How to Make Small Language Models Work. Yejin Choi Presents at Data + AI Summit 2024

Speaker: Yejin Choi, Professor and MacArthur Fellow at the University of Washington, and Senior Research Director for Commonsense AI at AI2

[@user-wr4yl7tx3w]

anyone has the link to the paper on archive?

[@BeginnerAlchemist]

Impossible Distillation for Paraphrasing and Summarization: How to Make High-quality Lemonade out of Small, Low-quality Models

[@BeginnerAlchemist]

I have a question: why we try to research Small-LM just to avoid using GPUs? If we want to save the money for training, we can do the research for how to make GPU or model more effectively, not to avoid using higher techs.

[@DamaruM]

GPU= power consumption

[@tulikabose5120]

It's not just for GPUs...Small-LM has its own market for on-device or on-edge processing, where there are concerns of privacy and customers would not want their data to go to clouds, and secondly in many industrial use-cases where internet and cloud access isn't accessible due to the remote nature of the use-case, and model inference needs to be done on device...The demand for SLMs is increasing in such use cases...Many big tech companies are not just working on LLMs but also on SLMs under the hood as both of them have to co-exist to cater to different user requirements.

[@BeginnerAlchemist]

@@tulikabose5120 Thank you, I see. It is useful for small devices with limited calculation hardware and the privacy. That's true. So many LLM need a huge data to train and it should collect people's private info to become stronger. That's hated by most of people.