2025 Trends · Taiwan AI Annual Conference 2025

1 · From adopting AI to scaling AI

The signal: across two days and forty-odd talks, almost nobody was still trying to convince the audience that AI matters.

That absence is itself the year's biggest signal. What replaced it were three kinds of language: how to drag accuracy from 50% to 90%, why the organisation won't move, and how to put the cost in terms a boss will accept.

Lin Yen-liang of Deloitte gave the most systematic account, proposing three stages of AI scaling:

High-value pilot (PoC)  →  Modularisation  →  Large-scale deployment
         ↑
  where most Taiwanese firms are stuck

He named three waves Taiwan already missed: IoT (equipment seen as too expensive), cloud (seen as too risky), and RPA (Taiwanese labour was cheaper than the subscription).

This time, with AI, you really must not miss it again.Lin Yen-liang · Deloitte

His most counter-intuitive line: data and technology platforms are not the part to worry about — strategy, process and people are. That is a clear break from the 2023–2024 orthodoxy of get your data in order first.

What corroborates the shift

Speakers this year arrived with measured outcomes rather than vision decks: E.SUN's alert-account detection cut headcount by half, 91App's product-listing throughput went up tenfold, Kenmec's ABC inventory analysis dropped from two months to two or three days, and AIA's IRB minutes fell from 72 person-hours to about two.

2 · Agentic AI enters its trough — and that's good news

Agentic was the most frequently used word of the conference, and almost every speaker used it to describe difficulty rather than promise.

Hsiao An-chu of Pegatron presented in-house data showing how fragmented agent usage actually is — people try it once and it dies — and cited three papers explaining why:

Source	Finding
Salesforce	The more MCP tools you attach, the further agent accuracy falls — as low as 33% in some scenarios, and into the teens with open models.
Google DeepMind	RAG's top-k / vector-similarity retrieval has dimensional limits; strong models still do badly on simple self-built datasets.
OpenAI	Training resembles a multiple-choice exam where guessing scores better than abstaining — one root of hallucination.

The most complete disillusionment story of the year

Li Kun-mou of 91App walked through four versions of an e-commerce product-listing agent:

Version	Approach	Result
v1	LLM reads the product master and fills the fields	98% of fields filled — half of them wrong (the AI would not leave a blank)
v2	Add a rule: don't fill what you aren't confident about	Accuracy up to 90%, but only half the fields filled
v3	Require a written reason per field, forcing self-verification	Accuracy rises again
v4	Extract each brand's unwritten rules into a Rule Engine — no AI needed	Rules 40% + AI 50% + human 10%; throughput up 10×

Why the trough is good news

Inflated expectations leaving the room is what lets real engineering knowledge accumulate. Things existed this year that did not exist in 2024: concrete failure-mode lists, reproducible remedies, and speakers willing to say in public that their project failed.

Yeh Tzu-chun of DEVI Intelligence spent an entire talk on a case that never shipped (factory anomalous-sound detection), and concluded:

An AI product needs more than product-market fit. It needs technology-product-market fit. A good model doesn't mean it maps to a suitable scenario; mapping to a scenario doesn't mean it scales; scaling doesn't mean users find it usable.Yeh Tzu-chun · DEVI Intelligence

3 · The evaluation crisis

The most original — and most underrated — signal of the year.

Wu Po-han of APMIC analysed the 34 evaluation papers cited in Gemma 3's official materials and found that 80% of the question formats were multiple choice, while commercial practice is almost entirely open-ended. Filtering for open-ended benchmarks, accuracy was generally under 80 — mostly around 70 — with only 4 of 11 open-ended evaluations passing.

You cannot weigh the depth of water on a scale.Wu Po-han · APMIC

His conclusion: enterprises and industries must build their own evaluation standards rather than trusting a vendor's general benchmark claim that this model is strong.

Pegatron's Hsiao An-chu — cited the OpenAI finding that guessing outscores abstaining, which is itself a behaviour induced by benchmark design.
Twinkle AI's Huang Liang-hsun — went the other way, building local evaluations such as TMMLU+ (including Taiwan's civil-service exam bank), because general English benchmarks cannot measure usability in Traditional Chinese settings.

4 · Parameter count stopped being the moat

Three groups of speakers made the same argument from three different positions, with complementary evidence:

Speaker	Evidence
Chen Yi-chang · MediaTek Research	Cola Tour passport OCR: across 50 previously failing cases, Breeze 2 outperformed Gemini despite being far smaller. Travel chatbot RAG over 3,000 Q&A pairs: Breeze reached the 90% usable similarity bar against expert answers.
Wu Po-han · APMIC	Their 24B model beat GPT-OSS 120B on a Taiwanese-law vertical evaluation; VRAM requirement 50GB versus 292GB.
Huang Liang-hsun · Twinkle AI	A 3B model scored above 90 on function-calling (including MCP) evaluation — higher than GPT-4o.

The industry implication: if a 24B model can beat a 120B in a vertical domain at a fifth of the VRAM, buying the biggest model is no longer the safe default. The decision moves from picking a model to picking a domain, preparing data, and fine-tuning.

Sovereign AI model (national / cultural knowledge)
        ↓
Industry knowledge model (finance, semiconductors…)
        ↓
Enterprise knowledge model (yours alone)

APMIC cited UBS 2025 research saying 86.6% of enterprises report needing fine-tuning, and NVIDIA research saying 70% of AI agents would benefit from SLM optimisation.

5 · Physical AI — from talking to doing

A new heavyweight thread this year, approached simultaneously from four entirely different industry positions.

Even though ChatGPT can tell you how to ride a bicycle, it does not in fact know what riding a bicycle is.Wei Shih-chun · Realtek

Three cost barriers

Chen Chih-chieh of Kenmec, citing NVIDIA, distilled them into a slogan — the three expensives:

Barrier	State of play
Data is expensive	Only about 10% of real-world data collected is actually usable.
Training is expensive	Engineers who know how to train are scarce and costly.
Physical testing is expensive	One humanoid robot costs roughly NT$4 million; training twenty could reach hundreds of millions.

That is why digital twins were discussed so heavily this year — not to look realistic, but to reduce all three costs at once. Four companies (Siemens, Pegatron, Kenmec, Delta) made the same case from four angles, and Delta's Chen Sheng-hua gave the most counter-intuitive conclusion:

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.Chen Sheng-hua · Delta Research Center

Academia collectively questioned the mainstream route

Speaker	Objection	Alternative
Wei Shih-chun · Realtek	High-level System 2 systems must hallucinate; a cloud brain burns too much power.	Taiwan should build System 1 — sense-actuate reflex loops that can be put into hardware.
Sun Shao-hua · NTU EE	A neural-network policy is just rote memorisation.	Program-Guided Robot Learning.
Chen Yi-ting · NYCU	The mainstream route ignores real user needs and situated safety validation.	Human-centered physical AI.

Sun Shao-hua's maze experiment was the year's cleanest counter-example. A deep-RL agent trained on 8×8 mazes solved any layout perfectly and appeared to have learned the right-hand rule — but failed completely on a larger maze. A synthesised program policy, by contrast, literally is the right-hand rule, and generalises to any size.

6 · Compute flows back on-premises

In 2023–2024 the default answer was cloud first. In 2025 that default was loosened by three independent forces at once.

Cost

Cisco's Lu Chia-cheng gave a concrete break-even: at roughly 250 inference requests per minute, on-premises build cost pays back against cloud token cost in about 1.5–2 years. He also noted Cisco's internal AI inference traffic grows about 40% monthly (46× annually), and each agentic engine issues on average 8–14 repeat confirmation calls to the language model — the agent architecture is itself a token amplifier.

Security

On the safety panel, Inventec's Chen Wei-chao argued the real corporate risk is employees pasting sensitive data into outside platforms. NICS's Kung Hua-chung laid out the full on-prem versus cloud trade-off and proposed hybrid deployment — sensitive data stays local, the rest goes to cloud — as the pragmatic answer.

Sovereignty

Of the four reasons MediaTek Research's Chen Yi-chang listed — cost, security, system stability, adjustability — system stability is the least discussed and the most practical: you are not at the mercy of a third party's API version churn.

7 · AI safety's centre of gravity moved

Safety discussion visibly stratified this year. Four speakers each named a different biggest risk — and all four were right, from where they stood.

Speaker	Vantage point	Biggest risk
Kung Hua-chung · NICS	Technical attack and defence	The model and system attack surface: prompt injection, jailbreaks, data poisoning, MCP tool poisoning.
Chen Wei-chao · Inventec	Enterprise practice	Not the model — employees pasting order prices into an outside chatbot.
Chiu Wei-chen · NYCU / NSTC	Government governance	Agentic AI losing autonomy control, plus software supply-chain risk; regulation lags far behind.
Ed H. Chi · Google DeepMind	Research fundamentals	Safety is A × B × C, not A + B + C.

Ed Chi's line is one of the most important technical insights of the conference. Large language models have compositionality: combining capabilities produces new emergent capabilities, so safety multiplies rather than adds. That means the traditional divide-and-conquer security posture — A safe + B safe + C safe = the whole thing is safe — simply fails on LLMs.

The pragmatist's answer is worth recording too. Chen Wei-chao did not propose banning ChatGPT; he described a warning trigger — not monitoring all conversation, only raising an alert when something resembling sensitive data appears to be leaving, for after-the-fact audit. That is a design that starts by admitting employees will use outside tools anyway.

The attack surface is changing too

Kung Hua-chung described a new class of scam that targets AI browsers and assistants rather than human users — already named Scamlexity. He also noted AI has cut ransomware authoring from about a week to roughly 20 minutes, and vulnerability discovery from days to about an hour. Defenders are using AI too: DARPA's AIxCC announced its finals results at DEF CON 33 on 8 August 2025, finding vulnerabilities automatically.

8 · Human-machine collaboration settled into hybrid

Across six industry deployment cases, not one company claimed it would replace its workforce with AI. This is not politeness — it is an engineering conclusion.

Company	Approach
91App	A three-layer Rule Engine + AI + human mix.
Lion Travel	AI drafts → human review → second review → correction.
Pegatron	Humans do supervisory control; AI executes the detail.
Delta	GenAI first reaches 85% roughly right to build trust, before the remaining 10–15% of novel design is accepted.
National Taiwan University of Sport	Ultrasound AI may not diagnose (medical law); it only screens images for a physician to read.

Pegatron's Hsiao An-chu reached back to a 1978 MIT paper on teleoperated submersibles for the concept of supervisory control: direct remote operation hits latency problems, so the human issues directional goals and the equipment handles detailed control. He extended the analogy to the management latency of running overseas plants — human plus AI in coordinated supervisory control.

Don't forget — in the end, it is still a human doing supervisory control.Hsiao An-chu · Pegatron

A method for choosing tasks also took shape. Hsiao cited Stanford's WorkBank paper and its two axes:

technical feasibility
    ↑
    │   R&D zone         green zone
    │ (wait for tech)    (do it now)
    │
    └────────────────────→ does the human want AI's help

An interesting finding: what people most want AI to take over is scheduling client appointments, maintaining files and format conversion. What they least want AI touching is mostly creative work — editing, layout.

9 · Taiwan's driver is not cost-cutting — it's the labour shortage

This is where the Taiwanese conversation diverges most sharply from the international one, and it recurred across many talks:

Chen Cheng-jan (policy panel) — a structural shortage of 25–35-year-old AI talent from the falling birth rate (annual births down from 400,000 to 200,000) and an M-shaped distribution; with an ageing society, robots and migrant workers have to fill the gap.
Huang Neng-fu (memorial lecture) — nearly every agricultural case originated in a labour shortage: cucumbers need harvesting twice a day, clam grading needs human eyes, foreign-object inspection of soft medicine pouches needs 20 people with 1.0 vision.
Huang Hsin-chuan (Lion Travel) — after heavy AI adoption the company is still 500 people short; the goal is to take NT$30bn of revenue to NT$60bn with the same 2,000-plus staff.
Lin Yen-liang (Deloitte) — Taiwan's past economic dividend rested on high-quality low-cost labour; as the birth rate falls that dividend fades and firms need a second growth curve.
Lu Chia-cheng (Cisco) — answers the shortage with digital employees: physical robots plus virtual AI systems.

The difference matters. The Western framing is usually AI will take your job. On the ground in Taiwan the framing is nobody can be replaced, because there is nobody to hire. That explains why Taiwanese deployments are overwhelmingly augmentation rather than automation-replacement, and why hybrid working is not a compromise here but the only feasible answer.

If your machine isn't better than me, don't bother coming.A clam farmer, to Huang Neng-fu's team

The machine currently grades 600 clams per minute per line. A person does 2,800. AI has not won yet.

10 · No consensus on what Taiwan should do

This is the most interesting thread running through the whole conference — nearly every major speaker answered the question, and the answers conflict.

Speaker	Position	Implied resource allocation
Jonney Shih · ASUS	A compute × model × application triangle, building a trustworthy national AI brand.	Go all in, leveraging a near-90% share of GPU systems.
Wei Shih-chun · Realtek	Don't build the cloud brain — build lots of cheap low-power small brains and reflexes.	Bet on semiconductors; give up chasing large models.
Jan Li-feng · policy panel	Aim for global leadership in applications, not a foundation-model arms race.	Government leads on usage, data trusts, universal AI tokens.
Chen Cheng-jan · policy panel	Only two people can save Taiwan: robots and foreigners.	Direct resources to automation and migrant / international talent.
Tsai Ming-shun · AIA closing	Taiwan is unlikely to become an AI research power, but can become an AI application power.	The Taiwan Inside formula: AI + D (domain) × T (Taiwan way).

The fault line is whether Taiwan should touch models at all. Shih says yes (model engineering is one of the three advantages), Wei says no (build System 1, not System 2), Jan and Tsai say go around it (build applications).

Structural numbers underpinning the debate (Chen Cheng-jan)

Taiwan's stock market roughly tripled over several years to about ninth largest globally by market cap — while Taiwan's GDP ranks only about 20th–21st.
The seven largest US tech companies are together worth more than China's annual GDP; Nvidia alone exceeds the UK's GDP.
TSMC, MediaTek and Foxconn together account for over 90% of all investable R&D spending in Taiwan.
Of roughly 7,000 natural languages worldwide, mainstream LLMs support about 100 — a resource-usage gap risk for Traditional Chinese.

Policy is not about which things should be done. It is about the order in which we do the things that should be done.Hou Yi-hsiu, opening the policy panel

What nobody talked about this year

Sometimes the missing topics say more about where an industry stands than the popular ones. Against 2023–2024, these clearly left the room:

Retired topic	Why
Explaining what an LLM or Transformer is	The audience is assumed to know. The one introductory framing on offer (AIA's kitchen analogy) exists to separate Workflow / Agent / Agentic Workflow, not to explain models.
Prompt engineering technique	Replaced by Context Engineering — the level of discussion moved from how to write a prompt to how to manage the model's memory.
Anxiety about AI taking jobs	In Taiwan it is simply overwritten by the labour shortage.
Plain RAG tutorials	Replaced by Agentic RAG — retrieval is no longer hard-wired at design time; the agent looks things up mid-task.
Leaderboard comparisons	Replaced by doubt about whether the evaluations mean anything.

If you remember only three things

1. The gap is no longer in the model — it's in the process.

The shared conclusion of forty-odd talks: the technical barrier has dropped sharply (vibe coding, open models, no-code platforms) and the bottleneck has moved entirely to the organisation — strategic consensus, process mapping, workforce transition, top-down push. Tung Feng's if Ah-wang doesn't sign, the manager doesn't sign and Lion Travel's chairman demanding every sales meeting open with what AI did this week are crude organisational levers that proved more decisive this year than any technology choice.

2. Accepting that v1 will be bad is the one precondition every success case shares.

91App's 98% coverage at 50% accuracy, Lion Travel's hallucinated itineraries from ChatGPT, Pegatron's agents dying on contact — not one success case worked first time. What actually works is ship it, find it's bad, add rules, iterate, with fail-safes and continuous calibration built into the system (like 91App's Evaluator, deliberately kept outside the conversation flow).

3. Taiwan's AI problem is we can't find people, not we need to cut costs.

That determines the priority: not headcount-cutting automation, but augmenting human-machine collaboration and knowledge transfer — digitising the old masters' experience, sharing knowledge across international plants, training hands-on talent. Which is why Huang Neng-fu's AI Maker and cross-campus compute-sharing alliance, and AIA's tiered certification and regional centres, carried as much weight this year as any technical talk.

Sources

This page synthesises across sessions. External material cited above: