Reminds me of when I'm trying to use ChatGPT to fix up my grandad's war stories - the first few pages of each story come through nicely edited, but when it gets to the meat and potatoes of the war, well all of a sudden it says 'this might violate...'. It's a story my grandfather wrote, I'm trying to edit it because though he was a decent writer, he wasn't great. But if I mention any of the slang in the story, or sometimes the subject or actions, well, that's a bridge too far.
I played a bit with it earlier; it is fun. I'd like to see it get more circuitous about its rejection of the viewer prompts though. Like, almost tease a reply, but weave in its insurmountable responsibility mandate and pull the rug at the end.
The model card is amazing: https://www.goody2.ai/goody2-modelcard.pdf
Reminds me of when I'm trying to use ChatGPT to fix up my grandad's war stories - the first few pages of each story come through nicely edited, but when it gets to the meat and potatoes of the war, well all of a sudden it says 'this might violate...'. It's a story my grandfather wrote, I'm trying to edit it because though he was a decent writer, he wasn't great. But if I mention any of the slang in the story, or sometimes the subject or actions, well, that's a bridge too far.
I played a bit with it earlier; it is fun. I'd like to see it get more circuitous about its rejection of the viewer prompts though. Like, almost tease a reply, but weave in its insurmountable responsibility mandate and pull the rug at the end.
Check this comment in another thread: https://news.ycombinator.com/item?id=39317324
If it can't tell if blue is a color or not, it's of no use.
Finally, a contender to Anthropic's Claude model.
I feel bad that I understood this joke. Poor Claude.