Day 1 — 9 September

01 · Jonney Shih — ASUS

KEYNOTE — Taiwan's AI competitiveness: strategic position and practice

Frames Taiwan's sovereign AI strategy as a compute × model × application triangle, arguing Taiwan neither can nor should chase the US–China hyperscale platform race head-on, and should instead become a trustworthy, deployable, scenario-led national AI brand on the strength of engineering optimisation and credible neutrality.

• Three years of AI 2.0 — 2023 the year of AI shock, 2024 the year commercial value showed up, 2025 the turning point for AI sovereignty and digital governance (mature open models + AI phones and PCs shipping + geopolitical supply-chain reshuffling).
• Four world routes — the US holds model and platform hegemony; China runs closed plus open-model tracks in parallel; Europe leads with compliance; Japan plays field applications. Taiwan's position: engineering optimisation plus credible neutrality.
• Compute engineering — near 90% global share of GPU system products, integration from data centre to edge, already building and operating AI supercomputers and starting to build sovereign AI infrastructure for overseas customers.
• Model engineering — local-context optimisation and industrial fine-tuning that turn open models into controllable, trustworthy sovereign AI.
• Application deployment — a dense ISV/SI ecosystem able to land quickly in manufacturing, healthcare, finance, retail and logistics.
• Hakka sovereign model — the Hakka Affairs Council has budgeted a Hakka corpus since 2021; ASUS trained a Hakka AI model that understands and speaks the language on Taiwania 2 (NCHC's 9-PFLOPS supercomputer).
• Smart healthcare CPS — OmniCare Plus with Taipei Veterans General Hospital and NTU Hsinchu Hospital: a 100% SaaS community-care prototype combining Zenbo robots with generative AI, including a mixed Mandarin/English/Taiwanese home voice assistant.
• Taiwan AI Cloud (an ASUS-affiliated company, unrelated to TSMC) runs an AI Foundry — GPU VM and bare-metal model fabrication, with the oneAI platform for training and fine-tuning; the core competency is optimising MFU (Model Flops Utilization) to each customer's workload.

Taiwan is not trying to go head to head with the US, China, Europe or Japan on hyperscale platform competition. It is going to build a trustworthy, deployable, scenario-led national AI brand on the compute–model–application triangle.

02 · Chen Ling-jyh — AIA

Building the AI ecosystem: AIA's year and what comes next

A review of AIA's year in policy advocacy, certification and international links, closing with the formal announcement of three regional talent-certification centres in the north, centre and south.

• The keyword for 2025 is tired — LLM release cadence went from a model every few months to one every month, while capability rose and model size fell, pushing AI onto edge devices.
• Model cost keeps falling — the newest, smartest models now cost less than first-generation ChatGPT did in 2022–2023, so there is no excuse left for refusing to use AI.
• DeepSeek lit up the global open-model conversation over Lunar New Year, proving open models are not less intelligent than closed ones.
• Alumni passed 13,000, with an emphasis on quality over quantity and more deeply customised corporate programmes (understand the company culture first, then build the curriculum).
• Principal H. T. Kung ran a National AI Policy Dialogue workshop over four consecutive weeks in May–June, each session assigning a dozen or twenty papers — sometimes whole books — to professors and senior officials.
• Two forums this year: a spring forum on DeepSeek and open models, and a summer forum held in Tainan for the first time, answering southern industrial clusters' demand. Both sold out.
• Announced: three regional certification centres with National Chengchi University (Taipei–Keelung), Feng Chia University (central Taiwan) and Southern Taiwan University of Science and Technology (Yunlin–Chiayi–Tainan), so candidates no longer have to travel to Taipei to sit the exam.

As long as you want to use AI, you can have your own AI right beside you.

03 · Hou Yi-hsiu — AIA

International collaboration and the responsible-AI curriculum

Announces on stage that she stepped down as AIA Secretary-General on 1 September to take a government post — confirmed during the later panel as Deputy Minister of Digital Affairs — and reports on AIA's international work, including selection as Meta's strategic partner for the AI Opportunity Fund in Asia-Pacific and a jointly developed responsible-AI literacy curriculum.

• The foundation's origin: the earliest related email was sent by Dr Chen Sheng-wei on 21 August 2017, drafting the mission and articles. Eight years on, it has accumulated 400–500 instructors and 13,000 alumni.
• Meta AI Opportunity Fund — 8 countries, 10 strategic partners and 49 local training partners across Asia-Pacific. AIA is Taiwan's sole strategic partner, working with OneForty, the Frontier Foundation and National University of Tainan, targeting more than 15,000 beneficiaries.
• The programme extends AIA's audience from research-industry linkage to migrant workers, southern and central labour, university students, educators, NGOs/NPOs and SME workers — positioned as AI for all.
• The responsible AI use and development curriculum is developed agile-style and is already on version 9, each version reviewed by the community and experts. It ships as slides, teaching handbook and scenario quizzes, and includes Taiwanese cases.
• Three graded versions, 14 partner training institutions, targeting 20 million reach — she estimates around 1% of Taiwan's population.
• Her argument: whatever legal framework a country builds (EU, US, China, Japan, Korea), without baseline literacy in responsible use and development, the statutes are hard to implement.

I'm worried you work too hard.Meta's feedback on the AIA team's workload

04 · Panel — AI policy: Taiwan's choices

Three speakers representing government, diplomacy/think-tank and industry debate how to set AI policy priorities under scarce resources.

Su Chun-jung — the government view

• In July the President ordered a civil-service training centre whose core task is driving AI adoption.
• Adoption in the public sector should put down roots, not just break ground. Named concrete targets: minutes generators, a common smart-customer-service system (led by the Ministry of Digital Affairs) and official-document drafting aids — stressing that agencies must not each rebuild the same thing.
• On tax and subsidies: benefits should be granted proactively by the state on the basis of identity, so citizens need not apply and chase verification across agencies. His analogy — trimming ear hair — process optimisation should be something the public never feels.
• Has promoted AI applications in 22 counties and cities, and made whether AI innovation is used a performance indicator for public agencies.

Chen Cheng-jan — the international / industry view

• Only three stock markets have visibly benefited from AI: the US, (parts of) China, and Taiwan. Taiwan's market roughly tripled over several years to about ninth globally by market cap — while GDP ranks only about 20th–21st. He calls the gap very unusual.
• Quotes Eric Schmidt's 2024 Stanford remark that Taiwan is a fantastic country, whose software is terrible while its hardware is amazing — and relays his private view that even if every Taiwanese AI and software engineer went to TSMC it would not be enough, because Intel, Samsung and Huawei are all strong at AI.
• The invisible AI thesis: the world's most influential AI models are concentrated in nine giants across the US and China. The seven largest US tech firms are together worth more than China's annual GDP and make up 25–30% of the S&P 500. He proposes a new G11 — two countries plus seven US and two Chinese companies deciding the direction of global AI.
• Nvidia alone exceeds the UK's annual GDP. Companies can redeploy market capitalisation 10–100× faster than governments can move GDP or defence budgets.
• TSMC, MediaTek and Foxconn together account for over 90% of Taiwan's investable R&D spending — an economy unusually concentrated in a few firms, which policy must account for.
• Three opportunities: a once-in-a-generation window to automate as 600–700 TSMC suppliers relocate; a supply-chain reshuffle handing Taiwan's ICT sector the key components of robots and drones; and an ageing-society golden decade using robots and foreign labour, learning from Japan and Germany.
• Three challenges: a resource-usage gap (mostly domestic-demand SMEs for whom a 5–10% productivity gain is imperceptible), under-optimised Traditional Chinese models, and a structural shortage of 25–35-year-old AI talent.

Jan Li-feng — three policy proposals

1. Government should become the country's most fluent AI user — set up a cross-agency AI adoption committee. Two concrete levers: evidence-based administrative decisions assisted by AI (he asked whether the meteorological bureau's new supercomputer and micro-climate stations were actually feeding disaster-resource dispatch and automated insurance settlement during the recent southern storms), and fully one-stop government services, which forces cross-agency regulatory alignment.
2. Build a Taiwanese data-trust regime — comparing the UK Open Data Institute's bottom-up model with Japan's legislated trust-bank top-down model, and arguing Taiwan can find a third way. Government-funded research should be required to submit raw data (open scientific data); self-driving traffic data, for example, also serves urban planning.
3. Universal AI tokens — if AI tokens become a kind of currency, government could hand them out in phases (schoolchildren first, then citizens and SMEs), cultivating an intermediary platform startup ecosystem and logging usage to refine policy. Citing Andrew Ng: people who use AI will replace people who don't.

Taiwan is the only one of these 198 countries with an invisible AI advantage — and that is hardware.Chen Cheng-jan

05 · Lu Chia-cheng — Cisco Taiwan

Inner and outer strength: mastering enterprise digital kung fu

A martial-arts framing — inner strength is AI compute infrastructure, outer strength is application governance and security resilience — arguing that scaling agentic AI requires an AI-ready data centre and dynamic defence first.

• Every successful technology wave (internet, mobile, cloud, AI) has to deliver five things: lower cost, higher efficiency, less inventory, better user experience, new business models. AI already clears at least three.
• Evident AI Banking Index (AI capability across the world's 200 largest financial groups): JP Morgan first, Morgan Stanley second, with double the gap in score; DBS is the only Asian entrant.
• Citigroup's 2024 annual report: across four business lines and seven business types, 28 are already served entirely by AI engines, moving toward agentic AI.
• Three stages of enterprise adoption: cloud PoC → shift on-prem as token costs rise → at 250 inference requests per minute, on-prem build cost breaks even against cloud token cost in about 1.5–2 years.
• Internal Cisco figures: AI inference traffic grows about 40% monthly (46× annually); each agentic engine averages 8–14 repeat confirmation calls to the language model.
• Five data-centre pain points: compute cost, power shortage, network management as compute goes distributed (rented compute centres, telco colo), security, and data as AI's fuel.
• Security governance must evolve from pre-defined segmentation to dynamic segmentation to behaviour-based segmentation — isolate the affected applications the moment a model is poisoned.
• Cisco redesigned its AI server mechanicals for full air cooling, so enterprises are not forced to rebuild for liquid cooling; and partnered with AMD on DPU-based dynamic behaviour micro-segmentation (also deployable on NVIDIA BlueField DPUs).
• Agentic Ops, announced June 2025: turn every Cisco software function into an AI engine, giving each operator a personalised dashboard with multi-person collaboration.

AI is no longer a nice-to-have option. This shift is basically unavoidable now.

06 · Frank Grunert — Siemens Taiwan

Accelerating Taiwanese smart manufacturing with digital twins and industrial AI

Real cases from Heineken's breweries, Siemens' net-zero campus in Zug, a US grid operator, and Merck / PlayNitride / Hota in Taiwan, showing that digital twin plus industrial AI cuts energy and emissions, shortens commissioning and lifts yield — with a call for Taiwan to move faster.

• Emissions urgency: manufacturing is 30% of global CO2, buildings 37% (humans spend 90% of their time indoors), energy 75%. Taiwan's three largest sources — energy, residential/commercial, manufacturing — together make up 84% of national emissions.
• Research suggests digital twin plus industrial AI could cut 12 billion tonnes of CO2 globally — over 20% of the 57 billion tonnes CO2-equivalent total.
• Heineken — a full production-line twin revealed 70% of energy use concentrated in specific process stages; optimisation cut plant energy 20% and CO2 50%, now rolled out to 15+ sites worldwide.
• Siemens Zug campus — a net-zero building designed and simulated entirely as a twin (HVAC, heating, airflow) before ground was broken; houses 1,700 people with zero operational carbon.
• AEP (largest US transmission operator) — a twin broke down data silos between planning, operations and maintenance, strengthening grid resilience and renewable/storage integration.
• Siemens' AI history is long: a ChatGPT-like natural-language Q&A prototype in the 1970s (hardware of the day killed it), image recognition in mail sorters in the 1980s; today 1,800 of its 42,000 patents are AI-related, with three AI labs and 1,500 AI specialists.
• Spent US$15bn in 2025 acquiring Altair and Dotmatics to strengthen materials-mechanics and dynamics databases and raise twin simulation fidelity.
• Taiwan cases: Merck's new Kaohsiung plant cut dry-run time by up to 50%; PlayNitride (Micro LED) adopted Simcenter/Simatic simulators for yield and early anomaly detection; Hota Industrial and Goodway Machinery in Taichung deployed adaptive control that adjusts tool parameters from live sensor data.
• Grid case: a German customer's Dynamic Line Rating uses machine learning to set line current limits from live weather instead of a conservative annual-maximum-temperature constant — 20% more throughput on the same line with no extra hardware.

Beyond the hype, we need to ask ourselves which problems do we actually want to solve.

We need to become more faster. We are maybe not fast enough.

07 · Lin Yen-liang — Deloitte

Agentic AI driving enterprises toward scalable AI

Uses the history of industrial revolutions to argue that adoption cannot stop at the high-value pilot, and that the real bottleneck is strategic consensus, process mapping and workforce transition rather than data or platforms.

• The first three industrial revolutions (steam, electricity, computerisation/automation) each took 70–150 years to mature. Industry 4.0 has existed since 2011 — about 15 years — and AI itself only a decade or so. It is early.
• Technology serves the business: what matters is not which model or GPU, but the operating process and management logic. Map and tidy the process before adopting AI — buy a garment and you either alter it or build the body for it.
• Deloitte's quarterly global survey: enterprises expect efficiency and productivity gains, but early costs actually rise (more compute platform, more senior talent). It does not save money immediately.
• WEF report: new job functions will appear fast; recruiting will increasingly target AI-enabled employees.
• Data is the biggest obstacle to scaling — especially in data-sensitive settings, moving from narrowly authorised access to a company-wide usable data-protection mechanism.
• Boards increasingly care about AI (good) but generally lack AI expertise, communication training and metrics — brief the board early.
• AI governance still uses the three-lines-of-defence structure; it does not become four lines. What is genuinely scarce is people who can govern agentic AI.
• Deloitte's April 2025 talent report: 80% of companies want agility, 80% of employees want stability.
• Case: Deloitte US analysed GenAI's impact on education, manufacturing and legal services for the US Department of Labor. Teachers now use GenAI to write syllabi rather than merely banning students; manufacturing human-machine gains are modest but risk exposure drops (with drones and robot dogs); legal services cut search and retrieval headcount but still need professionals for final analysis.

This time, with AI, you really must not miss it — this is basically the single most critical point for enterprise development over the next fifty years.

80% of companies want to be agile and fast, and 80% of employees want stability.

08 · Kung Hua-chung — NICS

Shaking hands with the unknown: building trustworthy AI in the generative era

Uses Hogwarts' Defence Against the Dark Arts as a frame for generative AI security, unpacking concrete escape techniques and countermeasures across three phases: training, use and systems integration.

• Three reasons to worry — AI is driving new forms of crime, the models and systems themselves are risky, and progress is too fast: a three-year-old suddenly 180 cm tall.
• AI-driven crime — authoring ransomware went from about a week to an estimated 20 minutes; finding a software vulnerability from days to about an hour. He likens AI attacks to drones on the security battlefield: cheap, effective, asymmetric, mass-deployable, precise. The lowered barrier has produced hacking as a service.
• Data poisoning — not just mislabelling (a cat tagged as a dog); the dangerous case is deliberate mislabelling, such as slipping another person's face into an access-control system's allow list.
• Backdoor triggers — by analogy with the 2024 Lebanon pager attack, a model with a planted passphrase executes hidden behaviour when the trigger appears, amplified badly if agentic AI or MCP connects it to physical systems.
• Hallucination in practice — asked ChatGPT to tabulate the OWASP LLM Top 10 and only the first two entries were right; the rest stitched together different years' versions plus invented items, coherent enough to pass. Wire that into a bank's customer service quoting a wrong rate and you have legal, brand and regulatory exposure.
• Indirect prompt injection is the real risk, not the direct kind — hidden instructions inside email, web pages or documents (white text on white) that the LLM reads and obeys. Main exposure is LLMs wired into mail summarisation, document systems and customer service.
• Four classes of systems-integration risk — cloud versus on-prem, RAG (leakage and permission tiering), MCP (tool misuse and tool poisoning in both directions), and agentic AI. Give an LLM a command line and you have given it the ability to do anything.
• Guardrails — filter input and output, monitor behaviour and policy, enforce least privilege, support human-in-the-loop. He recommends starting from Meta's Llama Firewall, which integrates Prompt Guard (a small model that runs on CPU), Llama Guard (~12B) and MCP/plugin control.
• Governance principles — transparent, explainable, and stoppable: there has to be a red emergency stop button.
• NICS offers an AI evaluation service, currently free, with test items designed for Taiwanese contexts.

Jailbreak techniques, ranked by measured effectiveness

AI is like magic — after the class you go from Muggle to wizard. But the most important subject at Hogwarts is Defence Against the Dark Arts, because used carelessly, AI becomes dark magic.

09 · Wei Shih-chun — Realtek

The evolution of autonomous systems: key architecture and Taiwan's advantage

Argues Taiwan's robotics opportunity is not a smarter cloud brain (System 2) but low-power, low-latency, hardware-implementable reflex loops — System 1 — built on its semiconductor strength.

• Reframes the question — instead of can we build something as intelligent as a human, ask can we build an AI being that can survive autonomously. Not intelligence first; staying alive and completing the expected task first. That is Taiwan's opening.
• Two core technical problems, each estimated at five years — autonomous proprioception, and a generalized reflexive system.
• Evolutionary analogy — biological intelligence grows through neural connection complexity, but evolution is constrained by balance: lose balance and you go extinct. Current AI training lacks that constraint. We are transplanting carbon's evolutionary experience onto silicon and should borrow its balance mechanisms, not ignore them.
• The transformer as a high-dimensional manifold — an LLM projects poetry, music, law and everything else into the same manifold, making it something like a multiple-personality individual. The next token depends on which forces in the manifold pull hardest on token space, so controlling that influence controls the output.
• Embodiment needs a feedback loop to stay stable (basic control theory), and today's foundation models do not have one — it has to be built separately.
• VLA history — Google RT-1 (2022) encoded actions as meaningful special tokens in a transformer; then trajectory information; then training across 22 embodiments, open-sourced; most recently NVIDIA Isaac GR00T N1, whose System 1 / System 2 split echoes Thinking, Fast and Slow. Its weakness: System 2 must fully understand context before producing a latent vector for System 1, so it reacts half a beat late — no good when a ball is flying at your head.
• Learning methods — imitation learning picks up common high-probability situations quickly but handles rare events badly; reinforcement learning handles surprise but adapts very slowly alone. Blending them is a major open problem: too much RL drags the whole system away, too little and it falls back on default reflexes.
• Proposes a System 1.0–1.x layer beneath today's robotic foundation models, doing nothing but the sensor-actuation loop, implementable in hardware or half-hardware, so a reflex like dodge the incoming ball is fast and extremely power-efficient.
• Distillation matters — start from a large cloud system, find balance and achieve the purpose there, then distil down: compressing a smart brain to something the size of a bee's (tens of millions of neurons) that still flies steadily and forages. Top-down, not bottom-up assembly.
• Two unsolved problems for humanoids — energy consumption and autonomy. Not autonomous enough and grandparents cannot use it, so volume stays low and cost stays high; add a huge brain for autonomy and power draw rises, hence Tesla Optimus demos where it thinks for 3–4 seconds before moving.

Even though ChatGPT can tell you how to ride a bicycle, it does not in fact know what riding a bicycle is.

Any high-level transformer system — System 2 — will hallucinate, because it is responding to something that has not happened yet. It is filling in the blanks by nature.

If you're worried AI robots will hurt humans, there's a problem to solve first: the robot has to be able to feel pain. AI does not know pain right now, and without pain there is no empathy.

10 · Roger Jang — E.SUN Financial

Using AI against financial fraud

How E.SUN uses AI — alert-account prediction, ATM mule-image recognition, credit-card fraud detection — to fight scams, and a look ahead to a cross-bank financial clean room for data collaboration.

• Models are retrained weekly, with champion and challenger models running in parallel for a week before switching over.
• Tooling — XGBoost, LightGBM, community detection and centrality on transaction graphs, Gemini 1.5 Pro.
• The financial clean room concept and the FinTech industry alliance (CTBC / E.SUN / KGI / Cathay) point toward cross-bank collaboration without exposing raw data.

Using AI against fraud is necessary — and it is the only means.

This is not only a race against time, it is a war of spear against shield.

11 · Chen Tun-ho — IISI

GenAI and AI agents: the arc of development, with cases

A systematic review of GenAI's development across theory, models and applications, three trends for 2023–2025 (AI agents, AI coding, workflow automation with n8n), and the company's RAG and agent deployments in government, finance and healthcare.

• Deployed public-sector and healthcare cases — the Central Weather Administration's weather Q&A service, a Public Construction Commission regulation knowledge base, semantic search over Ministry of Digital Affairs open data, smart-city traffic recognition, and a nursing handover assistant. No quantified outcomes were given.
• Tool landscape covered — word embedding, transformers, RAG, MCP, Google Agent-to-Agent, function calling, n8n, Dify, LangChain, text-to-SQL, GPTs, Ollama, Open WebUI, Perplexity, and the Taiwanese models TAIDE, MediaTek Breeze and Foxconn FoxBrain.

Talking to a model is really the same as talking to a person.

12 · Liang Wen-lung — Jorjin Technologies

The cutting-edge technology behind AI + AR glasses

Argues near-eye display is the most important display revolution in human history, unpacks three bottlenecks — optics (geometric versus waveguide), chips, and the human interface — and describes the AR supply-chain alliance with Foxconn.

• Demonstrated glasses — first-generation all-in-one at 47 grams; the demo unit at 43 grams with 720P colour display, supporting translation, navigation and spatial image recognition, using a wirelessly connected phone as the edge computer.
• Targets — light engine under 0.5cc, whole unit under 40 grams.
• Waveguide optical efficiency is under 1%, far below geometric optics — the core trade-off in choosing an optical architecture.
• Technology stack discussed — LCOS, silicon OLED, microLED, Lumus, DigiLens, Seiko, Qualcomm AR1, eye tracking, hand tracking, IMU, 6 DoF.

AR and AI are like a digital twin — two sides of one thing, impossible to separate.

Whoever solves the friendly UI/UX becomes the next Apple. Apple didn't invent touch either — it just did the UI/UX well.

13 · Hsiao An-chu — Pegatron

AI agents and digital twins driving manufacturing autonomy

From doing camera autofocus with Q-learning over a decade ago to training robots in a digital twin today, arguing that agent adoption should prioritise green-zone tasks — high willingness and technically feasible — and that only sustained infrastructure investment accumulates digital assets.

• Presents in-house data showing agent usage is highly fragmented — people try it once and it dies — and cites three papers on why: Salesforce on MCP tool count versus accuracy, Google DeepMind on RAG's dimensional limits, OpenAI on guessing beating abstaining.
• Task selection uses Stanford's WorkBank two axes — does the human want AI's help (x) against technical feasibility (y). Only the top-right green zone gets built; the top-left R&D zone is parked until the technology matures.
• Results — an issue-consolidation agent at 97.7% accuracy with only five simple tools attached, and a defect root-cause agent at 94% on an IC substrate case.
• Supervisory control — reaches back to a 1978 MIT paper on teleoperated submersibles: direct remote operation hits latency, so the human sets directional goals and the equipment handles detail, extended by analogy to the management latency of overseas plants.
• World-model routes surveyed — generative-based (NVIDIA Cosmos, Google Genie), Meta's JEPA, and Meta's recent video-plus-LLM world model.

Don't forget — in the end, it is still a human doing supervisory control.

Intelligence is the computational part of the ability to achieve goals.John McCarthy, quoted in closing

14 · Chen Chih-chieh — Kenmec

Systems integration through AI, digital twins and Physical AI

Thirteen years from 2D drawings to NVIDIA Omniverse and OpenUSD, arguing that physical correctness is what makes robot training affordable — the three expensives — and showing the German and Japanese humanoids the company now distributes.

Physical AI is born inside simulation. The virtual environment is the training ground for the Physical AI robot.

Physical AI is waiting for its software moment.

15 · Chen Sheng-hua — Delta Research Center

Enabling the AI factory with digital twins

Redefines the AI Factory paradigm shift — from smart manufacturing toward AI production intelligence — and shows how twins act as an accelerator across Smart Design Copilot, NPI pilot-line tuning and data-centre energy management.

• A twin plays four roles — data multiplier, algorithm sandbox, virtual-physical link, and feedback engine.
• Delta positions itself as a power-solution supplier for AI data centres, not a GPU or CPU vendor.
• In generative design the human-machine collaboration target is to first reach about 85% matching what the human imagined, and only then introduce the 10–15% of surprising, novel design.
• Techniques covered — surrogate models, NPI pilot-line generator/adjuster, black-box versus white-box tuning strategies, machine networking, and solving PDEs with deep learning (physics-informed neural networks).
• No quantified energy-saving percentage or pilot-cycle reduction was given.

Every company is an AI factory.Jensen Huang, quoted by the speaker — who noted the subtext is that every company should buy his GPUs

In mass production a digital twin can't do very much — if you want something that looks like the real thing, just look at the real thing. Where it really earns its keep is trial production.

16 · Tsai Cheng-lin — Taiwan AI Academy

Enabling SMEs: how industrial AI can ignite Taiwanese business innovation

Four analogies — brain model → assistant (Workflow) → agent → collaborating team (Agentic Workflow) — clear up the concepts in one go, followed by low-cost cases any SME can copy.

Whether to use AI is no longer a multiple-choice question. For an SME it is a question of survival.

The data inside your company is the single most important fuel there is.

17 · Lin Ming-wang — Tung Feng Fiber

Agentic AI in a textile plant

Tung Feng — a long-standing Nike and Adidas supply-chain partner — pairs nearly ten years of accumulated IoT, SCADA and big data with a textile-fluent AI startup to build an LLM agent, aiming to let a plant manager ask for last night's top three anomaly causes and get an answer within three minutes, while capturing the old masters' experience and cross-border (Taiwan / Thailand / Indonesia / Vietnam) troubleshooting knowledge.

• Plant scale — around 1,000 looms; the goal is to compress anomaly hunting from half a day (10 a.m. to noon) to under three minutes.
• Industry context — tariffs and mainland order suppression have pushed Taiwanese textiles toward small-batch, high-variety production; volume orders no longer carry the business.
• Adoption method — from 1 September all ISO forms went into the system: if Ah-wang doesn't sign, the manager doesn't sign. There was resistance at first; 3–5 months later staff refused to go back to paper.
• Stack — IoT, SCADA, big data, an on-premises LLM engine, NL2SQL.

Find the point that hurts most, and just play with it. If it works we take the next step; if it fails, well, that's the worst it gets.

If a really solid old hand in textiles can be combined with a PhD-level AI partner, how many people would the sparks from that one-plus-one knock out?

18 · Li Kun-mou — 91App

Three applications of AI agents in retail

Three directions for retail agents — internal efficiency (hire more agents), a customer-facing sales interface (build agent service), and getting ready for the agents consumers will send to buy on their behalf (ready for agent) — unpacked through two real builds: automated product listing and an AI customer-service store manager.

• The bottleneck — 15,000 fields ≈ 150,000 seconds ≈ 2,500 minutes ≈ about 11 working hours a week just filling in product listings.
• Evolution — v1 filled 98% of fields at about 50% accuracy; adding don't fill what you aren't sure of dropped coverage to 50% but lifted accuracy to 90%; requiring a written reason per field raised it further; extracting each brand's unwritten rules into a Rule Engine that runs on CPU with no AI completed it. Final mix: Rule 40% + AI 50% + human 10%, throughput up 10×.
• Architecture — Auditor gates input and output, Operator splits intent and dispatches, Composer assembles the reply, Context Manager maintains the system prompt and conversation summary. Critically, the Evaluator sits outside the conversation flow entirely, batch-scoring after the fact for human supervisor review.
• GEO (Generative Engine Optimization) is the emerging counterpart to SEO — if an AI-first world cannot find you, that is as serious as not existing in Google search.

AI did it for you, and you still have to go check the listing? Come on — I'd have laid it out faster myself.on why too-low accuracy actually reduces efficiency

We are already in an era where everyone will hire an AI assistant to do their shopping.

19 · Chen Yi-chang — MediaTek Research

Building industry assistants: AI agents on open models

Real results for the open Breeze family of Traditional Chinese models at Cola Tour and LINE, a systematic breakdown of Context Engineering for building agents on open models, and a full walkthrough of a multi-agent itinerary planner.

• Breeze ASR reached 120,000 downloads within two to three months of release.
• Cola Tour passport OCR — across 50 previously failing cases, Breeze 2 outperformed Gemini despite being far smaller.
• Cola Tour travel chatbot RAG — over 3,000 Q&A pairs, Breeze hit the 90% usable similarity bar against expert answers; Gemini actually did worse because it surfaced so much retrieved material that users were left confused.
• Four Context Engineering operations — write context (to-do lists on disk, long-term memory), select context (Agentic RAG), compress context, isolate context (sub-agents with independent windows).
• The multi-agent case — one main agent orchestrating five sub-agents (flights, hotels, attractions, incidental costs, report generation), each with isolated context, with chain-of-thought integration at the end.
• Four reasons enterprises pick open models — cost at volume, security (hosted in your own machine room), system stability (no third-party API version churn), and adjustability (fine-tune for a differentiated experience). LINE Taiwan switched marketplace search to Breeze semantic search precisely because query volume made commercial API cost unbearable.

An AI agent, put plainly, is letting the chatbot treat the act of talking as a line of thought, and then start operating things to finish more tasks.

20 · Huang Hsin-chuan — Lion Travel

Upgrading travel, innovating with AI

From August–September 2024 the chairman personally led the AI push, asking at every meeting what did you do with AI. Within a year: smart collection lockers, the AI agent Lily, Canva as a company-wide requirement, and digital humans 1.0/2.0. The core idea is to use AI to make service warmer, not to replace the warmth.

• The Canva rollout used inter-department competition (the winning team got a three-day trip to Japan) to create voluntary uptake and defuse this isn't my job resistance. The 60 existing designers moved up to overall visual planning and template iteration.
• Itineraries were first drafted with ChatGPT directly, but hallucinations were frequent and the model's judgement about experience fell short of a designer with 20 years behind them. The current flow is AI drafts → human review → second review → correction.

Technology is great, but people are greater — because people move with the times.

What my customers need is warmth. I should use AI and digital tools to make my service warmer.