How this comparison is built
This page reads the 2025 notes on this site against a separate set of notes for the 2024 conference. Both sets took the same line on what counts as a trend: a thread only earns a heading if at least two speakers raised it across at least two rooms.
That matters for how you read the gaps. When a topic is missing from one year, it means it did not become a main line that year — not that nobody in the building mentioned it. Where a 2024 talk touched something only in passing, this page says so.
1 · The two events, side by side
The event got smaller and more concentrated: five parallel rooms became three. That is the one figure on this table that is directly comparable, and it lines up with what the programme did — fewer rooms, longer threads, more talks that answer each other.
2 · What became of 2024's six axes
The 2024 notes organised the conference into six axes. Not one of them survived into 2025 untouched. Three kept the topic but moved the sore point, two flipped to the opposite conclusion, and one moved up a level. Four more threads appeared with no 2024 ancestor at all.
- Model strategy: RAG or fine-tuning?Focus movedThree-layer verticalisationThe two camps stopped arguing. APMIC's sovereign / industry / enterprise stack turns the question into which layer you fine-tune, and whether you build it or buy it.
- Localisation and sovereign AIFocus movedStill here — the bottleneck moved to evaluationIn 2024 the shortage was licensed data. In 2025 the shortage is an evaluation that reflects whether a model is commercially usable.
- Inference cost and hardware strategyReversedCompute moves back on-premisesThe 2023–24 default answer — cloud first — was loosened by three independent forces at once: cost, security and sovereignty.
- AI agents: deployment in sightReversedAgentic AI enters its disillusionment phaseFrom roadmaps to failure-mode lists. The 2024 notes score their own prediction that agents would go mainstream as not borne out.
- Cyber offence and defence co-evolvingFocus movedRisk shifts from the model to the person using itThe biggest risk named on stage is no longer a model going off the rails — it is an employee pasting order prices into an outside chatbot.
- Treating LLM work as engineering managementLevelled upFrom engineering discipline to scalingThe question moved up a level — from how do we build this properly to why can't we get it adopted.
- No 2024 axisNew this yearPhysical AI and digital twins2024's six axes contain no physical world at all. The closest thing was one line from Qualcomm that on-device LLMs would eventually drive robotic action.
- No 2024 axisNew this yearThe evaluation crisis2024 complained there was no evaluation data. 2025 argues the evaluations we do have have stopped measuring anything useful.
- No 2024 axisNew this yearThe labour shortage as Taiwan's first driverNot cost-cutting. There is simply nobody to hire, and that changes which AI projects are worth doing.
- No 2024 axisNew this yearOpen disagreement over Taiwan's route2024 discussed sovereign AI without a confrontation. In 2025 five senior speakers gave five conflicting answers to the same question.
3 · Three outright reversals
Focus moving is normal between two conferences. These three are different — the 2025 conclusion contradicts the 2024 one.
Where AI agents stand
2024Everyone was drawing roadmaps. E.SUN Bank laid out AI Embedded → AI Copilot → AI Agent as three stages with the agent as the end state; Far EasTone was building an agent platform as core infrastructure; MediaTek described agents pairing with each other like Bluetooth devices. ITRI showed the most concrete gain: sending an X-ray straight to a multimodal model carried a 50–60% error rate, while decomposing it into a chain — assess the clinical situation, then the image type, then the findings — reached over 90% accuracy.
2025Almost every speaker talked about the difficulty instead. Pegatron used its own telemetry to show agent usage is highly fragmented — people try it once and it dies — and cited Salesforce research that accuracy drops sharply as more MCP tools are attached, to 33% in some scenarios. 91App's product-listing pipeline went through four versions and landed on Rule Engine 40% + AI 50% + human 10%.
The default answer on cloud versus on-premises
2024The discussion was about pushing compute cost down wherever you ran it. Phison used NAND SSDs to extend GPU memory, taking on-premises training of a 90B model from NT$30m to under NT$1m; the National Center for High-performance Computing argued for buying across vendors — H100, A100, MI300, Gaudi 3 — so as not to bet everything on one option.
2025Cisco put a break-even point on it: at roughly 250 inference requests per minute, on-premises build-out pays back against cloud token spend in about 1.5–2 years. It also noted each agentic engine averages 8–14 repeated confirmation calls to the model — the agent architecture is itself a token multiplier. LINE Taiwan moved marketplace search to a self-hosted open model because commercial API cost would not carry the volume.
Whether a benchmark score persuades anyone
2024The number was the argument. TAIDE cited EduTAIDE passing at 76% against GPT-4o's 51%; MediaTek Research built a Traditional Chinese BFCL evaluation set for Breeze 2.0. Leaderboard position was the common currency for model capability.
2025The currency itself came under attack. APMIC went through the 34 evaluation papers Gemma 3 officially cites and found 80% of them are multiple choice, while commercial work is almost entirely open-ended; filtered down to open-ended evaluations, scores sat mostly around 70, with only 4 of 11 passing. My model ranks Nth on this leaderboard stopped being a commercial argument.
You can't weigh the depth of water on a scale.APMIC, on benchmark validity — 2025
4 · Four threads that were not there in 2024
Physical AI and digital twins
The single largest addition, and it arrived from four different industrial positions at once. The framing question was put most bluntly by Realtek: even if ChatGPT can tell you how to ride a bicycle, it does not know what riding a bicycle is.
The reason digital twins came up so often is cost, not fidelity. Kenmec framed three barriers: usable real-world data is only about 10% of what is collected, engineers who can train these models are scarce and expensive, and physical testing runs to roughly NT$4m per humanoid robot — training twenty of them could reach hundreds of millions. Delta gave the least intuitive read on where twins actually pay off:
In mass production a digital twin can't do much — if you want something that looks like the real thing, just go look at the real thing. Where it genuinely earns its keep is trial production.Chen Sheng-hua, Delta Electronics · 2025
There is also a supply-chain signal with no 2024 equivalent: Kenmec noted TSMC now requires vendors to deliver USD assets on some piping projects, and robot suppliers are starting to be asked for URDF. Digital twins are turning from an internal tool into a deliverable specification.
The evaluation crisis
2024 said Taiwan lacked evaluation data — domain experts had to write the prompts, because engineer-written prompts capped models around 70, and iKala substituted national civil-service exams where official evaluation data did not exist. 2025 makes a stronger and stranger claim: the evaluations that do exist have decoupled from commercial usability.
The constructive half of that thread is local evaluation sets built on purpose: Twinkle AI built TMMLU+ around Taiwanese national-exam material precisely because general English evaluations cannot tell you whether something works in a Traditional Chinese setting. APMIC's conclusion is that industries have to build their own standards rather than accept a vendor's claim that a model is strong.
The labour shortage as the first driver
This is where the Taiwanese conversation diverges most sharply from the international one, and it has no 2024 counterpart at all. The Western framing is AI will take your job. On the ground here the framing is nobody can be replaced, because there is nobody to hire.
- Policy panel — a structural shortage of 25–35-year-old AI talent, with annual births down from 400,000 to 200,000.
- Lion Travel — still 500 people short after heavy AI adoption; the goal is to take NT$30bn of revenue to NT$60bn with the same 2,000-plus staff.
- Memorial lecture — nearly every agricultural case originated in a labour shortage, and the farmers' bar is brutal: machine grading runs 600 clams per minute per line against 2,800 by hand. AI has not won yet.
That explains why Taiwanese deployments are overwhelmingly augmentation rather than automation-replacement — and why hybrid human-machine working is not a compromise here but the only feasible answer.
Open disagreement over Taiwan's route
2024 did discuss sovereign AI, and one sovereign-AI session already argued that models are temporary, data is permanent — opposing training frontier models from scratch on a budget under NT$200m with 72 H100s against Meta's roughly 24,000 GPUs. But it stayed a single position rather than a confrontation.
In 2025 five senior speakers answered the same question incompatibly. The fault line is whether Taiwan should touch models at all: ASUS says yes, a compute × model × application triangle; Realtek says no, build System 1 reflexes in silicon rather than a cloud brain; the policy panel and AIA's closing say go around it and lead on applications. One panellist put the resource argument at its bluntest:
Only two people can save Taiwan: robots and foreigners.Chen Cheng-jan, policy panel · 2025
5 · What 2024 talked about and 2025 did not
Retired topics say as much about where an industry stands as popular ones. Each row below was a substantial part of the 2024 notes and is either absent or demoted in the 2025 ones.
| Topic | What it looked like in 2024 | Where it went |
|---|---|---|
| RAG versus fine-tuning as a binary | Two camps facing off across the programme — eLand, Gss and Nexcom on the RAG side; TAIDE, Temple AI and Phison holding out for fine-tuning, with quantified results on both sides. | The binary dissolved. Fine-tuning is now a question of which layer (sovereign / industry / enterprise), and retrieval moved to Agentic RAG — no longer hard-wired at design time. |
| Licensed data and corpus scarcity | An axis of its own: roughly 1,300 legally licensed Taiwan-related images on Wikipedia, Traditional Chinese at about four parts per million of open-source LLM corpora, and government data requests answered with crawl it yourself. | Still true, but demoted to background. The sore point on stage is now evaluation, not corpus size. |
| No-code and low-code AI platforms | A visible 2024 wave — Quanta's QOCA compressing medical imaging model building from two weeks to two hours, Prophet Tech on manufacturing defect detection, MSI's AI Artist — all aimed at SMEs and hospitals with no AI staff. | No longer a thread of its own. It survives as one footnote in the observation that the technical barrier has dropped — alongside vibe coding and open models. |
| Leaderboard comparison | Scores were the argument, quoted freely across tracks as evidence of capability. | Replaced by doubt about whether the evaluations measure anything commercially meaningful. |
| AI governance as a dedicated programme block | Heavily staffed: a governance panel filling the Day 2 morning, a dedicated workshop room that gave its Day 2 afternoon to data governance, and a talk applying the ISO 27014 management-review mechanism directly to AI adoption. | Folded into the AI security forum instead of standing alone. Taiwan's Basic Act on AI was still in committee during the conference and only passed on 23 December 2025. |
6 · 2024's predictions, checked against the room in 2025
The 2024 notes did something unusual: they listed eleven forward-looking claims made at that conference and scored each one against later evidence. Roughly half held up. The pattern is that the direction was right more often than the timing or the magnitude.
- Borne out · 4
- Partly borne out · 3
- Not borne out · 2
- Still open · 2
- Not borne outAI agents would go mainstream in enterprisesThe 2025 conference is the on-site confirmation: Pegatron's own telemetry showed fragmented, abandoned usage, and 91App's shipped architecture ended up Rule 40% + AI 50% + human 10%.
- Still openEnterprises would stop fine-tuning and switch wholesale to RAG2025 supplied evidence the other way: APMIC cited UBS research putting 86.6% of enterprises as needing fine-tuning, and NVIDIA research that 70% of AI agents need small-model optimisation. The move is not away from fine-tuning but towards smaller, more vertical fine-tuning.
- Borne outPhison's aiDAPTIV+ approach — extending GPU memory with storage to cut on-premises training cost2025 promoted the idea to a main line of its own: compute moving back on-premises, with Cisco putting the break-even at about 1.5–2 years.
- Borne outThe Breeze series would keep shippingSeen on stage in 2025: on 50 previously failing passport-OCR cases, Breeze 2 beat Gemini despite being far smaller, and a travel chatbot on 3,000 Q&A pairs reached 90% similarity to expert answers.
- PartlyA Basic Act on AI plus a standalone data governance actThe Basic Act passed on 23 December 2025 with 20 articles, overseen by the National Science and Technology Council — but no standalone data governance act followed.
- PartlyTAIDE would release a multimodal modelThe text model kept iterating but no multimodal release arrived. On the 2025 stage, the local-model conversation had visibly moved towards Breeze 2 and purpose-built local evaluations such as TMMLU+.
- Not borne outFrontier models would converge to two or three providers — the most dangerous momentMore than a dozen frontier providers remain. The 2025 small-model and verticalisation thread runs the opposite way — dispersion, not convergence.
7 · Topic coverage, both years
One grid for the whole comparison. This scores how much weight each topic carried in that year's trend synthesis — not how many times the words were said in the building.
| Topic | 2024 | 2025 |
|---|---|---|
| RAG versus fine-tuning | A main line — has its own axis or section | Present, but attached to another thread |
| Sovereign AI and local models | A main line — has its own axis or section | A main line — has its own axis or section |
| Licensed data and corpus scarcity | A main line — has its own axis or section | Present, but attached to another thread |
| Evaluation validity | Present, but attached to another thread | A main line — has its own axis or section |
| Inference cost and hardware | A main line — has its own axis or section | A main line — has its own axis or section |
| On-premises versus cloud | Present, but attached to another thread | A main line — has its own axis or section |
| AI agents | A main line — has its own axis or section | A main line — has its own axis or section |
| Small models and verticalisation | Present, but attached to another thread | A main line — has its own axis or section |
| Robotics and embodied AI | Present, but attached to another thread | A main line — has its own axis or section |
| Digital twins, OpenUSD, URDF | Not in that year's trend synthesis | A main line — has its own axis or section |
| Cyber offence and defence | A main line — has its own axis or section | A main line — has its own axis or section |
| AI governance and regulation | A main line — has its own axis or section | Present, but attached to another thread |
| Engineering method and maturity models | A main line — has its own axis or section | Present, but attached to another thread |
| Organisational scaling | Not in that year's trend synthesis | A main line — has its own axis or section |
| No-code / low-code platforms | A main line — has its own axis or section | Not in that year's trend synthesis |
| Human-machine hybrid working | Present, but attached to another thread | A main line — has its own axis or section |
| The labour shortage as a driver | Not in that year's trend synthesis | A main line — has its own axis or section |
| Context engineering | Not in that year's trend synthesis | A main line — has its own axis or section |
- A main line — has its own axis or section
- Present, but attached to another thread
- Not in that year's trend synthesis
The one-line version
2024 was a conference asking whether AI could be used in my industry. 2025 is a conference asking why it still won't stick after we used it.
Not one of the six 2024 axes survived unchanged — three moved their sore point, two reversed, one moved up a level. And everything genuinely new points the same way: off the screen and onto the factory floor, the machine and the shift that nobody can be hired to cover.