Kakao’s multimodal LLM Honeybee: Korea’s answer to Gemini, GPT-4
Dong-jin Hwang
Jan 19, 2024 (Gmt+09:00)
Kakao Corp., South Korea’s top mobile platform operator, said on Friday it has developed a multimodal large language model (MLLM) with the latest image-to-text technology actively pursued by Big Tech firms worldwide.
Honeybee, the Kakao MLLM, is a multimodal large language model that enables reasoning across text, images, video, audio and coding.
During a government-private sector artificial intelligence strategy meeting hosted by the Ministry of Science and ICT, Chung Sina, Kakao’s chief executive nominee, said the company recently completed the development of Honeybee – a project led by Kakao’s AI research affiliate Kakao Brain Corp.
She said Kakako has shared the source code that enables inference of Honeybee on GitHub, an open-source platform, to allow other developers to utilize it in their AI model research.
Chung, formerly CEO and managing partner of Kakao Ventures Corp., was tapped to lead Kakao Corp., Korea’s dominant mobile messaging app operator, in December. Her appointment will be approved at the tech giant’s annual general meeting in March.
KOREA’S ANSWER TO GEMINI, GPT-4
An LLM is a deep-learning algorithm that can mimic human intelligence using an extensive language dataset – a core AI technology at the center of generative AI like ChatGPT or Bard.
Kakao’s Honeybee works like Google DeepMind’s Gemini and OpenAI’s GPT4 – multimodal LLM models that perform massive multitask language understanding.
Built on an MLLM foundation, Honeybee understands images and text simultaneously.
For example, if a question like “How many times did the left player win?” is presented with a photo of two basketball players in action, Honeybee creates an answer that often outperforms that of a human, according to Kakao.
Prompts, or questions, must be entered in English on the Honeybee platform.
Kakao unveiled KoGPT, a Korean LLM, in 2021 and developed an upgrade, KoGPT 2.0, late last year, although the company hasn’t yet officially launched the latter.
KoGPT, short for Korean generative pre-trained transformer, is a super AI conversation app, based on OpenAI’s GPT-3. Kakao said KoGPT is the most powerful Korean language chatbot, trained primarily on Korean text.
CALL FOR GOVERNMENT SUPPORT
At Friday’s government-private sector AI meeting, industry leaders called for active government support in Korea’s AI technology innovation.
“If AI is available throughout society, it will affect people’s way of thinking and behavior. This is why we’re working laboriously for and investing heavily in AI models,” said Choi Soo-yeon, chief executive of another Korean tech giant Naver Corp.
Participants at the meeting included LG Electronics Inc.’s LG AI Research chief Bae Kyung-hoon, Kim Seung-hwan, co-CEO of Amorepacific Group, Doosan Robotics Inc. CEO William (Junghoon) Ryu, Kim Young-shub, CEO of Korea’s telecom giant KT Corp., and Samsung Electronics Co.’s TV business chief Yong Suk-woo.
“Public and private sectors must become one across industries and make a concerted effort to build national capabilities for AI-based growth. We need AI-based innovation,” said Science Minister Lee Jong-ho at the meeting.