Sandhini Agarwal: We did find that it generated certain undesirable outputs, but they have been all things which GPT-3.5 also generates. So with regard to risk, for a study preview—for the reason that that’s what it https://youtu.be/aGubCnA5Qyk