GAI → AIA 2025

How this comparison is built

This page reads the September notes on this site against a separate set of notes for the 2025 Generative AI Conference, held on 23–24 May 2025. Both sets are field notes taken in the room and checked afterwards, and both mark anything they could not verify rather than filling the gap.

Unlike the 2024 page, this is not a year-over-year comparison. These two events ran in the same year, and they differ in kind as much as in date: one stage against three parallel rooms, individual practitioners against corporate and academic institutions, a paid two-day double bill against a foundation's annual conference. Neither is the better conference. They were built for different rooms.

• 2025 Generative AI Conference — field notes
• The 2024 comparison on this site

1 · The two events, side by side

The shape of the room explains most of what follows. One event put everyone in a single hall and ran the speakers one after another; the other ran three parallel tracks and asked you to choose. A single stage means every attendee hears the same twenty-four talks, so the programme can afford a primary-school student between a bank and Google DeepMind. Three rooms means the programme can carry drones, policy and sports science at once, and nobody hears all of it.

2 · What moved in the three and a half months between them

One hundred and eight days separate the two events. That is short enough that the same tools, the same models and largely the same problems were in play — and long enough that several of them changed sign.

3 · The same people, two rooms

Three speakers appeared at both events, out of roughly seventy people on the two programmes. That overlap is small, which is what makes it useful: when the same person says a different thing in May and in September, the room is the variable you can actually see.

• Ed H. Chi · Google DeepMind — keynote at both

  May

  The talk was built on convergence: functions that used to be separate systems — translation, summarisation, question answering — folded into one model, and the same force now extending to robots. He walked the Project Astra demo, named three robotics bottlenecks (generality, interactivity, dexterity), and closed with directions he thinks Taiwan can still take without stacking up enormous numbers of chips: multi-step reasoning, physical tool use, self-improvement, multimodal reasoning and personalisation. He also said outright that Taiwan put its resources into chips from the 1990s and missed the software wave.

  east

  September

  Same Astra, different frame. September went back to first principles: thirty years of the ranking economy in search and recommendation, why a transformer is really a transducer, and the claim that next-token prediction is association over existing knowledge rather than intelligence — real intelligence has to predict a thought pattern, which is why training needs explanation and reasoning in between. And the line the rest of the conference kept borrowing: the safety of a large language model is not A plus B plus C, it is A times B times C, because combining capabilities produces new emergent ones.

• Li Kun-mou · 91App — the same case study, twice

  May

  At the May main conference he told the product-listing case as a chief product officer: accuracy and coverage pulling against each other, then a formula for splitting the work between human brains, a rules engine and an AI agent. Half the session was about software development itself — who a PRD is written for once an AI reads it, and what a 300-person team with 300 AI assistants costs.

  east

  September

  In September the same case came back stripped of the software half and hardened into an argument about agents in retail. It gained an outcome multiple that was not stated in May (throughput up tenfold), a note that the rule engine runs on CPU with no AI at all, a named architecture with the evaluator deliberately sitting outside the conversation flow, and GEO — being findable by other people's agents — as the next thing to worry about.

• Wu Po-han · APMIC — the constructive half, then the destructive half

  May

  May was entirely constructive. He explained distillation with a central kitchen steaming lunchboxes, described compressing terabytes of enterprise knowledge into a 6GB model you can talk to, laid out a four-tier product ladder by data volume and budget, and open-sourced three things on stage: Traditional Chinese reasoning datasets, an evaluation tool, and a distilled 3B reasoning model integrating MCP.

  east

  September

  September was the demolition. He audited the 34 evaluation papers Gemma 3 officially cites and reported that 80% use multiple-choice formats while commercial work is almost entirely open-ended — and that filtering for open-ended evaluations, only 4 of 11 passed. His conclusion was that industries have to build their own standards rather than accept a vendor's claim that a model is strong.

You can't weigh the depth of water on a scale.Wu Po-han · APMIC — September, four months after open-sourcing an evaluation tool in May

The May releases did not stay in May. In September the founder of Twinkle AI — who is also APMIC's AI engineering director — gave a lightning talk in which a 3B model scored above 90 on function-calling evaluation including MCP, ahead of GPT-4o, and built TMMLU+ around Taiwan's civil-service exam bank precisely because general English benchmarks cannot tell you whether something works in a Traditional Chinese setting. The artefacts announced from one stage in May turned up as evidence on another stage in September.

4 · One project, three tellings

91App's automated product-listing pipeline is the single most-told case across the two conferences: twice in May, by a chief architect on the developer day and a chief product officer on the main day, and again in September. Three tellings of one project over three and a half months is a rare thing to be able to line up, so here it is lined up.

The end state is stable across all three tellings — the 40 / 50 / 10 split does not move — and so is the principle underneath it: coverage is negotiable, accuracy is not. It is the opening numbers that wobble, which is what you would expect from different evaluation runs or different sampling batches. Neither pair is adjudicated here.

5 · What only the May conference had

These are not topics September covered less. They are topics September did not have at all — and every one of them follows from who was in the room.

6 · What only the September conference had

The reverse list is longer, which is mostly an artefact of three rooms against one. What is more telling is that almost all of it is heavy — physical, regulated, or national in scale.

7 · Where the two rooms independently agreed

Given how little the two programmes overlap — three shared speakers out of roughly seventy — the places where they land in the same spot are worth more than either conference's claim on its own. These are five different sets of people hitting the same wall.

8 · May's open questions, checked against the September room

The May notes closed with five things that edition left unresolved. Three and a half months is not long enough to settle any of them, but it is long enough to see which ones the industry picked up and which it walked past.

9 · Topic coverage, both conferences

One grid for the whole comparison. This scores how much weight each topic carried in that conference's own trend synthesis — not how many times the words were said in the building.

The one-line version

May was a conference about what one person can now do. September was a conference about what an industry still cannot.

The three and a half months between them did not change the tools much. What changed is the unit of account: in May the win was measured in a person's afternoon, in September in a payback period, a headcount and a supply-chain requirement. Both rooms reached the same engineering conclusions — coverage is negotiable, evaluation is the bottleneck, the organisation is the blocker, the human stays in the loop — and neither room could hear the other say it.

• 2025 Generative AI Conference — field notes
• Generative AI Conference — official site
• Official agenda — Taiwan AI Annual Conference 2025
• INSIDE — the 2025 Generative AI Conference feature