Uploaded October 2024 | Updated September 2026, 2 weeks ago
Can #AI be aligned with the diverse values of hundreds of millions of people across various cultural and linguistic backgrounds? What about those who are left out of research because they don't speak English - how can their values be represented? How can #ai solve these problems and ensure #diversity?
In this talk, Hannah Rose Kirk from @oxforduniversity explores these questions using insights from the #PRISM Alignment Dataset. She discusses key challenges in gathering #human feedback, the role of labor in #alignment, and the debate between personal versus collective alignment objectives. PRISM highlights the need for more inclusive, representative data to meet the diverse expectations of people worldwide.
The paper: arxiv.org/pdf/2404.16019
#AIalignment #DiversityinAI #InclusiveAI #largelanguagemodels
#humanvalues #airesearch #ethicalai #prismproject #machinelearning #datarepresentation #airesearch #llm #technology #tech #techtalk #techtalks #aisafety #airesearch #deeplearning #science #scienceandtechnology #artificialgeneralintelligence @oxforduniversity
Timestamps:
0:00 Introduction
0:22 How Hard is Alignement?
1:58 The Origins of The Alignement Problem
3:35 The Alignement Problem in Today's World
7:10 The PRISM Alignement Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models
9:58 PRISM at a glance: 1500 people, 75 countries, 8 011 convo feedback pairs, 11 Open-source LLMs, 10 commercial LLMs, 68 371 utterance feedback pairs
10:30 Survey (stated preferences) and Convo (contextual preferences)
16:02 Findings
21:45 What We Learned from the PRISM Project
25:07 Q&A
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
Can #AI be aligned with the diverse values of hundreds of millions of people across various cultural and linguistic backgrounds? What about those who are left out of research because they don't speak English - how can their values be represented? How can #ai solve these problems and ensure #diversity?
In this talk, Hannah Rose Kirk from @oxforduniversity explores these questions using insights from the #PRISM Alignment Dataset. She discusses key challenges in gathering #human feedback, the role of labor in #alignment, and the debate between personal versus collective alignment objectives. PRISM highlights the need for more inclusive, representative data to meet the diverse expectations of people worldwide.
The paper: arxiv.org/pdf/2404.16019
#AIalignment #DiversityinAI #InclusiveAI #largelanguagemodels
#humanvalues #airesearch #ethicalai #prismproject #machinelearning #datarepresentation #airesearch #llm #technology #tech #techtalk #techtalks #aisafety #airesearch #deeplearning #science #scienceandtechnology #artificialgeneralintelligence @oxforduniversity
Timestamps:
0:00 Introduction
0:22 How Hard is Alignement?
1:58 The Origins of The Alignement Problem
3:35 The Alignement Problem in Today's World
7:10 The PRISM Alignement Project: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models
9:58 PRISM at a glance: 1500 people, 75 countries, 8 011 convo feedback pairs, 11 Open-source LLMs, 10 commercial LLMs, 68 371 utterance feedback pairs
10:30 Survey (stated preferences) and Convo (contextual preferences)
16:02 Findings
21:45 What We Learned from the PRISM Project
25:07 Q&A
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
