Day 1 · R0 Intelligent Transformation (Afternoon)
Compute, chip, and enterprise leaders from Qualcomm, Hon Hai Research Institute, Chunghwa Telecom, NCHC, China Steel, and MediaTek share strategies for driving intelligent transformation across edge AI, autonomous driving, HPC infrastructure, and generative AI platforms.
At a glance
| Speaker | Talk | In one sentence |
|---|---|---|
| Liu Szu-tai (劉思泰), Vice President and President of Qualcomm Taiwan, Southeast Asia & ANZ | Intelligent Computing X Edge AI in Driving Innovation in the AI Era | He argues that only 'hybrid AI' — pushing compute/inference down to the device — can truly unlock AI's potential, using tools like AI Hub to lower the development barrier. |
| Li Yung-hui (栗永徽), Director, AI Research Center, Hon Hai Research Institute (HHRI) | From Prediction to Simulation: HHRI's Breakthroughs in Autonomous Driving | HHRI's three generations of self-developed models — QCNet, QCNeXt, and BehaviorGPT — have won world championships in a row at the CVPR Argoverse and Waymo Sim Agents challenges. |
| Wang Ching-hung (王景弘), Deputy Director, Chunghwa Telecom Laboratories | The strategy and development of AI 2.0 in the ICT Industry | Centered on 'Sailing Toward AI 2.0', he showcases applications like smart customer service, AI anti-fraud phone screening, and smart judicial systems, emphasizing that this wave of AI technology is mature enough to directly generate business value. |
| Chang Chao-liang (張朝亮), Director, National Center for High-Performance Computing (NCHC) | High-Performance Computing Infrastructures for Wide Adoption of AI in Taiwan | A comprehensive review of Taiwan's compute infrastructure buildout (the Taiwania series, the Jing Cui host), AI cloud services, and AI sovereignty strategy. |
| Hsu Chao-yung (許朝詠), Researcher, China Steel Corporation | The Journey of Digital Transformation in the Steel Industry Driven by AI | A shift from 'single-technology deployment' to a 'benefit-metric-driven' digital transformation methodology: human-machine collaboration, digital twins, hybrid cloud, and generative AI. |
| Yeh Chia-shun (葉家順), Associate Vice President, MediaTek | Boost Productivity & Innovation – MediaTek DaVinci GenAI Platform | Unveils the security-layered architecture, model/plug-in ecosystem, and the next step toward multi-agent for 'DaVinci' (達哥), MediaTek's enterprise-grade generative AI assistant platform. |
6 of 6 talks
1Intelligent Computing X Edge AI in Driving Innovation in the AI EraLiu Szu-tai (劉思泰), Vice President and President, Qualcomm Taiwan, Southeast Asia & ANZ
Qualcomm argues that AI's potential can only be truly realized by pushing compute and inference down to the device (Hybrid AI), while tools like AI Hub lower the barrier for developer adoption.
Key points
- Core thesis: truly unlocking AI's potential requires pushing compute/inference down to the device (On-Device/Edge), forming a three-tier 'Hybrid AI' architecture — 'device → Edge AI Server → cloud'; if all compute were sent to the cloud, the costs in power consumption, latency, and privacy would be enormous.
- Citing an estimate: if current AI development trends continue, by 2030 AI will consume about 3.5% of global electricity — equivalent to 30 years of electricity usage for the city of Shanghai (population 18 million).
- Citing industry estimates: the economic value generated by Generative AI is estimated at between US$2.6 trillion and US$4.4 trillion (roughly equivalent to the UK's 2021 GDP); three other different institutions estimate a range from US$207 billion to US$7.9 trillion.
- Explaining on-device compute requirements: to output 'acceptable' generative AI results, smartphones need about 7–10B parameters, PCs about 13–20B, automotive about 20–60B, and IP cameras about 7–10B.
- Explaining Qualcomm's heterogeneous computing architecture: CPU + GPU + NPU (Neural Processing Unit) + Sensing Hub work together, meeting the extremely low-latency real-time decision-making needed for scenarios like autonomous driving.
- Snapdragon 8 Gen 3 case: its compute power surpasses a supercomputer from the year 2000 while consuming less power than an LED, with image-processing performance up to 30x or more compared to cloud-based solutions.
- Reveals three tools for developers/SMEs: ① Qualcomm AI Hub (a cloud service with 100+ built-in quantized/optimized models, allowing simulation and validation in a virtual environment before deployment to physical hardware); ② an AI model optimization tool (making self-built models leaner and less compute-hungry before deployment); ③ the Qualcomm AI Stack (supporting multiple OSes, with built-in default libraries).
- Case study: the Brampton Cricket League in Canada used AI Hub to develop a cricket Decision Review System (DRS) umpiring-assistance system, cutting development time from 60 days to 1 day, and hardware benchmark testing time from several hours to 5 minutes.
- Automotive outlook: argues vehicles are moving toward the Software Defined Vehicle (SDV), with the in-car experience becoming like a 'PC on wheels'; illustrates with contextual AI — a vehicle proactively interpreting the meaning of a dashboard warning light, or proactively offering suggestions based on location data (e.g., asking if you'd like to take a photo when driving past the Taiwan Presidential Office).
- Depicts the evolution path of LLM device applications: simple chat-based LLM → multimodal input (voice/image) → environment-aware contextual AI → eventually evolving into robot forms with the ability to act.
Tech, products & figures
- Snapdragon 8 Gen 3
- Qualcomm's flagship mobile computing platform, with image-processing performance up to 30x that of cloud-based solutions.
- Qualcomm AI Hub
- A cloud model marketplace with 100+ built-in optimized/quantized models for developers to deploy and test.
- Software Defined Vehicle (SDV)
- The concept of vehicles defining user experience and continuous updates through software.
- 5G connectivity
- The speaker emphasized that Taiwan deployed 5G early, five to six years ago, as a foundational advantage for developing on-device AI applications.
Notable quotes
You have to bring this AI compute and functionality to the very front end... the device is the front-most piece of equipment to receive these things.
Large models are getting smaller and smaller, and yet extremely capable.
Q&A
- No Q&A (the moderator asked if there were questions from the floor, but since each talk was only 30 minutes, there wasn't enough time, so it was skipped and the moderator moved directly to introducing the next speaker).
Fact-check notes
- The speaker's self-introduced English name sounded approximately like 'STLU', possibly a spoken or transliteration gap; the specific full name could not be verified, flagged (unverified).
- Verified via WebSearch: Liu Szu-tai, Vice President and President of Qualcomm Taiwan, Southeast Asia & ANZ, previously served as President of Acer's Smart Products Business Group, and before joining Acer worked at Motorola for over 18 years — consistent with his own account that "previously I also had a lot of experience at places like Acer or Motorola."Sources：工商時報電子工程專輯
2From Prediction to Simulation: HHRI's Breakthroughs in Autonomous DrivingLi Yung-hui (栗永徽), Director, AI Research Center, Hon Hai Research Institute (HHRI)
Hon Hai Research Institute redefines autonomous-vehicle trajectory prediction and simulation using Transformer architectures, with three generations of models — QCNet, QCNeXt, and BehaviorGPT — winning multiple world championships in succession.
Key points
- Speaker background: originally taught in the Department of Computer Science and Information Engineering at National Central University, engaged in AI research; after Hon Hai chairman Liu Young-way took office in 2021 and drove the group's transformation from manufacturing to a 'technology-based' industry, he was invited to organize the Hon Hai Research Institute and serve as Director of its AI Research Center.
- Hon Hai Group's transformation strategy is '3+3': three new business directions (electric vehicles, digital health, and robotics, all high-CAGR industries) plus three foundational technologies (AI, semiconductors, and next-generation communications).
- The AI Research Center's team has only a dozen-plus members, yet has published 57 papers and filed 91 patents (including granted and pending) since 2021.
- Breaks down autonomous driving into five key technical problems: perception (identifying moving vehicles/pedestrians/bicycles and static objects like road markings/traffic lights/crosswalks), prediction (predicting the next move of surrounding vehicles), planning (planning one's own response); two other less visible but equally important problems are 'lowering the cost of training-data collection' and 'building a safe simulation training environment'.
- QCNet (CVPR 2023): reformulates trajectory prediction as an NLP-like 'next-token prediction' task — given a starting position, speed, and direction, a Transformer progressively predicts the position and speed for the next 0.5 seconds; the architecture includes scene encoding (using a Local Spacetime Coordinate System to improve data reuse efficiency) plus three layers of Cross/Self-Attention (time, surrounding vehicles, map) plus a final fine-tuning attention layer. It won first place worldwide in the Argoverse 1 and Argoverse 2 single-agent trajectory prediction challenges, and held the top spot on that leaderboard for over one year and four months before being surpassed (competitors included Uber, SenseTime, Peking University, and Waymo); the paper was published at CVPR 2023 and will be resubmitted to the PAMI journal.
- QCNeXt (CVPR 2024 Argoverse 2 Multi-Agent Challenge): expands from 'predicting one agent at a time' to 'jointly predicting the interacting trajectories of multiple traffic participants simultaneously' — a harder task — and likewise won the world championship.
- BehaviorGPT (Waymo Open Sim Agents Challenge, CVPR 2024): aims to build a 'simulation environment' to replace real-world road testing, letting autonomous-driving models train in simulation and greatly lowering training costs; it continues to use a Transformer architecture (agent/map data tokenization → triple Cross/Self-Attention: temporal, agent-to-map, agent-to-agent → decoding position/speed/angle); it proposes a 'next patch prediction' training paradigm (predicting an entire path 'patch' at once, rather than one time step at a time), improving encoding efficiency and decoding accuracy. The model won the challenge's world championship with only 3M parameters (the smallest among the competing entries in that event), echoing the importance of 'edge inference efficiency'; the paper has been accepted at NeurIPS 2024 and is expected to be published in November–December.
- Application outlook: the simulation scenarios built by BehaviorGPT can be used for autonomous-driving module R&D, as well as for developing traffic planning and management algorithms such as traffic-signal control and lane closure/dispatch.
- Track record over the past two years: 1st place at the 2024 CVPR Waymo Sim Agent Challenge, 2nd place at the Waymo Motion Prediction Challenge, and 1st place at the Argoverse 1/2 challenges.
Tech, products & figures
- QCNet / QCNeXt
- Hon Hai Research Institute's self-developed trajectory-prediction Transformer models, champions of the CVPR Argoverse challenges.
- BehaviorGPT
- Hon Hai Research Institute's self-developed traffic-simulation (Sim Agents) Transformer model, with only 3M parameters, champion of the Waymo Sim Agent Challenge.
- Argoverse 1 / 2
- Public autonomous-vehicle trajectory-prediction competitions and datasets, with teams including Uber, SenseTime, Peking University, and Waymo competing.
- Waymo Open Sim Agents Challenge
- An autonomous-driving simulation challenge launched by Google's Waymo.
Notable quotes
We turned the trajectory-prediction problem into a natural-language-processing problem... give it a starting position, plus speed and direction, so the Transformer can do a good job predicting the position for the next second or the next 0.5 seconds.
AI is a chain of strength building on strength — it keeps getting stronger and stronger, and Hon Hai Research Institute has gotten hold of that key.
Q&A
- No Q&A (during the transition, the moderator shared personal observations and reflections rather than posing formal questions).
Fact-check notes
- Li Yung-hui (栗永徽): verified that he currently serves as Director of the AI Research Center at Hon Hai Research Institute, formerly taught in the Department of Computer Science and Information Engineering at National Central University, and holds a Ph.D. in Computer Science from Carnegie Mellon University — fully consistent with his self-described background (formerly taught at National Central University, invited to organize Hon Hai Research Institute in 2021).Sources：鴻海研究院官網TEDxTaichung 講者介紹
- Claims that QCNet/QCNeXt won the CVPR Argoverse challenges and that the BehaviorGPT paper was submitted to NeurIPS were verified as true via WebSearch; however, whether the first author's affiliated institution is listed as 'Hon Hai Research Institute' on every public paper/GitHub page could not be fully cross-checked, flagged (partially unverified).Sources：QCNeXt 論文QCNet GitHub
3The strategy and development of AI 2.0 in the ICT IndustryWang Ching-hung (王景弘), Deputy Director, Chunghwa Telecom Laboratories
Chunghwa Telecom Laboratories, centered on 'Sailing Toward AI 2.0', showcases application cases such as smart customer service, AI anti-fraud phone screening, and smart judicial systems, emphasizing that this wave of AI technology has matured to the point of directly generating business value.
Key points
- The talk's core theme, 'Sailing Toward AI 2.0': argues that today's AI is completely different from the AI the speaker 'saw in university 30 years ago', with the key difference being that this wave of technology has matured enough to be 'directly applied and generate value', no longer confined to lab-stage research.
- The rollout framework splits into two tracks, '+AI' and 'AI+': '+AI' leans toward adding AI to existing businesses for efficiency gains (business needs driving new technology), while 'AI+' develops entirely new products and internal value with AI at the core.
- Showcased AI applications across multiple domains, two of which were especially impressive: AI anti-fraud phone screening (using voice analysis to prevent scam calls), and a smart judicial system (able to precisely cite relevant statutes when answering legal questions).
- Case: the smart customer-service system KM Copilot — based on the enterprise's internal knowledge base, letting users ask questions in natural language; before adoption customers faced long wait times, and after adoption answers arrive within 5 seconds, cutting wait time by at least 10%.
- Case: a medical AI assistant — application scenarios cover before, during, and after a patient's outpatient visit, with features including AI audio-to-text transcription (further convertible into multiple languages) and LLM-based diagnostic-record summarization.
- Technology maturity is directly driving a surge in compute demand: citing an example where the Director of the National Science and Technology Council's 'Department of Natural Sciences and Sustainability Research' (referred to on-site as the 'Natural Sciences Department') relayed academia's needs — where a PC cluster used to suffice, now GPUs are wanted, showing a qualitative shift in compute demand.
- Using this compute gap as a segue, he welcomed the next speaker — NCHC Director Chang Chao-liang — to the stage, foreshadowing Chang's subsequent talk that approaches AI from the angle of compute infrastructure.
Tech, products & figures
- KM Copilot
- Chunghwa Telecom's smart customer-service system, a natural-language Q&A assistant based on the enterprise knowledge base.
- AI anti-fraud phone screening
- An application that identifies and prevents scam calls through voice analysis.
- Smart judicial system
- An AI application that can precisely cite relevant statutes when answering legal questions.
- Medical AI assistant
- An application covering before/during/after outpatient visits, combining speech-to-text with LLM-based diagnostic summarization.
Notable quotes
The AI of today is very different from, say, the AI I saw in university 30 years ago; the main reason is... this technology has really matured — it's something that, in terms of value, you can pick it up and put it to use right away.
Q&A
- No Q&A.
Fact-check notes
- Wang Ching-hung currently serves or previously served as Deputy Director of Chunghwa Telecom Laboratories, and has held positions including Deputy General Manager of Chunghwa Telecom's Information Technology Division, Information Security Division, and Enterprise Customer Branch; verified via WebSearch as consistent in direction with public information, but no official press release was found to verify word-for-word his current 'Deputy Director' title and first-hand information about this talk, flagged (partially unverified).
- The definitions of '+AI' and 'AI+' are a reasonable interpretation based on the literal wording; the speaker did not specifically elaborate on the definitions, flagged as interpretation.
- Chunghwa Telecom established its 'AI 2.0 Enterprise Strategy Committee' in August 2023 (under then-Chairman Kuo Shui-yi), with public messaging focused on three aspects: what AI capabilities Chunghwa Telecom itself should have, how to help customers use AI, and how to move from 'AI technology empowering existing operating models' further toward 'AI-driven innovative business models' — echoing this talk's 'Sailing Toward AI 2.0' theme; the talk appears to be elaborating on this committee's strategic framework.Sources：中華電信研究院官網中華電信研究院｜人工智慧研發方向中華電：推動 AI 2.0 策略布局 打造創新雨林生態系
4High-Performance Computing Infrastructures for Wide Adoption of AI in TaiwanChang Chao-liang (張朝亮), Director, National Center for High-Performance Computing (NCHC)
NCHC Director Chang Chao-liang starts from the Gartner Hype Cycle to give a comprehensive review of Taiwan's compute infrastructure buildout (the Taiwania series, the Jing Cui host), cloud services, and AI sovereignty strategy.
Key points
- Citing Gartner Hype Cycle analysis: Generative AI and LLMs have passed the peak of inflated expectations and entered a phase of pragmatic development; but technologies such as RAG (especially Graph RAG), Multimodal, and AI TRiSM (AI Trust, Risk & Security Management) are still developing, and will be the focus over the next few years.
- Breaks Gen AI development into four aspects: ① model development (even when adopting open-source models, one must add proprietary data for customization); ② AI engineering tools (e.g., RAG needs to be deployed in an engineered way); ③ application use cases (the key to convincing management to invest); ④ enablement techniques and infrastructure — this talk focuses on the last one, i.e. 'compute' and 'specialized AI chips'.
- Observes the trend of major companies developing their own AI chips: Meta, Google, and others are investing in self-developed ASICs, partly because they don't want NVIDIA to dominate the entire ecosystem, and partly because general-purpose AI chips can't meet their power-efficiency and training-time requirements; but dedicated chips also carry the inherent risk that 'technology iterates too fast, and a chip may become outdated as soon as its design is finished'.
- Proposes the concept of 'AI + Simulation' (echoing the previous Hon Hai speaker): a pure AI model cannot predict physical environments (such as how a factory responds to natural conditions or production conditions), and must be combined with simulation; cites Jensen Huang's talk on building virtual worlds with NVIDIA Omniverse as an example.
- Current state of compute buildout: Taiwania 2 (launched in 2019) and Taiwania 3 have taken on important compute roles since AI took off in 2022; an additional 72 H100 GPUs (about 3.8 Petaflops) were procured to advance the TAIDE model; illustrating compute scale using a GPT-3-sized model: about 1,680 H100s can train a 175B-parameter model in 9.8 days.
- Reveals expansion plans for this year and next: a 16-Petaflops machine this year (some online in November, some in early next year), and next year a roughly 100-Petaflops-class 'Jing Cui host' (under the Executive Yuan-backed 'Chip-Driven Taiwan Industrial Innovation' program); also planning to introduce quantum-computing resources and to build a next-generation AI supercomputer center in the south (Shalun, in the Greater South).
- World supercomputer Top500 ranking: Taiwan currently ranks 17th (including the combined compute of Taiwania 1 and the newly built weather-forecasting computer).
- Emphasizes that 'cloud services' are just as important as hardware: NCHC is pushing a Container-based AI cloud-service portal, routing data into different protection tiers by sensitivity via a 'blue line/red line' (general data/sensitive data) split.
- Emphasizes AI Sovereignty: argues that AI development involving national cultural elements or sensitive data should not rely entirely on international cloud services, and needs a self-reliant, trustworthy cloud environment; related capabilities also extend to land governance, defense technology (such as satellite data from the Space Center), and data protection in the biomedical field.
- TAIDE model update: originally developed (in its second year) by a team led by Academia Sinica professor Lee Yuh-jye, now taken over and continuously maintained/updated by NCHC; data collection faces difficulties while trying to respect copyright as much as possible; currently prioritizing the digital-transformation needs of government agencies.
- Reveals three AI-developer service modes to launch in mid-October: ① one-stop (bring your own model onto the NCHC platform, covering everything from asking a question to getting the inference result); ② API (developers call NCHC's HPC resources themselves for inference, offering maximum flexibility); ③ a lightweight front-end interface (usable directly by those without front-end development skills).
- Hardware planning: currently has Grace Hopper, H100/A100/V100, plus L40S purchased for inference, and is in the process of procuring AMD MI300 and Intel Gaudi 3.
Tech, products & figures
- Taiwania 2 / Taiwania 3
- NCHC's existing supercomputers; Taiwania 2 went online in 2019 and now handles part of the AI compute workload.
- Jing Cui Host (精粹主機, 16 PF this year, ~100 PF next year)
- NCHC's newly built AI supercomputer, under the Executive Yuan's 'Chip-Driven Taiwan Industrial Innovation' program.
- TAIDE
- A Traditional Chinese sovereign large language model led by NCHC (originally by Academia Sinica professor Lee Yuh-jye's team).
- NVIDIA H100/A100/V100/Grace Hopper, L40S, AMD MI300, Intel Gaudi 3
- AI computing hardware NCHC currently owns or plans to procure.
- AI TRiSM
- An AI Trust, Risk & Security Management framework proposed by Gartner.
- Greater South (Shalun) AI supercomputer
- An AI computing center the government plans to build in the south.
Notable quotes
If you're not ready for AI, you need to be AI ready.
Not everyone is... but when everyone's scrambling to buy something, there must be a reason for it.
Q&A
- No Q&A.
Fact-check notes
- Chang Chao-liang: verified via WebSearch that his current position and background are accurate: currently Director of the National Center for High-Performance Computing, holds a Ph.D. in Mechanical Engineering from Pennsylvania State University, formerly a Senior Research Scientist at NASA Langley Research Center, and took over as NCHC Director in April 2022.Sources：自由時報國網中心官網
- The 'Chip-Driven Taiwan Industrial Innovation Program' has a ten-year budget of NT$300 billion, with NT$4 billion allocated in fiscal year 114 (2025) for NCHC to build out compute capacity.
5The Journey of Digital Transformation in the Steel Industry Driven by AIHsu Chao-yung (許朝詠), Researcher, China Steel Corporation
China Steel shifts from a 'single-technology deployment' to a 'benefit-metric-driven' digital transformation methodology, using four pillars — human-machine collaboration, digital twins, hybrid cloud, and generative AI — to address the labor shortage.
Key points
- Speaker background: worked at Academia Sinica for 13 years, then moved to China Steel in 2014 and has remained there since, a period that happens to overlap with a decade of booming AI development; mentions Liao Hung-yuan (Deputy Executive Director, Academia Sinica / Chip-Driven Taiwan Industrial Innovation Program, also a speaker on the following day's agenda) as an important collaborator in the field of imaging technology.
- China Steel has adopted multiple AI technologies on its production lines and sales side since 2014: taking the 'unmanned overhead crane' as an example, it is China Steel's third best-selling product outside its core steel business (the top two being mooncakes and dumplings), with more than a dozen units already exported to mainland China.
- Review of past problems: early AI projects were mostly targeted at 'a single technology/single production line' (e.g., 'building an unmanned overhead crane' was itself the goal), lacking a quantifiable end-benefit metric — so even though AI had been deployed, it was hard to reflect this in the company's overall operating performance.
- Starting in 2023, the focus shifted to a 'digital transformation' goal, establishing a Digital Transformation Task Force (formerly the 'Intelligent Production & Sales Committee'), newly incorporating the IT and administration departments — because unblocking data flow and getting employees accustomed to working with generative AI are both key to digital transformation.
- Commissioned ITRI to survey the digital-transformation directions of the world's top seven steelmakers, distilling four main themes: improving production efficiency, intelligent operations (continuing existing Intelligent Production & Sales Committee work), low-carbon manufacturing, and new-material development (already handled by the company's ESG team, so the technology team does not duplicate this effort); the technology team therefore focuses on 'improving production efficiency' and 'intelligent operations'.
- Identifies the core challenge as labor shortage: Taiwan's demographic dividend is expected to disappear by 2028, with the labor force shrinking by 2 million; in recent years China Steel has also had to compete with the tech industry (such as TSMC) for talent, hiring in the south has grown increasingly difficult, and this has also given rise to an 'experience transfer' problem (senior employees' experience is hard to hand over to new hires).
- To address the two goals of 'accelerating experience transfer' and 'improving work efficiency', four technologies were introduced: human-machine collaboration, digital twins, hybrid cloud, and generative AI.
- Case 1 (human-machine collaboration — billet overhead crane): a purely automated crane, because AI can only operate 'two axes simultaneously' (a human can operate three axes simultaneously), reduced efficiency by about 30%; switching to a human-machine collaboration model was likened to the difference between Google's robotaxi and Baidu's 'Apollo Go' (人機協作、遠端接管) approach, with the latter reaching application maturity faster.
- Three levels of environment design for AI deployment (an important methodology proposed by the speaker): ① pure AI technology: often inaccurate because the environment isn't controllable (example: a camera directly counting the number of billets being moved fails to recognize them because the entire background is billets); ② AI within a restricted environment: adding a laser line for foreground/background separation (triangulation laser ranging) improved things, but was still affected by environmental variation (too small an angle between the laser line and camera); ③ precisely designed environment conditions: optimizing the laser line's mounting position, adding electromagnetic-chuck detection to exclude out-of-range laser lines, ultimately achieving an accuracy rate above 99%.
- Case 2 (remote collaboration — coil overhead crane): a 5G base station was set up in the crane storage area to bring the control signal down to a ground-floor control room, where operators use the same console as on the crane to operate it remotely; combined with PLC adjustments, 'gripping/releasing' (better done by humans, requiring three simultaneous axes) was divided from 'personnel detection, collision avoidance, and coil serial-number recognition' (better done by AI via computer vision), cutting a single crane's operating labor-time to 20% of the original and ultimately achieving 'one person operating multiple cranes'.
- Case 3 (digital twin — crane-operator training + LiDAR virtual warehouse): ① a real warehouse was built into a virtual warehouse with virtual cameras, letting crane operators train in a virtual environment with a view matching the real crane, speeding up training; ② slab billet crane case: LiDAR lead time once stretched to 6 months to a year, so a virtual storage area was first built with Point Cloud (billets simplified into 5–6-face geometric shapes) and the object-detection system was developed before the hardware even arrived — once the hardware arrived, only minor angle adjustments were needed to go live.
- Case 4 (digital twin — blast furnace burden-distribution simulation): adopted NVIDIA Omniverse to build a blast-furnace burden-distribution simulation, visualizing burden conditions that were previously 'invisible inside the furnace', used for real-time comparison and simulation-based judgment.
- Generative AI cases: ① querying an SQL Server database using natural language (no need to write SQL syntax), addressing queries such as the connection-cable length of sintering equipment; ② the maintenance-guidance system 'Engineering Little Housekeeper' (工程小管家): turning maintenance documents left by senior employees into a Q&A system for new employees to learn equipment-maintenance experience.
- The speaker candidly admits: it's hard for an enterprise's in-house generative AI applications to outpace the iteration speed of native large players like OpenAI, so China Steel's positioning is 'digital empowerment' — embedding capabilities/knowledge employees lack directly into the system and letting a large number of employees find applications that fit their own work contexts through use, rather than trying to overtake others with self-developed model capability.
Tech, products & figures
- NVIDIA Omniverse
- The digital-twin platform China Steel uses for blast-furnace burden-distribution simulation.
- Point Cloud + LiDAR
- Virtual-storage-area modeling technology used to detect billet stacking height/width for overhead cranes.
- Triangulation laser ranging
- A physical method using a laser line to separate foreground/background to detect the number of billets being moved.
- 5G base station
- Used for transmitting remote-control signals for overhead cranes.
Notable quotes
When it comes to deploying AI, most of the issues are environmental... pure AI technology, AI within a restricted environment, and AI with a precisely designed environment — these are three levels.
You can think about the difference between Google developing robotaxis and mainland China's Apollo Go (蘿蔔快跑) catching up in a very short time — that's the difference between pure AI and human-machine collaboration.
Q&A
- No Q&A.
Fact-check notes
- Hsu Chao-yung (China Steel Researcher): verified via WebSearch that he currently serves as a Researcher at China Steel's Green Energy & System Integration R&D Department, and previously presented 'Applications of Artificial Intelligence in the Steel Industry' at the 2020 Taiwan AI Academy Conference, consistent with his self-described background (13 years at Academia Sinica, moved to China Steel in 2014).Sources：2020 AI Academy 議程頁iThome 報導
- The 'Professor Liao Hung-yuan' mentioned by the speaker was confirmed to be Liao Hung-yuan, Deputy Executive Director of the Chip-Driven Taiwan Industrial Innovation Program — consistent with his identity as a speaker on the following day's agenda, with no contradictions found.
6Boost Productivity & Innovation – MediaTek DaVinci GenAI PlatformYeh Chia-shun (葉家順), Associate Vice President, MediaTek
MediaTek unveils the security-layered architecture, model/plug-in ecosystem, and next step toward multi-agent for 'DaGe' (MediaTek DaVinci), its enterprise-grade generative AI assistant platform.
Key points
- Opening framing: the two major business objectives for enterprise AI adoption — cost down (raising employee productivity) and value up (monetizing added value on existing products/services).
- 'DaGe' (達哥, MediaTek DaVinci) is an enterprise-grade generative AI assistant platform MediaTek launched in 2023; nicknamed 'DaGe' from its Chinese translation 'Da Vinci' (達文西), it has now become one of MediaTek's official product names; the target is benchmarked against the L1–L5 autonomous-driving grading scale, with the ideal end state being a 'Jarvis'-level super-intelligent assistant (L5), currently sitting at roughly L2.
- Three core principles of the platform: Security, Flexibility, and Simplicity — only with these can a network effect form.
- Security-layered architecture: three sets of environments split by department sensitivity — non-core business (can use external cloud models such as GPT-4o, Claude, Gemini, connected via a Virtual Private API), core business (such as IC design RTL code, 6G R&D — cannot access the external internet, only on-premises models such as Llama, Mistral, DeepSeek, requiring a self-built inference server); all access goes through a unified API layer for partitioning and information filtering.
- The Model Hub comes with dozens of built-in models; there are over a hundred plug-ins, covering integrations with systems such as Wiki, SharePoint, ERP, and Oracle, all with default templates for quick configuration.
- Emphasizes Prompt Engine and log-data governance: logs are used for auditing (checking for data leakage), billing/chargeback (quota and pricing across different units and models), and feedback for training (optimizing the Prompt Engine, RAG plug-ins, or future model fine-tuning).
- Rollout path (enterprise-adoption methodology): recommends 'using AI' first (with ready-made scenarios such as low-sensitivity work in finance, HR, and legal) rather than rushing to 'build AI' (training your own models); recommends setting 'at least a 50% productivity increase for a unit' as the first checkpoint, then using that successful case unit to persuade the whole company to roll it out.
- Case (finance): using a 'visibility × impact' (plus feasibility) four-dimensional quadrant to screen for pain points, they developed an 'expense-reimbursement customer-service assistant', which automatically understands questions from employees in different locations/languages based on permissions and calls a backend knowledge base; field testing showed 80–90% of the work could be handled by AI, and a team originally of 5 could potentially shrink to 1–2 people maintaining it after 3 months.
- More cases: a peer/competitor analysis assistant (automated web crawling + database queries), a CAPEX/OPEX query assistant for legal/HR/finance (combined with chart generation), a 'strategy consultant' assistant built by turning a retired NCCU professor's open 'strategy matrix' knowledge base into a RAG-plus-knowledge-base system (able to query in real time and analyze the latest industry developments from Microsoft/Mistral/OpenAI, etc.), and an internal 'Scholar Master' paper assistant (automatically fetching papers from sources like arXiv and comparing/analyzing them).
- Value-up case: embedding 'DaGe' capabilities into MediaTek's existing ICs/products (smartphones, smart speakers, white goods, automotive, kiosks, glasses, etc.), giving end devices a personalized assistant capability that 'understands human language, understands you' — likened to Apple Intelligence's 'cloud large model + on-device small model' collaboration architecture; looks ahead to agent-to-agent collaboration (similar to today's Bluetooth pairing mechanism), while also cautioning that if the interoperability design between on-device and cloud agents is done improperly, there is a risk of being hacked.
- Ecosystem status: currently a three-digit number of clients are testing it, and a two-digit number are already live; partner system integrators include Cloudmosa (雲豹), Systex (精誠), Hongchi (鴻基), Taiwan Mobile, and NTT; on the education side, an education cloud has been launched in partnership with the Taiwan AI Academy, the Taiwan Institute of Economic Research, and Microsoft; the model library covers the self-developed Breeze (BreeXe, a Traditional Chinese LLM), Llama, Mistral, DeepSeek, and NVIDIA NIM; also showcased integration cases with ecosystem partners such as Metaage (virtual humans, integrated with the LINE and Messenger APIs), Nengdian Animation (能點動畫, virtual human 'DaMei' 達妹), the Zhihui Huisu Exhibition Center (知識會肅展中心, AIOT development boards), and Yili Information (以利資訊, rapid API deployment).
Tech, products & figures
- MediaTek DaVinci (達哥)
- MediaTek's enterprise-grade generative AI assistant platform, launched in 2023.
- Breeze (BreeXe)
- A Traditional Chinese large language model released by MediaTek Innovation Base (聯發創新基地).
- Apple Intelligence
- A reference case the speaker used to draw an analogy to a 'cloud large model + on-device small model' collaboration architecture.
- NVIDIA NIM
- An inference microservice supported by the DaGe model library.
- Metaage, Nengdian Animation (能點動畫), Yili Information (以利資訊), Zhihui Huisu Exhibition Center (知識會肅展中心)
- DaGe ecosystem partners (API integration, virtual humans, AIOT development boards, etc.).
Notable quotes
Are you actually trying to build AI, use AI, or govern AI? (echoing a view from Provost Tsai Ming-shun (蔡明順))
For this assistant to be effective, it has to understand human language and get things done.
Q&A
- No Q&A (after the moderator's closing remarks, the end of the day's agenda was announced directly, with see-you-tomorrow).
Fact-check notes
- The speaker's title is Associate Vice President Yeh Chia-shun of MediaTek; verified via WebSearch that he is 'Associate Vice President of MediaTek's AI and Data Engineering Division', leading the team that built the DaGe platform, consistent with this talk's content.Sources：工商時報天下雜誌MediaTek 官方新聞稿
- An official press release reveals that 'the DaGe ecosystem already has 20+ high-tech companies, 6+ financial companies, 10+ consumer brands, 3 telecom operators, and 5 traditional-industry companies participating', consistent in direction with the claim of 'a three-digit number of clients testing and a two-digit number already live' (the latter being a more recent figure the speaker verbally supplemented on-site at the conference, for which no separately verifiable public report was found; flagged as the speaker's own verbal disclosure, not fully verified).