Overview

AI is changing rapidly with the rise of self-supervised models (e.g., BERT, GPT-3, DALL-E) trained at immense scale. We see this as the beginning of a sweeping paradigm shift for AI, which we call the era of universal models. By universal model, we mean a general-purpose model that is trained on raw data, potentially across multiple modalities, which can be usefully adapted to a wide range of tasks. These models represent a dramatic increase in capability in terms of accuracy, generation quality, and extrapolation to new tasks, but they also pose clear risks such as use for widespread disinformation, potential exacerbation of historical inequities, and problematic centralization of power. Further, we lack a clear understanding of when these models work and when they fail, or even what they are capable of due to their emergent properties. Given that future AI systems will likely rely heavily on universal models, it is imperative that we, as a community, develop more rigorous foundations for universal models and guidance for their responsible development and deployment.

We are convening the First Workshop on Universal Models on August 23rd and 24th to bring together the broad community (academia and industry) to discuss universal models. We have specifically designed the program to reflect a diverse array of perspectives and disciplinary backgrounds (e.g. social science, economics, computer science, law, philosophy, information science) that will provide vital expertise on the many dimensions of universal models. Broadly, we will address the opportunities, challenges, limitations, and societal impact of universal models. Specific points of emphasis include:

1. What applications and communities might benefit the most from universal models and what are some of the unique application-specific obstacles?
2. How do we characterize and mitigate the disparate, and likely inequitable, effects of universal models?
3. How do multimodal methods and grounding impact conversations around meaning and semantics in universal models?
4. When universal models are used in applications that cause harm, how do we handle matters of responsibility, accountability, and recourse?
5. What should be the professional norms and ethical and legal considerations around the release and deployment of universal models?
6. How should various groups (e.g. academia, industry, government), given their complementary strengths, productively collaborate on developing universal models?
7. Given universal models must be adapted for specific tasks, how do we evaluate them in ways that capture the needs of diverse stakeholders?
8. Universal models generally coincide with the centralization of power: how do we reason about this centralization, and its potential harms, and build ecosystems that better distribute the benefits of universal models?
9. Data plays a central role in universal models: how do we think about data sourcing, selection, documentation, and how do we build principles to guide how data shapes universal models?
10. The scale of universal models complicates principled scientific study: how do we build universal models in a sound manner given the potential inability to run comprehensive experiments, and how do we reaffirm our commitments to open and reproducible science in spite of this scale?

Schedule

Coming soon!