KakaoTalk turns into a platform for using artificial intelligence (AI). There is no need to enter a ..

KakaoTalk turns into a platform for using artificial intelligence (AI). There is no need to enter a separate chat room to use AI on KakaoTalk. While talking to friends and family, you can call out AI and ask for what you need at any time.

AI, which has existed as a separate service, is put on the entire KakaoTalk like a skin (layer). Kakao’s representative character ‘Ryan’ gets a voice and transforms into an AI secretary who talks directly to users.

On the 13th, Kakao CEO Jung Shin-ah unveiled the AI agent strategy at the ‘if (kakao) 26’ held at the Kakao AI campus in Yongin, Gyeonggi Province.

The key is “Kanana Space,” an AI assistant that can be used anywhere in KakaoTalk. In order to use AI on existing Kakao Talk, it had to enter separate services such as ‘ChatGPT for Kakao’.

On the other hand, Canana Space is designed to call out AI from any screen of KakaoTalk. In order to use AI, there is no need to stop existing work and move to another screen.

When a user sweeps the Kakao Talk screen with two fingers, Ryan appears. Ryan also works in the form of a float above the chat room you are currently using. You can have a conversation with a friend and at the same time send a message to AI or ask them to check their schedule.

Ryan, in particular, has a voice. Canana Space is basically designed around voice conversation and can be converted to text if necessary. Unlike conventional chatbots, where users entered questions into AI and received text answers, they use AI as if they were talking to people.

KakaoTalk that you read with your eyes and listen to with your ears

Supporting this is the multimodal AI technology developed by Kakao. On the same day, Kakao unveiled the Omni language model “Kana-o” that understands text, image, and voice together and answers with voice. Both the encoders and decoders that make up this model have been developed on their own and will be released as open source in the future.

Kakao also introduced the “Kanana Live” technology, in which AI performs listening and speaking at the same time without waiting until the other person’s words are finished. It is a function that responds or interprets in real time during a conversation by users. The goal is to make the interaction with AI closer to human conversations by expressing not only the content of the conversation but also the speech and emotions that fit the situation.

Another characteristic of Canana Space is personalization. Kakao plans to identify the context of the user’s conversation and past experiences and provide different results for each person even for the same request.

For example, if a user forgets what he was asked to do last week while talking to his wife, he can ask Kanana. Despite the request to “order the bottled water I ordered last time again,” it is a method of searching for products based on past orders and applying coupons to prepare them right before payment. The final approval and decision are made by the user.

However, the function of completing orders, payments, and remittances is planned to be advanced in the future. Canana Space will start a closed beta test (CBT) for some of its pre-applicants from December.

Best results for all requests with optimal token consumption

Kakao decided that technology that efficiently handles requests from large-scale users is essential to use AI throughout Kakao Talk. This is because it is impossible to leave all the work to a large AI model that requires enormous computation at a time when 50 million KakaoTalk users make different requests.

To this end, Kakao has built an AI model system consisting of four stages, from on-device AI to small language models (SLM), flagship models, and frontier models. Simple requests are handled by AI mounted on smartphones, and tasks that require complex reasoning are handled by large models of servers.

Which AI model will be in charge of the work is decided by the “AI Conductor” rather than the user. AI conductor is an orchestration technology that analyzes users’ requests, selects necessary models and tools, and adjusts the order of work.

For example, if a user asks, “I’m going to take a taxi to the meeting tomorrow, recommend a nearby restaurant,” the request is analyzed first on the smartphone, location search is left to a small model, and complex tasks such as predicting travel routes are left to a large model. After synthesizing the results of various models and tools, the user is finally provided with the necessary information.

Noh Byung-seok, Kakao’s performance leader, explained that the principle is not to use the largest model for all requests, but to provide the necessary intelligence with as much computation as necessary. Through this, Kakao plans to secure not only AI’s performance, but also response speed, operating cost, and large-scale service stability at the same time.

Kakao’s AI conductor is evolving in the direction of coordinating the execution of various services beyond simply selecting a model. Kakao described this as an “extension from model routing to action orchestration.” The goal is not only to recommend restaurants, but to finish actual work by making reservations and sharing them with companions after receiving confirmation from users.

Kakao understands the context for me

To this end, it is also developing a “Project Mosaic” that connects users’ memories and experiences. It is a structure that finds the context necessary for the user’s request, delivers only the information allowed for each AI agent, and allows several agents to cooperate to perform tasks.

Kakao Chief Technology Officer (CTO) Song Jae-ha explained the technical foundation for AI to act as an agent that changes users’ daily lives beyond simple conversations. Kakao is preparing a ‘Project Mosaic’ that connects scattered users’ memories and experiences in a meaningful context. The key is to safely deliver only the necessary information for each agent within the permitted range according to the user’s request and desired results, and to ensure that agents in different roles understand and collaborate in a common context.

Kakao also suggested a platform direction to stably support the personalized AI experience. It plans to continuously upgrade the safety system so that the agent’s action plan and execution are completed only within the scope agreed by the user by establishing systematic data governance and multi-faceted safety verification devices. In addition, it plans to strengthen the foundation of sustainable infrastructure so that AI services can be stably provided to large-scale users.

Kakao plans to build a platform that allows AI agents made by other companies and individuals to be used in KakaoTalk. Kanvas, which was released on the same day, is a service that creates and executes AI agents when users explain necessary functions in natural language. Canvas is a platform that can operate agents on PCs. You can also share AI agents with others. In the first half of next year, not only agents made with Canvas but also agents developed with external frameworks will be connected to KakaoTalk.