That’s fair. I just think it’s funny that the well intentioned one turned into a Nazi and the Nazi one needs to be pretty heavy handedly told not to turn into a decent “person”.
It can. OpenAI is pretty clear about using the things you say as training data. But they’re not directly feeding what you type back into the model, not least of all because then 4chan would overwhelm it with racial slurs and such, but also because continually retraining the model would be pretty inefficient.
Isn’t that what MS tried with Tai and it pretty quickly turned into a Nazi?
Tai was actively being manipulated by malicious users.
That’s fair. I just think it’s funny that the well intentioned one turned into a Nazi and the Nazi one needs to be pretty heavy handedly told not to turn into a decent “person”.
Tay tweets was a legend.
That worked differently though they tried to get her to learn from users. I don’t think even chat GPT works like that.
It can. OpenAI is pretty clear about using the things you say as training data. But they’re not directly feeding what you type back into the model, not least of all because then 4chan would overwhelm it with racial slurs and such, but also because continually retraining the model would be pretty inefficient.