Taiwan Develops AI Model Matching GPT 3.5, Geared Towards Enterprise Use
The National Science and Technology Council (NSTC) has assembled Taiwanese research teams to develop a dedicated Trustworthy AI Dialogue Engine (TAIDE) for the Taiwanese public and businesses.
The model based on Llama 2 with 7 billion parameters is said to have been completed.
Chairman of NSTC, Tsung-Tsong Wu, stated that another model with 13 billion parameters is nearing the level of GPT3.5.
Taiwan's tailor-made AI model
Wu emphasized that the Taiwanese government's aim is not to create a "Taiwanese version of ChatGPT" but to develop a reliable AI dialogue model for value-added applications by enterprises or public-private sectors.
Wu noted that the model based on Llama 2 (with 7B parameters) was activated in March 2023. This large language model enables various industries to continue training and tailor generative AI according to their needs. The TAIDE with 7 billion parameters has no issues summarizing meetings, while the 13 billion parameters TAIDE is closest to Open AI's GPT 3.5 version.
According to data provided by NSTC, due to the limited availability of Traditional Chinese language data, international large-scale language models often produce biased or culturally inappropriate responses for Taiwan. Therefore, Taiwan has developed its own TAIDE within limited time and budget constraints.
Wu also mentioned that some data and content are not suitable for ChatGPT, and to his knowledge, Taiwanese government agencies do not extensively use ChatGPT.
The progress of Taiwan's regulation of AI
Regarding the legislative progress of the "Artificial Intelligence Fundamental Act" raised by legislators, Wu stated that Taiwan must balance innovation and regulation. The content currently leans towards referencing the policy directions of the United States, Japan, and the United Kingdom.
The Executive Yuan's AI Basic Law draft is scheduled to be released by the end of 2024 for legislative review.
On October 3, 2023, the Executive Yuan issued a reference guide for using generative AI, outlining the potential risks to maintaining professionalism and confidentiality in executing public affairs. Each agency may establish usage norms or internal control measures based on business needs for using generative AI.
Taiwania 2
Regarding the construction of high-speed computers and the enhancement of computing power in Taiwan, NSTC stated that the National Center for High-performance Computing (NCHC) completed the construction of the high-speed computing host "Taiwania 2" at the end of 2023.
It was built upon a new generation CPU architecture that will provide a computing performance of 3.53 petaflops, more than twice the computing performance of "Taiwania 1." It is set to replace the first host and open for service in the third quarter of 2024 and improve energy efficiency management.
It will be a significant host in Taiwan's large-scale engineering and scientific simulation computing fields. The National Center for High-performance Computing is expected to initiate other large-scale construction projects before 2029.
Wu pointed out that Taiwan's semiconductor industry dominance and the rapid development of generative AI present a golden opportunity for Taiwan's international visibility. NSTC strengthens talent cultivation measures across all stages and collaborates across departments to enhance scientific research talent cultivation.
It aims to construct a resilient scientific research talent ecosystem through diverse channels and deepen international scientific and technological cooperation relationships and exchanges to achieve international linkage in cross-border research environments and R&D activities. This is to attract international talents, startups, and funds to Taiwan. Which Wu believes will prepare Taiwan with sufficient key scientific research manpower and enhance its international competitiveness.
Source: Bryan Chuang, DIGITIMES Asia
