Themes across the tracks

1 · Agentic AI: the real difficulty of PoC → production

The year's highest-frequency keyword, used almost entirely to describe difficulty rather than promise. That is the biggest change in mood from last year.

Hybrid, not all-AI

Three speakers from three industries reached the same conclusion — the best answer is neither 100% human nor 100% AI:

• 91App (Li Kun-mou) — four iterations ending at Rule Engine 40% + AI agent 50% + human 10%, throughput up 10×.
• Pegatron (Hsiao An-chu) — picks tasks with Stanford's WorkBank two-axis method (does the human want help × technical feasibility), doing only the top-right green zone. Results: issue-consolidation agent 97.7% accurate (only five simple tools attached), defect root-cause agent 94%.
• Lion Travel (Huang Hsin-chuan) — first tried ChatGPT directly for itineraries, found too many hallucinations and judgement inferior to a designer with 20 years of experience; now runs AI drafts → human review → second review → correction.

Architecture: Context Engineering

Chen Yi-chang of MediaTek Research borrowed Andrej Karpathy's analogy — an LLM is a new kind of operating system, the model is the CPU, the context window is RAM — and broke out four operations:

His full case was Cola Tour's itinerary planning system: one main agent orchestrating five sub-agents (flights, hotels, attractions, incidental costs, report generation), each with isolated context, with the main agent doing chain-of-thought integration at the end.

91App's AI store-manager architecture divides labour differently. The notable move is keeping the Evaluator outside the conversation flow entirely, scoring in batch after the fact and feeding a human supervisor review loop:

Auditor (gates input/output) → Operator (splits intent, dispatches) → Composer (assembles reply)
                                   ↑
                        Context Manager (system prompt, history summary)

            Evaluator (outside the flow, batch-scores after the fact)

The four-layer kitchen analogy

Tsai Cheng-lin of AIA used a kitchen to separate four words that get used interchangeably — the most practical introductory frame of the conference:

2 · Digital twins: the bridge into the physical world

Four companies made the same case from four angles — the most architecturally complete thread of the year.

Why a twin at all: the three expensives

Kenmec's Chen Chih-chieh, citing NVIDIA, summarised Physical AI's three cost barriers: data is expensive (only ~10% of collected real data is usable), training is expensive (Tesla Optimus's brute-force video training appears to have hit a wall; engineers who can train are scarce), and physical testing is expensive (~NT$4m per humanoid; training twenty could run into hundreds of millions).

Three routes to a world model

• Hsiao An-chu (Pegatron) — generative-based (NVIDIA Cosmos, Google Genie), Meta's JEPA, and Meta's recent video-plus-LLM world model.
• Wei Shih-chun (Realtek) — criticises purely text-described world models for losing critical latent-space information; prefers visual world models, citing DeepMind Genie generating interactive game frames in real time and Meta JEPA predicting a ball rolling over a bump.
• Chen Yi-ting (NYCU) and Sun Shao-hua (NTU) — question, from an academic standpoint, the generalisation of the mainstream collect lots of data and train a foundation model route.

3 · Open models and Traditional Chinese sovereign AI

Four reasons enterprises choose open

1. Cost — once amortised, self-hosting beats per-call commercial APIs, especially at volume. LINE Taiwan switched marketplace search to Breeze semantic search precisely because the query volume made API cost unbearable.
2. Security — the model is hosted in your own machine room; data does not leave the building.
3. System stability — you are not exposed to a third party's API version churn.
4. Adjustability — fine-tuning produces a differentiated experience, which becomes competitive advantage.

GRPO, explained with a study analogy

Huang Liang-hsun of Twinkle AI explained the difference between GRPO and PPO clearly:

Measured: their F1 model went from 50 to 64 on MMLU after GRPO training, using only two reward functions (a format check and a length check). Before training the model would loop in self-reflection without answering; after, its reasoning converged. The limitation: if the base model never saw the knowledge during pre-training, GRPO cannot train it in either.

4 · Governance, security and trustworthiness

Defence against the dark arts

Kung Hua-chung used Harry Potter to unpack generative AI's attack surface — the most structurally complete security briefing of the conference. NICS grouped roughly ten jailbreak techniques into three strategies, with measured effectiveness:

Four classes of systems-integration risk: cloud versus on-prem deployment; RAG (leakage and the hard problem of permission tiering); MCP (tool misuse and tool poisoning, in both directions); and agentic AI. His definition of the last has a certain sting to it:

A good AI agent can set its own goals, plan and execute, interact with the outside world, and keep learning and adjusting — which matches Skynet's capabilities exactly.Kung Hua-chung · NICS

• Arup's Hong Kong office lost about US$25m to a deepfake video-conference scam in 2024.
• Samsung employees leaked confidential material through ChatGPT in 2023.
• AI has cut ransomware authoring from about a week to roughly 20 minutes, and vulnerability discovery from days to about an hour.
• Recommended defence: something like Meta's Llama Firewall (Prompt Guard, Llama Guard, MCP/plugin control). NICS offers an AI evaluation service — currently free — with test items designed for Taiwanese contexts.

Exploration first versus explainability first

• Ed H. Chi argues exploratory first — let's find out what the possibilities are — understand the capability boundary before handling safety. Google's internal practice is to open access widely to employees for red teaming. His reasoning: the cure for anxiety is not abstinence, it is knowing where the possibilities lie.
• Chiu Wei-chen supplies the other half: beyond exploration, critical domains need exploitation — understanding why the AI decided what it did and what follows (explainability and accountability). He holds that law is society's last line of order, and the pragmatic move now is bottom-line legislation for high-risk domains with soft guidance elsewhere.
• Chiu also proposed a concrete governance idea: if a single government portal has AI agents executing cross-agency queries automatically, each agent needs its own identity and access entry point so responsibility can be traced.

5 · Drones and embodied intelligence

Three talks assemble a complete industry picture

How bad is no GPS? Cheng Teng-hu: like finding the exit with your eyes closed. On IMU integration alone, error compounds exponentially to hundreds of metres within about 40 seconds; pure visual relocalisation is off by around 20 metres.

• A contrast case: US police drones since 2014 average 71 seconds to scene; Phoenix PD flies over 10,000 missions a year — drone as first responder is now a mature commercial model. The FAA spent 3–4 years negotiating route approval down from 6–12 months to as little as 6 hours.
• A live battlefield case: Ukraine's Operation Spiderweb (1 June 2025, 18 months in planning) crippled Russian strategic bomber capacity with small drones at roughly 10,000× leverage. Counts differ by source — Ukraine says 117 drones, The Economist estimates about 150; Russian losses are estimated at about US$7bn.
• An accidental finding: Cheng Teng-hu's team misclassified an eagle as a drone in field testing (birds were not in the training data), which inspired a commercial idea for airport bird dispersal.

Embodied intelligence: three speakers, three objections

Sun Shao-hua also ran an interpretability experiment: undergraduates at USC (paid US$10/hour) were asked to debug the generated program, and needed to change only 3 to 5 tokens to improve performance substantially. Neural-network weights, by contrast, are essentially impossible to hand-tune.

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.Wei Shih-chun · Realtek

He half-jokingly floated a System 0: let the AI enter a forever loop when it meets a problem it cannot solve — the machine equivalent of getting stuck in your own head — noting that smaller language models do in fact fall into repetition loops in front of hard problems.

6 · Sports science: from scouting to movement training

The whole of R0's second afternoon — the track furthest from the AI industry, and the one with the most solid methodology.

A counter-intuitive conclusion: scouting matters less than you think

Scouting is a multiplier on the result — roughly ×0.9 to ×1.1. What actually decides performance is the athlete's underlying baseline score.Huang Chih-hao · National Taiwan University of Sport

Which is why he moved his research focus from scouting to raising that baseline through movement training.

lab ready ≠ game ready

• Huang Chih-hao — a lot of sports-science demos look beautiful, but lab ready to game ready is a very long road. Rain, light, protective gear and jersey colour all interfere. An electronic strike zone took 1–2 years to reach 90% accuracy; 90% to 99% may take three more.
• Hu Min-chun (NTHU) — the NBA has had multiple ceiling-mounted cameras since 2012, but Taiwanese teams mostly rent venues (even the floor logo has to be installed and removed, around NT$200,000 per game), so that setup cannot be copied. Her answer is to allow a single movable phone or camera and still do real-time fine-grained action classification.
• Yi Chih-wei (NYCU) — TrackNet proved feasibility on tennis first (bright colour, slower) before taking on badminton (small object, extremely fast, few visual features).

7 · Talent and getting the organisation to move

Lin Yen-liang (Deloitte) stressed within his five factors of scaling that data and technology platforms are not the worrying part — strategy, process and people are. Common symptoms of being stuck at the pilot stage: fragmented projects, ROI that resists quantification, no governance standard, KPIs disconnected from business goals, on-prem architecture that is hard to replicate, and a culture that has not caught up.

Top-down is the common answer

Lion Travel's mandatory-Canva case is worth copying: all 2,000-plus employees (only the chairman and president exempt) had to learn Canva, driven by 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.

Learn AI by playing with agriculture

The Chen Sheng-wei Memorial Lecture, given by Huang Neng-fu (President of STUST, and Chen Sheng-wei's master's supervisor at NTHU), had the highest case density of the conference — a dozen agricultural and fisheries deployments: cucumber-harvesting robots, coffee-picking AR glasses, banana grading, shrimp-larva counting, dengue egg monitoring, an AI pineapple tapping stick, an AI clam grader, and solar AI drying equipment (a student team took the 2025 Hult Prize Taiwan championship with it).

• AI pineapple tapping stick — a red/amber/green light within one second of the tap, because farmers said they cannot wait three seconds.
• AI clam grader — currently 600 clams per minute per line against 2,800 by hand, so the plan is to expand to ten lines. The farmer's own words: if you're not better than me, don't bother coming.
• AI Maker — a no-code platform running collect data → label → train → download and deploy → run → evaluate → iterate, with no programming at any step; he is also convening a distributed compute-sharing alliance across universities so high schools and vocational schools without compute can take part.

Three gaps in education and talent

1. People with AI or security expertise do not want to join individual industries, and people already in those industries lack AI and security skills — the two pools do not mix.
2. Academia habitually simplifies problems into familiar algorithmic forms, drifting away from practice.
3. Government once opened about 80 security-specialist teaching posts with salary top-ups to help universities recruit from industry — actual take-up was only about 25%. The main cause was that some departments applied even higher standards to applicants, screening out exactly the practitioners with potential.

Chiu Wei-chen called on academia to be more open-minded about candidates without a PhD but with strong hands-on ability, praising NYCU's recruitment of Jserv (Huang Ching-chun).

8 · Six deployment cases, one shared pattern

Drawn from the R2 track on day one (SMEs, textiles, retail, marketing, travel), but it holds across the whole conference:

1. Data first, then AI — everything that produced fast results was built on years of accumulated structured or semi-structured data (IoT, call-centre recordings, transaction logs). AI squeezes value out of existing data; it does not conjure it.
2. Start from the sharpest small pain point and move in short steps — not one success case began with a large all-encompassing system.
3. Human-machine collaboration, not human replacement — no company claimed full replacement; the consistent framing is freeing people for higher-value work.
4. Failure rates and hallucination are normal, and rules and validation have to cover for them — agent systems need fail-safes and continuous calibration both before and after go-live.
5. Culture and top-down push are the decisive variable — the technical barrier has dropped sharply; the real bottleneck is whether the organisation will change.
6. Open-source and domestic models are becoming the pragmatic option for traditional industry and SMEs — cost, security, stability and adjustability all favour them.