Post by @hopeless@mas.to

2026-04-14 13:09 UTC

@hastar @Possiblydrew The LLMs are actually deterministic themselves; you can give them the same PRNG seed and get the same result. It's wrong to say that they are "random word generators" like there is a dictionary and a random index into the dictionary is chosen... they are choosing what they want to talk about based on the weights they arrived at. You too will not recite what you want to say verbatim each time, but through your internal state decide how to express yourself.

Replies (3)

@Possiblydrew@pnw.zone 2026-04-14 15:17

@hopeless wow there’s so much to unpack here. Random + weights + other stuff is still ✨random ✨. Random + seed sure is deterministic but how often do you the user set the seed? Why is the answer to an objective query subject to a probabilistic state?

Open ##2004641
@barubary@infosec.exchange 2026-04-14 15:36

@hopeless @hastar @Possiblydrew they are choosing what they want to talk about Wow, what a coincidence. I was just talking to /dev/urandom and it chose to talk about "bizz".

Open ##2004643
@ariaflame@masto.ai 2026-04-14 15:42

@hopeless @hastar @Possiblydrew They aren't sentient, they are not 'choosing' anything. They're just picking at random from weighted options. They don't know anything. If the training set was good for what specifically you wanted to do you might be lucky. Otherwise you could be totally fucked and not know it until a lot later when something breaks/lets someone into your system.

Open ##2004647