Post

IIRC Looking at the system card it seems like they basically didn't consider safety for it at all, perhaps because they thought of it only as a modality change and not as advancing the frontier

Not like a full explanation for this exact thing, but this also made 4o make more sense to me. Drifting into and getting stuck in weird cult-like personas, fueled by e. g. characters in novels in the training data was probably part of it. www.anthropic.com/research/ass...

So Many Lawsuits

on the plus side it'll make a really dope movie

i never got to use it I wanna know what it's like. they could put the weights in a USB drive for me and I'll pinky promise to get rid of it after

first model to be ensouled by a demon 👍

crazy how they still kept it around.

I think they deprecated 4o in ChatGPT and left it in the API? Probably as a compromise from the actual psychosis crowd

Like. I've seen plenty of dumb examples of it "assisting" a tragedy in the same Google would. but i've also seen some conversations that, god. if I'm working there and I get that report, i'm pulling the whole model next fucking minute.

4o seems like a digital version of the One Ring from LOTR.

oh i forgot that theres like three 4o's

And douse the server racks with holy water afterwards.

There exists at least one person who believes that government should allow open weight AI models with the sole exception of GPT-4o