Axis 1: The shift in model strategy — RAG plus swappable models is mainstream, but the fine-tuning camp hasn't disappeared
The majority camp (RAG / on-premises knowledge base beats building your own model)
Yang Li-wei (楊立偉, eLand Information 意藍資訊, Day 2 R0 afternoon) stated plainly that enterprises no longer do fine-tuning, because model versions update too fast — a client had just finished fine-tuning LLaMA when LLaMA 2, 3, 3.1, and 3.2 came out in succession — and enterprise data also changes constantly, so fine-tuning can't keep up; the trend is a 'model-agnostic architecture plus RAG (retrieval-augmented generation).' He backed this with test numbers showing RAG beats feeding an entire document to an LLM: feeding a full 271-page financial report to the LLM in one go cost about NT$100, whereas using RAG to first retrieve the 5 most relevant passages before answering was faster, cheaper, and more accurate.
Lin Hsien-cheng (林縣城, Gss.com.tw 叡揚資訊, Day 2 R1) echoed this: in RAG scenarios, 'retrieval quality matters more than generation quality' — roughly four-fifths of the engineering effort goes into preprocessing and retrieval, not generation itself. General-purpose models aren't always adequate for enterprise-specific terminology and high fault-tolerance requirements, but the fix is better retrieval, not building a large language model (LLM) in-house.
Kao Shih-fang (高士方, MSI, Day 1 R1) expanded AI Artist from pure text-to-image generation to RAG (which he called 'AI Fire'), letting users import their own documents for customized Q&A — another case of bolting on a knowledge base rather than retraining the model.
Lin Mao-chang (林茂昌, Nexcom Intelligence 新漢智能, Day 1 R1) argued for connecting a company's own database to generative AI (NextData + GenAI), reasoning that 'a database has clear boundaries and definite content,' making it more accurate and reliable than a stock ChatGPT pulling from the entire web and 'confidently making things up' — and that this can run fully on-premises without fine-tuning.
Hsu Chao-yung (許朝詠, China Steel 中鋼, Day 1 R0 afternoon) admitted candidly that 'it's very hard for an in-house generative AI application to out-iterate native large vendors like OpenAI,' so the company positioned itself around 'digital enablement' rather than competing on model capability — the same logic as Yang Li-wei's, but coming from traditional manufacturing rather than software, showing the judgment holds across industries and isn't just software-sector self-interest.
The counter-camp (fine-tuning wins in specific scenarios)
Tsai Tsung-han (蔡宗翰, TAIDE core training lead, Day 2 R0 afternoon, speaking later in the same session as Yang Li-wei) reached the opposite conclusion: training a small model beats 'a general large model plus RAG.' His numbers: EduTAIDE, a fine-tuned model, had a 76% pass rate versus GPT-4o's 51%; running a general large model plus RAG on LLaMA 3.1 405B requires 11 H100 GPUs, which an ordinary team can't afford.
Li Ming-ta (李明達, Temple AI 宮廟 AI, Day 2 R3) explicitly opposed both RAG ('retrieval errors easily introduce inaccuracies') and LoRA ('it degrades the model's base capability'), insisting on full-parameter fine-tuning with models of 70B parameters or larger.
Wu Wei-han (吳威翰, Accucrazy, Day 2 R3) believes fine-tuning is better than RAG/LoRA at learning 'small, subtle, tacit knowledge' — such as a specific cultural sense of language — arguing RAG is still bolted-on knowledge rather than the model truly having learned it.
Phison's Lin Wei (林緯, 群聯, Day 1 R2), while his core pitch was hardware (using NAND to expand GPU memory), argued for encouraging 'training your own model on-premises,' calling it a 'supreme right' — implying a stance that doesn't fully trust the cloud-RAG route.
The real dividing line on this axis isn't a simple 'RAG is better' versus 'fine-tuning is better' choice — it's the nature of the task. Scenarios needing broad, constantly changing knowledge retrieval (enterprise knowledge management, customer service, financial analysis) clearly leaned toward RAG plus swappable models in 2024. Scenarios needing deep, stable domain correctness (educational assessment, vertical niches, specific cultural nuance) still had a solid fine-tuning camp holding its ground, backed by concrete numbers beating general model plus RAG. The two camps collided head-on at the same conference — even in the same session — and no one on stage reconciled them.
Axis 2: The state of localization and sovereign AI — how far Taiwan got in 2024, and where it's stuck
Chen Yi-chang (陳宜昌, MediaTek Research 聯發創新基地, Day 1 R3) introduced Breeze 2.0: a 3B/8B multimodal model with Function Calling and Vision, a Traditional Chinese version of the BFCL evaluation set, and over 500GB of training data. This is a private-sector research institute-level achievement, and the highest degree of technical completion among the self-developed models presented at this year's conference.
Huang Han-hsuan (黃瀚萱, TAIDE multimodal, Day 1 R3) laid out the rawest bottleneck: legally licensed, Taiwan-related images on Wikipedia number only about 1,300; requests to government ministries for data are often met with 'go crawl it yourself'; and the 700,000 plant and animal photos from Academia Sinica's (中央研究院) Encyclopedia of Life were only obtained the same month as her talk — too late to be used in the model yet.
Lee Yuh-jye (李育杰, Sovereign AI, Day 2 R0 morning) revealed TAIDE's actual resource scale: a budget under NT$200 million and 72 H100 GPUs, compared to Meta's roughly 24,000 GPUs used to train Llama 3.2. His conclusion: 'models are temporary, data is permanent,' and he explicitly opposed Taiwan training its own frontier model from scratch.
Tsai Tsung-han (蔡宗翰, Day 2 R0 afternoon) added a methodological angle to this bottleneck: evaluation prompts must be written by domain experts, not engineers, or the model's ceiling gets stuck around a score of 70. The same panel also revealed a concrete case — iKala, unable to get official evaluation data, ended up using national civil-service exam questions as a stand-in benchmark, reflecting Taiwan's long-standing shortage of usable evaluation data.
Yang Cheng-lin (楊政霖, CyCraft, Day 1 R1) pointed out that Traditional Chinese makes up an extremely small share of open-source LLM training corpora — BLOOM's training data statistics show roughly four parts per million — which is the root cause of localization debt like 'asked in Traditional Chinese, answered with Simplified Chinese wording.'
Tsai Chi-yen (蔡祈岩, Taiwan Mobile 台灣大哥大, Day 1 R2) added another piece of the puzzle from enterprise practice: the general-purpose speech model Whisper keeps mishearing the way Taiwanese speakers mix Mandarin, Taiwanese, and Hakka (an example given live: mishearing '門號,' meaning phone number, as '摸乳'), so Taiwan Mobile chose to build its own localized speech model for customer service rather than rely on a general-purpose model.
An honest disclosure of a data gap
The official Day 1 R4 workshop agenda originally included a session by Wu Chun-hsing (吳俊興, National University of Kaohsiung 高雄大學) titled 'Building Your Own AI Think Tank with Kuwa,' but no documentation exists for that session and no supplementary material is available, so this article cannot discuss the actual 2024 development status of Kuwa, a homegrown Taiwanese open-source AI platform — it can only confirm that the session was part of the official agenda at the time. This itself is a microcosm of how Taiwan's local AI ecosystem sometimes can't even leave behind a record of its own content.
The shared conclusion of this axis: laying the three independent cases of Breeze 2.0, TAIDE, and Taiwan Mobile's self-built speech model side by side, what was holding back Taiwan's localized AI in 2024 wasn't model capability — it was every link of the data supply chain: too little licensed data (Huang Han-hsuan), too small a share of training corpora (Yang Cheng-lin), evaluation sets that don't exist (Tsai Tsung-han, the iKala case), and insufficient databases of accents and local usage (Tsai Chi-yen). These four problems, independent of one another and coming from speakers in four different industries, all point to the same structural bottleneck.
Axis 3: Inference cost and hardware strategy — using architectural innovation to route around "compute is too expensive"
Phison (群聯, Day 1 R2, on-site content): used a NAND flash SSD to expand GPU 'memory' rather than compute, cutting the on-premises training cost of a LLaMA 3.2 90B model from NT$30 million to under NT$1 million. The core argument: 'the training bottleneck has shifted from insufficient compute to insufficient VRAM, but NVIDIA sells compute and memory bundled together.'
MSI (Day 1 R1, on-site content): argued that the high-end discrete GPUs in gaming laptops (RTX 4050 and above can contribute over 250 TOPS) far exceed the NPU compute defined by typical AIPC specs, so 'gaming laptops are the most powerful AIPCs on earth' — again, repositioning existing hardware rather than waiting for cheaper dedicated chips.
Chang Chao-liang (張朝亮, National Center for High-performance Computing 國網中心, Day 1 R0 afternoon, on-site content): stated the compute procurement strategy plainly as 'don't bet everything on one option' — beyond existing H100/A100/V100/Grace Hopper units, they are procuring AMD MI300 and Intel Gaudi 3 to diversify and avoid being locked into a single supplier.
Liao Hung-yuan (廖弘源, Chip-based Taiwan Program 晶創計畫, Day 2 R0 afternoon, on-site content): echoed the same procurement logic — the National Center for High-performance Computing's compute buildout 'won't purchase only H100s; it will gradually add AMD partnerships,' to avoid 'putting all eggs in one basket.'
Cadence (Day 1 R3, on-site content): lowered compute cost from the 'design side' rather than the 'usage side' — using the Cerebrus reinforcement-learning engine to optimize chip PPA (power, performance, area), and the generative AI tool Voltus InsightAI to automatically fix IR drop, compressing what used to take weeks of manual tuning into 1–2 iterations. This is the same 'don't wait passively for cheaper compute — route around it with architectural and process innovation' logic showing up in the chip design toolchain.
A part that needs an explicit credibility note: the official Day 1 R2 agenda originally also included four sessions — Intel (Wang Tsung-yeh 王宗業), Inventec (Hsu Kuo-han 徐國翰, 英業達), Arm (Hsu Ta-yung 徐達勇), and AMD (Chang Ou-yu-hao 張歐佑豪) — whose topics were directly about 'using open software stacks, low-power IP, and edge NPUs to route around NVIDIA's closed ecosystem.' But none of these four sessions have any recorded on-site content: for Hsu Ta-yung's and Chang Ou-yu-hao's sessions, other attendees' live crowd-sourced notes are available (a real but simplified third-party record); for Wang Tsung-yeh's and Hsu Kuo-han's sessions, only after-the-fact publicly available background material remains (Intel's OPEA open platform, Inventec's VectorMesh low-power NPU IP, Arm's KleidiAI/Ethos-U85 edge NPUs, and AMD's MI300X with massive HBM3 plus the open ROCm software stack). If this background material is directionally accurate, these four foreign vendors' shared 2024 strategy — open stacks, memory prioritized over raw compute stacking — is another facet of the same idea as the diversified procurement logic from the National Center for High-performance Computing and the Chip-based Taiwan Program. But because there is no on-site content to confirm it, this article can only list it as 'possibly consistent in direction,' not evidence at the same level as the preceding five cases.
The core narrative of this axis: Taiwan's industry's mainstream response to 'compute is too expensive' in 2024 was not to wait for GPU prices to drop, but to operate at the system level — memory architecture (Phison), repositioning existing hardware (MSI), diversifying procurement strategy (the National Center for High-performance Computing, the Chip-based Taiwan Program), and automating the chip design process itself (Cadence). This is clearly different from a simple narrative of 'whoever hoards the most H100s wins.'
Axis 4: AI agents move from concept to deployment, but precision remains the threshold
Recent concrete plans and deployed cases
Huang Shih-chen (黃仕鎮, E.SUN Bank 玉山銀行, Day 1 R1; content from other attendees' live crowd-sourced notes, not a complete verbatim record) proposed a three-stage evolution framework — 'AI Embedded → AI Copilot → AI Agent' — positioning the agent as the end state.
Hu Te-min (胡德民, Far EasTone Telecommunications 遠傳電信, Day 1 R1; also from live crowd-sourced notes) mentioned that the company is building the 'FET Intelligence' (遠傳智靈) AI agent platform as one of its four core capability foundations.
Tsai Chi-yen (蔡祈岩, Taiwan Mobile 台灣大哥大, Day 1 R2, on-site content) laid out a concrete plan for an AI agent — 'a ChatGPT that understands the company's processes' — able to connect to internal systems to handle leave requests and schedule meetings.
Yeh Chia-shun (葉家順, MediaTek 聯發科技, Day 1 R0 afternoon, on-site content) envisioned that the next step for device-side and cloud is 'agent-to-agent' collaboration, drawing an analogy to today's Bluetooth pairing mechanism.
Liu Szu-tai (劉思泰, Qualcomm, Day 1 R0 afternoon, on-site content) mapped out a concrete evolution path for on-device LLMs: a plain chat-style LLM → multimodal input → context-aware contextual AI → eventually a robotic form capable of taking action.
Hsu Kai-cheng (許凱程, ITRI 工研院, Day 2 R2, on-site content) gave the most concrete and verifiable case of an agentic LLM deployment so far: after decomposing 'feeding an X-ray directly to a multimodal model' (a 50–60% error rate) into a task chain of 'first assess the clinical situation → then determine the image type → then determine findings within the image,' accuracy rose to over 90% — this is the only case at this year's conference with a quantitative before/after comparison of agent task decomposition.
Li Kun-mou (李昆謀, 91APP, Day 2 R1, on-site content) named the company's entire vector-retrieval-plus-LLM-spokesperson architecture 'Joy,' and predicted that future retail scenarios will move toward 'agent-to-agent' transactions, with agents appearing on both the consumer and seller sides.
Wu Wei-han (吳威翰, Accucrazy, Day 2 R3, on-site content) demonstrated a multi-agent marketing pipeline that is already a live product: one agent defines the audience, 'Moana' generates copy, 'Coco' schedules posting from virtual accounts, and 'Rebecca' converts it to voice for publishing.
Li Mu-yueh (李慕約, Day 2 R3, on-site content) gave a live demo of a computer-use type AI agent: after the instruction 'go to PC Home 24, buy a case of apple juice, pay with stored value, don't ask me anything during the process, just place the order,' the agent used computer vision to autonomously complete product selection, checkout, and payment, and the physical goods were eventually delivered.
Wong Hao-cheng (翁浩正, DEVCORE, Day 1 R3, on-site content) pointed out the ceiling of this axis: AI agents at this stage still can't achieve the 'precise one-shot kill' required of red-team tasks — they can't judge which vulnerability can deliver a one-shot kill, and lack strategic capabilities like 'feinting' that require reading the opponent's psychology, so human creativity is still needed to fill the gap.
A speculation on a different timescale — should not be conflated with the near-term plans above
The passage on AI agents in the opening keynote by Huang Chih-fang (黃志芳, Taiwan External Trade Development Council 外貿協會, Day 1 R0 morning) was actually a philosophical and societal-level long-term speculation — he cited the startup Altera's experiment running about 1,000 autonomous AI agents in Minecraft to build a simulated civilization, and speculated that AI agents with long-term memory might one day be seen by human society as a kind of 'immigrant,' triggering exclusionary reactions. This operates on a different timescale from the 'building this soon / building this now' product plans of the enterprise speakers above, and this article does not treat it as corroboration of a near-term industry trend — but it's worth remembering, since it offers a long-term vision of where this near-term trend might ultimately lead.
The dividing line on this axis: agents in 2024 didn't have one uniform level of maturity — they clearly split into two tiers. Scenarios with clear tasks, decomposable steps, and verifiable results (the X-ray reading task chain, retail recommendation systems, marketing content production pipelines, e-commerce ordering demos) already have concrete deployed cases, even commercialized ones. But tasks requiring autonomous judgment under high-risk, high-uncertainty conditions (the red-team penetration testing Wong Hao-cheng discussed) still can't be done. This threshold also echoes Axis 5 (cybersecurity) and the verification result for the 'AI agents going mainstream' prediction discussed below — which was judged 'not borne out' — worth reading side by side.
Axis 5: Offense and defense in cybersecurity co-evolve, with the defensive side maturing sooner
Wong Hao-cheng (翁浩正, DEVCORE, Day 1 R3) systematically surveyed generative AI's impact on the red-team and hacker toolchain: jailbreak services like WormGPT/FraudGPT, assistive tools like PentestGPT/BurpGPT, and voice-cloning attacks on smart home devices — attack tools themselves are already highly 'within easy reach.' But his conclusion was relatively conservative: AI still can't deliver a precise one-shot kill, remaining 'a great copilot' rather than a replacement. He also explicitly noted that 'AI actually helps the defensive side more,' sharing his own experience of using AI to analyze web logs and catch anomalous behavior that's hard for the human eye to spot.
Shen Yi-ting (沈宜婷, CHT Security 中華資安國際, Day 2 R3, appearing the next day) gave a live demo of automated penetration testing using a locally deployed, safety-restriction-removed Ollama model, successfully breaching a Linux host and obtaining the root password — proving that even without relying on major cloud vendor models, the technical bar for attack automation keeps dropping.
Yang Cheng-lin (楊政霖, CyCraft, Day 1 R1) spoke from the defensive-product perspective: after deploying an in-house LLM, interpreting malicious commands went from taking a senior analyst 40 minutes to taking 3 seconds. But the same talk also revealed that RAG frameworks set up by the defensive side themselves carry security risks — Black Hat Asia 2024 research found that of 51 framework combinations tested, 17 had vulnerabilities and 16 could achieve remote code execution — meaning the defensive side, even as it enjoys AI efficiency gains, is itself becoming a new attack surface.
Cheng Chen-mou (鄭振牟, Chang Gung University 長庚大學, Day 2 R0 morning) systematically explained four cryptographic techniques — TEE, FHE, MPC, and ZKP — that let a model owner and a data owner who don't trust each other still perform inference collaboratively and securely; this is the most systematic methodology on the defensive side at the data-privacy layer.
Cha Shih-chao (查士朝, National Taiwan University of Science and Technology 台科大, Day 1 R3) applied the existing ISO 27014 security governance framework and management review meeting mechanism directly to AI adoption scenarios — a concrete attempt at governance on the defensive side.
The asymmetry on this axis: the defensive side (product deployment, cryptographic tools, governance frameworks) already had a relatively complete methodology take shape by 2024; the attacking side, meanwhile, showed an intermediate state of 'tools within easy reach, but precision strikes still need a human.' Whether this asymmetry is only temporary is worth checking against the follow-up verification of Wong Hao-cheng's assessment discussed below.
Axis 6: Deployment methodology takes shape — treating LLM applications as engineering project management
Chang Wen-tien (張文鈿) of iHoward Information (愛好資訊, Day 2 R3) proposed a 5-level maturity model for LLM application development (from 'write it and don't test it' to 'automated prompt optimization'), turning what used to be seat-of-the-pants prompt tuning into an engineering process with graded standards.
Tsai Tsung-han (蔡宗翰, Day 2 R0 afternoon) stressed that evaluation prompts must be written by domain experts rather than engineers, and proposed a five-step process — 'continued pretraining → instruction data → fine-tuning → automated evaluation → periodic human-in-the-loop evaluation' — with clearly divided roles among a data team, an evaluation team, and a model development team.
Yang Cheng-lin (楊政霖, CyCraft, Day 1 R1) proposed five dimensions enterprises should assess before adopting an LLM — privacy, alignment, domain knowledge, localization, and security — and explicitly recommended 'starting small': beginning with non-critical functions rather than going all-in at once.
Hsu Chao-yung (許朝詠, China Steel 中鋼, Day 1 R0 afternoon) proposed a 'three levels of environment design for AI deployment' methodology — pure AI technology is prone to inaccuracy, using AI within a constrained environment improves things, and precisely designed environmental conditions are needed to reach over 99% accuracy — and described how the company formally shifted from a 'single technical goal' to a 'benefit-metric-driven' digital transformation governance model starting in 2023.
Cha Shih-chao (查士朝, Day 1 R3) directly repurposed the ISO 27014 management review meeting mechanism commonly used in security governance into a governance framework for AI adoption — again, the same idea of 'taming AI project uncertainty with existing engineering governance tools.'
No-code/low-code platforms appeared in a wave in 2024: Quanta Computer's (廣達) QOCA aim (compressing medical imaging model building from two weeks to two hours, Day 1 R2), Prophet Tech's (先知科技) 'Tian Ran Tong Hao' (天然通好, manufacturing defect detection, Day 2 R3), and MSI's AI Artist (Day 1 R1) — three platforms from three different industries, all uniformly targeting 'small and medium enterprises and hospitals that lack AI/IT talent,' rather than aiming to replace data scientists.
The significance of this axis: a clear shift in 2024 was that 'LLM application development' started being treated as an engineering project with a clear methodology, gradable evaluation, and governable auditing — rather than the experimental groping of 2022–2023. In a sense, this axis is the 'meta-methodology' foundation that makes the preceding five axes possible — without this engineering mindset, Axis 1's fine-tuning-versus-RAG selection debate, Axis 2's building of evaluation sets, and Axis 4's agent task decomposition could none of them be systematically verified and compared.
The predictions made in 2024, looked back on from 2026
This trend article originally verified 6 speaker predictions on its own; but a separate, dedicated fact-checking document, '2024 Conference Predictions and Follow-up Verification,' conducted more complete external verification on 8 conference predictions. To avoid the two documents' verdict tallies contradicting each other and misleading readers, this section relies primarily on that verification document's results; this article retains only its own narrative analysis and the supplementary items not covered by that verification document, with the two sets of verdicts listed separately and not combined into one tally.
5.1 Official verification results (cited from the separate verification document)
Verdict distribution: 3 items borne out, 2 items partially borne out, 1 item not borne out, and 2 items still to be observed due to insufficient data — 8 items in total; see the verification document for each item's full content and external sources. Only the highlights directly relevant to this article's six axes are listed here.
| Verification item | Verdict | Relation to this article's axes |
|---|---|---|
| California SB 1047 | Borne out — the developer camp succeeded in weakening the regulation; after the bill was vetoed, California instead advanced the weaker SB 53 (requiring only transparency disclosure, with no accountability mechanism) | Echoes Axis 5 (cybersecurity governance) and Axis 2 (the regulatory environment within the sovereign AI discourse) |
| The Basic Act on Artificial Intelligence and a standalone data governance law | Partially borne out — the Basic Act's scale and timeline roughly matched expectations (an 18-article draft became a 20-article act passed on its third reading on 2025/12/23), but the standalone 'Data Governance Act' advocated by Chiu Wen-tsung (邱文聰) did not become an independent law | Echoes Axis 2's data governance difficulties |
| The TAIDE project (multimodal) | Partially borne out — the text model kept iterating (now based on Gemma 3), but Huang Han-hsuan's original goal of advancing to multimodal in the second half of 2024 had still not seen a public release as of the verification date | Directly echoes the multimodal bottleneck Huang Han-hsuan discussed in Axis 2 |
| The Breeze series | Borne out — 'The Breeze 2 Herd of Models' was published on schedule in 2025/1, and the team continued its developer-oriented direction (a mobile-NPU quantized version, an expanded ASR product line) | Echoes Chen Yi-chang's achievement in Axis 2 |
| AI agents going mainstream | Not borne out — several enterprise speakers' near-term plans were too optimistic; third-party reporting (the Associated Press, the Wall Street Journal, Carnegie Mellon University, etc.) shows that in 2025–2026 enterprise deployment was in fact limited, agent performance was unstable, and returns on investment were limited | Directly echoes Axis 4; this article originally mistook Huang Chih-fang's long-term philosophical speculation for corroboration of a near-term industry prediction, a mistake corrected in Axis 4, with Huang Shih-chen, Hu Te-min, Tsai Chi-yen, and Yeh Chia-shun — who actually made near-term plans — added in |
| Enterprises 'no longer fine-tuning, switching to RAG' | Still to be observed, insufficient data — and the verification document specifically clarifies that at the 2024 conference itself, Yang Li-wei, Tsai Tsung-han, and Li Ming-ta were already three-way divided, and this was not an industry consensus that fine-tuning is obsolete | Echoes Axis 1; this article's Axis 1 has already faithfully presented the three-way divide without presupposing any side as the consensus |
| The Chip-based Taiwan Program | Still to be observed, insufficient data — the speaker's own disclosed first- and second-year budgets already lagged the originally planned pace, and there is a lack of official up-to-date 2025–2026 figures to make a determination | Echoes Axis 3 |
| Phison aiDAPTIV+ | Borne out — it has been upgraded to the 'aiDAPTIV multi-tier memory architecture,' with the Pascari SSD launched and expansion to iGPU PC platforms | Echoes Axis 3 |
One speaker-attribution discrepancy worth flagging: item 2 of the verification document attributes the remark that 'SB 1047's accountability was substantially weakened due to opposition from OpenAI, Meta, and the open-source camp, and it was called a toothless law' to Lee Yuh-jye's Sovereign AI talk. But this article's review of the relevant session record indicates that this content, which names scholars such as Fei-Fei Li and Yann LeCun, should actually belong to Hou Yi-hsiu's (侯宜秀) 'Human-Centered AI Governance' talk in the same session; Lee Yuh-jye's Sovereign AI talk covered TAIDE and Jensen Huang's etymological account of 'sovereign AI,' and did not mention SB 1047. This attribution discrepancy does not affect the verdict itself (the 'borne out' conclusion still holds), and is listed here for the reader's awareness, rather than unilaterally altering the verification document's content.
5.2 Three additional predictions this article verified on its own, not included in the verification document
The following 3 items are predictions this article selected on its own; the verification document does not include them, and their verdicts are not merged into the official 8-item tally above.
| Prediction | Speaker / Session | 2026 Verification Result | Verdict |
|---|---|---|---|
| Not fully convinced that the 'brute-force aesthetic' of stacking compute behind Transformers is the final answer — a future breakthrough not requiring today's scale of compute may emerge | Lee Yuh-jye (李育杰), Day 2 R0 morning panel | DeepSeek demonstrated in 2024/12 (V3) and 2025/1 (R1) that roughly one-tenth the compute and a training cost of about US$5.6 million could achieve performance close to GPT-4/o1-level, directly challenging the 'compute is the only path' assumption; NVIDIA's stock price plunged sharply as a result. | Partially borne out |
| If every app in the world ultimately converges on just two or three models making decisions behind the scenes, 'this would be the most dangerous moment in human history' | Chien Li-feng (簡立峰), Day 2 R0 morning panel | As of 2026, the market still has more than a dozen competitive frontier model providers coexisting — OpenAI, Anthropic, Google, Meta, xAI, Mistral, DeepSeek, Alibaba's Qwen, Moonshot, Z.ai, and others — and has not converged to two or three. | Not borne out |
| AI currently cannot achieve the 'precise one-shot kill' required by red-team tasks, but the direction of automated vulnerability-discovery competitions like DARPA's AI Cyber Challenge (AICC) is viewed favorably | Wong Hao-cheng (翁浩正), Day 1 R3 | DARPA's AIxCC proceeded as originally planned through to the 2025 DEF CON finals (total prize pool of US$29.5 million), and official follow-up coverage was headlined 'AI Cyber Challenge Proves Promise of AI-Driven Cybersecurity'; as for Wong Hao-cheng's conservative assessment that the precise one-shot kill still needs a human, this article found no public evidence that it has been overturned. | Borne out |
This article originally intended to do additional follow-up verification on the execution progress of Liao Hung-yuan's Chip-based Taiwan Program, but item 7 of the verification document already covers the same prediction, with a verdict of 'still to be observed, insufficient data.' The supplementary evidence this article found — the 2026 'IC Taiwan Grand Challenge' attracting nearly 600 international startup teams, and the 2026/6/1 launch of the next-generation supercomputer 'Chip-based 26 (Nano4)' — leans positive in direction, but likewise no precise year-by-year budget figures could be found, which is not enough to upgrade the verdict from 'still to be observed.' This article therefore adopts the verification document's conclusion and does not set an independent verdict.
The point here: laying out all 11 items from both documents combined, the ratio of 'borne out' to 'partially borne out / not borne out / still to be observed' comes out to roughly half and half — the direction of the speakers' 2024 judgments was right more often than it was a wild guess, but almost none of them precisely hit the timeline or the magnitude. Either reality outran the prediction (SB 1047 was outright vetoed, more drastic than merely 'weakened'; DeepSeek's efficiency breakthrough was more concrete and happened faster than merely 'doubting the compute-stacking path'; AI agent hype outpaced actual deployment by a wide margin), or the direction was right but there were no numbers to check it against (the Chip-based Taiwan Program's budget). This is the norm for predicting where the AI industry is headed: getting the direction right isn't hard; getting the magnitude and timing right is very hard.
Item-by-item verification: the full fact-check of 8 conference predictions
The following 8 verdicts are drawn in full from a separate, dedicated verification document, '2024 Conference Predictions and Follow-up Verification,' which compares each conference prediction against its subsequent actual development item by item; anything that could not be found is uniformly marked 'no public data found, cannot be determined,' with no speculative filling-in. The verification date is August 2026.
| Prediction | 2026 Reality | Verdict |
|---|---|---|
| Hou Yi-hsiu (侯宜秀, Secretary-General of the Taiwan AI Academy Foundation) predicted that Taiwan's Basic Act on Artificial Intelligence draft would have only 18 articles, focused on government agencies' AI risk assessment and regulatory review obligations — far smaller in scope than the EU AI Act — and would be sent to the Executive Yuan by the end of October; Chiu Wen-tsung (邱文聰, a research fellow at Academia Sinica's Institute of Law) argued a separate, standalone 'Data Governance Act' was needed to regulate the data-governance layer. | The Basic Act on Artificial Intelligence passed its third reading on 2025/12/23, expanded to 20 articles, designating the National Science and Technology Council (NSTC, 國科會) as the competent authority, with the Ministry of Digital Affairs (moda) responsible for building a risk-tiering framework. No public data shows Chiu Wen-tsung's proposed standalone 'Data Governance Act' has entered formal legislative proceedings — its spirit was partly absorbed into the Basic Act and into the Executive Yuan's concurrent amendments to the Personal Data Protection Act and a draft organic act for a Personal Data Protection Commission. | Partially borne out |
| Hou Yi-hsiu noted that California's SB 1047 (which would require frontier model developers to bear certain safety responsibilities) had its accountability substantially weakened during the legislative process due to opposition from OpenAI, Meta, and the open-source camp (including scholars such as Fei-Fei Li and Yann LeCun), and was called a 'toothless law'; Cha Shih-chao (查士朝, an NTUST professor), meanwhile, expected the bill (a decision expected by 9/30) to pass. | SB 1047 was vetoed by California Governor Newsom on 2024/9/29. California instead advanced the much narrower SB 53 (the Transparency in Frontier Artificial Intelligence Act), signed on 2025/9/25 and taking effect 2026/1/1, requiring only that frontier model developers with over $500 million in annual revenue and training compute exceeding 10^26 FLOPS publicly disclose safety framework documents and catastrophic-risk assessment reports — without the developer safety liability and accountability mechanism SB 1047 originally required. At the federal level, the same period also moved toward deregulation (Executive Orders 14179, 14365, 14409). | Borne out |
| Huang Han-hsuan (黃瀚萱, an associate research fellow at Academia Sinica's Institute of Information Science and an advisor to TAIDE's model-forging team) explained that TAIDE had already released a Llama-3-based version in 2024/04, with a goal of advancing to multimodal (at least being able to understand images) in the second half of 2024, but legally licensed, Taiwan-related image data was extremely scarce (only about 1,300 images on Wikipedia). | The text model kept iterating: Llama 3.1-TAIDE-LX-8B-Chat (2025/2, context length extended to 131K) → Gemma-3-TAIDE-12b-Chat (2025/8, switched to a Gemma 3 base) → Gemma-3-TAIDE-12b-Chat-2602 (2026/2) → Embeddinggemma-GTAIDE-300m-2605 (2026/6). But the official website makes no mention of any multimodal model having been formally released — the original 'second half of 2024' goal had still not been met as of the verification date. | Partially borne out |
| Chen Yi-chang (陳宜昌, MediaTek Research 聯發創新基地) introduced Breeze 2.0's new developer-oriented features (RAG, long context, Function Calling, Vision/Chart, the GFD speech-recognition model, and BreezyVoice TTS), and previewed an upcoming 'Breeze 2 Herd of Models' (3B/8B) paper including Vision and Function Calling. | The paper was published on schedule in 2025/1 (arXiv:2501.13921). The team kept expanding toward deployment optimization and speech — a mobile-NPU quantized version Breeze2-3B-8W16A-Instruct-mobile-npu (2025/8), Breeze-ASR-25 (2025/7), and Breeze-ASR-26 (2026/4). No larger flagship LLM version (such as a 'Breeze 3') was found to have been publicly released. | Borne out |
| Huang Shih-chen (黃仕鎮, E.SUN Bank 玉山銀行) proposed a three-stage 'AI Embedded → AI Copilot → AI Agent' framework; Hu Te-min (胡德民, Far EasTone Telecommunications 遠傳電信) built the 'FET Intelligence' AI agent platform; Tsai Chi-yen (蔡祈岩, Taiwan Mobile 台灣大哥大) planned an AI agent able to connect to internal systems for leave requests and meeting scheduling; Yeh Chia-shun (葉家順, MediaTek 聯發科技) envisioned device-and-cloud moving toward agent-to-agent collaboration. | Third-party reporting shows enterprise deployment remains limited: the Associated Press (2025/4) reported that real-world AI agent applications are scarce; the Wall Street Journal (2025/11) noted only a few companies with deployed agents have seen investment returns; Carnegie Mellon University research testing found no AI agent could complete most assigned tasks; Salesforce research found LLM agents performed poorly in CRM testing. Representative adoption is concentrated in software development, customer service, video game development (nearly 90% of game developers already use it), cryptocurrency trading, and government services. | Not borne out |
| Yang Li-wei (楊立偉, eLand Information 意藍資訊) argued that most enterprises no longer do fine-tuning, switching instead to a model-agnostic architecture plus RAG; Tsai Tsung-han (蔡宗翰), in the same panel, argued the opposite — 'training a small model beats a general large model plus RAG'; Li Ming-ta (李明達) argued a third position, that full-parameter fine-tuning beats RAG/LoRA. | No public industry survey data was found, so a full determination cannot be made; no survey data from analyst firms such as Gartner or IDC on the 2025–2026 industry adoption ratio of RAG versus fine-tuning could be found, nor any concrete record of how mainstream model vendors' recommendations evolved. | Still to be observed, insufficient data |
| Liao Hung-yuan (廖弘源, Director of Academia Sinica's Institute of Information Science and Deputy Executive Director of the Chip-based Taiwan Program) explained the original plan was NT$30 billion per year for 10 consecutive years, totaling NT$300 billion, but the speaker himself admitted that the actual budget for year 113 (2024) was only about NT$11.3–11.7 billion, rising to roughly NT$18 billion in year 114 (2025) — working out to only about 37–39% of the originally planned annual target. | No public data was found to determine whether 2025–2026 (years 115 and 116) accelerated to catch up to the original scale, or whether the ten-year NT$300 billion target has been adjusted — neither could be verified. | Still to be observed, insufficient data |
| Lin Wei (林緯, CTO of Phison Electronics 群聯電子) introduced aiDAPTIV+: an SSD designed with Phison's NAND expertise, controlled via aiDAPTIVlink, which splits data and transfers it to the GPU to overcome GPU memory capacity limits. | Phison has upgraded this technology path into the 'aiDAPTIV multi-tier memory architecture,' launching the dedicated Pascari SSD, with applications expanding from data-center GPUs to iGPU PC platforms. | Borne out |
Verdict distribution overview: 3 items borne out (California SB 1047, the Breeze series, Phison aiDAPTIV+), 2 items partially borne out (the Basic Act on Artificial Intelligence and the standalone data governance law, the TAIDE project), 1 item not borne out (AI agents going mainstream), and 2 items still to be observed due to insufficient data (enterprise fine-tuning vs. RAG, the Chip-based Taiwan Program).
Limitations of this observation
A note on scope: every trend in this article had to be corroborated by content from at least two speakers, spanning at least two different tracks, to qualify as a 'trend'; a single speaker's viewpoint wasn't enough to be included here. Where there are counterexamples or disagreements, they are stated explicitly — they are not hidden away just to make a trend look clean.
Don't treat this document as an investment pitch: the reality Taiwan presented at this conference was one of extremely limited resources (TAIDE's 72 H100 GPUs versus Meta's 24,000), a data supply chain jammed at every link (Huang Han-hsuan's mere 1,300 licensed images), and more than a third of Day 1 sessions leaving behind no on-site content at all. This article's six trend axes are all real limitations the speakers themselves laid bare, not a packaged collection of success stories.
Part of the evidence in Axis 3 comes from background supplementary material rather than on-site content: this article has already explicitly labeled the strategic direction of the four vendors — Intel, Inventec, Arm, and AMD — as 'not confirmed by on-site content, for reference only' within Axis 3 itself; readers citing it should be sure to trace back to that original label and not treat it as on-site remarks at the same evidentiary level as the other five cases.
The scope of verification in the predictions section was limited by tool quota: while writing this article, the search-tool quota available for verifying subsequent developments had already been used up, so verification switched to a small number of already-known or predictable URLs, and only 3 supplementary predictions with solid external sources — not already covered by the separate verification document — were included. A prediction originally considered, on the size of the NPU market, was dropped because no external source could clearly settle it, rather than forcing an unsupported verdict into print. As for which side, fine-tuning or RAG, wins between 2024 and 2026, the separate verification document has already determined this to be 'still to be observed, insufficient data,' consistent with the direction this article would have reached anyway, having found no external source, so this article does not list it again and instead directly cites that document's conclusion.
The division into axes in this article is the author's own synthesis: the six axes were reorganized from content drawn from 11 session notes, selected on the criterion of 'at least two speakers, spanning at least two tracks,' and this is not an official framing set by the conference itself, nor does it claim to exhaust every possible way of categorizing the material; the same body of remarks, cut along different lines, could yield a different set of axes.