Laura Weidinger: Ethical Risks, Harms, and Alignment of Large Language Models
The Gradient: Perspectives on AI
00:00
GPT 4chan - What Are the Main Harms of Training GPTJ on Hate Speech?
Yana Culture trained GPTJ, fine-tuned it on the very toxic and hateful data set of forechan. The main harms that we see here are, of course, about offensive speech. And so it is, should be expected, and indeed it does output content that creates these kinds of language harms too. But making that kind of model available for someone to maybe just expand their trolling onto a new asset isn't really a research goal.
Transcript
Play full episode
