2025 Generative AI Trends

1. These two days landed on a very crowded week

In the days either side of the conference, the model vendors kept shipping: the night before 保哥 went on stage, Anthropic released Claude Opus 4 and Claude Sonnet 4 (official announcement, dated 22 May 2025, the day before the developer conference); the universal-assistant prototype 紀懷新 demonstrated had been announced three days earlier; Davis Chang mentioned Veo 3, the video generation model Google had just released (Google announcement). The curator also mentioned a few of these in passing during the opening.

That density is itself part of the trend. The curator said in the opening that every time they run the conference a pile of new things appears beforehand and the speakers have a hard time of it. In practice: one speaker heard someone else's session the day before and changed his topic on the spot.

2. From "answers you" to "does the thing"

If last year's keyword was conversation, this year's is agents. The curator put the difference most plainly: it used to be that you spoke and it replied with a paragraph of text; now you speak and it takes action.

What is striking is how ordinary the actual tasks are. A logistics company has it download files, merge them and reconcile accounts every day; a software company has it go straight from a requirements form to opening a branch and writing code; an NGO worker has it check books and fill in forms. Nobody is talking about who it will replace. Everybody is talking about which stretch of their day it removed.

The other thing brought up repeatedly is sign-off. 李慕約's line — "people need to be good clients, and a good client knows how to write a brief and how to sign off" — and 保哥 deliberately displaying an inflated per-run cost on his intake form to force people to state their requirement properly, are two versions of the same point. The more an agent can do, the further forward the human's work shifts, toward defining the problem and accepting the result.

3. The protocol layer started to take shape this year

A year earlier, everyone wired up their own tools. This year a common layer appeared. 張文鈿 spent his whole session on MCP: the tool vendor writes a server once and every supporting client can use it.

That session also shows the real state of protocol maturity: the specification itself is usable, but remote deployment and authentication were both still being filled in at the time, and even the official debugging tool had version-dependent bugs. This is the classic shape of a protocol that has just won consensus while the implementations are still catching up.

The rate of spread says as much. In a single event, MCP appeared as a spec walkthrough (張文鈿), as an engineering implementation (保哥 wiring it into his own coding agent), and as a bank's medium-term plan (黃仕鎮 listing MCP in E.SUN Bank's next round of agent integration). 張文鈿 also mentioned Google's A2A: how agents talk to each other is the next layer down, and it is not settled.

4. Models are moving in two opposite directions at once

This is the clearest tension of the two days.

One direction is convergence. 紀懷新's whole session is built on it: translation, summarisation, question answering — functions that used to stand alone are folded into a single model, and even for robot-arm control he stresses one model rather than several.

The other direction is distillation. 吳柏翰's whole session is about condensing a giant model's domain knowledge into a small model an enterprise can afford, then using inference-time techniques to pull the quality back up. His customers want something that runs in their own server room.

The two lines do not conflict, because they serve different people: the platform wants a model that can do everything, the enterprise customer wants a model whose data never leaves. What changed this year is that the second one became genuinely viable, and the driver is compliance and confidentiality, not cost. Both E.SUN Bank sessions repeat the same sentence: customer data cannot leave, so it has to be built in-house.

5. Evaluation has become the new bottleneck

Once model capability rises, "how do I know it is good enough" becomes a harder question than "how do I make it stronger". Three people in completely different positions raised this independently across the two days.

林彥廷 comes at it from research: the benchmark determines which mountain the whole field decides to climb, and current public benchmarks are losing their meaning fast — the entire exam set he designed for Traditional Chinese models two years ago has been all but solved by every vendor. He also names the structural problem of models trading accuracy for agreeableness.

吳剛志 comes at it from product: if public metrics are not enough, define accuracy and coverage yourself, then take the team through dozens of rounds. His comparison is a strong one — a prompt you use internally can be checked case by case across a hundred runs; once it ships and runs a million times, case-by-case checking is impossible.

李昆謀 comes at it from organisation: turn the evaluation result directly into a split of the work, with the AI doing only what it is confident about and the rest going back to the rules engine and to people.

All three land in the same place: better low coverage than low accuracy. That is probably the single most usable sentence of this edition.

6. Image generation crossed the "is it usable" threshold

None of the three AI imagery sessions this year demonstrated how strong a model is. All three were about everything except the model.

陳志信 argues that what decides whether a video is any good is the compositional design done before you generate anything, and his method is deliberately built to not depend on any one piece of software. Davis Chang argues that what actually changed is who is able to start creating; once a production team compresses from hundreds of people to a handful, the scarce thing becomes aesthetic judgement.

Running the other way, 李怡志 spent his whole session on the other face of the same technology. His argument is counter-intuitive: whether a fake image fools you has little to do with how real it looks, and much more to do with whether it lands on your position, your emotions, your knowledge gaps and your social trust. As the technical barrier drops, the focus of literacy moves from "look for the flaws in the picture" to "examine why you want to believe it".

7. The people in the room changed

The clearest structural change this year is not technical, it is the speaker list. The conference deliberately programmed a whole strand for non-engineers: an advertising salesperson, an NGO worker, a primary-school student. The curator also mentioned in an interview that the gender split of speakers had gone from male-dominated to even (INSIDE interview).

Peggy Lo's observation carries the most weight in that whole strand: 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. Put that next to the other speakers' "you cannot push this through without management support" and you get an uncomfortable but possibly correct inference — the people who need AI most and the people with the power to decide whether to adopt it are not the same group.

8. Where Taiwan sits

紀懷新's advice was concrete: multi-step reasoning, tool use in the physical world, self-improvement, multimodal reasoning and personalisation are all directions that do not require stacking up enormous numbers of chips, and Taiwan has a foundation in combining hardware and software. He also said outright that Taiwan put its resources into chips from the 1990s onward and, as a result, missed the software wave.

From the two days of enterprise cases, Taiwan's actual position looks like "solid application domains, thin model layer": banking, retail, logistics, printing and entertainment can all produce quantifiable adoption cases, but apart from a few teams working on Traditional Chinese models and enterprise distillation, autonomy at the model layer remains thin.

9. Still unresolved

• How agents interoperate. MCP solved models reaching tools. Agents reaching agents has no consensus yet.
• Remote deployment and authentication. The pothole 張文鈿 named directly from the stage.
• The cost of evaluation. Annotation costs for high-quality benchmark data are high enough to steer research direction.
• Outcome numbers are hard to verify. Almost every quantified result from these two days of enterprise cases is the speaker's own account with no external audit, which is clearest in the summary table in Cross-session themes.
• Literacy is not keeping up with generation. Generation advanced a long way in a year; ordinary people's methods for telling real from fake barely moved at all.

10. Primary sources