South Korean telecommunications company KT Corp. has introduced its new large language model (LLM) called Mi:dm, as part of its venture into the thriving artificial intelligence (AI) market. Mi:dm aims to cater to corporate clients by enabling them to build their own next-generation AI services.
Choi Joon-ki, the head of the AI and big data business unit at KT, highlighted the importance of feedback from enterprise customers in shaping the development and trajectory of Mi:dm. The name Mi:dm is derived from the Korean word mideum, meaning belief.
Mi:dm offers four different service plans with varying computational capacities, ranging from under 10 billion to over 200 billion parameters. These plans allow companies to develop AI-based products, especially generative services like ChatGPT.
The key feature of Mi:dm is its reliable package, which combines Document AI, Search AI, and FactGuard AI to enhance the model’s efficiency and accuracy. FactGuard AI, in particular, aims to minimize the generation of false information, known as hallucination, by utilizing fine-tuned data from expert domains. KT claims that this technology can reduce hallucination by up to 70% compared to a Mi:dm model without it.
In addition to the models, KT has developed Mi:dm Studiov, a user-friendly interface for enterprises to evaluate various Mi:dm services.
KT aims to offer its AI solutions to corporate clients at prices 20 to 30 percent lower than its domestic competitors. Leveraging its existing customer base of 650,000 companies through KT Cloud, the company believes it has a competitive advantage in providing tailored LLM-based services.
During the announcement, KT also discussed its partnerships with domain-specific startups like Upstage and educational AI applications such as Qanda. Qanda, in particular, is focused on addressing challenges in creating a reliable AI math tutor by overcoming latency, costs, and accuracy issues.
KT and Qanda plan to collaborate by specializing the model in the Korean language and high school mathematics. This collaboration aims to leverage Qanda’s educational data and finely-tuned algorithms to develop an AI tutor capable of understanding individual students’ learning patterns.
Bae Soon-min, senior vice president at KT and head of the AI2XL research lab, highlighted real-life applications and potential uses for Mi:dm. For example, small businesses in areas like interior decoration can utilize the basic model to summarize client calls through a customized AI model. Gaming companies can adopt the standard model and train it further to create unique non-player characters. Local governments can utilize the efficient premium model to assist citizens in navigating a wide range of welfare benefits.
Regarding funding, Choi mentioned that while the initial investment in GPU farms is substantial, the company anticipates reaching a break-even point within approximately three years, drawing from the experiences of implementing large-scale IT system services in other Korean enterprises.
On an international scale, KT recently partnered with Thailand’s Jasmine Group to develop a Thai-based large language model. This move aligns with the company’s strategy to create AI solutions tailored for languages currently underrepresented in existing large language models.
According to Bae, Mi:dm has been designed to be versatile and scalable, catering to a wide range of businesses and their specific AI needs. The company aims to provide solutions for both small businesses requiring basic AI support and large corporations in need of high-performance models.
As KT Corp. makes its foray into the AI market with the launch of Mi:dm, it seeks to leverage its expertise, partnerships, and competitive pricing to establish a strong presence in the corporate AI services sector.