56 PERSPECTIVES ON BUSINESS AND ECONOMICS | VOL 43 | 2025 Current AI involvement and initiatives As Taiwan comprises essentially the world’s entirety of advanced chip manufacturing, its involvement in the AI market has mainly been in chip making for AI technologies. The demand for AI chips exists because training a leading AI algorithm using typical or general-purpose chips can require a month of computing time and cost upward of US$100M. AI chips can be tens to thousands of times quicker than typical chips when implementing AI algorithms and are significantly more cost-effective given their lower use of computing power (Khan & Mann, 2020). Current applications for AI chips include high-performance AI processors, such as graphics processing units; consumer electronics, such as smartphones and smartwatches; autonomous vehicles; cloud infrastructure; and quantum computing. However, production of these applications still constrains Taiwan to semiconductor manufacturing. In 2024, President Lai’s interest in AI quantum computing, robotics, precision medicine, and other advanced technologies (Sharwood, 2024) indicated a desire to reconfigure Taiwan’s role in the emerging AI industry from a supplier of physical hardware to a supplier of services. A recent AI “grand strategy” initiative was developed from the National Science and Technology Council 2017–2021 project, which invested US$517.5M to build an AI local ecosystem (Yau, 2024). The project supported infrastructure, such as centers for sufficient data and processing power, and set up R&D hubs to foster innovation in AI services. The project also provided funds for AI competition-based events to inspire and cultivate AI talent. The Science and Technology Advisory Group of the Executive Yuan (Taiwan’s cabinet) also established the AI Taiwan Action Plan to broaden AI capacity, demonstrating an awareness of the importance of entering the industry (Yau, 2024). The Taiwan Action Plan 1.0 has sought to further current industries by increasing financial support for R&D activities in AI. By 2021, this plan resulted in the Taiwania series of supercomputers and an AI real-time object detection system—the YOLOv4. It also contributed to the founding of organizations, such as the Taiwan AI Academy Foundation, aimed at educating youth about AI (Yau, 2024). AI sectors To explore where Taiwan should focus future efforts in AI, it is necessary to understand the different sectors of the field. AI can be divided into three camps: symbolic, statistical, and neuro-symbolic. Symbolic AI generates outputs using logic, such as algorithms and proofs (Smolensky, 1987), whereas statistical AI learns from large amounts of data and artificial neural networks to model predictions based on massive amounts of data beyond what a human could individually analyze (Domingos et al., 2016). Neuro-symbolic AI acts as a hybrid, generating explainable results from limited data and complex topics (Sheth et al., 2023). A notable type of neuro-symbolic AI, discussed in the following section, is discovery-based AI, an unsupervised type of machine learning that draws patterns and inferences from unlabeled data (Deng et al., 2022). AI investors tend to disfavor oversaturated, finite, and costly sectors. Open language models (OLMs), such as ChatGPT, Meta, and Gemini, are statistical AIs that raise such concerns. Although popular at the moment, OLMs rely on large amounts of novel human data, which are quickly reaching a plateau, and require astronomical amounts of computational power (Hadi et al., 2023). Alternatively, emerging neuro-symbolic AI requires far fewer data and has lower operational demands (Susskind et al., 2021). Compared to large language models, of which OLMs are a publicly available subset, neuro-symbolic AI utilizes structured knowledge representations to condense datasets, generalizes from fewer examples, computes more efficiently using smaller neural components integrated with symbolic modules, and has more targeted processing (Bhuyan et al., 2024). Recommendations Given economic and political constraints, there is promise in discovery-based AI (i.e., AI that contributes to scientific discovery), especially within the neuro-symbolic camp. Examples are deep representation learning, which simulates brain-like layers of neural networks, to resolve complex tasks (Wang et al., 2023), and drug discovery, which efficiently predicts drug candidate effectiveness (Blanco-Gonzalez et al., 2023). Discovery-based AI is focused on incorporating scientific knowledge, such as molecular restraints, into mathematical formulas. With that, AI models can compute and form connections between concepts not yet unearthed or even comprehensible by human capacity (Wang et al., 2023). Discovery-based AI An example of discovery-based AI is Robot Scientist Adam, a hardware and software system capable of developing functional genomics hypotheses for yeast (King et al., 2009). Robot Scientist Adam combines the neuro-symbolic AI of machine learning in pattern recognition with symbolic AI, simulating the scientific method in discovery-based AI generating novel knowledge on genomics. One output was Ro-
RkJQdWJsaXNoZXIy MTA0OTQ5OA==