1. These are not the same event
Start with the basics, so neither is mistaken for a previous edition of the other:
| Item | 2024 Taiwan AI Academy Conference | 2025 Generative AI Conference |
|---|---|---|
| Organiser | Taiwan AI Academy Foundation (seventh edition), co-organised with Academia Sinica's Institute of Information Science and CITI | Generative AI Conference Production Committee (another account has it co-organised with TNL Mediagene) |
| Dates | 2024-09-27–28 | 2025-05-23–24 |
| Venue | Academia Sinica, Nangang campus | Howard Civil Service International House, Excellence Hall |
| Theme | AI for Every Industry | A double bill: developer conference plus main forum |
| Structure | Two days, five parallel tracks (R0–R4) | Two days, single track, one stage in one hall |
| Sessions | 58 talks | 24 talks plus an opening |
| Speakers | 71 (including hosts, panellists, flash talks and workshop speakers) | 23 |
| Sponsors | 13, including Chunghwa Telecom, Qualcomm, AMD, Intel, MediaTek and MSI | The official press release named NVIDIA and Appier; the 2025 tiered list page has been taken down |
| Attendance | No official figure published; reporting described a packed opening, with Vice President Hsiao Bi-khim attending | Press afterwards reported close to a thousand on site and over two thousand replay purchases |
| Ticket price | No public pricing obtained | Main conference standard NT$6,000, developer conference standard NT$3,000 |
The same 2024 notes give two different speaker counts on different pages: the speaker index says 71 (covering every role), while the event background page says "55 speakers across a two-day agenda with 4 keynotes, 1 panel, and nearly 50 industry-application sessions". The two counts rest on different bases; both are listed here and neither is adjudicated.
2. Five tracks against one: two opposite curatorial assumptions
The difference in scale is not only a matter of size. It reflects two opposite assumptions about what a conference is for.
2024 ran five parallel tracks and 58 talks across two days. That structure assumes an audience that picks a track by its own industry — industry applications, deep technology, platform services and smart transformation each running their own way, five things happening in any given slot, and nobody able to hear them all. It serves "every industry finds its own piece".
2025 ran a single track and 24 talks, everyone in the same hall hearing them in order. That structure assumes every attendee should share one narrative line: agents, coding with AI, trends, enterprise cases, and finally AI imagery. It serves "everyone sees the same panorama". This is also why the 2025 notes can draw cross-session comparisons while the 2024 notes had to organise by trend lines instead — an event with five parallel tracks has no path that everyone walked.
One direct consequence is a different distribution of depth. 2024 cut its time into half-hour slots, which fits 58 talks in at the cost of making any single one hard to develop; 2025 has only four-tenths as many sessions, but the MCP talk could run from the principles of the protocol all the way down to which version of a debugging tool has an authentication bug. Breadth traded for depth is what the single track buys.
3. Five speakers appeared at both
This is the most valuable part of the comparison: five speakers appeared at both events, so you can watch the same person eight months apart.
| Speaker | The 2024 session | The 2025 session | What changed in eight months |
|---|---|---|---|
| 黃仕鎮 (E.SUN Bank) | "Generative AI applications in E.SUN Bank" (R1, industry applications), as Director of Data Science, proposing a three-stage framework: AI Embedded → AI Copilot → AI Agent | 玉山銀行生成式AI服務的革新之路, by then as Chief Engineer | From a three-stage roadmap about where to go, to laying out that only 17.8% of eligible staff had ever used the internal platform and how each problem was patched. From vision to post-mortem |
| 李昆謀 (91APP) | "Retail AI: from recommendation to agent" (R1, industry applications), deploying a vector-retrieval architecture called "Joy" and predicting retail would move toward agent-to-agent transactions | 把 AI 融入零售業的日常流程 | From a roadmap running "recommendation toward agents" to one concrete project auto-filling product specification fields, landing on how accuracy and coverage pull against each other |
| 張文鈿 (愛好資訊, ihower) | "Eval-Driven-Development" (R3 flash talk), proposing a five-level maturity model | 淺談模型上下文協定MCP應用開發 | The topic moved from "how to evaluate an LLM application" to "how to connect tools to a model". Notably, the eval-driven development he presented in 2024 appears in 吳剛志's 2025 slides as a methodological reference — the same idea crossing two conferences |
| 李慕約 | The workshop "用 AI 寫程式！工作流程自動化！" (R4) and "非工程師如何用 AI 寫程式、提早五分鐘下班" (R3 flash talk), demonstrating an agent completing an online purchase by itself | Curator of the 2025 conference, and speaker on 可以知道自己可以怎麼樣馬上入門 AI Agent | 2024 demonstrated that it really can finish the purchase; the 2025 theme became "people need to be good clients, and a good client knows how to write a brief and how to sign off". The emphasis moved from whether it can, to how you know it did it right |
| 吳振和 (cacaFly) | "在多模態影像生成架構中實作 RAG" (R3, deep technology) | 多模態與RAG碰撞AI新視界 | Almost the same topic at both editions. The 2025 session pushes it further: use a multimodal model to produce the description in reverse, then use RAG to retrieve prompt structures that work, solving the concrete problem that one prompt gives wildly different images across models |
The row that says most is 黃仕鎮's. In 2024 he offered a three-stage framework — a direction. In 2025 his first slide of numbers is a 17.8% usage rate — what he hit on the way. Within a year, the same bank on the same road shifted its narrative from roadmap to post-mortem, which is probably the best evidence that the adoption genuinely happened.
The other row worth keeping is 李慕約's. In 2024 he gave a flash talk on "how non-engineers can code with AI" in the fourth of five tracks; in 2025 he curated the conference and promoted that exact topic into a whole programme block, with an advertising salesperson, an NGO worker and a primary school student. A small session one year became a curatorial argument the next.
4. How the centre of gravity moved
The 2024 notes organise the whole event into six trend lines, which makes the comparison unusually clear:
| The 2024 trend line | What it maps to in 2025 | Direction of travel |
|---|---|---|
| Model strategy: RAG or fine-tuning | Convergence or distillation (紀懷新 against 吳柏翰) | The argument moved up from "how do we connect knowledge to a model" to "should the model itself get bigger or smaller". RAG is barely an issue in 2025; it has become standard equipment |
| Localisation and sovereign AI: TAIDE's budget and compute bottlenecks, the scarcity of Traditional Chinese corpora | Almost gone. Only 林彥廷 noting that the benchmark he designed for Traditional Chinese two years earlier has been solved by every vendor | From "we need to build our own" to "the one we built has already been overtaken". Taiwan's 2025 position is solid application domains and a thin model layer |
| Inference cost and hardware: Phison using flash storage to extend GPU memory, cutting training cost from NT$30 million to under NT$1 million | Hardware barely discussed; cost is now counted in tokens — 吳剛志's "your prompt will be run 1,000,000+ times", 保哥's US$1.26 per work item | The cost question moved from the server room to the API bill. Not whether you can afford a GPU, but what one prompt times a million costs |
| AI agents: from concept to deployable, but not for high-uncertainty tasks | The main theme of the whole event, with cases that are uniformly unglamorous chores: reconciliation, form filling, opening branches and writing code | From "can it be done" to "what is it for". What landed is smaller and more menial than 2024 expected, but it is genuinely running |
| Cybersecurity offence and defence: jailbreak services, automated penetration testing, four cryptographic techniques | Almost entirely absent; only a short stretch on tool poisoning and naming attacks in the MCP session | This is where the two sets of notes differ most. 2024 has a full security trend line; 2025 has no corresponding session |
| Deployment methodology: five-level maturity, a five-step process, ISO 27014 governance | Evaluation becomes the new bottleneck (林彥廷, 吳剛志 and 李昆謀 raising it from three different positions) | From "how do we do it with discipline" to "how do we know it is good enough". The methodology remains, but the weight shifts from process to acceptance |
Compressing all six lines into one sentence: 2024 asked whether to do it and how to afford it; 2025 asked how you know it did any good. The line between them is evaluation.
5. The predictions the 2024 notes tracked, one year on
The 2024 notes carry a section tracking eight predictions made at the time, marking how each turned out: three borne out, two partly, one not, and two with insufficient data. The one worth setting against 2025 is the one marked as not borne out:
- "AI agents going mainstream" was marked as not borne out, on the grounds that third-party reporting showed limited enterprise deployment, unstable performance and poor return on investment.
- 2025 looks like the opposite at first glance: agents are the main theme. But laying the cases out, the two do not conflict — the agents actually running in 2025 download files and reconcile accounts daily, open a branch and write code from a form, check books and fill in forms. Not one is the "enterprise-grade autonomous decision" that 2024 had in mind.
- Put together, these are two statements of the same conclusion: agents did not become mainstream in the form 2024 expected. They landed in a smaller, more constrained, more checkable form instead.
- Two other items are also instructive: the Basic Act on AI that 2024 predicted passed at the end of 2025 (marked as partly borne out in those notes, because the article count differed from the prediction and the accompanying data governance act never appeared); and TAIDE's text model kept iterating while its multimodal goal went unmet. Taiwan's progress at the model layer lags what was set out at the time, which is consistent with the "thin model layer" observation at the 2025 conference.
6. The speaker list changed
This is the biggest difference in feel between the two, and it is the other face of the shift in subject matter.
The 2024 speaker list groups by kind of organisation: government, academia, semiconductors and hardware, telecoms and internet services, finance and traditional industry, cybersecurity, healthcare and startups, community and non-profit. It contains a vice president, the chairman of TAITRA, the director of the National Center for High-performance Computing, the chairman of Pegatron, senior executives from Qualcomm and MediaTek, and more than a dozen professors and research fellows. It is a list organised by institution, and what it represents is "industry and government are both here".
The 2025 list has no government officials, and no academic representation beyond one university teacher — 李怡志 of NCCU, whose subject is information literacy rather than technology. In their place are an advertising salesperson, an NGO officer from a four-person foundation secretariat, and a primary school student. The conference deliberately programmed a whole strand for non-engineers, and the curator mentioned in an interview that the speaker gender split had gone from male-dominated to even.
That turn can be summed up by one line from each set of notes. 2024's theme is "AI for Every Industry", with the weight on industry. The heaviest observation of 2025 comes from Peggy Lo: AI coding is a lever for the people lowest in the organisation with the most drudgery, precisely because managers have no use case of their own and therefore never learn it. From every industry to every person; from institutional representatives to individual practitioners.
7. How the two sets of notes differ
The notes themselves are built differently, and it helps to know how:
| Item | The 2024 notes | These 2025 notes |
|---|---|---|
| Organisation | Eleven session blocks by track and time slot, plus six trend lines | One page per session, 25 in total, plus a cross-session themes page |
| Coverage | Covers all tracks across both days | A single-track event, so every session is present |
| Prediction tracking | Yes, with each outcome marked | None |
| Sourcing | Sources attached per page | A source table at the end of every session, plus a standalone evidence page and a list of items with no public record |
| Times | Per-session start and end times | Per-session times have no public record; only a morning or afternoon block is given |
That last row is worth explaining: for 2025 the organiser published only the overall window for each day, the per-session schedule appears on the ticket page as an image, and no other official text source provides it. So the two day pages on this site go only as far as a morning or afternoon block, with the split following the agenda map the curator set out during the opening. Times that do not exist have not been filled in.
8. Primary sources
| Item | Source |
|---|---|
| 2024 sessions, speakers, track structure, six trend lines and prediction tracking | 2024 Taiwan AI Academy Conference notes |
| 2024 organiser, co-organisers, sponsors and attendance | Same site, event background page |
| The 2024 speaker list and its grouping | Same site, speaker index |
| The 2024 per-session agenda and tracks | Same site, agenda page |
| 2025 sessions, speakers and conflicting accounts | Agenda |
| 2025 cross-session comparison and quantified results | Cross-session themes |