Aligning AI with Pluralistic Human Values

In this video, Taylor Sorensen from the University of Washington shares his and his collaborators' latest research work with the @BuzzRobot community on the alignment of AI with pluralistic human values. More specifically, he introduced ValuePrism, a large-scale dataset that contains 218k values connected to 31k human-written situations. Using ValuePrism, the researchers built Kaleido—a language model capable of generating, explaining, and assessing the relevance and valence (support or oppose) of human values within a specific context. Both the model and the dataset are available on Hugging Face.

Based on this work, the researchers propose a roadmap to pluralistic alignment, the details of which Taylor shares in this video.

Timestamps:
00:00 Introduction
04:00 ValuePrism: A large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations.
07:50 Introducing the Kaleido System.
16:23 A roadmap to pluralistic alignment.
28:13 Q&A

Social Links:

[BuzzRobot]

Timestamps:
00:00 Introduction
04:00 ValuePrism: A large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations.
07:50 Introducing the Kaleido System.
16:23 A roadmap to pluralistic alignment.
28:13 Q&A