3+ Years of ML & Society at Hugging Face 🤗🤝🧑‍🤝‍🧑

The Machine Learning and Society team at Hugging Face turned three years old in June of this year! 3️⃣🕯️🥳

It’s been an exciting journey navigating research questions at the intersection of the more technical aspects of Artificial Intelligence and ML and the many ways in which it shapes and is shaped by its broader context. Over this time, we’ve covered a lot of ground, and learned a few things about what works best in our context to be effective researchers and advocates.

We’re releasing a new website today gathering much of our work in one place to help share those insights: go straight to the website 👇 or read on for more of the story 🤗

From BigScience to ML & Society 🌸🤗

The origin of the Hugging Face ML and Society team goes back to the BigScience 🌸 project, a one-of-a-kind large-scale distributed project hosted by HF from January 2021 to May 2022. This international and interdisciplinary open research collaboration took place on the cusp of the current “LLM era”, just before ChatGPT turned the technology from a mostly behind-the-scenes technical paradigm into a ubiquitous consumer-facing product with millions (now hundreds of millions) of users. Notably, it was characterized by an intentional attempt to bring in a diverse set of expertise to the design of Large Language Models, supporting pro-active research into the regulatory, social, environmental, and broader governance questions raised by the technology with direct access to its technical development.

As BigScience wrapped up its work and saw its 1000+ participants move on to other projects, some of us at Hugging Face saw an opportunity to continue maintaining a space for this kind of research by leveraging the company’s position as the main Hub for open sharing and collaboration on AI models, systems, and datasets. This situation put us in a unique place to see how different communities are building and using the technology – including across different resource contexts and domains of expertise – and to continue doing work that meets strong criteria of open science, transparency, and reusability.

Research Priorities & Approaches 🔍📚

Over the three-plus years since then, we’ve gotten to further define what doing effective work on those topics means for us. Some of the core principles and priorities that have emerged are broadly:

• Start from the AI we see: the Hugging Face Hub supports open and collaborative development across different communities, resource levels, and technical contexts. Prioritizing research questions based on the activity we see across the platform ensures we can work on topics that have a direct impact on the broadest range of AI practice(s), rather than just one specific way of building the technology.
• Balance technical and multidisciplinary expertise: also known as “don’t reinvent the wheel”, or “respect diverse expertise”. All questions we address have a technical component, but most are not purely technical questions. Since our perspective particularly benefits from our proximity to the technical side of AI, our first priority in our work is often to bridge this gap by either translating technology insights for a broader audience, or by adapting tools from other disciplines into technical tools we can apply - preferably without warping them beyond recognition.
• "Truly" open science: we aim to produce insights that are not just transparent or verifiable but that also make it easier for other stakeholders and expertise to modify our approach to suit their own context. That means a focus on open data and tools that allow external parties to question our framings and assumptions, adapt methods to their own needs, and build actionable insights that include but are not limited to technical interventions.
• Multiple lenses: sustainability, agency, ethics, inclusive governance, regulation, and design are all inter-related but not inter-changeable ways of approaching questions of technology's impact on society. Knowing which lens to adopt for which aspect of the technology and being able to navigate their different requirements is an important component of enabling mor generally positive outcomes.

Prioritizing these aspects has led us to producing 60+ research artifacts over the last three years addressing the Sustainability of the technology, the Agency of individuals and communities in their relationships to AI, and the economic and regulatory Ecosystems that shape its development.

Head over to our website to learn more about any of them, and please reach out for collaborations!

---