How Interpretability Research Helps Build Better Models

*From Fully Connected 2023*

Join Stella Binderman, Executive Director of EleutherAI and Head of Interpretability Research, as she delves into how interpretability research can help build better large language models.

EleutherAI is a non-profit research institute known for its open-source work on datasets like the Pile, large language models such as GPT-J and GPT-NeoX, and the VQGAN-CLIP text-to-image synthesis technique.

In this insightful talk, Stella covers three case studies: debiasing large language models, improving rare fact learning, and defending APIs …

apis apps attacks case case studies clip datasets development director eleutherai gpt gpt-j head image institute interpretability interpretability research join language language models large language models profit research studies synthesis talk text text-to-image the pile work