"Why Timnit Gebru Wants Tech Giants to Start Small in Artificial Intelligence"

Timnit Gebru used to believe that it was possible to curb the worst impulses of tech giants from the inside out.

She experienced this firsthand at Microsoft in 2018 when she co-authored a groundbreaking paper that revealed how commercial facial recognition systems, including Microsoft's, were terrible at detecting the faces of Black women. While the paper did prompt Microsoft to adjust the accuracy of its system, Gebru believed that the company could have completely halted the development of facial recognition technology.

She attempted to bring this level of accountability to Google when she joined the search giant in the fall of 2018 as the co-lead of its Ethical AI team. It was there that she co-authored another paper outlining the dangers of large language models - the potential for bias, the environmental costs of training them, and the labor exploitation required to build them. Google tried to force Gebru to remove her name from the paper, and this confrontation ultimately led to her sudden public dismissal from the company in December 2020.

Witnessing the limitations of major tech companies in effecting meaningful change from within, Gebru founded her own independent research group - the Distributed Artificial Intelligence Research Institute (DAIR). Sponsored by the nonprofit organization Code for Science & Society, the institute has since attracted other former Google employees, including Alex Hanna, the current Director of Research at DAIR. Since its establishment in 2021, Gebru has been dedicated to creating a safe haven for independent AI researchers to honestly and openly scrutinize their industry without fear of crossing any trillion-dollar company's bottom line. Over the past year, as companies like OpenAI, Meta, Google, and Microsoft have increasingly demanded more from large language models, Gebru has become a unique and trusted critic, calling out the industry's overpromises and the harm caused by the recent AI hype.

One of Gebru's biggest concerns is that these companies are attempting to build very large language models without considering a clear purpose. She says, "When you try to build a model that's supposed to be applicable to all sorts of scenarios, you've already lost safety." "You can't even ask the question: What is this for? What shouldn't it be used for? What are the inputs? What are the outputs? You haven't properly scoped it."

Gebru believes that any language model claiming to be applicable to all domains will almost certainly perform worse than a smaller model specifically built for a particular task or community. One example she gives is machine translation, where smaller models trained specifically for a certain language often outperform huge models claiming to translate hundreds of languages but ultimately perform poorly in non-mainstream languages. Gebru says, "When you understand your users and plan your dataset accordingly, you can build small datasets." She recently co-authored a conference paper that exposed flaws in Meta's machine translation claims for certain Eritrean and Ethiopian languages.

As Gebru puts it, "Small-scale building" benefits not only the end users of AI technology but also the communities around the world and the companies developing these tools for those communities. This is the harm behind the hype: Gebru believes that when a few Silicon Valley companies overpromise beyond their technological capabilities, it makes it harder for smaller companies to compete.

"I want to show people that there is another way," Gebru says.