←back to thread

236 points inesranzo | 1 comments | | HN request time: 0s | source
Show context
giancarlostoro ◴[] No.46231943[source]
This will not end well for Disney, there were certain historical characters removed from Sora 2 because people kept making racist videos that are hard to censor, and it became increasingly unhinged. This feels like another circular investment where Disney is hoping to make money back I'm sure. On the other hand, assuming they do the freemium stuff, I look forward to making a few videos of my daughters favorite Disney princesses "talking" to her.
replies(21): >>46231963 #>>46232027 #>>46232049 #>>46232134 #>>46232217 #>>46232316 #>>46232450 #>>46232484 #>>46232594 #>>46232634 #>>46232732 #>>46232770 #>>46232998 #>>46233138 #>>46233293 #>>46233477 #>>46234040 #>>46234843 #>>46235615 #>>46235942 #>>46236727 #
impulser_ ◴[] No.46232217[source]
I guess it will depend on how good their security is, because I'm assuming Disney is entering this deal with hard guidelines on what will be allowed.
replies(2): >>46232401 #>>46232756 #
bilbo0s ◴[] No.46232401{3}[source]
This.

Don't believe for a second that Sora will allow you to make racist content with Disney characters.

That said, there are a lot of other models out there that care about neither licensing nor alignment. So those will allow you to make racist content. Then you can do whatever you like with that generated content.

A lot of IP owners will learn that there is more than one way to skin a cat. It's easier than people think to turn children's characters, like say, Hermoine, into a raging racist. And there's very little technically speaking that they can do to stop it.

But yes, on OpenAI specific properties, they can definitely stop it dead in its tracks. They can even get better at stopping it over time. In fact, the more users try to generate it, the better the system will get at stopping it.

replies(11): >>46232460 #>>46232464 #>>46232466 #>>46232469 #>>46232486 #>>46232641 #>>46232648 #>>46233618 #>>46233677 #>>46233712 #>>46233869 #
1. corobo ◴[] No.46233618{4}[source]
How are they going to stop it?

A certain combination of nonstandard characters will make an AI character drop an n-word no problem

I guess they could chuck the output through whisper or something to see if it transcribes back to anything dodgy?

LLM security feels very ball of sand held together with duct tape haha