Corporate clients can select and learn the language model on a website called KT Mi:dm Studio. Mi:dm is categorized into four types. The smallest model has 7 billion parameters, while the super large one has more than 200 billion parameters, exceeding GPT-3’s 179 billion, KT said. The telecom giant has also strengthened Mi:dm’s learning potential to minimize hallucination, a phenomenon in which an LLM generates nonsensical or unfaithful content. Mi:dm is equipped with AI services that help it to understand diagrammatic and complicated documents, identify information optimized for target content and generate responses based on original sources. These technologies reduce hallucinations by up to 70% compared with existing generative AI services, KT said.  The company will provide the LLM and its cloud management service in a package. The cloud’s neural processing units (NPUs) are operated by Rebellion Inc., a Korean fabless backed by KT. The NPUs can cut the LLM’s costs for service and inference by up to 50%, KT explained. The mobile carrier will focus on globalization, manufacturing, finance, public services and education during its LLM service expansion. It will strengthen collaboration with Upstage, Quanda, Enuma and BI Matrix — all Korean AI startups it has invested in or forged tech partnerships with. KT has discussed usage of Mi:dm with some 100 companies, including state-run Industrial Bank of Korea (IBK), which has already launched a product information provider based on the LLM. The telecom giant is advancing its LLM business in the sectors of internet protocol television (IPTV), wireless service, vehicle infotainment systems, robotics and customer service. The company said that voice recognition by call center bots can improve by 5%, while post-call processing and knowledge base building have been improved by 20% and 30%, respectively. KT said it will focus on work process improvement in finance, telecom, IPTV, senior-care and child care counseling with the LLM. The company is intensifying AI as its new growth driver. In June, the company said it will spend about 7 trillion won by 2027 to become the country’s largest AI service provider.   (ICE SEOUL)

Fonte notizia: The Korea Economic Daily