- South Korean companies are competing in the development of Small Large Language Models (sLLM), joining global tech giants.
- Meta’s official blog featured a case study on South Korean startup Upstage and Mathpresso’s ‘MATH GPT,’ highlighting its mathematical prowess.
- sLLMs are emerging as cost-effective alternatives to Large Language Models (LLMs), attracting major tech players like Google, Microsoft, and Meta.
- These lightweight models offer improved response speeds and efficiency, making them suitable for on-device AI applications.
- South Korean companies like Naver, Upstage, and Fasoo have introduced their own sLLMs, catering to various industries with cost-effective solutions.
Main AI News:
South Korean companies have entered the competition to develop Small Large Language Models (sLLM), joining the ranks of big global tech players. A Large Language Model (LLM) is an artificial intelligence (AI) program with the ability to understand and generate text, among other functions.
Meta recently showcased a case study on South Korean AI startup Upstage and Mathpresso’s ‘MATH GPT’ on its official blog. This is a departure from Meta’s usual content, which predominantly focuses on its own products or research endeavors.
While MATH GPT’s parameters are smaller compared to GPT-4 (1 trillion), standing at 13 billion, its mathematical prowess is considered top-notch. It outperformed OpenAI’s GPT-4 in the ‘Math Benchmark,’ scoring 0.488 out of a perfect 1 in a set of 12,500 challenging math problems.
According to an Upstage official, MATH GPT is specialized in mathematics, leveraging advanced math data held by Mathpresso. This niche specialization sets it apart from broader models like ChatGPT, which have been trained on a relatively smaller volume of mathematical data.
The emergence of sLLMs has attracted attention as a cost-effective alternative to their larger counterparts, which demand substantial investment for both training and operation. Major tech companies such as Google, Microsoft, and Meta have all unveiled their versions of sLLMs this year.
For example, Google introduced ‘Gemini Nano,’ Microsoft launched ‘Phi-3 Mini,’ and Meta unveiled ‘Llama 3.’ These sLLMs have parameter ranges from hundreds of millions to tens of billions, in contrast to LLMs with parameters exceeding 100 billion.
Sam Altman, CEO of OpenAI, has previously highlighted the exorbitant costs associated with running LLM-based models like ChatGPT. sLLMs aim to address this issue by improving response speeds and efficiency, making them more accessible for various applications.
These lightweight models can be directly embedded into devices like smartphones and laptops, enabling on-device AI capabilities. Moreover, their cost-effectiveness makes them particularly appealing for the business-to-business market, where efficiency and affordability are paramount.
In line with this trend, South Korean companies have also begun rolling out their own sLLMs. Naver recently unveiled ‘HCX-Dash,’ a lightweight model of HyperClovaX, offering a cost-effective solution for tasks such as generating sentences, summaries, reports, or custom chatbots.
Upstage took a step further by releasing ‘Solar Mini’ as open source on Amazon SageMaker JumpStart and AWS Marketplace. This allows clients to customize the model for their specific needs, particularly in the edutech industry where tailored AI solutions are in demand.
Fasoo, known for its expertise in data and application security, introduced ‘Ellm,’ an on-premises sLLM suitable for various professions and industries. Ellm enables additional training using small datasets specific to certain tasks or domains, catering to the diverse needs of organizations.
Conclusion:
The entry of South Korean firms into the global sLLM competition signifies a significant shift in the market landscape. With the emergence of cost-effective alternatives to traditional LLMs, businesses can expect increased accessibility to AI technologies, driving innovation and efficiency across industries. This trend underscores the importance of agility and specialization in the rapidly evolving AI market, where tailored solutions and competitive pricing will shape future success.
