Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
AI & Enterprise
AI chatbot services that virtually recreate conversations with real idols are drawing attention, prompting interest in the underlying technologies. The key is advancing persona techniques long used to improve large language model (LLM) response quality by combining them with information about real people and long-term memory technology into a service.
On Nov. 2, the AI industry says Mindlogic has drawn attention for its fan communication platform BloomingTalk, an AI chat service based on real idols. Users can chat and make voice and video calls with AI avatars of 24 members of A2O Soul and Light, idol groups introduced by Lee Soo-man (이수만), founder of SM Entertainment. Mindlogic developed the app, while Estsoft produced the AI video avatars. The company says about 1 million conversations took place in the first five days after the two groups joined, and payments totaled tens of millions of won.
The first step in creating a persona for a real person is collecting and analysing that person’s characteristics. Personality, tone of speech, preferences and interests need to be organised into information an AI can use. Interviews, profiles, past messages, video and audio are used for this. As the number of targets grows, time and costs for gathering and analysing information also rise.
To streamline the work, Mindlogic has the AI talk directly with artists to collect and analyse personality traits, tone of speech and recent interests. The company calls it an AI interviewer. It uses the information it secures together with existing materials to reduce the manual workload.
The next hurdle is how faithfully the collected information is reflected in actual AI answers. The AI must generate responses that match an artist’s preferences and background knowledge, not simply mimic a certain way of speaking. The company uses information about the artist as the standard for AI answers. Information that changes frequently, such as performance schedules, is entered separately by management to improve accuracy.
Long-term memory technology is also used to maintain conversation context for each user. If the persona keeps an artist’s distinctive personality and conversation style, long-term memory is used to continue personalised conversations based on information accumulated for each fan.
For example, if a fan says they are preparing for a job, the next conversation might ask about the interview result. If the AI forgets past conversations or mentions incorrect information, immersion similar to communicating with the real artist can fall.
The key is to cut computing costs while selecting what to remember. Feeding the AI the entire past conversation every time increases the amount of information to process and raises costs. Mindlogic chose an approach that selects only important information from conversations and structures it.
It is a different approach from retrieval-augmented generation (RAG), which supplements answers by searching external materials. Mindlogic organises information extracted from conversations, such as a user’s name, preferences and job, into a profile and reflects it in the system prompt. It aims to keep context for each individual without processing the full chat history each time.
When selecting information, it prioritises reliability. It distinguishes between information users enter themselves and content the AI infers from conversation. If the two conflict, it prioritises the former. The step is intended to reduce the problem of incorrect AI inferences being remembered as fact.
A Mindlogic official said, "Like a person’s memory, AI also chooses what is important and what is not," and added, "By distinguishing and remembering information needed to understand the other person, such as a user’s name, hobbies, job and preferences, it reflects individual context as conversations accumulate."
This content was produced with the assistance of AI and reviewed by our editorial team. You can read the original version in Korean here.
