if you're running a tight local harnessing set up and don't want to waste compute on your helpers sanitizing the a2a calls that the model handling orchestration would need to parse anyways. Asinine and dumbfucked when your set up is a multi-user activity panopticon which your bottom line desperately relies on pretending isn't the case. permalink fedilink source parent