Glossary

An AI system that sets its own goals, plans and executes, interacts with the outside world and keeps adjusting. The conference's most-used phrase — and almost always used to describe difficulty rather than promise. NICS noted the four capabilities match Skynet's exactly.

Several specialist agents collaborating. In AIA's kitchen analogy: an LLM is a learned expert who can only talk; an AI Workflow is an assistant following a fixed SOP; an AI Agent is a goal-driven proxy that plans and evaluates; an Agentic Workflow is a Michelin kitchen with head chef, sous chef and saucier.

Retrieval decided at run time by the agent rather than hard-wired at design time. Named this year as what replaced plain RAG tutorials.

Huang Neng-fu's no-code platform for building vision and hearing models: collect data → label → train → download and deploy → run → evaluate → iterate, with no programming at any step. Paired with a distributed compute-sharing alliance so schools without GPUs can take part.

The two flight-control software camps. ArduPilot began on a 2009 Arduino Mega, is now STM32/ARM, and is fully open source under GPL. PX4 is foundation-backed.

Open-source autonomous driving stacks. NYCU bridges its OpenSCENARIO scenario templates into them to run simulation-based safety validation.

MediaTek Research's open Traditional Chinese model family. Breeze 2 beat Gemini on Cola Tour's passport OCR failures despite being far smaller; Breeze ASR reached 120,000 downloads within two to three months of release.

Putting explanation and reasoning between input and output during training, so the model predicts a thought pattern rather than only the next token. Ed Chi co-authored the January 2022 paper; the underlying idea of schemata he picked up proofreading his mother's doctoral thesis nearly thirty years earlier.

Running a new model in parallel with the incumbent for a period before switching over. E.SUN retrains weekly and runs both for a week before promoting the challenger.

Combining an LLM's capabilities produces new emergent capabilities — so safety multiplies rather than adds. Ed Chi's point that safety is A × B × C, not A + B + C, and that divide-and-conquer security thinking therefore fails on LLMs.

Managing what is in the model's context window, borrowing Karpathy's analogy that the LLM is an operating system, the model is the CPU and the window is RAM. Four operations: write context (to-do lists on disk, long-term memory), select context (Agentic RAG), compress context, isolate context (sub-agents). This is what replaced prompt engineering this year.

Working out which part of a solution deserves the credit or blame. Sun Shao-hua found LLMs lack it — they cannot tell which part of a generated program did well or badly — so search or RL is still needed to refine a program policy.

Simulating a situation virtually, tuning parameters and finding the best answer before deploying to the real world. Framed this year as the bridge carrying AI into the physical world — not to look realistic, but to cut Physical AI's three costs at once. Delta's counter-intuitive point: it earns its keep in trial production, not mass production.

Setting a transmission line's current limit from live weather with machine learning, instead of a conservative constant based on the annual maximum temperature. A German Siemens customer got 20% more throughput on the same line with no extra hardware.

New behaviour appearing that was not in the training data. Accu Crazy's fine-tuned model invented internet slang the speaker verified had never appeared online before.

91App's design choice: keep the evaluator entirely outside the conversation flow, batch-scoring after the fact for a human supervisor to review, forming an optimisation loop rather than an inline check.

The emerging counterpart to SEO for an AI-first world. 91App's framing: if the AI entry point cannot find you, that is as serious as not existing in Google search.

A training method that replaces PPO's separate value model with a hand-written reward function, saving substantial GPU. Twinkle AI's analogy: PPO is a private tutor marking every step, GRPO is a workbook with worked solutions the student marks themselves. Their F1 model went 50 → 64 on MMLU using two reward functions. Limitation: if the base model never saw the knowledge in pre-training, GRPO cannot train it in either.

A protective framework around a model — filtering input and output, monitoring behaviour and policy, enforcing least privilege, supporting human-in-the-loop. NICS recommends starting from Meta's Llama Firewall.

A human decision point inside an automated flow. Taiwan AI Labs' point: a real procurement process has many such points, and designing them is what AI governance actually means.

The standard for validating whether a black-box AI system understands the scene, operates safely and completes its task within its design domain. Used by NYCU for scenario-based autonomous-driving validation.

Yann LeCun's world-model architecture at Meta. V-JEPA 2 was released in June 2025, trained on a million hours of video and supporting zero-shot robot control. Cited at the conference for predicting a ball rolling over a bump.

Meta AI

Meta's integrated protection framework combining Prompt Guard (a small model that runs on CPU), Llama Guard (about 12B parameters) and MCP/plugin control. NICS's recommended starting point for a development team.

The core protocol between a drone and its ground station — heartbeat, status reporting, command issuing. Named as the natural entry point for someone who knows AI but not drone hardware.

The protocol connecting a model to external tools. Two risks run in both directions: prompt injection tricking the LLM into misusing a connected tool, and the LLM connecting to a malicious or compromised MCP server that feeds it poisoned data. Salesforce's finding that accuracy falls as tool count rises was cited repeatedly.

How efficiently a training or inference run actually uses the available floating-point throughput. Named by ASUS as Taiwan AI Cloud's core competency — optimising it per customer workload.

The image generation model that came up repeatedly as the year's cost-collapse example — about NT$10 to NT$1.17 an image depending on the case, against NT$500,000 for a traditional shoot.

NVIDIA's multi-GPU communication library. The practical landmine flagged at the conference: its version has to match the card tier — a lower-tier new card such as the RTX 6000 series needs a manual upgrade (2.26.5) or the official vLLM image, built for B200/H100 class cards, will not run.

Open-source musculoskeletal simulation software. Used at NTU of Sport for joint angles at key moments, kinetic-chain defect detection and per-joint loading — which revealed that UCL load in the follow-through exceeds that at maximum shoulder external rotation.

The scene description format for digital twins. The industry signal this year: TSMC has begun requiring USD assets as a deliverable on some piping projects, turning the twin from an internal tool into a delivery specification.

The reference framework for LLM security risks. NICS mapped its three-phase model (training, use, systems integration) onto it, naming prompt injection, sensitive information disclosure and supply-chain risk as the highest impact.

A system that perceives, reasons and acts, interacting with the physical world through sensors and actuators. The layer stack given at the conference: Generative AI → Agentic AI (plus digital tool use) → Physical AI (plus physical tool use).

Sun Shao-hua's route: before acting, synthesise a human-readable, verifiable, locally debuggable program in a domain-specific language as the decision logic, instead of training a black-box neural policy. On the maze task it produces literally the right-hand rule and generalises to any size.

Encryption designed to survive quantum computers, whose ability to factor the mathematical problems classical cryptography rests on will break current methods. The US government has set 2030 as the deadline; NIST provides a risk-management framework for migration.

Hidden instructions planted in external material — email, web pages, documents, white text on white — that the LLM reads and obeys. NICS's point: direct injection is self-inflicted and not the real concern; indirect injection is, especially where an LLM is wired into mail summarisation, document systems or customer service.

Guardio Labs' term (August 2025) for the new scam surface created by agentic AI browsers — where the AI, not the human, is the target.

Guardio Labs research report

A small model, usually fine-tuned for a vertical domain. NVIDIA research cited at the conference says 70% of AI agents need SLM optimisation to be usable; APMIC's 24B beat GPT-OSS 120B on Taiwanese law at a fifth of the VRAM.

See ISO 21448.

From a 1978 MIT paper on teleoperated submersibles: direct remote operation hits latency problems, so the human issues directional goals and the equipment handles detailed control. Pegatron extended the analogy to the management latency of running overseas plants.

Borrowed from Thinking, Fast and Slow. In NVIDIA Isaac GR00T N1, System 2 does long-horizon planning and System 1 reacts fast. Realtek's argument: System 2 must hallucinate because predicting the unhappened is filling in blanks, so Taiwan should build System 1 — sense-actuate reflex loops that can be put into hardware, with no hallucination and very low power.

The National Center for High-performance Computing's supercomputer, 252 nodes at 9 PFLOPS, 20th on the global TOP500 in 2020. Half its capacity serves academia and half is operated commercially. ASUS trained a Hakka sovereign AI model on it.

TechNews

Taiwan AI Cloud Corporation, an ASUS-affiliated company running an AI Foundry (GPU VM and bare-metal model fabrication, plus the oneAI training and fine-tuning platform). Unrelated to TSMC — the similar-sounding Chinese names cause confusion.

Deep Wave's extension of product-market fit, drawn from a case that never shipped: a good model does not mean it maps to a suitable scenario; mapping to a scenario does not mean it scales; scaling does not mean users find it usable.

A Traditional Chinese general-knowledge benchmark including Taiwan's civil-service exam bank — nearly 10,000 questions. Built because general English benchmarks cannot measure usability in Traditional Chinese settings.

NYCU's deep-learning ball-tracking model, validated on tennis (bright, slower) before badminton (small, extremely fast, few visual features). Its multi-frame design learns flight-trajectory features across consecutive frames; later versions run multi-in multi-out (10-in-10-out) for throughput.

How long before the model emits its first token. A key benchmark metric, with a fixed trade-off: shortening TTFT lengthens inter-token latency, since total time is fixed. Tunable via max_model_len and max_num_seqs.

The robot description format. Like OpenUSD, it is moving from internal tool to contractual deliverable — robot supplier transactions are starting to require it.

Models that map vision and language to robot action. The lineage given at the conference: Google RT-1 (2022) encoded actions as special transformer tokens, then trajectory information was added, then training across 22 embodiments and open-sourcing, and most recently NVIDIA Isaac GR00T N1.

An inference framework suited to multi-GPU and multi-node deployment; the enterprise default. Ollama is the quick local-test alternative (Docker-CLI-like, with a ready model repository), alongside LM Studio.

Inventec's pragmatic answer to data leakage: do not monitor all employee conversation, only raise an alert when something resembling confidential data appears to be heading to an external model, for after-the-fact audit. A design that starts by admitting employees will use outside tools anyway.

The Stanford paper Pegatron used to pick tasks, with two axes: does the human want AI's help (x) against technical feasibility (y). Only the top-right green zone gets built. Its finding: people most want AI to take over scheduling, file maintenance and format conversion, and least want it touching creative work.

A model of how the world behaves, used to predict what happens next. Three routes surfaced: generative-based (NVIDIA Cosmos, Google Genie), Meta's JEPA, and video-plus-LLM. Realtek's criticism: a purely text-described world model loses critical latent-space information and is inferior to a visual one.

Kenmec's slogan for Physical AI's three cost barriers, citing NVIDIA: data is expensive (only about 10% of collected real data is usable), training is expensive (engineers who can train are scarce), and physical testing is expensive (about NT$4m per humanoid). This is what digital twins exist to reduce.