Kakao Unveils Kanana Space AI Assistant for KakaoTalk, Plans December Beta

The service, which will be able to understand users' verbal requests, was introduced at Kakao's annual developers' event at its AI campus in Yongin, Gyeonggi Province, south of Seoul.

Kakao CEO Chung Shin-a said the company intends to move from a graphical user interface to a language user interface that can understand the context and purpose of the speaker. "We will transition from a graphic user interface ... to a language user interface which can understand the context and purpose of the speaker, to create a user experience where requests are made seamlessly and tasks are performed swiftly," Chung said.

Kanana Space is built on a layered AI architecture, according to a demonstration at the event. The service will be able to recommend workout clothes based on weather and users' workout data, or plan schedules by understanding a user's message history on the KakaoTalk app.

Kakao also outlined its technological strategies for its foundation AI model, Kanana, which powers the new service. "Our principle is we provide the best outcomes for all requests but do not use our biggest model for all of the requests," said a company executive overseeing the development of the foundation model. Kakao applied Multi-Teacher On-Policy Distillation, a type of AI post-training technique, to maintain the quality of its smaller language models, the executive added.

Chung said the key metric for assessing generative AI quality is shifting from whether it can produce good answers to whether it can complete requested tasks. "We plan to create an AI that provides palpable change to people's daily lives, by adding intelligence that understands individual differences on top of a service that connects everyone," Chung said.