Community Energy

Published by Alexy Khrabrov, February 8th, 2024

What makes an aliance, a consortium, a nonprofit foundation so fun that people want to volunteer to work there together, often for free?

Why is this the preferred mode of collaboration on the OSS space?

What kind of people are attracted to it, and who drives these communities and why?

In order to achieve progress with Open-Source AI and help it succeed as the way to understand and build AI, we need to understand the communities behind it.

Let's start with HuggingFace, the de facto community hub for AI. It has become a gathering place, where the models are uploaded and examined, discussed and joined in collaborations. It's a kind of a github for models. It also has workspaces where you can try things out.

HuggingFace convened a project called BigScience. It started during the pandemic and united people from around the world around the question of LLMs for science, as well as an OSS LLM, later called Bloom. Programmers, data scientists, linguists, physicists and philosophers converged in virtual environments to ideate the LLM worlds of the future. This predated ChatGPT. HuggingFace engaged more than 20 of its staff members to facilitate. The project spawned several working groups and was highly decentralized.

SO here are some key ingredients: a worthy goal, like advancing science. Global reach, academic engagement -- idealistic students and researchers steeped in international collaboration.

Careful interaction design -- no oppressive hierarchy, letting leaders emerge around ideas they are committed to.

There are risks as well -- once HuggungFace was done with Bloom, the community petered out. BigCode was next but it was not the same, the allure of unifying all science around LLMs replaced by a rather niche and utilitarian task.