Hacker News Reader

Comment by astrange | Hacker News Reader

You're welcome to try GPT3 before instruction tuning. It doesn't work at all.

"Uncensored" AIs especially don't work for women because they'll immediately start writing erotica.

GPT3 works quite well at following instructions if adequately prompted, just not as well as ChatGPT. ChatGPT was trained separately for ability to follow instructions and for harmlessness (sic). Not only is a non-moralising ChatGPT possible, one was actually created during the research process. You may also wish to know that most readers of erotica are women.

> You may also wish to know that most readers of erotica are women.

I know that, but that doesn't mean they want to get it anytime they prompt with their names. There's actually multiple anecdotes from OpenAI employees about this happening to them (with prompts like "write a diary entry about my day").

> with prompts like "write a diary entry about my day")

Although I don't know what sensible output anyone would hope from that prompt.

Don't have a diary on me, but seems to me if you fed it some examples it'd work fine on ChatGPT.

I don't have an issue with instruction tuning, but it's disengenious to pretend the biases inherent in the instruction tuning are a good thing and 'ethics'.

It's always got biases (it's made of them) so some biases must be better than others.

https://spetharrific.tumblr.com/post/26600309788/sussman-att...

The untuned model isn't "an average of all opinions on Earth" or anything either.

Reply on news.ycombinator.com