The Taiwan government is developing its own LLM called TAIDE to support the growth of Taiwan’s AI industry. TAIDE aims to generate fluent traditional Chinese conversations tailored to Taiwan’s culture, geography, history, and customs. The model is being developed through a collaboration between government agencies, academic experts, and industry partners. Taiwanese companies such as MediaTek, Asus, and Innolux are also developing traditional Chinese AI models.

Source: TAIDE
However, TAIDE faces challenges in acquiring sufficient high-quality training data. Due to uncertainties around copyright infringement in using web data for AI training, the TAIDE team is cautiously relying almost only on government agency data. This limits the diversity and scale of the training dataset compared to industry-developed models that leverage vast web-scraped data. Although the TAIDE team is in discussions with news media and publishers to expand its training data, progress has been slow. As a result, TAIDE’s performance currently lags behind industry models. For example, ASUS’s FFM-Llama3–70B has surpassed GPT-4 in traditional Chinese capabilities, while TAIDE can only catch up with GPT-3.5. This performance gap may hinder adoption of TAIDE by Taiwanese companies. Instead of directly developing AI models, the Taiwanese government should consider focusing on facilitating access to local training data. Major tech companies like OpenAI, Meta, and Apple are already forming data licensing agreements with news agencies and publishers. Taiwanese AI developers often lack the resources and experience to navigate such content industry partnerships. The government could play a key role by mediating licensing negotiations and establishing a comprehensive traditional Chinese dataset for AI training. This would provide a valuable backbone for Taiwan’s AI ecosystem to build upon. Overcoming the training data bottleneck will be crucial for Traditional Chinese AI models to realize their full potential.