Will a Korean LLM-Dedicated Chip Emerge? Industry Anticipation for Samsung's 'Maha Series'

Simultaneous Development of Mach 1 and Mach 2... Details Remain Confidential
No AI Yet Perfectly Implementing Hangul
"Language-Centered LLMs Are the Trend... Hangul Inference Chips Possible"
Attention on Whether Hangul-Based AI Will Resolve Corporate Inconveniences

Samsung Electronics has reportedly begun developing the AI inference chip ‘Maha Series (Maha1·Maha2),’ raising expectations within the industry that a proper ‘Korean language-based AI’ will emerge. Since the development of AI and large language models (LLMs) is centered around the language used, it is highly likely that Samsung Electronics will create a chip based on the Korean language.

According to industry sources on the 9th, Maha is being developed secretly with about 40 semiconductor engineers involved, without disclosing details such as development and investment directions or the language to be used. An industry insider said, "Considering that LLMs are being developed centered on the language used, the Maha series is very likely to be made based on the Korean language."

So far, AI systems commercialized by global companies have all been based on English, which is the global lingua franca and because the U.S. has been at the forefront of AI development. However, Samsung Electronics declined to comment on whether the Maha series would be based on Korean, stating, "It is difficult to disclose at this time."

If Samsung Electronics releases the Maha series based on Korean, it is expected to alleviate the challenges faced by domestic companies. Currently, most generative AI used by Korean companies is based on English. To use them in Korean, other programs must be integrated. Even if AI has learned Korean, the translation and usage are not perfect, leading to inaccurate conveyance of meaning and causing inconvenience for companies.

In February, when Mark Zuckerberg, CEO of Meta, visited Korea, our AI startup Upstage proposed developing a Korean version of ‘LLaMA 3.’ LLaMA is an LLM developed by Meta’s AI group. The previous model, LLaMA 2, received harsh criticism among Korean companies for its poor Korean language performance. Google’s AI chatbot service ‘Bard,’ launched in May last year, also offered a Korean version from the start but has been evaluated as still lacking in information accuracy and reliability.

The fact that Maha is being developed as an inference chip further raises expectations for the birth of a proper Korean language-based AI. According to linguists, Korean has a more complex grammatical structure and much more diverse expressions than other languages, making it difficult for AI models to learn. The ‘tokenization technology’ used by current AIs is said to be unsuitable for learning Korean. Tokenization is a technique that breaks text into smaller units such as words or characters for learning. Using this, it is difficult to perfectly learn the more diverse and complex Korean consonants and vowels compared to the English alphabet. Therefore, the industry consensus is that significant time and money are required for AI to learn Korean. On the other hand, inference is considered relatively more likely to succeed. It only needs to analyze learned information in English and generate new information in Korean. If inference succeeds, it is expected to provide solutions for Korean language learning methods in the future.

The partnership with Naver is also attracting significant attention. After development, Maha is scheduled to have its performance verified by Naver as early as the end of this year. Naver has built a large-scale infrastructure based on the Korean language, making it an ideal debut stage if Maha is created based on Korean. Naver’s release of ‘HyperCLOVA X,’ which has learned Korean 6,500 times more than other generative AIs in August, and its efforts to build a large-scale AI ecosystem seem to align well with Samsung Electronics’ goals. At last month’s regular shareholders’ meeting, Naver CEO Choi Soo-yeon stated, "The commercialization of Maha1 and the scale at which Naver will use it are basically undecided," adding, "We expect to conduct performance verification and stabilization tests this year to determine whether it can be used as a chip after going through programmable logic devices (FPGA)." Recently, the demand for K-content such as Korean movies and dramas has surged in Southeast Asia, China, and Japan, increasing the need for Korean language-based AI development. There is also analysis that Maha is likely to attract attention in these countries.

Editor's NoteMaha☞ Maha is a general-purpose AI (AGI) semiconductor specialized in AI inference. AI functions are divided into learning and inference, and Maha focuses on inference. Samsung Electronics’ System LSI division is developing two versions, Maha1 and Maha2, simultaneously.