Glossary

TAIDE (Trustworthy AI Dialogue Engine, 可信任 AI 對話引擎) is a Traditional Chinese large language model program led by Taiwan's National Science and Technology Council (NSTC), built via continued pre-training (CPT) on Meta's open-source Llama 3 to create a trustworthy dialogue engine aligned with Taiwan's language context and cultural values; it released the 8-billion-parameter Llama3-TAIDE-LX-8B in April 2024. Yuh-Jye Lee (李育杰, researcher at Academia Sinica's Center for Information Technology Innovation) leads the program, and Richard Tzong-Han Tsai (蔡宗翰, researcher at Academia Sinica's Institute of Sociology / professor at National Central University) convenes the core model training and RLHF team. It appeared in Day 1's Opening & Keynote session (Tung Tzu-hsien (童子賢) introducing the sovereign AI context), Day 2's AI Governance panel (Yuh-Jye Lee discussing program resources, timeline, and lessons learned), and Day 2's Smart Transformation & Memorial Lecture session (Richard Tzong-Han Tsai, the EduTAIDE case).

Sovereign AI (主權 AI) is a core concept that recurred throughout the conference without a single agreed definition: Tung Tzu-hsien's (童子賢) opening keynote took a defensive, nationalistic tone, arguing for building a "sovereign AI with defensive capability"; Yuh-Jye Lee (李育杰, TAIDE program lead), in the governance panel, explicitly opposed Taiwan "training from scratch," arguing the priority order should be "compute > data > talent"; Jane Yung-jen Hsu (許永真) corrected herself mid-panel, revising "we need our own large language model" to "we don't necessarily need to build our own — let the world help build it for us, but the 'data' should be contributed by us." These three positions correspond to a "national narrative," a "pragmatic path under resource constraints," and "collaborative co-building" respectively, marking the key fault line to watch when interpreting this conference's sovereign AI discourse. It appeared in Day 1's Opening & Keynote (Tung Tzu-hsien) and Day 2's AI Governance panel (Yuh-Jye Lee, Jane Yung-jen Hsu, and Lee-Feng Chien (簡立峰) in panel discussion).

DIU / DIO (國防創新單位／國防創新辦公室, Defense Innovation Unit / Defense Innovation Office) are dual-use reference cases cited by Yuh-Jye Lee (李育杰) in the governance panel: DIU (Defense Innovation Unit) is a unit established roughly seven or eight years ago by the US Department of Defense to match Pentagon needs with mature Silicon Valley startup technology; DIO (國防創新辦公室, Defense Innovation Office) is Taiwan's corresponding unit, established by the Ministry of National Defense under Minister Chiu Kuo-cheng (顧立雄). Lee used this to encourage Taiwanese academia and industry to consider dual-use applications such as drones and decision-support systems, rather than avoiding them entirely out of aversion to the "military-industrial" label. It appeared in Day 2's AI Governance panel (Yuh-Jye Lee).

Human-centered AI governance (以人為本的 AI 治理) is a governance framework proposed by Hou Yi-hsiu (侯宜秀), built on the core metaphor of "first do no harm, then seek benefit": doing no harm means safety (AI Safety / value alignment), while seeking benefit means establishing a new order for wellbeing and happiness. It covers four areas: AI Safety alignment issues; a comparison of copyright rulings (the US and China take markedly different stances on copyright for AI-generated content); allocation of infringement liability; benefit distribution (the Hollywood writers' strike); and a comparison of global regulatory approaches (Taiwan's draft Basic Law on AI, the EU AI Act, the US NIST framework, and China's censorship mechanisms). It appeared in Day 2's AI Governance panel (Hou Yi-hsiu).

The Chip-based Industrial Innovation Program (晶創臺灣方案, officially 晶片驅動臺灣產業創新方案, CbI Program) is a cross-ministry program led by Taiwan's National Science and Technology Council (NSTC), launched in 2024 (ROC year 113). It originally planned NT$300 billion over ten years across four pillars: "AI + chip-driven industrial innovation / recruiting talent / accelerating technology integration / attracting international startups," and commissioned the National Center for High-performance Computing (NCHC) to build national-level compute infrastructure; Hong-Yuan Mark Liao (廖弘源, Director of Academia Sinica's Institute of Information Science) serves as Deputy Executive Director. It appeared in Day 2's Smart Transformation & Memorial Lecture session.

The Traditional Chinese AI Open Source Grant (繁體中文 AI 開源實踐計劃) is a grant program hosted by the Frontier Foundation (開拓文教基金會) and executed by g0v Lingtsan School (g0v 零時小學校), running from January 2023 to April 2024 across three application rounds, encouraging developers to release Traditional Chinese corpora, benchmarks, and open-source models. The Legislative Yuan open-data project presented at this conference by Wang Hsiang-Jung (王向榮) was one of the grant recipients. It appeared in Day 2's "Where Does Good Data Come From" workshop.

The Draft Data Governance Act (資料治理法草案) is a policy proposal put forward by scholars including Chiu Wen-Tsung (邱文聰) and Ching-Yi Liu (劉靜怡) of Academia Sinica's Institute of Law, under the NSTC's AI Humanities and Rule of Law research project. It advocates a "rights management system" that lets data subjects dynamically exercise consent/opt-out rights, replacing the approach of vaguely labeling personal data as "non-personal data." As of the verification date (August 2026), it had not yet entered formal review by the Legislative Yuan. It appeared in Day 2's workshop session.

The Basic Law on Artificial Intelligence (人工智慧基本法, Taiwan) passed its third reading in the Legislative Yuan on December 23, 2025 (20 articles in total), with the NSTC designated as the competent authority; "privacy protection and data governance" is one of its seven guiding principles. At the time the conference was held (September 2024), it was still at the draft stage, addressing only application-layer rules. It appeared in Day 2's workshop session (noted as a post-hoc verification).

The Privacy-Enhancing Technology Application Guidelines (隱私強化技術應用指引, PET Guidelines) is a policy document published in 2024 by Taiwan's Ministry of Digital Affairs (moda), promoting application scenarios for privacy-enhancing technologies (PETs) such as differential privacy and synthetic data; Yu Chia-Mu (游家牧, associate professor at National Yang Ming Chiao Tung University) participated in drafting it. It appeared in Day 2's workshop session.

The Ministry of Health and Welfare's "Three Arrows" policy (衛福部「三支箭」政策) comprises three complementary measures to drive real-world adoption of smart healthcare: (1) an AI Trust Center (opening the AI black box); (2) a Certification Center (helping hospitals obtain TFDA certification for AI models); and (3) an NHI (National Health Insurance) sandbox (evaluating benefits and pursuing NHI reimbursement after medical AI goes live). It appeared in Day 2's Platform Services / Smart Healthcare session.

The Taiwan Smart Healthcare Alliance (台灣智慧醫療聯盟, TSHA) is a cross-hospital AI model validation/certification program led by the NSTC, running as a three-year project from 2021 to 2024, with Taichung Veterans General Hospital serving as the overall coordinator, integrating AI models from medical centers nationwide and promoting cross-hospital validation via federated learning. It appeared in Day 2's Platform Services / Smart Healthcare session.

Taiwania / Chingtsui (台灣杉／精粹主機, National Center for High-performance Computing supercomputers) are the primary compute facilities of the National Center for High-performance Computing (NCHC): Taiwania 2 (launched 2019) and Taiwania 3 handle existing AI compute workloads. To meet the training demands of sovereign AI projects like TAIDE, NCHC additionally procured 72 H100 GPUs (about 3.8 petaflops) and plans to build a new generation of "Chingtsui" (精粹主機) machines at the 16 PF and 100 PF tiers during 2024–2025, part of the compute layout under the Chip-based Industrial Innovation (CbI) Program. It appeared in Day 1's Smart Transformation session (Chang Chao-Liang (張朝亮), on compute build-out details) and Day 2's Smart Transformation & Memorial Lecture session (Hong-Yuan Mark Liao (廖弘源): the CbI Program's FY2025 (ROC year 114) budget of about NT$4 billion entrusted to NCHC for compute build-out, mixing H100 and AMD procurement).

Breeze 2.0 is a Traditional Chinese multimodal language model (3B/8B parameters) continually pre-trained on Llama 3.2 by MediaTek Research (聯發創新基地), strengthening Traditional Chinese language and cultural understanding while adding visual understanding and function-calling capabilities; the same family also includes the larger 45-billion-parameter BreeXe model, which is adopted by the MediaTek DaVinci platform. It appeared in Day 1's Deep Tech session (Chen Yi-Chang (陳宜昌)).

GFD (Generative Fusion Decoder) is a training-free shallow fusion framework developed by MediaTek Research that fuses the outputs of an LLM with a speech recognition (ASR) or OCR model directly at inference time, using "contextual prompting" to make recognition results better match domain-specific terminology without retraining the model. It appeared in Day 1's Deep Tech session (Chen Yi-Chang (陳宜昌)).

BreezyVoice is a Taiwanese-accented text-to-speech (TTS) model co-developed by MediaTek Research and National Taiwan University, capable of voice cloning from just about 10 seconds of recording; it adopts the CosyVoice architecture combined with Zhuyin (Bopomofo) input to disambiguate polyphonic (heteronym) characters. It appeared in Day 1's Deep Tech session (Chen Yi-Chang (陳宜昌)).

The TAIDE multimodal model (TAIDE 多模態模型) is a sub-project that expanded TAIDE from text-only to multimodal (image understanding) starting in the second half of 2024, adopting a LLaVA-like "late fusion" architecture (image features are projected into the existing LLM's semantic space via a projector). Its biggest bottleneck is the extreme scarcity of legally licensed, Taiwan-relevant training images (Wikipedia has only about 1,300). It appeared in Day 1's Deep Tech session (Hen-Hsen Huang (黃瀚萱)).

QCNet / QCNeXt / BehaviorGPT (鴻海研究院自駕模型三部曲, Hon Hai Research Institute's autonomous-driving model trilogy) are Transformer-based trajectory prediction/simulation models for self-driving cars developed in-house by Hon Hai Research Institute's AI Research Center: QCNet (single-agent trajectory prediction, CVPR 2023) and QCNeXt (multi-agent joint trajectory prediction, CVPR 2024) successively won the Argoverse Challenge world championship; BehaviorGPT (only 3M parameters) won the Waymo Open Sim Agents Challenge, aiming to replace real-world road testing with simulated environments. It appeared in Day 1's Smart Transformation session (Yung-Hui Li (栗永徽)).

EduTAIDE is a Mandarin-teaching-material authoring copilot built on TAIDE technology, developed by Richard Tzong-Han Tsai's (蔡宗翰) team in collaboration with the Ministry of Education; it can write lesson texts, generate vocabulary lists, design grammar exercises, and auto-generate test questions. Officially published figures show a 76% pass rate, outperforming GPT-4o's 51%. It appeared in Day 2's Smart Transformation & Memorial Lecture session.

Kuwa (GenAI OS) is an open-source generative AI operating system/platform developed by the Computer Science and Information Engineering team at National University of Kaohsiung, letting users connect to multiple LLMs (GPT, Gemini, LLaMA, etc.) without writing code to build personalized knowledge bases and AI applications; Associate Professor Wu Chun-Hsing (吳俊興) taught "Building Your Own AI Think Tank with Kuwa" in a Day 1 workshop.

VeriOsteo (智骨篩) is an AI developed jointly by Acer Medical (宏碁智醫) and Taichung Veterans General Hospital that performs opportunistic osteoporosis screening from a single chest X-ray; it has obtained TFDA clearance, with publicly validated clinical sensitivity of about 88.7% and specificity of about 89.4% (the speaker's own account during the talk cited an accuracy of about 98%, which diverges somewhat from the published data). It appeared in Day 2's Platform Services / Smart Healthcare session.

BOBI is a medical speech-transcription LLM tool developed by ITRI (Industrial Technology Research Institute) and Hsu Kai-Cheng's (許凱程) team, used for speech-to-text of initial outpatient consultation notes and nursing shift-handover records; after training on professional medical corpora, its recognition error rate dropped from 30–40% to below 6%, the threshold needed for clinical usability. It appeared in Day 2's Platform Services / Smart Healthcare session.

QOCA aim (Quanta Open Care AI) is a no-code/low-code medical AI cloud platform from Quanta Computer. Version 1.0 focused on medical imaging and structured-data modeling (including explainability heatmaps), while 2.0 added LLM/RAG/fine-tuning capabilities; it has partnered with hospitals including Taipei Medical University, Veterans General Hospital, Cathay, and Fu Jen Catholic University hospitals, compressing modeling time from two weeks to under two hours. It appeared in Day 1's Platform Services session.

aiDAPTIV+ is a GPU memory expansion solution from Phison (群聯電子) that uses NAND flash SSDs plus system DRAM as an extended memory tier for GPUs, breaking through VRAM capacity limits; it claims to cut the on-premises training cost of a 90B-parameter model from NT$30 million to less than one-tenth of that. Phison also uses this same on-premises compute stack to build its internal employee AI assistant, "Javis." It appeared in Day 1's Platform Services session.

The MediaTek DaVinci GenAI Platform (聯發科技達哥, nicknamed "Dage/DaVinci") is an enterprise-grade generative AI assistant platform that MediaTek began developing in 2023 and formally launched externally in April 2024, named after its Chinese nickname "達哥" (Dage). Depending on a department's sensitivity level, it layers in either external cloud models (GPT-4o, Claude, Gemini, etc.) or on-premises models (Llama, Mistral, DeepSeek, and MediaTek's own Breeze/BreeXe, etc.); modeled on autonomous-driving levels, it measures intelligence maturity from L1 to L5, with the ultimate goal being an L5 super-assistant comparable to "Jarvis." It appeared in Day 1's Smart Transformation session (Yeh Chia-Shun (葉家順)).

Cadence Cerebrus / Voltus InsightAI are two AI-assisted chip design tools from Cadence, an IC design EDA software company: Cerebrus is a reinforcement-learning-driven PPA (power/performance/area) optimization engine that automatically explores design parameters and iteratively reruns to converge on an optimal solution; Voltus InsightAI is a generative-AI-driven root-cause analysis and auto-fix tool for IR drop (voltage drop), which the vendor claims can fix up to 95% of violations before signoff. It appeared in Day 1's Deep Tech session (Huang Yen-Hsiang (黃彥翔)).

GENIE is E.SUN Bank's (玉山銀行) internal generative AI platform (an office-automation smart assistant), which entered pilot operation in Q4 2023, launched formally in early 2024, and released version 2.0 in September 2024. Because customer data in banking is sensitive, the bank chose to build in-house rather than connect directly to external big-vendor LLMs; initial adoption was poor (only 17% unique users), but after optimization, monthly active users grew 200%. The official agenda listed this as speaker Huang Shih-Chen's (黃仕鎮) topic. It appeared in Day 1's Industry Applications session.

The AI Air Pollution Source-Tracing Platform (AI 空污溯源平台, CAMEO × Environmental Protection Administration) is a collaboration between CAMEO (卡米爾) and Taiwan's Environmental Protection Administration (now the Ministry of Environment), deploying about 8,000 IoT sensors nationwide to catch factories illegally discharging exhaust gases, helping issue over NT$200 million in fines; because Python's preprocessing speed was insufficient, the pipeline was rewritten in Julia, achieving roughly a 10x speedup.

OPEA (Open Platform for Enterprise AI) is an open-source "composable generative AI systems" framework established on April 16, 2024 by the LF AI & Data Foundation, part of the Linux Foundation, with Intel as one of the leading founding vendors; it focuses on standardized architecture for RAG systems and enterprise-readiness evaluation. It appeared in Day 1's Platform Services session.

VectorMesh (Minima / Parva / Magna) is an in-house NPU IP brand from Inventec (英業達), divided into three tiers by target device compute needs: Minima targets ultra-low-compute applications under 32 GOPS (launched August 2023), while Magna targets scenarios with greater on-device compute; a dedicated LLM IP is also planned for mass production in 2025. Specific power-consumption and PPA figures are unverified. It appeared in Day 1's Platform Services session.

Arm KleidiAI is a compute-kernel library that Arm provides to AI framework developers, letting them achieve optimal inference performance across various Arm CPU devices; it has already been integrated with PyTorch, TensorFlow, and MediaPipe, targeting acceleration of models such as Meta's Llama 3 and Phi-3. Arm's concurrent product lineup also includes Arm CSS for Client (an AI PC/smartphone compute subsystem) and Ethos-U85 (an edge NPU supporting Transformer architectures). It appeared in Day 1's Platform Services session.

AMD Instinct MI300X / ROCm refers to AMD's data-center GPU (announced December 2023, mass-produced Q1 2024, with 192GB HBM3 and over 5,000 TFLOPS of FP8 compute) and its open-source AI software stack ROCm, which differentiates itself from NVIDIA's CUDA ecosystem via an open-source community strategy; concurrently, AMD also has the Ryzen AI PRO 300 series on-device NPU (50 TOPS, for Copilot+ PCs). Public information about the speaker themself is unverified. It appeared in Day 1's Platform Services session.

XCockpit is CyCraft's (奧義智慧科技) AI-autonomous Threat Exposure Management platform. It appeared in Day 1's Industry Applications session.

iRPA2000 / iAT2000 (with companion platforms NextData / eSafe / DVOI) form NEXCOM's (新漢智能系統) AIoT product line: iRPA2000 is a next-generation robot co-creation platform that uses EtherCAT to connect heterogeneous equipment such as CNC machines, robotic arms, and AMRs; iAT2000 is a cloud-intelligence monitoring system connected to generative AI for Q&A-style decision-making (the core of an enterprise "war room"); NextData is a manufacturing data middle platform, eSafe is an OT security platform, and DVOI is a smart imaging platform. It appeared in Day 1's Industry Applications session.

Moana is a word-of-mouth (buzz post) generation model fine-tuned by Accucrazy (肖準行銷集團) on its own historical marketing data, now in its fourth generation; it can generate social posts in the tone of Taiwanese internet forum users ("鄉民"), tailored to a given persona and product angle, and once generated a Dcard post that drew over 1,300 comments without anyone realizing it was AI-generated. It appeared in Day 2's Deep Tech Flash Talk session.

MSI AI Artist is MSI's (微星科技) on-premises AI assistant, whose functionality has evolved through four stages: "text-to-image → image-to-image / image-to-text → language model + hardware control → RAG (AI Fire)," positioning gaming laptops' high-end GPU/NPU compute as making "the most powerful AI PC on Earth." It appeared in Day 1's Industry Applications session.

Tookie is an AutoML plus generative AI integrated platform from Chimes AI (詠鋐智能), converting equipment sensor data into "equipment lifespan indicators" for predictive maintenance, and using generative AI to translate predictions into trustworthy energy-saving/maintenance recommendations; it has been deployed in 50 factories with 2,500 AI applications in production. It appeared in Day 2's Deep Tech Flash Talk session.

TianRanTongHao (天然通好) is a no-code image-defect classification/localization platform from Prophet Tech (先知科技), covering the entire pipeline from data collection, labeling, and training to evaluation and deployment in one stop; it uses generative image augmentation to solve manufacturing's "high-variety, low-volume" pain point of insufficient training samples. It appeared in Day 2's Deep Tech Flash Talk session.

The Jubo product family (Vital Sign AI Alert / Wound Care AI / FamilyLine family contact notebook / N-Copilot) comprises AI features developed by Jubo Health (智齡科技) for long-term-care institutions: personalized vital-sign anomaly alerts, tablet-photo-based wound size estimation, AI-generated family-readable care summaries, and a RAG-based Q&A assistant built on nursing records (N-Copilot, launched in 2025). It appeared in Day 2's Platform Services / Smart Healthcare session.

Joy is 91APP's internal codename for an architecture that integrates vector retrieval (a three-tower model of people/products/tags) with an LLM spokesperson agent (named for its phonetic resemblance to "91"); it can be composed into agents for retail customer-service Q&A, product recommendations, and similar tasks. It appeared in Day 2's Industry Applications session.

RAG (Retrieval-Augmented Generation, 檢索增強生成) first retrieves the most relevant data snippets, then hands them to an LLM to generate an answer, reducing hallucination and making answers traceable to sources; multiple sessions at this conference (eLAND Information, Phison, Quanta, etc.) used this approach in place of direct fine-tuning or feeding entire documents to the LLM. It appeared in Day 2's Industry Applications session and several other sessions.

LoRA (Low-Rank Adaptation) is a lightweight fine-tuning technique that only adjusts an added low-rank parameter matrix without touching the original model weights, making it far cheaper than full fine-tuning. It appeared in Day 2's Industry Applications session.

CPT (Continued Pre-Training, 接續預訓練) sits between full fine-tuning and pre-training from scratch: it continues pre-training an existing model on additional language- or domain-specific corpora. TAIDE adopts this approach. It appeared in Day 2's Industry Applications session.

RLHF (Reinforcement Learning from Human Feedback) trains a reward model on human feedback, then uses reinforcement learning to adjust an LLM's outputs so they better match human value judgments; it is one of TAIDE's core training steps. It appeared in Day 2's Industry Applications session and Day 2's Smart Transformation & Memorial Lecture session.

Chain of Thought (CoT, 思維鏈) is a prompting technique that guides a model to "think step by step" before answering; a case presented at this conference showed CoT is not a universal fix — for tasks that are already simple enough, it may not improve accuracy. It appeared in Day 2's Deep Tech Flash Talk session.

Agentic LLM refers to an LLM application pattern with autonomous task decomposition, tool selection, and result integration capabilities, distinct from traditional single-turn instruction-following LLMs; ITRI's Hsu Kai-Cheng (許凱程) used the analogy of "Iron Man's butler Jarvis": given a goal, it automatically breaks it into subtasks, finds tools to complete each one, and finally integrates the results for human confirmation. It appeared in Day 2's Platform Services / Smart Healthcare session.

Eval-Driven Development (評估驅動開發) is an LLM application development maturity framework proposed by Chang Wen-Tien (張文鈿) of iKala (愛好資訊), spanning Level 0–4: no evaluation → eyeballing → manual test sets → automated evaluation → automated prompt optimization. It argues that prompt engineering should be evaluated objectively, following the same discipline as traditional software testing. It appeared in Day 2's Deep Tech Flash Talk session.

MetaPrompt is a tool, open-sourced from Anthropic's (Claude) backend, for generating "a prompt that generates prompts": given a brief task description, it produces a complete prompt including few-shot examples. It appeared in Day 2's Deep Tech Flash Talk session.

DSPy / TextGrad are two automated prompt-optimization frameworks that let different models be assigned to synthesize training data, generate candidate prompts, and make final predictions respectively, iteratively producing hundreds of candidate prompts and automatically selecting the best one. It appeared in Day 2's Deep Tech Flash Talk session.

LLM-as-judge uses a separate LLM as a judge that scores a model's outputs against defined rubrics, solving the automated-evaluation problem for tasks with no single correct answer; it requires human calibration of the judge's scores to avoid the "who watches the watchmen" problem. It appeared in Day 2's Deep Tech Flash Talk session.

The Two-Tower Model (雙塔模型) is an architecture commonly used in recommendation systems such as YouTube, Spotify, and Netflix, encoding users and content into separate vectors for similarity-based retrieval; 91APP extended it into a three-tower model of "people / products / tags." It appeared in Day 2's Industry Applications session.

SaMD (Software as Medical Device, 軟體醫材) is the TFDA's certification classification for smart medical devices, with CADx (computer-aided diagnosis) and CAD-triage (computer-aided triage) as representative subcategories; it forms the regulatory foundation for real-world adoption of smart healthcare. It appeared in Day 2's Platform Services / Smart Healthcare session.

Differential Privacy (差分隱私) adds noise to query results so that whether or not a database contains any specific record does not affect the query outcome, mathematically guaranteeing that individual records' presence cannot be inferred; a rule of thumb is that the dataset needs at least tens of thousands of records to remain usable. It appeared in Day 2's workshop session.

TEE / FHE / MPC / ZKP (隱私保護推論四件套, the "four pillars" of privacy-preserving inference) are four cryptographic/hardware techniques, systematically explained by Chen-Mou Cheng (鄭振牟), that let mutually untrusting parties (a model owner and a data owner) still collaborate on AI inference or training: TEE (Trusted Execution Environment, e.g., Intel SGX) uses hardware to carve out a protected memory region that even system administrators cannot read, with minimal overhead but limited capacity and GPU support; FHE (Fully Homomorphic Encryption) can compute directly on encrypted data, offering the highest level of privacy, but is often hundreds to tens of thousands of times slower; MPC (Secure Multi-Party Computation, commonly via Garbled Circuits) lets two parties exchange messages to jointly compute without either learning the other's input, with performance clearly better than FHE but requiring large amounts of communication; ZKP (Zero-Knowledge Proof) lets one party prove that "a result was faithfully computed according to a given model" without revealing the input. The same set of techniques was also cited by Chen Wei-Chao (陳維超) for protecting federated learning and model inference privacy, and listed by Yu Chia-Mu (游家牧) as members of the Privacy Enhancing Technologies (PET) family. It appeared in Day 2's AI Governance panel (Chen-Mou Cheng, full systematic teaching), Day 2's Smart Transformation & Memorial Lecture session (Chen Wei-Chao), and Day 2's workshop session (Yu Chia-Mu).

New York City's Local Law 144 (紐約市 Local Law 144, Automated Employment Decision Tools Law) is a regulation requiring businesses that use AI to screen job applicants to proactively disclose this and undergo an annual bias audit; Hou Yi-hsiu (侯宜秀) misspoke during her talk, calling it a "New York State" regulation, when the correct jurisdiction level is New York City. It appeared in Day 2's AI Governance panel (Hou Yi-hsiu).

California's SB 1047 (加州 SB 1047, the frontier AI model safety bill) was a California bill requiring developers of frontier large models to bear certain safety responsibilities; opposition from OpenAI, Meta, and the open-source camp substantially weakened its liability provisions during the legislative process, earning it the nickname "the law with no teeth." It was vetoed by California Governor Gavin Newsom on September 29, 2024 (during the conference, the day after Cha Shih-Chao (查士朝) and Hou Yi-hsiu (侯宜秀) gave their talks) — neither speaker knew the final outcome at the time of their talks. It appeared in Day 1's Deep Tech session (Cha Shih-Chao) and Day 2's AI Governance panel (Hou Yi-hsiu).

Federated Learning (聯邦學習) keeps data on each local device or institution for training, sending only model parameters (not raw data) back to a central server for aggregation, balancing data privacy with model performance; the Taiwan Smart Healthcare Alliance used this method for cross-hospital model validation. It appeared in Day 2's Platform Services / Smart Healthcare session and Day 1's Opening & Keynote session (where Tung Tzu-hsien (童子賢) referred to it as "Federal Training").

K-anonymity (K-匿名) is a traditional de-identification technique and the only privacy-protection method explicitly recognized by Taiwan's Personal Data Protection Act / GDPR, but its protective strength and resulting data usability are both weak, with numerous re-identification incidents reported over the past decade or more. It appeared in Day 2's workshop session.

Randomized Response (隨機應答) is a noise-adding technique originating from a 1965 US method for surveying sensitive questions: respondents first flip a coin to decide whether to answer randomly or honestly, giving every answer "plausible deniability" while the true proportion can still be statistically inferred; a case at this conference used it to apply differential privacy to non-numeric fields. It appeared in Day 2's workshop session.

SMILES (Simplified Molecular Input Line Entry System) is a simplified notation that represents chemical molecular structures as text strings, allowing compound data to be processed as sequential input similar to language models; it is applied to drug molecule generation and multimodal medical models. It appeared in Day 1's Industry Applications session and Day 2's Deep Tech Flash Talk session.

EtherCAT is an industrial Ethernet real-time communication standard that NEXCOM (新漢智能系統) uses to connect heterogeneous equipment such as CNC machines, robotic arms, and AMRs, serving as the underlying communication foundation of its edge AIoT architecture. It appeared in Day 1's Industry Applications session.

Sim2Real is a technical approach of first training AI/robots in a simulated environment, then transferring them to real-world deployment; it is used to improve a robot's robustness when switching between different gaits. It appeared in Day 2's Smart Transformation & Memorial Lecture session.

MMLU is a large multiple-choice knowledge benchmark; in essence it measures a model's probability distribution over answer-choice strings (a likelihood estimator), making it well suited to evaluating the knowledge content of pretrained models but poorly suited to evaluating models that have already undergone instruction fine-tuning. It appeared in Day 2's Smart Transformation & Memorial Lecture session.

Function Calling is a mechanism that lets an LLM decide, based on semantic understanding, to call external tools/APIs to fetch real-time data before answering; a case at this conference involved connecting to Bloomberg/Reuters financial databases for advanced analysis. It appeared in Day 2's Smart Transformation & Memorial Lecture session.

HDBSCAN is a density-based hierarchical clustering algorithm; paired with Multilingual BERT, it is used for semantic-level clustering of music genre tags, consolidating duplicate or similar categories. It appeared in Day 2's Industry Applications session.

BFCL (Berkeley Function-Calling Leaderboard) is an internationally used benchmark standard for evaluating LLMs' function-calling capability, dividing tasks into five categories by number of functions/calls: simple, multiple, parallel, parallel multiple, and irrelevant; MediaTek Research localized and translated it, releasing a Traditional Chinese version of the benchmark set. It appeared in Day 1's Deep Tech session.

AI TRiSM (AI Trust, Risk & Security Management) is a framework proposed by Gartner for AI trust, risk, and security management, used to assess an enterprise's governance maturity when adopting generative AI. It appeared in Day 1's Smart Transformation session (Chang Chao-Liang (張朝亮)).

The Taiwan AI Academy (台灣人工智慧學校) is the host organization of this conference, founded in 2018 at the initiative of Chen Sheng-Wei (陳昇瑋); over seven years it has accumulated more than 11,000 alumni and partnerships with over 2,200 companies across 15 major industries, and since 2024 has been promoting a tiered AI-literacy certification for the general public. It appeared in the event background & media coverage material and Day 1's Opening & Keynote session.

The Chen Sheng-Wei Memorial Lecture (陳昇瑋紀念講座) is a fixed lecture established at every annual conference since 2021 to honor Chen Sheng-Wei (陳昇瑋, 1976–2020), founding CEO of the Taiwan AI Academy, who passed away from a cerebral hemorrhage caused by a sports accident; it invites his friends or academic peers to share the latest perspectives on AI. The 2024 speaker was Chen Wei-Chao (陳維超, Chief Digital Officer of Inventec). It appeared in Day 2's Smart Transformation & Memorial Lecture session (the lecture itself) and the event background & media coverage material (his full biography and past speakers).

Academia Sinica's Research Center for Information Technology Innovation (中央研究院資訊科技創新研究中心, CITI) is a cross-disciplinary research center at Academia Sinica focused on developing key information and communication technologies and cross-domain integration; it is one of this conference's co-organizers (alongside Academia Sinica's Institute of Information Science). It appeared in the event background & media coverage material.

The Taiwan Artificial Intelligence Association (台灣人工智慧協會, TAIA) is a platform for collaboration among industry, government, academia, and research, founded in 2020 to promote both the industrialization of AI and the AI-ification of industry; Chang Kang-Ling (張剛羚, TAIA Executive Director & CSO) served as moderator for Day 2's Deep Tech track. It appeared in the agenda overview.

g0v Lingtsan School (g0v 零時小學校) is the grant-execution unit under Taiwan's civic tech community g0v (零時政府), and it executed the Traditional Chinese AI Open Source Grant on this occasion; the Day 2 workshop moderated by Hou Yi-hsiu (侯宜秀, Secretary-General of the Taiwan AI Academy) grew out of this civic-tech/open-data context. It appeared in Day 2's "Where Does Good Data Come From" workshop.

The Association of Hackers in Taiwan (台灣駭客協會, HIT) is a Taiwanese information-security community organization that hosts the annual HITCON hacker conference; Chiu Ming-Chang (邱銘彰, also a co-founder of CyCraft), speaking in an advisory capacity, presented "The Esoteric Art Behind Successful AI Projects" in Day 1's Deep Tech track, using complex systems science to explain deep learning's generalization ability. It appeared in Day 1's Deep Tech session.

DEVCORE (戴夫寇爾) is Taiwan's first information-security company specializing in Red Team exercises, founded in 2012 and staffed by a white-hat team with a hacker mindset; CEO Weng Hao-Cheng (翁浩正) presented "The Evolution of Cybersecurity Attacks in the GenAI Era" in Day 1's Deep Tech track, surveying from a red-team perspective how generative AI is affecting social engineering and the penetration-testing toolchain. It appeared in Day 1's Deep Tech session.

CyCraft (奧義智慧科技) is a Taiwanese AI-security startup specializing in automated threat hunting and SOC analysis, with a product line that includes the XCockpit platform; speaker Yang Cheng-Lin (楊政霖) discussed four key considerations for building an in-house security LLM: privacy, alignment, domain knowledge, and localization. It appeared in Day 1's Industry Applications session.

iiiSI (資拓宏宇) is a large Taiwanese information and communications technology service company that has long assisted the government in building transportation, finance, and healthcare systems, and provides deep-learning-based AI product solutions such as chatbots, knowledge management, and content generation. It appeared in Day 2's Industry Applications session.

eLAND Information (意藍資訊, stock code 6925) is a Taiwanese information company with years of expertise in public-opinion analysis, which acquired Tornado Technology (龍捲風科技) in 2022; its clients cover over 80% of the top-ten brands in government, finance, and retail. Speaker Yang Li-Wei (楊立偉) shared a real-world case of applying RAG to enterprise knowledge management. It appeared in Day 2's Industry Applications session.

Jubo Health (智齡科技) is a long-term-care information platform startup founded in 2018 by Shih-Chung Kang (康仕仲, a Stanford PhD and former NTU professor); it holds the largest market share among long-term-care information platforms in Taiwan, serving over 1,200 institutional clients and about 130,000 elderly residents. It appeared in Day 2's Platform Services / Smart Healthcare session.

Harbor Technology Solutions (皓博科技) is a Taiwanese industrial AI startup founded in 2015 that started out in acoustic signal processing; its product line includes Vibration Guard (vibration diagnostics), Power Guard (partial-discharge monitoring), Sound Guard, and Vision Guard, with its core technology positioned as "A-PHM" (Active Prognostics and Health Management). It partners with Chimes AI (詠鋐智能) to supply equipment-anomaly early-warning data. The official agenda listed its CEO, Lu Hung-Yi (呂宏益), as a Day 2 speaker. It appeared in Day 2's Deep Tech Flash Talk session.

CAMEO (卡米爾) is a Taiwanese company founded in 2000 whose business spans big-data mining, AI, data visualization, AR/VR, and experience design, and which also has a generative AI product line called "BOTRUN." The official agenda listed founder and CTO Chiu Po-Huan (邱泊寰) speaking on "The Challenges of Government AI Adoption and an Unexpected Word-of-Mouth Success," with the flagship case being the AI Air Pollution Source-Tracing Platform built with the Environmental Protection Administration.

Chimes AI (詠鋐智能) is an AI energy-saving and carbon-reduction startup founded by Hsieh Tsung-Chen (謝宗震, a Statistics PhD from National Tsing Hua University known as "the father of the R language in Taiwan"); using its predictive-plus-generative-AI platform "Tookie," it helps energy-intensive factories cut power consumption, having served 50 factories and cumulatively reduced carbon emissions by over 40,000 tons annually. It appeared in Day 2's Deep Tech Flash Talk session.

Prophet Tech (先知科技) is a Taiwanese startup focused on smart inspection for manufacturing; founder Kao Chi-An (高季安) previously took part in TSMC-style smart manufacturing deployments. Its flagship offering is the no-code image-defect-inspection platform "TianRanTongHao" (天然通好), and it argues that generative AI can work alongside discriminative AI to solve the few-shot training problem. It appeared in Day 2's Deep Tech Flash Talk session.

NEXCOM (新漢智能系統) is a Taiwanese industrial-computer/AIoT group founded by Lin Mao-Chang (林茂昌), which argues that Taiwan should develop edge AI via open standards such as EtherCAT rather than competing with cloud giants on raw compute; its product line includes iRPA2000, iAT2000, and others. It appeared in Day 1's Industry Applications session.

OpenFun (歐噴有限公司) is a civic-tech company founded by Wang Hsiang-Jung (王向榮), long engaged in the open-government/open-parliament movement (projects such as political-donation transparency and job-search assistants). This time it presented a project that turns the Legislative Yuan's scattered, unstructured data into a structured API (v2.ly.govapi.tw) and an open dataset (data.ly). It appeared in Day 2's "Where Does Good Data Come From" workshop.

CHT Security (中華資安國際) is a cybersecurity subsidiary of Chunghwa Telecom; speaker Shen Yi-Ting (沈宜婷, SOC/Security Engineer) demonstrated automating penetration testing using a locally deployed, safety-restriction-removed Ollama model. It appeared in Day 2's Deep Tech Flash Talk session.