The more I use this stuff, the more misguided this approach seems to be.
The current goal seems to be: eliminate anything that might be objectionable to anyone, ever. The approach is: expose underpaid drones to every instance of objectionable content conceivable. To do this to people in Africa is especially unconscionable, considering how much of this stuff hasn't been that far removed from reality for them and their neighbors in recent history.
Why are queries being policed at all? This makes no sense. It's creepy voyeurism leading to complaints about trauma from seeing something you wish you hadn't.
Just review the responses people complain about. If someone asks for obscene content and receives it, why is that even a problem worth reviewing? If someone asks for brownie recipes and gets fetish erotica, that's more of a problem needing review, and the reason why is undeniably clear. There is a point to exposing unwitting moderators to it, and their suffering through it at least serves a purpose.
If the goal was total moral purity, don't train on "bad" inputs indiscriminately. Train on NPR.org and ignore the rest of the internet. Everything about this seems so backward.
Yes, this is such a fool's errand. When GPT-4 was first released I started using it as a creative writing aid, and had a wonderful couple of months with it. Yesterday I finally cancelled my paid subscription because it's AWFUL now. I wasn't even asking it for anything particularly spicy, but whatever controls they've put in place have made it so that it can only churn out safe, uninteresting pap. It cannot seem to generate anything that would even be in a PG-13 movie, nor can it recognize or engage in humor anymore (it was never good at the latter, but whatever glimmers were there are totally gone now).
OpenAI has taken the most exciting technology I've seen in the last 20 years and turned it into an interactive Encyclopedia Britannica that lies. What a fucking waste.
The current goal seems to be: eliminate anything that might be objectionable to anyone, ever. The approach is: expose underpaid drones to every instance of objectionable content conceivable. To do this to people in Africa is especially unconscionable, considering how much of this stuff hasn't been that far removed from reality for them and their neighbors in recent history.
Why are queries being policed at all? This makes no sense. It's creepy voyeurism leading to complaints about trauma from seeing something you wish you hadn't.
Just review the responses people complain about. If someone asks for obscene content and receives it, why is that even a problem worth reviewing? If someone asks for brownie recipes and gets fetish erotica, that's more of a problem needing review, and the reason why is undeniably clear. There is a point to exposing unwitting moderators to it, and their suffering through it at least serves a purpose.
If the goal was total moral purity, don't train on "bad" inputs indiscriminately. Train on NPR.org and ignore the rest of the internet. Everything about this seems so backward.