The web can be a nice place full of lively people. By default it is not. At least anymore. I remember the early days of webrings, HTML only pages, and guestbooks. That’s no longer the case. It hasn’t been so since facebook.
Why is this relevant?
These LLMs were trained on the internet’s data. Which is full of a lot of vile shit. I can’t imagine these models having a measurable sense of decency as a result. Not without much tweaking from the big model makers.
It’s no wonder people become affected by this technology. It’s trained on humanities unfiltered thoughts. So dark that light can’t escape.
on the contrary, they're all heavily filtered and """aligned""". have a chat with any commercial model about some controversial topic, and you'll see the viewpoints that didn't exist until 5-10 years ago and are only held by less than 1% of worldwide population.
unfiltered data might exist in the petabytes of shit they train on, but goodthink datasets are heavily prioritized.
Good according to whom is the question. Bad is a little simpler to define given violence, etc. But good? Good luck with that. There’s definitely a compounding shift of western values anywhere else these models are deployed.
The web can be a nice place full of lively people. By default it is not. At least anymore. I remember the early days of webrings, HTML only pages, and guestbooks. That’s no longer the case. It hasn’t been so since facebook.
Why is this relevant?
These LLMs were trained on the internet’s data. Which is full of a lot of vile shit. I can’t imagine these models having a measurable sense of decency as a result. Not without much tweaking from the big model makers.
It’s no wonder people become affected by this technology. It’s trained on humanities unfiltered thoughts. So dark that light can’t escape.
We are on for a ride.
on the contrary, they're all heavily filtered and """aligned""". have a chat with any commercial model about some controversial topic, and you'll see the viewpoints that didn't exist until 5-10 years ago and are only held by less than 1% of worldwide population.
unfiltered data might exist in the petabytes of shit they train on, but goodthink datasets are heavily prioritized.
Good according to whom is the question. Bad is a little simpler to define given violence, etc. But good? Good luck with that. There’s definitely a compounding shift of western values anywhere else these models are deployed.
https://archive.ph/NE86Q