Making language models accessible to all

Video by Sam Han | Text by Ignacio Lobos | UW-IT Communications & Engagement
Tim Dettmers, center. Artidoro Pagnoni, right.
Tim Dettmers, center. Artidoro Pagnoni, right.

Something that easily gets lost when people talk about working with large language models such as ChatGPT is cost and the unbelievable amount of power needed to run them.

That’s where Tim Dettmers, a Ph.D. student at the University of Washington, comes in. His work is focused on machine learning and large language models, with the ultimate goal of making them more widely accessible to a larger field of researchers.

There are plenty of bright people all over the U.S. and the world who have a lot to contribute to the emerging language model field, but often they lack access to computational resources to do their own research.

But Dettmers has found a way to allow researchers to work with these large models on a simple laptop — and perhaps not too distant into the future, allowing you and I to train our own language models on our smart phones.

Learn what he is doing on this video about his work, and the future of language model research.

“What our research is showing is that you don’t need the expensive servers that can be like $50,000 or more expensive” to work with the largest of the language models now available, Dettmers said. “You can use a consumer (computer) so people can set up at home and use these things at home.”

But their work also extends beyond merely using open source language models. “You can take it and make it your own. You can personalize it, you can fine tune it on your data. This is very powerful,” he said.

“I want to make our work as widely accessible as possible for the people with the least resources … and accelerate research so we can figure out more things about language models and how to use them well,” he said.

Partnering with UW-IT to make it happen

Dettmers and his colleagues have been doing their research on Hyak, the University’s own supercomputer, which is managed by UW-IT. Hyak, Dettmers said, has accelerated their work and made it possible to use large language models — and to personalize them, which is an even more difficult undertaking.

“Research computing at the University of Washington is just a game changer,” Dettmers said. “We need so much computational resources in our research. Hyak (makes it possible) to do the work that we need to do in order to stay at the cutting edge.”