Considerations for - Key Takeaways - Governing Open Foundation Models Introduction

Background Central to understanding foundation models is the a web interface, like Inflection’s Pi); cloud-based topic of release: To what extent and through what access (available via an API, like OpenAI’s GPT-4); mechanisms are foundation models made available to cloud-based fine-tuning access (both the model and entities beyond the foundation model developer? The the ability to fine-tune it are. [...] We use the notion of open foundation models: The release of foundation models is a gradient: models released with widely available weights, which Models can be fully closed (not available to anyone corresponds to the three rightmost categories in outside the developer organization, like Google the figure (modified with permission from Solaiman DeepMind’s Flamingo); hosted access (available via (20. [...] This Guterres proposed to “make transparency, fairness is doubly true for AI, because these new tools can be and accountability the core of AI governance … [and] developed for many different types of activities, with consider the adoption of a declaration on data rights the potential to spread rapidly in every sector of the that enshrines transparency.” The 2023 Foundation economy and in every asp. [...] While foundation models between open and closed foundation models is can be released openly without the release of the predicated on release, policies governing the usage underlying data used to build the model, some of foundation models are likely to have differential developers choose to release both the model weights impacts. [...] Of the 10 major foundation downstream usage could chill the open foundation developers assessed by the 2023 Foundation Model model ecosystem by exposing open foundation model Transparency Index, the two developers that released developers to severe liability risk.