9–10 September 2025 · Academia Sinica, Taipei

Taiwan AI Annual Conference 2025

Forty-plus talks across three parallel tracks, read as a single slice of an industry. This year almost nobody argued that AI matters — they argued about how to make it work at scale, and where Taiwan should place its bets.

2Days

4Keynotes

44Talks

1500Attendances

The event

Organiser: Taiwan AI Academy Foundation (AIA)
Co-hosts: Academia Sinica — Institute of Information Science; Research Center for Information Technology Innovation
Dates: Tue 9 – Wed 10 September 2025
Venue: Humanities and Social Sciences Building, Academia Sinica, Nangang, Taipei
Theme: Taiwan's AI Competitiveness
Scale: 4 keynotes, 2 panels, ~44 track talks, 10 lightning talks across 3 rooms; ~1,500 attendances
Official site: conf2025.aiacademy.tw
Official agenda: conf2025.aiacademy.tw/agenda

The thirty-second version

Seven things the two days were actually about.

The axis moved from whether to adopt AI to how to scale it — Almost nobody spent time selling the audience on AI. They talked about what happens after the PoC, how to drag accuracy from 50% to 90%, why the organisation won't move, and how to put a number in front of the boss.
Agentic AI was the most-used phrase, and almost every mention was about its difficulties — Pegatron cited three papers on why agents die on contact (accuracy drops to 33% once too many MCP tools are attached); 91App's answer was a Rule Engine + AI + human hybrid; MediaTek Research walked through four Context Engineering operations.
Digital twins were framed as the bridge that carries AI into the physical world — Siemens, Pegatron, Kenmec and Delta approached the same claim from four directions: data, training and physical testing are all too expensive for Physical AI, and a virtual environment is the only way out.
Open-source and Traditional Chinese models became the pragmatic option for SMEs — MediaTek Research, APMIC and Twinkle AI made the same argument from three positions: a vertical small model can beat a general model many times its size, at roughly a fifth of the VRAM.
Drones and robotics arrived as a headline topic — One of the two keynotes went to GPS-denied drones; the whole of R1's second afternoon was drones and embodied intelligence. The recurring frame: de-risking the red supply chain, open standards, and Taiwan's semiconductor edge.
AI governance and security moved from slogan to checklist — The cyber-security institute unpacked OWASP LLM Top 10 and three jailbreak strategies; the safety panel split risk into R&D, enterprise and government views, and pointed out that the real corporate risk is not a runaway model but an employee pasting confidential data into an outside chatbot.
The policy panel argued about priority, not direction — Three concrete proposals: government should become the most fluent AI user in the country, build a data-trust regime, and hand out AI tokens universally.

Where to start

The trends page is the one to read first; everything else is reference.

What 2025 actually signalledNot by agenda and not by topic — this reads the two days as one slice of an industry. When forty-odd speakers from model research, semiconductors, manufacturing, finance, retail, academia and government each talk about their own work, what do they all end up saying? And what has everyone quietly stopped saying?Themes across the tracksThe two days, three rooms and forty-odd talks taken apart and reassembled by topic, putting speakers side by side — with particular attention to where they echo each other and where they flatly disagree.Full agendaEvery slot across both days. R0 was the main hall; R1 and R2 ran in parallel, with an R3 workshop track on day one.Day 1 — 9 SeptemberOpening keynote and the policy panel in the morning; three parallel tracks all afternoon — enterprise transformation and trustworthy AI in R0, financial fraud and digital twins in R1, industry deployment cases in R2. Timecodes are relative to the start of each session.Day 2 — 10 SeptemberTwo keynotes and the AI safety panel in the morning; three parallel tracks in the afternoon — sports science in R0, drones and embodied intelligence in R1, nine lightning talks in R2 — closing with the Chen Sheng-wei Memorial Lecture. Timecodes are relative to the start of each session.By the numbersEvery figure quoted across the two days, gathered in one place — useful when you need to put a number in a proposal. Each row names the speaker or organisation it came from.GlossaryEvery proper noun, product name and acronym that came up across the two days, defined in the sense the speakers used it. Type to filter.Sources and verificationEvery externally verified claim, with its source link and a confidence rating. Filter by category, sort any column, or search. Where nothing could be confirmed, the row says so rather than guessing.

Lines worth keeping

The safety of a large language model is not A plus B plus C — it is A times B times C.

Ed H. Chi · Google DeepMind

You cannot weigh the depth of water on a scale.

Wu Po-han · APMIC — on judging commercial fitness with multiple-choice benchmarks

In mass production a digital twin can't do very much. Where it really earns its keep is trial production.

Chen Sheng-hua · Delta Research Center

If your machine isn't better than me, don't bother coming.

A clam farmer, to Huang Neng-fu's team

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 · chairing the AI policy panel

Only two people can save Taiwan: robots and foreigners.

Chen Cheng-jan · on ageing and the labour shortage