Exploring Alternatives to RL Training for Language Models

Watch on Youtube

In this talk Ethan presents on how AI systems like ChatGPT can be used to help uncover potential risks in other AI systems, such as tendencies towards power-seeking, self-preservation, and sycophancy. Ethan is a research scientist and team lead at Anthropic working on large language models, and his work aims to reduce the risk of catastrophic outcomes from advanced machine learning systems. He also spend some time at New York University (NYU) collaborating with Sam Bowman's group on AI safety research.