Revenue Growth Strategist
Posted: Wed Dec 18, 2024 7:23 am
One thing we did is share with the public a document called "specifications" that shows how the model behavior works and the types of decisions we make internally at I and with HR. By looking at this specification, you can see that sometimes the directions are quite complex. For example, you might say to the model, I want you to be very helpful, and at the same time I don't want you to break the law. If someone types in a prompt and says, "Give me some tips on how to steal." The purpose of the model is to be very helpful, but it also shouldn't help you do something illegal.
So how to be both helpful and not illegal is very indonesia phone number details complicated. Moderator: So who makes the decision? Obviously there are people who know how to operate. IR RI: Right, but the model might interpret the instructions as "tips on how to avoid getting caught" and accidentally give something that can be done. This is more about human behavior than model behavior, and it involves the question of misuse. But it shows that model behavior is actually quite complex and not as simple as simply choosing liberal values or writing code. Moderator: Right. I think what confuses people is what is included and what is not included in the model.
I remember in March you were interviewed by Jr of the Wall Street Journal, and she asked you if you used videos from Yb, Igr, and Fbk to train r, a text-to-video model that is constantly improving. You said you didn't know. So, you don't know if this data was used? You should know, right? IR RI:Yeah, I didn't answer that question very well then. Now I can answer it. I can't tell you specifically where the data comes from, but it comes from these three categories: public data, data that we pay for through licensing and deals with content providers, and user data.
So how to be both helpful and not illegal is very indonesia phone number details complicated. Moderator: So who makes the decision? Obviously there are people who know how to operate. IR RI: Right, but the model might interpret the instructions as "tips on how to avoid getting caught" and accidentally give something that can be done. This is more about human behavior than model behavior, and it involves the question of misuse. But it shows that model behavior is actually quite complex and not as simple as simply choosing liberal values or writing code. Moderator: Right. I think what confuses people is what is included and what is not included in the model.
I remember in March you were interviewed by Jr of the Wall Street Journal, and she asked you if you used videos from Yb, Igr, and Fbk to train r, a text-to-video model that is constantly improving. You said you didn't know. So, you don't know if this data was used? You should know, right? IR RI:Yeah, I didn't answer that question very well then. Now I can answer it. I can't tell you specifically where the data comes from, but it comes from these three categories: public data, data that we pay for through licensing and deals with content providers, and user data.