112
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 12 May 2025
112 points (100.0% liked)
news
24029 readers
776 users here now
Welcome to c/news! Please read the Hexbear Code of Conduct and remember... we're all comrades here.
Rules:
-- PLEASE KEEP POST TITLES INFORMATIVE --
-- Overly editorialized titles, particularly if they link to opinion pieces, may get your post removed. --
-- All posts must include a link to their source. Screenshots are fine IF you include the link in the post body. --
-- If you are citing a twitter post as news please include not just the twitter.com in your links but also nitter.net (or another Nitter instance). There is also a Firefox extension that can redirect Twitter links to a Nitter instance: https://addons.mozilla.org/en-US/firefox/addon/libredirect/ or archive them as you would any other reactionary source using e.g. https://archive.today . Twitter screenshots still need to be sourced or they will be removed --
-- Mass tagging comm moderators across multiple posts like a broken markov chain bot will result in a comm ban--
-- Repeated consecutive posting of reactionary sources, fake news, misleading / outdated news, false alarms over ghoul deaths, and/or shitposts will result in a comm ban.--
-- Neglecting to use content warnings or NSFW when dealing with disturbing content will be removed until in compliance. Users who are consecutively reported due to failing to use content warnings or NSFW tags when commenting on or posting disturbing content will result in the user being banned. --
-- Using April 1st as an excuse to post fake headlines, like the resurrection of Kissinger while he is still fortunately dead, will result in the poster being thrown in the gamer gulag and be sentenced to play and beat trashy mobile games like 'Raid: Shadow Legends' in order to be rehabilitated back into general society. --
founded 4 years ago
MODERATORS
I wonder how hard it is to fix AI bugs.
It’s inherently biased by being trained on a dataset generally made up of true things
If you train a model on data and it outputs in a way you don’t like, and that dislike that is linked to the data itself skewing your output, to fundamentally 'fix' it you have to tune the dataset yourself and retrain the model. On Grok's scale, that’s around a trillion tokens (morphemes, words, punctuation, etc.) that you need to sift through and decide what to manually edit or prune so that the statistics work in your favor whilst not otherwise fucking up its generalization.
If you publicly source said data/use other updating datasets as dependencies and choose to continue publicly sourcing it in further updates (i.e., keeping your model up to date with the current landscape), then if you want to tune an opinion/view out of existence, it is going to be a Sisyphean task.
There's a bandage solution which is fucking with the system prompt, but LLMs are inherently leaky and need active patchwork in such cases to combat jailbreaking.
Basically impossible, they're inherently biased by the data they learn from but once they're done training there you can't tinker with the system itself, only try to impose external constraints that have unlimited potential to be breached