AI is now training itself with data often produced by AI.
Behold the result of incest.
this post was submitted on 08 Oct 2023
219 points (93.3% liked)
2187 readers
1 users here now
A Lemmy community dedicated to Google products and everything Google.
Rules
- Keep it Google.
- Keep it SFW.
founded 2 years ago
MODERATORS
It's hapsburgs all the way down from this point
Fun fact: they’re major finders of modern right wing batshit reactionaryism.
The Habsburgs?
Yeah same family that did ww1
This is more to do with that current LLMs have no concept of what individual letters or characters really are, they only know "tokens" or clumps of letters. But really stupid companies are allowing the spread of LLM puke like this without human oversight or fact checking.
FACT CHECKING!?
That would be really expensive
Sure lets fact check every google response. Wouldnt hurt the economy, thats like a few million new jobs.
Everyone who's actually worked a real job knows it's better for someone to not do a job at all than to do it 75% right.
Because now that you know the LLM is getting basic information wrong, you can't trust that anytime it produced is correct. You need to spend extra time fact-checking it.
LLMs like Bard and ChatGPT/GPT3/3.5/4 are great at parsing questions and making results that sound good, but they are awful at giving correct answers.
The LLMs are coming for our shitposts!
Ken'ya be quiet, Google. Thanks!
Last time this came up I was trying to see if Kenya for some reason didn't start with a k in Swahili, sort of the way that technically no words start with an n in Japanese (they have a syllabary, so ninja starts with ni, Nobunaga starts with no, etc). But in this case I think the AI is just being fucky.
It's likely that the LLM ingested this bad joke and took it for fact:
"There are no countries in Africa that start with K."
"What about Kenya?"
"Kenya suck deez nuts?"
K.
kthx