Sasha Luccioni’s path to Hugging Face started with a deliberate rejection. After earning a Ph.D. in computer science and spending years in applied AI research, including a stint in finance, she quit. “I decided that no amount of money was worth working on AI for AI’s sake,” she says on the Machine Learning Experts podcast. The move led her to work with Yoshua Bengio on AI-for-good projects before landing at Hugging Face, where she now leads work on the ethical and societal impacts of machine learning.
Her current focus is the Big Science project, an initiative she describes as turning large language models from a “black box thing” into a “glass box.” For years, Luccioni watched massive models emerge without access to their code, data, or development processes. Big Science aims to change that, involving researchers from across disciplines to build a model transparently. She chairs the project’s carbon footprint working group, which goes beyond simply counting CO2 emissions during training. The group is investigating manufacturing costs and even the carbon impact of everyday digital activities like sending an email.
Data curation is another major concern. Luccioni points out the prevailing attitude has been to “get the most data we can, stuff it in a language model and it’s gonna be great.” But the internet contains “so much weird stuff” that you don’t necessarily want a model learning from. The project is emphasizing mindfulness in data selection and pushing for multilingualism so the resulting model isn’t 99% English.
The carbon angle is particularly relevant given the industry’s trajectory. As models like GPT-3 and PaLM have scaled up, so have their energy costs, and most companies have been reluctant to share those figures. Big Science’s decision to treat environmental impact as a core research question, not an afterthought, sets a precedent that commercial labs may eventually have to answer to.
