KT joins KAIST and others to create a hyperscale AI model

Representatives from KT, AI One Team, KAIST, Hanyang University and the Electronics and Telecommunications Research Institute (ETRI) pose for a photo after signing a joint research agreement on Wednesday. [KT]

KT joined forces with KAIST and two other institutions to develop and commercialize a large artificial intelligence (AI) language model by early 2022.

Being large, or hyperscale, means that the AI model has more parameters and a better capacity to efficiently process data. Unlike other AI models, it will have a more accurate text analysis algorithm and will provide speech-to-text and text-to-speech responses that are almost human-like.

KT announced Wednesday it signed a joint research agreement with KAIST, Hanyang University and Electronics and Telecommunications Research Institute (ETRI) to co-develop a hyperscale AI language model.

Although specific number of parameters of the language model weren’t disclosed due to security reasons, KT's spokesperson said it will be one of the top in the nation in terms of the number of parameters.

KT will be in charge of developing graphics processing unit (GPU) infrastructure. It will collect and analyze data to be put into a data set for machine learning and will oversee the overall research and development and commercialization process.

KAIST will optimize the machine learning algorithm of the language model. Hanyang University plans to develop a data filtering algorithm and help optimize the final model. ETRI will upgrade its Exobrain language for the joint team's language model.

Deep learning methods developed by the joint team will put an emphasis on filtering out harmful and toxic content, preventing the final language model from learning bias towards a certain class or group.

Machine learning developed will be completed by end of this year. The language model will be commercialized by the first half of next year, and be used on KT’s GiGA Genie smart speaker and its AI Contact Center.

KT and the three institutions that formed a partnership Wednesday are members of the "AI One Team” research alliance, which includes representatives from industry, academia and state-run institutions dedicated to developing AI.

Other industry members of the alliance will be able to use the language model in various fields, including finance, retail and logistics.

BY LEE TAE-HEE [lee.taehee2@joongang.co.kr]