More than most AI researchers, Sara Hooker is aware of the social implications of her work. Despite her warm disposition, she is often “grumpy” about the field.
“It stops me in my tracks,” she says. “Our work is not just used in conference centers anymore. It’s used all over the world, and so this means that we have to not only have the role of researchers, but also to have an understanding of that impact.”
[time-brightcove not-tgx=”true”]Her lab, Cohere for AI, a nonprofit that serves as the research arm of the enterprise AI startup Cohere, works on issues like increasing the efficiency of a given model and optimizing the quality of data used to train AI. This hybrid structure allows it to benefit from the computing power available to a private company while freely collaborating with academic, industry, and civil society institutions.
Since its launch in June 2022, it has published dozens of papers, and cultivated an expansive independent research community. One of its most notable projects is Aya, a set of language models and a dataset created in collaboration with over 3000 researchers across the world to support a diversity of languages. The project also aims to address the data scarcity issue in non-western languages. While the model is still bested by proprietary models such as GPT-4 and Cohere’s own Command R, Hooker says its value comes from allowing others to build atop it.
Having grown up in Mozambique and other countries across Africa, Hooker developed a “necessary humility about the world”, she says. She is conscious of the wealth of talent that exists outside of rich countries, and the relatively few opportunities for such people to enter the field. It’s part of why her lab has bet big on hiring global talent.
*Disclosure: Investors in Cohere include Salesforce, where TIME co-chair and owner Marc Benioff is CEO.
