演講摘要 Abstract

In recent years, increasingly powerful AI systems have gradually become our partners in both life and work. With the emergence of the AI Agent era, AI is evolving from an assistant to a true colleague. Human–AI collaboration brings numerous benefits and opportunities, yet it also introduces new challenges and risks.What does the ideal model of “human–AI collaboration” look like? And how should we respond to the arrival of this new era?

In an age when artificial intelligence is revolutionizing every facet of engineering, seismic hazard mitigation is no exception. This talk explores how cutting-edge machine learning models—ranging from deep neural networks to explainable AI—are transforming how we predict, assess, and respond to earthquakes in Taiwan and beyond. We begin with the use of models like XGBoost to develop advanced ground motion prediction equations (GMPEs), improving both the accuracy and interpretability of seismic hazard maps. We then introduce the Taiwan Transformer Shaking Alert Model (TT-SAM), a deep learning–based early warning system that delivers rapid and precise shaking forecasts using real-time seismic waveforms and physically meaningful features. Finally, we showcase how convolutional neural networks such as Mask R-CNN enable automated, large-scale building detection from satellite imagery—providing crucial exposure data for urban seismic risk assessment. Through these case studies, attendees will see how AI-powered innovations can help engineers, governments, and communities build a safer, more resilient future in the face of seismic threats.

The development of modern deep learning models faces critical bottlenecks, including the significant energy consumption associated with high computational loads, the substantial cost of large-scale data storage, increasingly complex and massive model architectures, and the scarcity of large-scale, adversarial-free training datasets. Consequently, low-cost model architectures that rely on limited training data offer the potential to lower the hardware entry barriers and provide robust, mature deployment solutions for a broad spectrum of users. Such models are well-suited for closed or edge-computing environments, including medical institutions and industrial fabrication systems. This presentation will introduce a series of lightweight topological learning models designed by our laboratory, built upon the unique mathematical frameworks of differential geometry and quantum mechanics. These models have been applied across diverse domains to support collaborative and generative tasks within contemporary human–machine interface environments.

1.Introduction: The evolution from machine operation and human-robot collaboration to true human-robot coexistence.

2.The Current Challenge: As robots step out of their safety cages, can AI anticipate and mitigate potential risks?

3.Traditional Mechanisms: The core problem of reactive safety- being rigid and passive.

4.The Generative AI Solution: Moving beyond content creation to embrace prediction and reasoning.

5.Core Technology I: Human Intent Prediction.

6.Core Technology II: Dynamic Path Planning and Proactive Response.

7.Core Technology III: Interaction and Communication.

8.Application Case Study.

As artificial intelligence advances at a breakneck pace, we face an urgent challenge: aligning these powerful new tools with the well-being of people and the planet. Since the 2017 Transformer breakthrough and the mass adoption of ChatGPT in 2022, systems such as large language models and agentic AI are reshaping every sector. Still, they also introduce new technical and ethical risks.

This presentation argues for a holistic framework that connects cutting-edge innovation to clear sustainability goals, such as reducing carbon footprints and ensuring equitable access to AI's benefits. This human-centered view is built on seven core dimensions of trustworthy AI: human agency and oversight, technical robustness, privacy and data governance, transparency, diversity and non-discrimination, societal and environmental well-being, and accountability.

We can achieve these goals through practical tools, such as explainable AI (XAI) techniques, bias detection tools, and inclusive design, with real-world examples from education, healthcare, and smart grid management. By focusing on fairness and explainability, we can move "from cool technology to warm humanity," ensuring automated decisions are understandable, contestable, and keep people in control.

AI plays a dual role in our climate crisis, both contributing to the problem and potentially mitigating it. The energy demand from information and communications technology is soaring, projected to hit 20% of global electricity use by 2030. To counter this, green AI strategies such as on-device inference, renewable-powered data centers, and disciplined model sizing are essential. At the same time, AI offers powerful environmental solutions, including precision agriculture, biodiversity monitoring, and disaster-response robotics.

Taiwan is making significant contributions through the Future Earth Taipei Working Group on Sustainability in the Digital Age. This group brings together academia, government, and industry to co-design AI solutions. Current projects include agent-based research assistants and public education on responsible AI. Furthermore, international workshops planned at Academia Sinica in November 2025 will help strengthen regional collaboration and align local regulations with global standards. 

Ultimately, intelligence is more than just computing power. Sustainable progress depends on embedding ethics, social dialogue, and ecological awareness into every stage of the AI lifecycle. By integrating advanced engineering with culturally grounded values, we can ensure that AI catalyzes inclusive and resilient development, rather than driving inequity. The goal is a shared agenda for human-centered sustainable AI that balances innovation with stewardship, turning technological breakthroughs into a force for our collective flourishing.

In an era of rapid AI development, this paper re-examines the nature of human learning and the direction of talent cultivation through epistemological and ontological perspectives. The central thesis asserts that human existence ultimately pursues freedom, ideality, and meaning. AI should be regarded as an extension of human capability—not a replacement. Its value lies not in substituting human faculties, but in helping us move closer to these existential goals. 

Although AI has transformed how knowledge is accessed and learning is conducted, human learning remains rooted in two irreplaceable capacities: tacit knowledge and meaning-making. Tacit knowledge refers to implicit understanding gained through practice and intuition, while meaning-making involves the critical process of transforming information into understanding and value.

There are varying discourses regarding whether Artificial Intelligence (AI) substitutes, enhances, or complements our work. AI can replace standardized jobs, leading to the substitution of employment opportunities; however, AI can also alter how tasks are performed, enhancing work output. Furthermore, AI possesses the potential for "complementarity" in work execution. The impact of AI on employment varies according to company size and the educational level of workers. Large enterprises, possessing more resources and greater data volumes, are better positioned to leverage the competitive advantages created by AI. Similarly, the higher a worker's education level, the greater the productivity gains derived from task automation. Conversely, AI reduces the demand for entry-level jobs and standardized tasks. Since administrative work and reporting can be generated in batches through AI application, junior workers or those with lower educational levels who originally performed such tasks may require retraining to learn how to utilize AI. However, manual labor that cannot be standardized—such as plumbing and welding—may actually make it easier for lower-skilled workers to enter the labor market. Under the AI trend of "skill-biased technological change," workers possessing specific skills benefit, while those without them cannot share in the technology's economic benefits. Consequently, the government needs to further consider retraining approaches to facilitate "occupational transition."

ChatGPT, short for Chat Generative Pretrained Transformer, is a tool characterized by the features of Generative AI. It can respond to users’ needs with remarkable speed and assist in accomplishing a wide variety of tasks, thereby demonstrating tremendous potential across diverse fields. However, the misuse of ChatGPT may bring about unexpected risks and challenges. Accordingly, there is a need to discuss how to use ChatGPT with right methods and attitudes. 

In this talk, I will begin by introducing the core concepts of ChatGPT and explaining why it has become an indispensable assistant. Through relatable, real-life analogies, I will help the audience better understand ChatGPT’s role and value. Subsequently, I will share real cases of students using ChatGPT, along with some painful lessons that both my students and I have encountered.

At the same time, I will present a systematic set of steps to demonstrate how ChatGPT can be used correctly and efficiently, ensuring that its potential is fully realized. Finally, I will explore how students, teachers, and parents can collaborate in the age of AI. In short, my hope is that this talk will not only help everyone learn how to coexist with ChatGPT, but also transform it into a trusted partner in both learning and daily life—one that empowers us to embrace the new challenges and opportunities brought by AI.

The rise of short-form videos, self-media, and generative AI—along with the rapid transformation of educational systems, politics, and society—has profoundly reshaped students’ learning behaviors and motivations. Traditional models of lecturing, assignments, and examinations are facing unprecedented challenges. Existing pedagogical problems have not diminished; instead, new ones continue to emerge, widening the gap between teachers’ expectations and students’ learning realities. Nevertheless, advances in digital tools and AI technologies also present unprecedented opportunities for educational innovation.

This presentation takes the undergraduate EMI (English-Medium Instruction) required course Electromagnetism as an example to examine typical challenges currently faced in higher education. It further demonstrates how Large Language Models (LLMs) can assist in addressing these challenges through various teaching applications, including course content restructuring, real-time tutorial interaction, classroom activity design, automatic problem generation and solution, and personalized feedback mechanisms.

In terms of professional knowledge and linguistic ability, LLM-based AI has reached a level comparable to that of a doctoral student, while offering exceptional patience and accessibility. It is thus capable of replacing traditional teaching assistants in many instructional tasks and serving as a learning companion that provides individualized guidance and practice for students. For instructors, AI can also act as an efficient teaching assistant that supports course design, material generation, and conceptual organization—significantly reducing preparation workload and enhancing classroom interaction quality.

However, both teachers and students must remain aware that AI-generated information and suggestions should only be adopted after critical evaluation and verification. It is recommended that LLM AI be used in conjunction with other AI tools and reliable databases or search engines to cross-check information accuracy and maintain academic integrity. Finally, this report will introduce Uedu platform of National Central University as an experimental environment for AI-assisted teaching, providing practical insights for integrating AI teaching assistants into future higher education contexts.

As Generative Artificial Intelligence (GenAI) advances rapidly, higher education is undergoing a paradigm shift from traditional knowledge transmission toward competency based and adaptive learning. However, textual interactions and conventional assessments alone are insufficient to capture learners’ cognitive, emotional, and physiological states comprehensively. This presentation introduces the "Educational Omics Data Lake" and the Uedu platform developed by the National Central University team. The platform integrates multimodal data sources—including speech, text, images, heart rate variability (HRV), and learning trajectories—and applies multimodal learning analytics to generate richer characterizations of learning processes and to inform instructional decision making. Uedu constructs student digital twin modules and personalized adaptive learning pathway blueprints. Validated across multiple course contexts, the platform supports real time instructional interventions and competency oriented learning; preliminary outcomes include: (1) empirical evidence that a visualization tool enhances comprehension of abstract concepts; (2) improved learning transfer and learner confidence in a Python programming course via integration of QuizGPT and UeduGPT; and (3) reconstruction of learning trajectories using a prompt based cognitive performance assessment approach. With strict safeguards for sensitive data privacy and adherence to academic ethics, Uedu translates educational big data and human–AI collaboration into an intelligent, innovative, sustainable, and resilient campus teaching infrastructure, thereby facilitating cross disciplinary human–AI co learning and promoting the responsible sharing and practical application of educational big data.