Day 2 · R0 Smart Transformation & Memorial Lecture
From LLM task design and Ministry of Finance tax-audit case selection to the Taiwan CbI Program and enterprise knowledge management, closing with the Chen Sheng-Wei Memorial Lecture and the Taiwan AI Academy's closing remarks — covering AI transformation across government and industry.
At a glance
| Speaker | Talk | In one sentence |
|---|---|---|
| Moderator | Opening remarks | Introduced Tsai Tzong-Han, the first of the afternoon's four speakers |
| Tsai Tzong-Han (蔡宗翰), Academia Sinica RCHSS / National Central University | Task Design, Dataset Creation, and Evaluation | The key to building a proprietary LLM isn't the model itself, but task design, data, and evaluation |
| Chang Wen-Hsi (張文熙), Director, Ministry of Finance Fiscal Information Agency | The Ministry of Finance's Experience Using AI for Business Transformation | Using invoices, cash-flow data, and AI case selection for tax audits — averaging over NT$50 million in back taxes per case |
| Liao Hong-Yuan (廖弘源), Director, Academia Sinica Institute of Information Science / Deputy Executive Director, CbI Program | The Taiwan CbI Program | The four strategic pillars and resource allocation of the ten-year, NT$300 billion CbI Program |
| Yang Li-Wei (楊立偉), Managing Director, eLAND Information | AI is revolutionizing knowledge management! | RAG lets enterprise knowledge management skip fine-tuning while staying traceable and model-agnostic |
| Chen Wei-Chao (陳維超), Chief Digital Officer & Senior Vice President, Inventec | Chen Sheng-Wei Memorial Lecture: Delivering Trust in the AI Era | Examines AI's privacy, robustness, fairness, transparency, and social responsibility through an engineering-specification mindset |
| Kuo Ping-Chen (郭秉宸), Industry-Academia Liaison, Taiwan AI Academy | Empowering and Balancing Taiwan with AI | Closing: 7 years, 11,000 alumni — calling on industry to jointly advance data governance and homegrown models |
6 of 6 talks
1Task Design, Dataset Creation, and Evaluation: Keys to LLM Deployment and Taiwan's AI FoundryTsai Tzong-Han (蔡宗翰), Research Fellow, Research Center for Humanities and Social Sciences, Academia Sinica; Professor, National Central University; Convener, Core Model Training & RLHF Group, TAIDE Program
The success of building an enterprise- or domain-specific LLM hinges not on the model itself, but on task design, data, and custom evaluation methods.
Key points
- Opened with the metaphor of "Kuafu chasing the sun" (誇父追日) to describe Taiwanese developers' anxiety facing the international LLM arms race, coining the term "LLM anxiety": unable to match international tech giants in funding, data, or compute.
- Argued that enterprises building proprietary models have two routes: (1) "general large model + RAG" or (2) "training a small model"; he personally prefers the latter, because a general large model is like "a smart fresh graduate" — able to answer anything, but untrained on internal knowledge, with unstable accuracy and hallucinations.
- Cited an example: when GPT-4o answered "why did Emperor Wanli not meet his officials for thirty years," laypeople and engineers thought the answer was good, but a National Taiwan University History graduate teaching assistant scored it only 78%, citing insufficient historical evidence and low credibility — illustrating that a general model's answer-quality ceiling is set by domain-expert perspective.
- Compute reality: running LLaMA 3.1 405B as a general large model + RAG requires 11 H100 GPUs for inference; even the 70B version needs 2 H100s — unaffordable for typical teams, and scaling cost rises steeply as user numbers grow.
- Proposed that building a proprietary model needs a mixed team of three roles: a model development team (training the model); a data team (inventorying, preprocessing, de-identification, task design, producing instruction-response pairs — per audience crowd-notes, one specific de-identification method is using BERT to find entities before processing, which is lower-cost); and an evaluation team (designing evaluation instructions and gold answers, which must involve the client/domain experts, not engineers alone).
- Criticized common general benchmarks (e.g., MMLU) as essentially a "probability estimator" evaluation method (comparing the model's probability for option strings A/B/C/D), which measures a pretrained model's knowledge content and is not suitable for evaluating instruction-fine-tuned models; he cited Google's official Gemini page, which likewise states MMLU is better suited for evaluating pretrain models.
- Classroom experiment: had computer science graduate students and history graduate students each revise the same "history Q&A grading prompt." The CS students changed not a word (believing it already perfect), while the history students substantially added a requirement to "provide research-usable evidence" — proving that evaluation prompts must be written with domain experts involved, otherwise the model's performance ceiling gets stuck around 70 points.
- Outlined a five-step process for building a private model (Continued Pretraining (CP) → instruction data → fine-tuning → automated evaluation → periodic human-in-the-loop evaluation), emphasizing that the human evaluation step must be done at irregular intervals until the model reaches an 80–90 score before evaluation can be fully handed to machines.
- Case 1: In collaboration with the Ministry of Education, trained "EduTAIDE," a Mandarin-teaching-material-creation copilot (can write lesson texts, extract vocabulary lists, generate grammar exercises, and auto-generate test questions). Pass rate: EduTAIDE 76% vs. GPT-4o 51%; level deviation between generated text and the specified proficiency level: EduTAIDE averaged 0.776 vs. GPT-4o's 1.31 — proving that clear task definition plus fine-tuning can beat a general model.
- Cited a September 26 case from an insurance-industry ESL company: after training an insurance-specific LLM, accuracy improved 30% over GPT-4o and cost dropped 30%, with better compliance and consistency (unverified — an industry case cited by the speaker, original source not confirmed).
- Cited a keynote view from Barbara Plank of LMU Munich at ACL (Bangkok): the key focus of LLMs has shifted from "modeling" to "data and evaluation" — consistent with Tsai's own experience.
- Called for Taiwan to pursue an "AI foundry" route (citing a view from Chen Liang-Gee): companies trust Taiwan enough to hand over data for training proprietary models without fear of leaks, and since every industry needs different AI, this suits customized services well.
Tech, products & figures
- TAIDE
- A Traditional Chinese trustworthy AI dialogue engine program led by the National Science and Technology Council (NSTC); Tsai Tzong-Han is convener of its core model training and RLHF group.
- EduTAIDE
- A Mandarin-teaching-material-creation copilot developed by Tsai's team in collaboration with the Ministry of Education.
- MMLU
- A multiple-choice benchmark measuring a pretrained model's knowledge content; not suitable for evaluating instruction-tuned models.
- BioCreative, BioASQ
- International biomedical text-mining/QA-system competitions hosted by the U.S. NIH; Tsai Tzong-Han has won first place in them.
Notable quotes
"I feel like Kuafu (誇父) chasing the sun — no matter how hard I chase, I can't catch up... it's given rise to a new psychological term unique to the AI field: LLM anxiety."
"Task, data, and evaluation are the keys to LLM development — data is never just raw text; it must be transformed into the optimal form for the task."
Q&A
- No independent Q&A session (this session consisted of four consecutive keynote talks, with only moderator transitions in between).
Fact-check notes
- Confirmed via WebSearch that Tsai Tzong-Han holds positions as Research Fellow at Academia Sinica's Research Center for Humanities and Social Sciences, Professor at National Central University, and convener of TAIDE's (a generative AI dialogue engine) core model training and RLHF group.Sources：社會創新平台專訪科技大觀園
- "Using BERT to identify entities for de-identification is cheaper" is a technical detail added via audience crowd-notes.
- TAIDE also came up in other sessions at this year's conference: in a talk on multimodal model development, and in another discussion of the etymology of "sovereign AI" and the scale of TAIDE's resources.
- Tsai Tzong-Han's view that "training a small model beats a general large model + RAG" contrasts with this session's Talk 4 (Yang Li-Wei), who argues "enterprise knowledge management should shift to RAG + swappable models"; yet another speaker elsewhere at this year's conference argues full-parameter fine-tuning beats RAG/LoRA, making this a three-way disagreement on the topic.
2The Ministry of Finance's Experience Using AI for Business TransformationChang Wen-Hsi (張文熙), Director, Ministry of Finance Fiscal Information Agency (財資中心)
The Ministry of Finance introduced AI centered on the "case-selection tax audit" scenario, using data governance plus risk scoring to sharply improve audit precision.
Key points
- Two major pain points for government agencies pursuing digital transformation: difficulty retaining talent (private-sector salaries are too high, leading to high civil-servant turnover within 3 years) and lack of incentives; the strategy is to start by adding AI features to existing systems, then gradually advance toward an AI platform.
- Governance structure: using a "shepherd/sheepdog" metaphor for division of labor — the Information Center director is the "lead sheep," the Deputy Minister overseeing business units is the "sheepdog"; below them are four groups (the "data collection group" is the most important, since nothing can be done without data), involving cross-ministry data coordination among five regional National Taxation Bureaus, roughly 70–80 branch offices, and local tax collection authorities across 22 counties/cities.
- Core tax categories were chosen by revenue share: income tax (business income tax + individual income tax) plus business tax plus customs duties together account for over 50% of tax revenue, so AI resources were prioritized there.
- Preparatory work (starting in 2020 / ROC year 109): tax registration and voucher data formats were inconsistent (e.g., addresses like "一二三段" [Section 123] mixed Arabic numerals and Chinese numerals), requiring data governance, normalization, a sampled database (about 1/10 of records) preserving the original distribution, and de-identification/masking to protect privacy.
- Concrete applications: mining invoices to detect B2B/B2C underreporting, cross-tax-category (income tax + business income tax + individual income tax) integrated auditing, and combining cash-flow/fund-flow analysis to trace asset transfers — catching loopholes invisible to single-tax-category audits, such as an influencer-agency "bulk operator" tax-evasion case.
- Case-selection results: manual selection previously had roughly a 1-in-10 hit rate; after switching to AI selection, about 90% of business-tax cases selected required back-tax payment, and about 60% for individual income tax; the average back-tax amount per case exceeds NT$50 million, with cumulative back taxes already exceeding NT$4–5 billion (the specific threshold design is not disclosed, as it involves audit-practice details).
- Smart customer service: the current rule-based chatbot gives legally correct answers but is "not user-friendly enough"; they are switching to generative AI to make answers closer to everyday language. This runs into the same problem Tsai Tzong-Han described — the public likes it, but tax officers feel the legal basis isn't precise enough; still being improved.
- Customs tariff classification: classifying imported goods is extremely complex (e.g., potato chips are classified differently depending on ingredient formula, and "notebook computer" isn't even a tariff-code name), so a language model assists with tariff classification lookup. For drug interdiction, beyond image recognition, association rules (behavioral conditions accompanying drug carriers) strengthen risk prediction, in collaboration with the National Applied Research Laboratories (referred to in the source as "科院").
- Pilot program "Business Tax Study Companion" (營業稅書童): the goal is to give every staff member an AI assistant; the core workflow first converts the public's colloquial questions into regulatory language before retrieval, and the system lists both "supporting references" and "opposing references" along with their counts — when opposing references outnumber supporting ones, the system suggests issuing a new ruling favorable to the public based on the opposing data. This presentation format is especially helpful to tax officers because it exposes viewpoints outside their own trained background.
- Cybersecurity: as cross-agency systems multiply and the attack surface expands, they are working (with external vendors) to give systems self-learning capability to defend against new types of hacking attacks.
- The nationwide tax-officer workforce numbers about 30,000; the goal is a unified platform to reduce staffing/deployment burden. The data is classified, with physical separation between internal and external networks; AI-adoption benefit assessments must be reported up for approval by the Vice Minister of Finance.
Tech, products & figures
- Case-selection risk scoring system
- Assigns a risk score to cases based on confidential business threshold conditions, used to determine audit priority.
- Customs tariff-classification AI
- Assists customs officers in mapping everyday product names to complex tariff classifications.
- Drug-interdiction association-rule model
- Combines image recognition with behavioral association rules to predict drug-carrying risk, developed in collaboration with the National Applied Research Laboratories (referred to in the source as "科院"; the institution's full name could not be fully verified).
Notable quotes
"When auditing, go after the big cases — investigating a case takes the same amount of time and manpower either way; only the data differs."
"This back-tax amount is money that was owed but not paid — it's not that we're collecting more than what was actually due."
Q&A
- No independent Q&A session (the moderator transitioned directly to the next speaker after this talk).
Fact-check notes
- Confirmed via WebSearch that Chang Wen-Hsi currently serves as Director of the Ministry of Finance's Fiscal Information Agency (財資中心), having transferred there in 2020 from Director of the Ministry of Labor's Information Office.Sources：中央社財政部財政資訊中心官網
- Government/public-sector AI-adoption pain points were also discussed in other sessions at this year's conference: several speakers there cited "convincing senior leadership" as a shared difficulty, while another speaker discussed a draft Data Governance Act from a legal-framework perspective.
3The Taiwan CbI ProgramLiao Hong-Yuan (廖弘源), Director, Institute of Information Science, Academia Sinica; Deputy Executive Director, CbI Program
The ten-year, NT$300 billion CbI (Chip-based Innovation for Taiwan) Program only just launched this year, aiming to use chips + AI to drive cross-industry innovation and support small and micro enterprises.
Key points
- Clarified his own positioning: his expertise is edge-side algorithms (working on handheld-device edge computing since 2019); he only joined as Deputy Executive Director this past March. The Executive Director role was originally meant to be held concurrently by the NSTC Minister, but since the Minister just took office, Chueh Chih-Ta (闕志達, Dean of NTU's College of Semiconductor Engineering) is acting Executive Director (interim) and concurrently Deputy Executive Director.
- Four program goals: harness the chip-and-AI revolution to drive cross-industry innovation; bring in semiconductor talent; lay the foundation for the nation's technological strength over the next decade; and promote industrial upgrading through deep integration of semiconductors and AI.
- Organizational structure (corrected against a slide photo provided via audience crowd-notes): the Chip-based Innovation for Taiwan (CbI) Promotion Office reports to the NSTC's Office of Science and Technology (at the Executive Yuan level). The office is currently led by Chueh Chih-Ta (闕志達, Dean of NTU's College of Semiconductor Engineering) as Executive Director (interim) and concurrently Deputy Executive Director, with Liao Hong-Yuan (廖弘源, this talk's speaker) as Deputy Executive Director. Under them are four directors, each covering one strategic pillar: Lin Hsuan-Tien (林軒田, Chief Strategy Officer, "industrial innovation"), Chen Ping-Hui (陳炳煇, Chief Investment Officer, "startup incubation"), Huang Chun-Lang (黃俊郎, Chief Technology Officer, "key technologies"), and Su Chao-Chin (蘇朝琴, Chief HR Officer, "international talent cultivation"). The four pillars, in order, are: industrial innovation, startup incubation, key technologies, and international talent cultivation (corresponding to budget pillars one through four described below).
- Budget scale: originally planned at NT$30 billion per year for 10 consecutive years, totaling NT$300 billion; but in ROC year 113 (this year, 2024) actual funding is only about NT$11.3–11.7 billion, rising to roughly NT$18-point-something billion ("180-plus 億") in ROC year 114 (next year, 2025).
- Budget allocation across the four pillars (ROC 113 → 114): Pillar 1 (AI+chip-driven industrial innovation) NT$2.53 billion → nearly NT$7 billion; Pillar 2 (strengthening the R&D environment, recruiting talent) NT$3.93 billion → about NT$5 billion; Pillar 3 (accelerating technology integration) NT$1.79 billion → NT$1.84 billion; Pillar 4 (attracting international startup investment) NT$3.51 billion → NT$4.52 billion. Industry contributes about NT$5.94 billion, academia about NT$3.4 billion, and the research sector (the "national team") about NT$20.7 billion.
- The single largest project: compute infrastructure is being built by the National Research Institute (國網中心 / NCHC), with a ROC year 114 budget of about NT$4.15 billion, of which NT$4 billion is dedicated to compute-infrastructure buildout. To avoid "putting all eggs in one basket," they will not procure H100s exclusively — they will gradually pair this with AMD partnerships and involve domestic vendors in the buildout.
- Admitted that Taiwan has long overemphasized hardware while neglecting software/algorithm investment; the "edge AI algorithms" area has been undervalued in the past and is one of the key directions this program aims to strengthen.
- A newly added focus for Pillar 4 (specifically requested by the new NSTC Minister): supporting digital transformation for traditional industries and small/micro enterprises — this is why this year's conference theme was set as "AI for Every Industry," echoing this direction; overseas bases (e.g., in the Czech Republic, Poland) are planned to connect with local semiconductor/AI talent resources.
- Mentioned during Q&A: the average per-project funding for the information and social computing discipline within the NSTC's Department of Engineering and Technologies is relatively low (about 50% less than the Department of Natural Sciences); Liao responded that compute should be regarded as a "consumable" (just as labs once used chemical consumables), and called for this to be included in program-report discussions, while admitting he "may not be the one who gets to decide."
Tech, products & figures
- CbI Program (Chip-based Innovation for Taiwan)
- A cross-ministry program led by the NSTC, planned at NT$300 billion over ten years, launched in 2024 (ROC year 113).
- NCHC compute-infrastructure buildout project
- Of the roughly NT$4.15 billion ROC-114 budget, NT$4 billion is earmarked for procuring/building GPU compute (planning a mixed procurement of H100s alongside partnerships with vendors such as AMD).
Notable quotes
"In our field, what used to get 'consumed' was... the liver — pushing yourself until your liver gave out; it's different now, because now what we need is compute."
Q&A
- Q: Is average per-project funding for the information and social computing discipline within the NSTC's Department of Engineering and Technologies too low, causing professors to give up applying for compute-related projects? (Asked by the moderator, relaying a morning discussion with two professors)
- A: Liao agreed this is a problem and proposed reflecting the "compute should be treated as a consumable" framework upward, but admitted he "may not be the one who gets to decide" and can only help relay it.
Fact-check notes
- Confirmed via WebSearch that Liao Hong-Yuan currently serves as Director of Academia Sinica's Institute of Information Science, is an IEEE Fellow and TECO Award laureate, and his team is internationally known for developing the YOLOv4/YOLOR/YOLOv7 object-detection systems.Sources：中研院資訊所首頁未來城市＠天下專訪
- The CbI Promotion Office's organizational structure was checked against a slide photo provided via audience crowd-notes: the names "Lin Hsuan-Tien" (Chief Strategy Officer) and "Huang Chun-Lang" (Chief Technology Officer) matched correctly; however, a name that also appeared, "Professor Chiu Wei-Cheng," could not be matched to any known role in the CbI program and is marked as unverified for now — readers should defer to official announcements on the CbI Program's website.Sources：cbi.nstc.gov.tw
- The point about NCHC handling compute buildout for the CbI Program in ROC year 114 can be cross-checked against another session at this year's conference, where a speaker from NCHC discusses the same national-level compute-infrastructure project from the implementation side.
4AI is revolutionizing knowledge management!Yang Li-Wei (楊立偉), Managing Director, eLAND Information; Assistant Professor, Department of Business Administration / Information Management, National Taiwan University
Generative AI + RAG lets enterprise knowledge management achieve, for the first time, a one-stop solution that is "knowledge in place, traceable, and model-agnostic."
Key points
- Company background: eLAND Information (意藍資訊) has done public-opinion/sentiment analysis for many years, serving over 80% of the top-ten brands in domestic government, finance, and retail sectors; it deploys millions of crawler bots daily, collecting 6–7 billion characters per day just in Taiwan — if all of that were fed directly to a GPT-class model, the estimated cost would be NT$3 million per day, which is why they had to develop their own model-efficiency technology.
- Four traditional bottlenecks in knowledge management: organizing knowledge is too labor-intensive to centralize; employees don't know what's in the system and so can't use it; knowledge sharing requires heavy manual reading and organizing; and almost no tools can "create new knowledge" from existing knowledge.
- The breakthrough of generative AI: it can process knowledge in place (Word/PDF/NAS/OneDrive, no need to move files), supports multiple sources and formats, and lets users get summaries with cited sources just by asking questions in natural language — addressing enterprises' three biggest concerns: explainability, verifiability, and internal auditability.
- Explained why RAG (retrieval-augmented generation) beats feeding an entire document to an LLM: the longer the input, the more likely an LLM is to forget the middle content (in a real test, feeding a 271-page, 140,000-character financial report into the model once cost about NT$100). Instead, first use traditional NLP methods to retrieve the 5 most relevant passages (about 1,500 characters), then hand those to the LLM to answer — faster, cheaper, and more accurate.
- Cited Google NotebookLM as an example of a personal-knowledge-management RAG product: a single notebook can hold up to 50 uploaded files, multiple notebooks can be created, and it features hierarchical summaries, auto-generated guiding questions, and source-cited answers.
- Why most enterprises no longer fine-tune: model versions update too fast (example: a client had just finished fine-tuning LLaMA when LLaMA 2 came out, followed by 3, 3.1, and 3.2 in succession); combined with the fact that enterprise data changes constantly and fine-tuning can't keep pace, the trend has shifted to a swappable, model-agnostic architecture paired with RAG.
- RAG isn't a cure-all either: answer scope must be constrained to avoid wild answers; if the retrieved passages themselves are incomplete, the answer will be "incomplete" rather than "wrong" (e.g., it should list ten methods but only gives seven or eight); enterprises also need cross-department "production-line control" (to avoid mixing up data and models).
- Three real-world cases:
- National disaster-prevention agency (the unit that sends earthquake alert texts): multi-stage question decomposition + evidence-based governance + real-time external data calls (e.g., reservoir water levels, rainfall observation stations); can answer questions about typhoon intensity, affected range, and currently hardest-hit areas, and even give response recommendations by analogy to historical typhoons.
- A large financial holding company: combines internal data with external data (market sentiment), restricting retrieval scope by role/permission (e.g., the securities department and futures department get different data for the same question), and uses function calling to connect to macro databases like Bloomberg and Reuters for advanced predictive analysis (e.g., impact of Fed rate-cut expectations, a stock's one-week price movement); completes in 15 minutes advanced analysis that fine-tuning previously couldn't achieve.
- A construction company: regulatory compliance checks (e.g., legal building materials for beams and columns), requiring the system to have "time awareness" to select the currently applicable version of regulations; also used for automatic translation of foreign-language reference materials and staff training.
- Cited a Harvard Business School and Boston Consulting Group (BCG) study of over 1,000 consultants: for knowledge workers using generative AI, 12.2% of daily work could be fully completed by AI; for the remaining tasks, AI sped things up by a quarter, with 40% of that work seeing quality improvements; lower-performing consultants saw a 43% performance boost after adoption, while originally top-performing consultants gained only 17% (diminishing marginal returns).
- Cited the well-known online case of "how many R's are in strawberry" (on Reddit, most large models other than o1 incorrectly answered "two"), illustrating that LLMs are "bad at arithmetic but good at language"; suggested having the LLM write code to perform the calculation and then correct its own answer is a more effective "leverage" strategy.
Tech, products & figures
- OpView (eLAND Information's official public-opinion monitoring product) / Typhoon Search Technology (龍捲風科技, a search-engine company eLAND acquired in 2022)
- Google NotebookLM
- A personal-knowledge-management RAG tool demonstrating hierarchical summaries and guided questions.
- Function Calling
- Used to connect to external databases like Bloomberg and Reuters for financial predictive analysis.
- RPA (Robotic Process Automation)
- Connects to LLM judgments to execute automated workflows (e.g., internal clearing and settlement at the stock exchange).
Notable quotes
"This tool isn't perfect, but it's like humans: I can't run fast, but I can get in a car that lets me move very fast."
"You don't need to dump the whole document in — first use traditional NLP methods to find the most relevant passages, then let the LLM answer. It's fast, low-cost, and accurate."
Q&A
- Moderator's follow-up remark: future enterprise management needs to adopt a dual-track mindset of "managing employees + AI," considering how to improve the performance of AI "proxies" and how to accomplish organizational goals.
- No independent Q&A session.
Fact-check notes
- Confirmed via WebSearch that Yang Li-Wei is the founder and Managing Director of eLAND Information (意藍資訊, stock code 6925), which acquired Typhoon Search Technology (龍捲風科技) in 2022 and listed on the Emerging Stock Board; he also teaches in NTU's Department of Business Administration/Information Management.Sources：維基百科經理人雜誌
- Yang Li-Wei's view that "enterprise knowledge management should shift to RAG + swappable models" contrasts with this session's Talk 1 (Tsai Tzong-Han), who argues "training a small model beats a general large model + RAG"; yet another speaker, in a temple-AI case study elsewhere at this year's conference, argues full-parameter fine-tuning beats RAG/LoRA — making this a three-way disagreement on the topic.
5Chen Sheng-Wei Memorial Lecture: Delivering Trust in the AI EraChen Wei-Chao (陳維超), Chief Digital Officer & Senior Vice President, Inventec Corporation; Founder & Chairman, Skywatch; Visiting Professor, Department of Computer Science and Information Engineering, National Taiwan University
Uses real-world engineering-specification failures — from aircraft to telescopes to chip design — to argue that AI needs an engineering mindset to examine five dimensions of trust: privacy, robustness, fairness, transparency, and social responsibility.
Key points
- Background: Chen Sheng-Wei (陳昇瑋, 1976–2020) was a well-known Taiwanese data scientist and AI expert, holding bachelor's and master's degrees from National Tsing Hua University and a PhD from National Taiwan University. He served as a Research Fellow at Academia Sinica's Institute of Information Science, Chief Technology Officer of E.SUN Financial Holding and E.SUN Bank, Chairman of the Taiwan AI Academy Foundation (財團法人人工智慧科技基金會), and President of the Taiwan Data Science Association. He founded the "Taiwan Data Science Conference" (the predecessor of this conference) in 2014, established the Taiwan Data Science Association in 2016, and founded the Taiwan AI Academy in 2018, serving as its CEO — training over 6,000 AI professionals for Taiwan within two years. On March 29, 2020, he suffered a head injury from a fall while exercising; a brain hemorrhage on the way home left him comatose, and he passed away on April 11, at age 44. Since then, the Taiwan AI Academy has held a "Chen Sheng-Wei Memorial Lecture" at every annual conference, inviting his close friends, mentors, and scholars who helped the Taiwan AI Academy to share their latest views on AI, in memory of his contributions to advancing AI education and data science literacy in Taiwan. This talk's speaker, Chen Wei-Chao (陳維超), was a close friend of Chen Sheng-Wei during his lifetime, and is also a visiting professor in NTU's Department of Computer Science and Information Engineering and Chief AI Advisor at Inventec; he previously worked as a systems architect at NVIDIA and a senior research scientist at Nokia Research Palo Alto; he holds a 1996 master's degree in electrical engineering from NTU and a 2002 PhD in computer science from UNC Chapel Hill, with research interests including GPU graphics hardware, computational photography, augmented reality, and computer vision. He opened his talk with personal photos and anecdotes remembering Chen Sheng-Wei (playing cards together, a photo from cycling around Taiwan, discussions about the founding idea behind the AI Academy), joking self-deprecatingly, "What kind of cram school is this — I'm an engineer, not an educator," leading into the talk's central metaphor: AI is engineering (engineered intelligence), and its trustworthiness must be examined with the mindset of engineering specifications and testing.
- Opened by quoting computer-science pioneer Fred Brooks (author of The Mythical Man-Month, who passed away in 2022) and his famous line, "The scientist studies in order to build; the engineer builds in order to study," offering a reverse reading of it — scientists build things in order to learn, engineers learn in order to build things; also cited his well-known laws: adding manpower to a late project only makes it later, teams should stay small (e.g., Amazon's "two-pizza team"), and the "Second-System Effect" (after a first system succeeds, a second system often fails because its creators get fixated on fixing bugs and forget the original intent).
- Used the Boeing 737 MAX crash case to illustrate flawed engineering specifications: a faulty nose sensor sent an erroneous dive signal, and the flight-control system, following "correct logic," kept pushing the nose down, ultimately causing the crash — a correct engineering process executing a wrong specification still leads to disaster.
- Used the case of the Hubble Space Telescope's "nearsightedness" (the lens-grinding company made the refractive-index specification extremely precise, but the spec's underlying math was wrong), and a graphics-card project he himself had worked on (engineers etched a second-order Taylor expansion into the chip but forgot to reduce the power term, causing the company's stock to plunge from NT$70–80 to NT$7–8, wiping out 90% of its value), to show that a "technically correct spec" is not the same as a "correct spec from the user's standpoint."
- Core argument: AI is "a code that writes itself" (the model's weights are the code); the traditional engineering mindset of "spec → implementation → test" still applies, except AI uses data as the specification and test data for validation — and "95% accuracy counts as passing," a standard that would be unacceptable in traditional engineering.
- Analyzed why AI is harder to trust than traditional engineering: data volume has exploded but quality is hard to guarantee (labeling errors, or even deliberate contamination by competitors); data distribution can be skewed (e.g., because different social-media platforms use different algorithms, the same news event can present completely different "facts" on different platforms; in a cat/dog classification dataset with too few cat photos, the model may misclassify all cats as dogs while overall accuracy still looks high); traditional "monkey testing" and human-in-the-loop testing struggle to exhaustively cover every scenario; out-of-distribution problems (e.g., a face-recognition model trained only on Asian faces loses accuracy on white, Latino, or Black faces in the US); and when reinforcement learning uses a simple cost function to guide behavior, it easily "breaks" along other dimensions not accounted for (his own team's experience training robots for a competition: a model learned to "refuse to climb" to save energy cost, or overfit to deformation features of the training enclosure due to insufficient training variance).
- Proposed a five-pillar framework for trustworthy AI: Privacy, Robustness, Fairness, Transparency, and Social Responsibility, and shared his own lab's preliminary results in each direction (offered as a starting point for discussion):
- Privacy: uses cryptographic techniques (multi-party secure computation methods such as coin tossing) for federated learning under low-trust assumptions, without needing to fully trust that a cloud provider honestly aggregates gradients; also uses homomorphic encryption to protect model inference content, but since nonlinear operations like ReLU are extremely slow in ciphertext space, switched to multiparty computation (e.g., garbled circuits) combined with "noise obfuscation," bringing performance down to only about 10x slower than plaintext computation (versus roughly 10,000x slower with the original homomorphic-encryption scheme).
- Robustness: uses Sim2Real (training in simulation then transferring to a real robot) to make gait transitions more stable for robots; uses an "anchor/scaffold" method to first frame trustworthy data and then correct noisy data, reducing the impact of noise on training.
- Robustness + Fairness (cross-domain): uses topological manifold methods to detect out-of-distribution test data (e.g., a deep-sea new-fish-species identification case: the model should report "I haven't seen this before" rather than forcing it into a known species).
- Fairness + Transparency (cross-domain): when data is required to be deleted, a learned "summary representation" (not manually engineered, but learned by the model itself as to what information to retain) replaces the raw data for training and model reconstruction, balancing usability with de-identification (analogy: storing only the relative-position ratios of facial features rather than the raw face photo).
- Social Responsibility: uses a large language model to analyze doctor-patient interview dialogue to detect early signs of Alzheimer's disease (the English-language version is more accurate than Chinese, due to less Chinese-language data); an Apple Watch app measures a 30-second ECG to predict five-year heart-attack risk (explainable, but hasn't yet addressed the psychological-support problem of "how to reassure someone who scores 70"); and developed an extremely low-power AI chip that detects whether a person is looking at the screen to automatically turn it off and save power — arguing this too is an extension of socially responsible AI.
- Cited The Wall Street Journal's investigative series "The Facebook Files" to warn of the social consequences of algorithmic objectives: Facebook's engagement-oriented reinforcement-learning optimization target was shown to negatively affect teenage girls' mental health (self-denial) and to create echo chambers (people with differing opinions unable to hear each other) — a case where "one small reinforcement-learning objective caused political turmoil across the entire world."
- Contrasted the bias in two narratives — Silicon Valley's (technologically optimistic: AI makes the world better) and Hollywood's (dystopian: AI destroys the world) — and cited the differing views of 2018 Turing Award co-recipients Yann LeCun and Geoffrey Hinton: LeCun believes LLMs merely use human-in-the-loop feedback to keep trimming disallowed content out of a vast space of possible outputs, a process that can never be finished and lacks core values; Hinton, by contrast, believes LLMs are more than "next-word prediction" and must possess some not-yet-understood form of intelligence that still needs continued research to clarify.
- Closed with Steve Jobs's metaphor of the computer as "a bicycle for the mind," arguing that AI's purpose should be to boost productivity, create new tasks, and improve human wellbeing (healthcare, education) — echoing the founding intent of the Taiwan AI Academy: social responsibility and fairness (education bringing democratization).
Tech, products & figures
- Homomorphic Encryption, Multiparty Computation / Garbled Circuit
- Cryptographic techniques used to protect privacy in federated learning and model inference.
- Sim2Real
- A technical approach where robots are trained in simulation and then transferred to real-world environments.
- Manifold-based OOD Detection
- Uses topological methods to detect out-of-distribution test data.
- Moral Machine
- A public experiment website developed by MIT on self-driving car ethical decision-making; cited by Chen Wei-Chao as a case of AI ethical dilemmas.
- Ship of Theseus / Alan Kay's (2003 Turing Award laureate) analogy that "all the atoms in the human body are replaced every seven years"
- Used to discuss the philosophical question of whether a model, after continual learning, is "still the same model."
- Euclid's five postulates, Asimov's Three Laws of Robotics
- Used as an analogy for "deriving complex system behavior from a small set of axioms," illustrating that AI alignment rules are bound to have exceptions and are hard to enumerate exhaustively.
- The Facebook Files (The Wall Street Journal investigative series)
- Exposed the negative effects of Facebook's engagement algorithm on teenage girls' mental health and on political polarization.
Notable quotes
"AI is artificial intelligence, and it's also engineered intelligence — it's something that writes its own code. The Python or PyTorch you write isn't the real code; the real code is the weights."
"Even if you have a very simple axiom to guide it, the model is still dangerous, because it's hard to balance all the costs involved."
"This one small (Facebook) reinforcement-learning objective caused turmoil in the political landscape across the entire world."
Q&A
- No Q&A (after Chen Wei-Chao's talk, the moderator thanked him directly and transitioned to the closing remarks).
Fact-check notes
- Audience crowd-notes provided a livestream replay link (Facebook) for this talk; testing confirmed the link resolves correctly, with the page title showing "[2024 Chen Sheng-Wei Memorial Lecture] Chen Wei-Chao, Chief Digital Officer," matching this session — but due to Facebook's restrictions on automated retrieval, it could not be further confirmed whether the video still plays correctly; readers are advised to click through and check for themselves.Sources：本場直播回放連結
- Confirmed via WebSearch that Chen Wei-Chao currently serves as Chief Digital Officer and Senior Vice President at Inventec, founder and Chairman of Skywatch, and a visiting professor in NTU's Department of Computer Science and Information Engineering; his background includes serving as a systems architect at NVIDIA and a senior researcher at Nokia Research Palo Alto; he holds a 1996 master's in electrical engineering from NTU and a 2002 PhD in computer science from UNC Chapel Hill.Sources：數位時代青創學院2022 台灣人工智慧年會講者頁
- Chen Sheng-Wei's (1976–2020) biography (NTHU bachelor's/master's, NTU PhD, Research Fellow at Academia Sinica's Institute of Information Science, CTO of E.SUN Financial Holding, founder and CEO of the Taiwan AI Academy, etc.) and the date and circumstances of his passing are all verified from public reporting.Sources：數位時代遠見雜誌維基百科
- "Dr. Chen Sen-Wei" (mentioned in Talk 4 as the source of an analytical framework, referenced at the end of its discussion of a four-level questioning framework) is suspected to actually refer to "Chen Sheng-Wei" (陳昇瑋 was indeed well known in his teaching and writing for a four-level data-science analysis framework: what / why / will happen / make happen), but the exact origin of that diagram could not be fully verified — noted here for readers' attention, and not included in the body of this Memorial Lecture section.
6Closing: Empowering and Balancing Taiwan with AI — Making Talent the Engine for Expanding and Deepening ImpactKuo Ping-Chen (郭秉宸), Industry-Academia Liaison, Taiwan AI Academy
The Taiwan AI Academy has trained over 11,000 alumni in seven years, calling on industry to jointly advance data governance and the development of homegrown domain-specific models.
Key points
- The Taiwan AI Academy has been operating for 7 years since its founding (co-founded by Chen Sheng-Wei and Kung Hsiang-Tsung), accumulating over 11,000 alumni and partnership engagements with more than 2,000 companies.
- Cited Gartner's generative-AI hype-cycle perspective: the market has shifted from the expectation that "generative AI alone can create the next generation" to recognizing that generative AI needs to be paired with discriminative AI and standardized processes to actually be deployed; also cited IoT Analytics' quarterly technology-heat tracking report and a Deloitte report, showing that generative AI has become a variable enterprises must factor into workforce planning.
- Reiterated the Taiwan AI Academy's three long-standing pillars: data, algorithms, and compute, emphasizing that the most fundamental foundation underneath is "talent" — which is also the shared identity of the alumni and prospective alumni in the room.
- Discussed the layered structure of the LLM industry (citing an article by Alex Rampell / a16z): mega companies mostly occupy the foundational tooling layers (Layers 1–4), while the survival space for startups and enterprises lies mostly in the application layer and user-interface layer (Layers 5–6) — Taiwan's opportunity isn't limited to foundry-style manufacturing, but also lies in finding a position within these application layers.
- Echoed the morning session's (the AI governance panel moderated by Hou Yi-Hsiu) discussion on the urgency of data curation/data governance: it shouldn't wait solely on government action — following the model of how industry once spontaneously drove the founding of the Taiwan AI Academy, industry associations should proactively organize, process, and preserve data.
- Called for AI transformation across every industry and small/micro enterprises to require domain-specific models for each field, which would in turn drive surrounding compute-supply services to form an ecosystem.
- Announced that the Taiwan AI Academy will continue promoting AI-related competency certification, and will roll out an "engineering aggregate certification" (the term as used in the source, without further elaboration of its specific name) by the end of this year or early next year.
- Proposed a "Smart Productivity Partner Program" (also mentioned a similarly named "Feng-Huo-Lin / Feng-Huo-Ren Program," possibly corresponding to an existing industry-partnership project of the Academy, though its official name could not be verified), advocating a human-centered, human-AI collaboration model to drive enterprise partnerships, and citing a line from Sun Tzu's Art of War — "marching day and night without rest, doubling the pace" — as a metaphor for the urgency of accelerating this push.
- Cited Jensen Huang's line, "Taiwan is the cornerstone of the world; everything at NVIDIA starts from Taiwan," along with two other recently popular images (specific content unclear, so not elaborated here), calling for Taiwan to be more than just the starting point of hardware manufacturing, and to expand its AI influence internationally.
- Closed by thanking all foundation-donating companies (Formosa Plastics, Chimei, AUO, Inventec, eLAND Information, MediaTek, etc.) and conference sponsors for supporting this year's conference, which drew over a thousand attendees, and expressed hope to meet again at the 2025 Taiwan AI Academy Conference.
Tech, products & figures
- Gartner Generative AI Hype Cycle, IoT Analytics quarterly technology-heat report, Deloitte workforce report
- Industry trend reports cited by Kuo Ping-Chen (the source text did not specify each report's full name and year, so individual source links could not be verified).
- a16z (Alex Rampell) article on LLM industry-chain layering
- Used to explain the market division of labor between mega companies and startups.
Notable quotes
"If we wait for the government to do this, it probably won't happen quickly; having enthusiastic industry players start organizing, processing, and preserving data would be a good approach."
Q&A
- No Q&A (the conference ended after the closing remarks).
Fact-check notes
- Kuo Ping-Chen holds the title "Industry-Academia Liaison" (產學長) at the Taiwan AI Academy, and his title and talk topic fully match the official agenda; however, specific project names mentioned in this section, such as "engineering aggregate certification" and "Smart Productivity Partner Program," could not be individually verified for their official names and content — readers who wish to cite these should separately check the Taiwan AI Academy's official website for the latest announcements.
- Chen Sheng-Wei's background and the founding of the Taiwan AI Academy have already been verified with sources in Talk 5 above and are not repeated here.