AI for Every Industry

簡立峰 argues sovereign AI is fundamentally a 'national identity problem' — the priority is clarifying risks around national security, legal jurisdiction, health, education, and cultural rights before rushing into strategy; his root solution is legislative reform, modeled on Japan's revised copyright law that exempts AI training use from infringement. 李育杰 instead stresses that a resource-constrained country needs clear priorities — compute (build it, but don't bet everything at once) > data ('models are temporary; data is permanent') > talent; he explicitly opposes Taiwan training a frontier model from scratch, citing insufficient data and compute, and states he isn't fully convinced that Transformer-style brute-force compute scaling is the final answer. 許永真 emphasizes 'imagination' and 'collaboration' — don't let scarcity limit imagination, and don't let Taiwan's multiple LLM projects compete against each other; mid-panel, she revises her own earlier position from 'we need our own large language model' to 'we don't necessarily need to build it ourselves — we'd rather have the world help build it, but we should contribute our own data.' The three also offer three different answers to 'where should the data come from': 簡立峰 proposes legislative reform, 許永真 proposes gamified crowdsourcing with a lower barrier to contribution, and 李育杰 proposes establishing a new 'data usage right' distinct from existing intellectual property rights. The three solutions aren't mutually exclusive, but they reflect genuinely different diagnoses of the root problem — legal obstacles, social-cultural shyness, or a missing ownership framework.

蔡宗翰, convener of TAIDE's core model training and RLHF effort, argues that training a small model beats 'a general-purpose large model plus RAG' — a general model is like 'a bright new graduate' who can answer anything but isn't reliably correct; concretely, EduTAIDE (a fine-tuned model) hits a 76% pass rate versus GPT-4o's 51%, and running a general model plus RAG on LLaMA 3.1 405B requires 11 H100 GPUs, out of reach for most teams. 楊立偉 of eLand Information, speaking later in the same session, reaches the opposite conclusion — most enterprises have stopped fine-tuning because model versions iterate too fast (clients barely finish fine-tuning LLaMA before LLaMA 2, 3, 3.1, and 3.2 arrive), and enterprise data keeps changing too, so fine-tuning can't keep up; the trend is 'model-agnostic architecture plus RAG.' A third position comes from 李明達 of Temple AI: he explicitly rejects RAG ('prone to retrieval errors that introduce inaccuracy') and LoRA ('degrades the model's underlying capability'), insisting on full-parameter fine-tuning with models of 70B parameters or more. The three come from TAIDE's academic camp, an enterprise knowledge-management camp, and a vertical niche-application camp respectively — three mutually exclusive answers to the same technical question, and no one at the conference stepped in to reconcile them.

翁浩正, CEO of DEVCORE (Day 1, Deep Tech), gives a relatively conservative assessment — AI still can't deliver a 'precise, one-shot kill,' can't judge which vulnerability would land such a kill, and zero-day discovery still needs human creativity; at this stage AI is, for red teams, 'a great copilot,' not yet a substitute for precision penetration. 沈宜婷 of CHT Security (Day 2, Flash Talks, the very next day) live-demos an unrestricted, locally-deployed Ollama model automating an attack on a Linux host, successfully escalating from ordinary privileges to root password and access, then continuing to automatically mine sensitive files like /etc/passwd and shadow. The two never directly engaged each other and their scenarios differ (翁浩正 discusses the 'precise stealth' demanded by real enterprise red-team engagements, while 沈宜婷's demo targets a deliberately set-up practice environment) — but placed side by side, the same conference produced two starkly different demonstrations of whether AI can now automate breaking into a system: one says 'not yet,' the other demonstrates 'yes, right now.'

游家牧, on privacy-enhancing technology (PET), argues that techniques like differential privacy and synthetic data allow 'privacy-enhance first, then link' — releasing high-value sensitive data for model training while still protecting personal information. 邱文聰 of the Institute of Law, Academia Sinica — who opened by calling himself 'the legal scholar here to crash the party' — pushes back directly: under the Personal Data Protection Act, as long as records can still be 'linked' afterward via a hash ID or code, they legally remain personal data even after de-identification; whether the process is 'enhance-then-link' or 'link-then-enhance,' if linkage remains possible, it's still personal data in substance. He also names and criticizes the Ministry of Digital Affairs' 'data for public good' concept for blurring the line between personal and non-personal data. Differential privacy, because it involves an intermediary 'curator' role, is one of the few exceptions that could plausibly turn personal data into non-personal data — but that stops being true the moment it's opened for public query outside the curator's control. The cross-discussion segment surfaces an even sharper split: 卞中佩 personally leans toward accepting 'link-first, then fully de-identify' as acceptable for enterprise data; 邱文聰 counters with Germany as an example — 'Germany strictly prohibits cross-agency data linkage' — and questions whether Taiwan's 'immediately anonymize after linking, but it can still be linked later' approach is even legal, saying 'whether enhance-after-link is lawful remains disputed.' What this clash means: most sessions at this conference assumed that technology (PET/differential privacy) can solve the challenge of opening up data, but 邱文聰's session explicitly points out that technical solutions can't route around the legal definition itself — the two camps give different answers to whether data can be opened up at all.