Day 2 — 10 September

01 · Ed H. Chi — Google DeepMind

KEYNOTE — The future world of intelligent personal assistants

Starting from thirty years of the ranking economy in search and recommendation, he explains how sequential transduction, chain-of-thought and post-training turned language models from word-completion into reasoning machines — and shows that future through Project Astra.

• Two foundations of information retrieval — indexing (mapping keywords to document positions) and the vector space model (cosine similarity over term-frequency distributions). These two ideas built Google search.
• Indexing, vector space and neural networks together underpin more than US$500bn a year of internet economy since 1995 — Google search about $200bn, YouTube about $50bn, Facebook and Instagram about $130bn, TikTok about $130bn. The mechanism is ranking and sorting, not generative AI.
• Why it is called a transformer — the correct Chinese translation is closer to transducer: like a microphone, a speaker or an eardrum, it converts one energy wave or information sequence into another. That is sequential transduction.
• The paper timeline — Dec 2014 Sequence to Sequence Learning with Neural Networks (Sutskever, Vinyals, Le), RNN-based, O(n), state of the art on English-to-German, and a 2025 NeurIPS Test of Time Award; 2017 the Transformer paper, matrix computation at O(n²), more accurate but far hungrier — essentially scaling up; Jan 2022 the chatbot paper he co-authored and the Chain-of-Thought paper in the same month; Feb 2022 the post-training paper.
• Where chain-of-thought came from — the core idea of schemata (step-by-step problem-solving patterns in learning theory) he picked up nearly thirty years earlier while proofreading his mother's doctoral thesis.
• The bottleneck that drove single multilingual models — in 2015 Google Translate had 200–300 people, and supporting pairwise translation across 40 languages meant 1,560 separate models.
• What intelligence actually is — next-token prediction (your brain auto-completing a sentence, Paris → Eiffel Tower) is association over existing knowledge, not intelligence. Real intelligence has to predict a thought pattern, which is why training needs explanation and reasoning between input and output. Hence: next-token prediction + chain-of-thought + post-training is the reasoning revolution.
• Project Astra demo — repairing a bicycle (reads the manual, finds a YouTube tutorial, digs through the inbox for part specs, phones the bike shop about stock, remembers the user's dog is called Zooka and recommends a dog basket) and helping a blind guitarist (reads the Wi-Fi password off a sign, locates the mic stand). He noted the video was Astra running live on a phone and smart glasses, not a pre-recorded advert.
• Current research focus — multi-step complex reasoning, tool use, synthetic data generation, multi-modality integration, and personalization: general intelligence is not enough, we want intelligence built for you.

Large language model is in fact a wrong term, because it was not born only for language — what we actually want to do is reasoning.

That's not intelligence. That's just next token prediction.

Anxiety is ignorance of the future — you feel it because you cannot predict what is coming.

02 · Cheng Teng-hu — NYCU / Tron Future

KEYNOTE — AI-driven drones in GPS-denied environments

How multi-sensor fusion of radar, cameras and IMU lets a drone detect, identify and intercept incoming hostile drones — soft-kill and hard-kill — with no GPS and no communications at all, extended to GPS-denied navigation competitions and indoor warehouse automation.

• The scenario — with no defensive equipment, spotting an incoming drone by eye at about 400 metres while it closes at 100 km/h leaves roughly 25 seconds to react (about 15 if you can see it is carrying explosives). Radar plus a high-zoom camera pushes detection out to 5–6 km.
• Losing GPS is like finding the exit with your eyes closed — a good airframe hovers until the battery dies; a bad one simply falls. On IMU integration alone, error compounds exponentially to hundreds of metres within about 40 seconds.
• Visual navigation helps but has no global map, so it only localises where it has already been — and swarms exchanging relative positions make it harder still. Ukrainian battlefield experience shows on-board visual feature matching against an offline local map (satellite imagery) gets relocalisation error to about 20 metres; the other common answer is a fibre-optic tether for live remote piloting.
• Four soft-kill components — radar (wide search to 5–6 km; small airframes have tiny radar cross-sections so a small radar is needed), camera (identification within 3 km but only a 2° field of view against radar's 120°), jammer (GPS and comms, 4–5 km), and a sensor to intercept the operator's control signal.
• Multi-sensor online calibration — customers may site radar and cameras anywhere, so alignment cannot be done in the factory. The team brought calibration of 3D relative position and rotation between radar and multiple cameras down from 2 hours to 15 minutes, with angular error from 45° to under 1°.
• An accidental finding — a field test misclassified an eagle as a drone (birds were not in the training data), which inspired a commercial idea for airport bird dispersal using radar plus camera to identify the species and play a sound it fears.
• The April 2025 GPS-denied navigation defence competition — about 20 teams entered, 6 reached the final, and only 2 could still fly once GPS was jammed (most lost control; one hit a nearby hospital, without injury). His team flew 13 minutes and 6 km with precise positioning but misclassified targets because the rules mandated a domestic camera around NT$10,000, ending with the Potential Award (NT$100,000) against first place's NT$1.5 million.
• Interceptor design — no GPS, no comms, tracking and identifying autonomously on an onboard RF or EO seeker; weight limits usually mean the seeker is rigidly fixed to the airframe, so the target must stay inside the field of view throughout, which makes the acceleration and inertia controller design critical. Demonstrated fully autonomous interception of a human-piloted target, repeatedly within a metre.
• Light shows versus defensive swarms — a light show can plan paths offline and run without comms once airborne; a defensive swarm has to detect and coordinate avoidance in real time over peer-to-peer edge communication, with no cloud.

If a drone has no GPS, it is like closing your eyes and being told to find the way out.

03 · Panel — Practising AI safety

Three panellists unpack real-world AI risk from the model-research, enterprise-application and government-governance vantage points — and converge on talent mobility and cross-domain social trust as the root solution.

Round one — what each of them fears most

• Ed H. Chi — recalls the 2022 Bard 100-day sprint: the LaMDA team was 3–5 people, the training data was full of Reddit-style invective, and the early chatbot's sensitivity was not good enough, causing an internal uproar and fears the CEO would be hauled before Congress. The hardest part of AI safety is that it is not universal: medicine, finance, education and manufacturing need completely different safety criteria, and no single standard covers them.
• Chen Wei-chao — argues the enterprise view should treat ordinary security and AI safety together, because the core is still databases getting breached and accounts stolen. He flags the IT/OT gap (a Windows box beside a production line may not have been patched in ten years). In his view the model itself carries limited risk today because it sits in a black box, not wired to embodiment or buttons. The real risk is employees pasting confidential data — order prices — into ChatGPT or Gemini.
• Chiu Wei-chen — names three core risks: privacy and security (MCP is easy to enter and describes tools in natural language, so malicious instructions slip in, and users are often unaware while the agent runs; software supply chains also lack the scrutiny hardware gets); loss of control (agentic AI plans and executes on chain-of-thought and can walk outside the range the user expected or can intervene in); and regulation lagging far behind. He also notes fact-checking today relies on cross-source comparison — effectively majority voting — which will fail fast as generated content floods in.

Round two — safety versus speed of innovation

• Ed H. Chi — exploratory first: let's find out what the possibilities are, then handle safety. Iteration is now so fast that exploration to shipped product can be weeks rather than years. Google's internal practice is to open access widely to employees for red teaming, then talk to society: we understand your anxiety, but not knowing the possibility makes you more anxious.
• Chen Wei-chao — on the ROI of safety investment, he says no investment maps cleanly to profit and AI is not special. He offers a diffusion → downstream frame: most enterprises are still stuck at diffusion (are employees actually using company-approved tools rather than something else?) and only then meet downstream consequences (AI emails a customer the wrong quote). Safety investment should follow the same logic as security investment, through the CSO structure — while keeping security separate from trust and privacy.

Round three — the government's role

• The DeepSeek episode early in the year forced the Ministry of Digital Affairs and NICS to mobilise to assess risk and set usage rules — and exposed how thin the government's combined AI-plus-security capability is: NICS is mostly traditional systems-security people. Government cannot do it all and must draw on the private sector.
• On legislation — law is society's last line of order, and the legislative process is too slow for the technology. The pragmatic move now is bottom-line legislation for high-risk domains with soft guidance elsewhere. Beyond exploration, critical domains need exploitation: understanding why the AI decided as it did and what follows.
• Three places government can act — partner with industry on data governance and security tooling frameworks and funding, to grow an AI safety industry; build traceability of transparency and responsibility (a future single government portal running cross-agency queries via AI agents needs a distinct identity and access entry point per agent); and push AI safety awareness down into civic education.

Round four — talent and education

Ed Chi's contribution here is the conference's sharpest technical point. Old systems thinking is divide and conquer — A safe + B safe + C safe = the whole is safe. But large language models have compositionality: combining capabilities produces new emergent capabilities (translation plus summarisation composes into something new), so safety multiplies rather than adds. Safety analysis is therefore far harder than before, and talent development has to transmit that shift in thinking rather than only technique.

• Chen Wei-chao's two layers — AI literacy (teach employees to use tools correctly; at minimum, that the query you send is remembered forever) and abstraction (stand up a company-approved internal system as a wall, while pragmatically accepting employees will still use outside tools). His trial: a warning trigger that does not monitor all conversation but raises an alert when something resembling confidential data appears to be heading to an external model, for after-the-fact audit.
• Chiu Wei-chen — the two talent pools do not mix, and government can broker. Industry problems should come into the classroom (joint labs, internships) instead of academia simplifying problems into familiar algorithmic forms. He urges academia to be more open to people without PhDs but with strong hands-on ability, praising NYCU's recruitment of Jserv — and discloses that of about 80 government-funded security teaching posts with salary top-ups, actual take-up was only about 25%, mainly because some departments screened applicants by even higher criteria.

The safety of a large language model is not A plus B plus C — it is A times B times C.Ed H. Chi, on compositionality

Law is society's last line of order.Chiu Wei-chen

04 · Huang Chih-hao — NTU of Sport

Moneyball 2.0: AI optimising sport in the age of large simulation

Markerless 3D motion capture, OpenSim musculoskeletal simulation and ultrasound AI screening, used to quantify pitching mechanics, prevent injury and train youth and professional players.

• Why he moved off scouting — scouting is a multiplier on the result (roughly ×0.9 to ×1.1); the baseline ability score in front of it matters more, so he shifted to raising that baseline through movement training.
• Why markerless — with reflective markers the athlete strips to underwear and the markers fall off above about 140 km/h, distorting the motion away from a natural competitive posture. The lab instead shoots at 1,050 frames per second with 4–6+ synchronised high-speed cameras, extracts joints with pose estimators (MediaPipe, OpenPose, Apple Vision, RTMPose) and smooths the jitter.
• Accuracy — the team's 2024 IEEE Access paper put markerless 3D joint reconstruction at about 3 cm against marker-based; 2025 optimisation brought it to about 2 cm.
• Three OpenSim analyses — joint angles at key moments (shoulder abduction at front-foot strike), kinetic-chain defects (a low elbow causing abnormal arm external rotation at release), and per-joint loading for injury prevention.
• The injury finding that matters — UCL (Tommy John ligament) loading during the follow-through exceeds that at maximum external shoulder rotation: a risk the naked eye cannot see.
• Pitch-side live system — every pitch is checked against the normative range and flagged red/amber/green (50–100% of norm green, 10% out amber, further out red), with AI-generated commentary.
• Ultrasound UCL screening — built with Dr Kuo Po-lin's team at NTU Hospital; the AI identifies humerus, ulna and joint space and returns the highest-scoring image for an orthopaedist to read. AI may not diagnose (medical law), only screen. Trialled at Pingjhen High School — zero pitcher injuries that academic year.
• Hardware economics — force plates are expensive (60×60cm about NT$600k, 100×100cm about NT$1m; the university's 10-metre plate cost about NT$10m). The team is developing a micro strain-sensing strip as a low-cost substitute.
• Older-adult spillover — strengthening dorsiflexors and hip abductors can drop knee loading from 1.7× body weight to 1.1×.

A lot of sports-science demos look beautiful. But lab ready to game ready is a very long road.

Ask an athlete if it hurts and it never hurts. It will never hurt.on athletes hiding injury to stay on the field

05 · Hu Min-chun — NTHU

Fusing sensors, AI and VR for athlete training

Working mainly in basketball, her team does fine-grained action and hotspot analysis from a single camera or phone, AI-assisted collaborative annotation, and affordable immersive tactical and decision training in XR/VR.

• The affordability problem — the NBA has had multiple ceiling cameras since 2012, but Taiwanese teams mostly rent venues (even the floor logo is installed and removed, about NT$200,000 a game). UWB-style local positioning runs about NT$3m a set, and consumer apps (Home Court, Zap) misclassify events such as a pass read as a shot.
• Their answer — allow a single movable phone or camera and still classify fine-grained events (layup, bank shot, dunk) in real time. Deployed at T1 League games and integrated with the league's manual tagging system (which needs six people annotating in sync) as an AI-first, human-confirms workflow, with AI tags marked as system predictive for later correction.
• PTZ auto-tracking camera — trained by imitation learning to mimic a professional camera operator's framing logic rather than simply chasing the largest cluster of movement, staying stable from the sideline or a corner and filtering out bystanders.
• 3D mesh reconstruction plus AR effects — reconstructs the ball-handler's 3D mesh so effects stick precisely to a specific player. A brand's social team hand-editing a ten-second effect in Adobe needs about six hours; the AI judges the play break and the ball-handler in real time.
• Visual attention prediction — working with an eye-tracking glasses team, they collected spectators' gaze plus footage to train a model predicting which player the audience will watch, which then decides who gets the effect.
• Virtual ad replacement — estimates per-pixel depth to swap billboards for virtual 2D/3D ads with correct occlusion.
• Coaching platform for beginners — born from COVID, a web platform where students upload footage and coaches annotate and compare skeletons. A 1.5-second shooting motion is decomposed into nearly 20 key steps as a personalised SOP; the system compares the student against the coach's template and generates a report.
• VR tactical training — the coach draws attacking routes on a tablet, AI generates the corresponding defender trajectories (auto-regressive / transformer generative model), and both go into the headset so the player feels, in first person, the space a screen opens up. Now multiplayer, with separate coach and player UIs and custom avatars.
• VR decision training — defensive scenarios are pre-recorded or animated and the player must respond in real time, with the system recognising about 20 actions and demanding a redo when wrong. An unexpected finding: because VR forces players not to look at the ball, post-training tests showed dribbling improved across the board.

Players hear the tactic and think they've got it — they memorise the lines off the whiteboard, then get on court and can't run it. Because on court there are real people defending you, and on the board you only see dots, and the dots barely move.

06 · Yi Chih-wei — NYCU

From CoachAI to a smart badminton venue platform

From the badminton strand of the National Science and Technology Council's precision sports science programme in 2017: a shot-type-recognising smart racket, the TrackNet shuttle tracking model, per-stroke micro-data annotation, and an affordable multi-camera smart venue platform.

• Smart racket — an IMU sensor embedded in the racket; early pure-inertial swing detection misfired (students hitting each other's rackets read as swings), so they added acoustic recognition (the racket striking the shuttle resonates like an instrument) timestamped against the inertial signal, with classical feature engineering extracting hundreds of features. Goal: recognise shot type and count training volume (did you finish 100 smashes and 50 clears?).
• Smart serving machine — friction-roller models (about NT$100,000) are mostly Chinese-made and the Taiwanese distributor would not open the controller interface, so the team reverse-engineered and replaced the control board for wireless remote control, filed a patent and won competition awards. Used in AI for Kids high-school badminton teaching.
• TrackNet — validated on tennis first (bright colour, slower) before badminton (small object, extremely fast, few visual features). The multi-frame idea lets the model learn flight-trajectory features across consecutive frames; later evolved into multi-in multi-out (10-in-10-out) for throughput.
• Why per-stroke — badminton is a net sport, so the only moment a player controls the shuttle is the strike. Data is therefore collected per stroke: posture (forehand/backhand, dominant hand and foot — badminton is a same-hand-same-foot sport), contact height relative to the net (above means initiative, below means disadvantage), and contact coordinates.
• Case analyses — loss-cause distribution by type (net, out, floor, misjudgement) with video corroboration; smash count versus effective smash rate showing different playing styles; movement heat zones, including Tai Tzu-ying, whose opponents deliberately played to the rear court out of respect for her front-court game, which shows up directly in her rear-court return and movement-speed data.
• Per-stroke win contribution — back out the probability of winning the rally after each stroke from historical positioning and stroke data, then use the difference between adjacent strokes to quantify how much that stroke benefited whom, letting a coach point precisely at the decisive rallies.
• Smart venue system — a modular, extensible distributed multi-camera system doing edge preprocessing per camera before merging for 3D trajectory analysis, deployed with Docker. One RTX 4090 AI PC handles 8 cameras in real time (measured ceiling around 12). Designed as an open IoT interface so third parties can build on the data.

Video can tell story — broadcast footage actually records every detail of the match.

07 · Li Chih-ching — Taiwan Hope Innovation

AI-enabled drones: from inspection to defence

A company that started in swarm light shows, showing how it worked AI into inspection (bridge crack detection, autonomous semiconductor-fab patrol) and defence (unmanned surface vessels, loitering munitions) — and revealing a domestic flight-computer module being built with Pegatron and MediaTek on the Dimensity 9400.

• Where to plug in if you know AI but not drone hardware — MAVLink, the protocol between drone and ground station (heartbeat, status, commands), is the natural interface. Flight-control software splits into ArduPilot (from a 2009 Arduino Mega, now STM32/ARM, 100% open source, GPL) and PX4 (foundation-backed).
• Sensor economics — inertial navigation components once cost about NT$300,000 and importing them required a signed undertaking (to prevent missile use); an equivalent component is now about US$3, though the noise has to be filtered.
• Bridge inspection drone — a Y6 configuration (six motors so the arms stay out of shot) with Intel RealSense and inertial navigation. There is no GPS under a bridge, so it flies on vision plus IMU, photographing every 50 cm at a resolution of 0.2 mm to catch the smallest cracks.
• Semiconductor fab — validated fully autonomous flight through a 1-metre tunnel, automatic obstacle avoidance and precision landing to charge, covering the blind spots an inspection AGV cannot see.
• Second-generation airframe — LiDAR plus NVIDIA Orin NX, automatic dock landing and charging, dual-spectrum camera (visible plus thermal); plus a four-fisheye 360° obstacle-avoidance rig (rectify and stitch the fisheye images, estimate depth with a neural network, build a point cloud), already tested for autonomous navigation along their Taipei office corridor.
• The localisation project — with Pegatron and MediaTek, replacing NVIDIA modules with the Dimensity 9400 (about 50 TOPS, enough for four-eye fisheye localisation, navigation and avoidance) on Linux plus ROS2 with in-house VIO. Expected price around US$600 at roughly 100 TOPS, comparable to Orin NX. On a drone, he notes, you count every gram and every watt.
• Defence — two fixed-wing types in the Armaments Bureau's public exercise footage are theirs. He describes separating sensing from striking: a scout returns coordinates and the loitering munition strikes, still confirming the target visually before impact. Fibre-optic drones have risen in Ukraine because of jamming, and AI's role is to keep locking on and complete the attack when comms and positioning are jammed.

I often say — everyone here is probably an AI expert. What am I an expert in? I'm an expert in messing about with aircraft.

In the end my AI will fly my aircraft at that target and attack it. This is a cruel thing, but it is happening right now.

08 · Chen Yi-ting — NYCU

Toward human-centred physical AI

Against the global mainstream of collecting mass data to train a general robot foundation model, he argues for a human-centred entry into physical AI, using assistive feeding and scenario-based autonomous-driving safety validation to show that real deployment needs three things: understanding user needs, situated safety validation, and cross-university talent development.

• Historical anchor — Shakey the Robot (1966–1972, Stanford Research Institute) already had the perceive-reason-act architecture, and one of the algorithms it produced, A* search, still routes Google Maps today.
• Assistive feeding — the commercially available Obi robot is designed for people with disabilities but its motions are pre-set (scoop from a fixed position) and cannot react to a changed environment. In a demo video a friend used it for ice cream; the robot did not know the texture had changed, scooped hard, and sprayed food.
• Their algorithm estimates the probability of spillage during the scoop in real time (visualised as red and blue risk curves) and corrects the motion accordingly, extending to compound sequences such as spearing a strawberry and dipping it in chocolate.
• The harder half is subjective — feeding is not just getting food to the mouth; bite size matters, and a demo where the mouthful was too large left the student visibly worn out. User interviews found assistive-device demonstration barely exists in Taiwan, devices are expensive, and using one easily costs the user autonomy over what they eat.
• Scenario-based safety validation — grounded in Taiwanese traffic (a scooter running a red light, a distracted driver) and in Cruise's fatal San Francisco crash, which nearly ended the company. ISO 21448 (SOTIF) is the standard for validating whether a black-box AI system understands the scene and operates safely within its design domain.
• Current methods do not scale — fixed-site and closed-road testing cannot cover Taiwan's uniquely dense, heterogeneous mixed traffic. Their simulation pipeline: fly a drone over accident-hotspot junctions for tens of hours → annotate specific hazardous scenarios (braking, cut-ins) → statistical analysis → convert to OpenSCENARIO templates → bridge into open autonomous stacks (Autoware, Apollo) → output a safety performance report with pass/fail rates and performance-boundary analysis.
• Data collection bottleneck — drone endurance is only about 20 minutes, so they are evaluating V2X or roadside monitoring for long-horizon collection.
• Talent — his lab spun out the student startup Twins Robotics; the NYCU Robotics Center works on robot as a team member for critical missions, combining the medical school and nursing with electrical engineering and computer science, in real care settings. He ran a cross-university co-teaching camp in summer 2025 with at least four schools.

I want this robot to come into my home and do all sorts of things for me. But if it does come into my home — my home is a mess. How is it supposed to figure out what my home even looks like?

Back then they were inside Stanford Research Institute. We don't need to be inside one institute — we can push this forward with the whole of Taiwan.

09 · Sun Shao-hua — NTU EE

Program-guided robot learning

Argues the mainstream black-box neural policy has two defects — generalisation and interpretability — proves it with a maze experiment showing the network only memorises rather than learning the right-hand rule, and presents Program-Guided Robot Learning: synthesise a human-readable, verifiable, locally debuggable program as the decision logic before acting.

• Industrial automation versus general robot intelligence — automation scenes are deliberately designed for reduced diversity in a closed, structured environment, which is exactly why they cannot breed real intelligence. A home robot faces three axes: a highly unstructured, changeable environment; unknown objects (it must estimate 3D structure, material, weight, grasp points and friction itself rather than look them up in a factory object library); and genuinely new tasks rather than pre-programmed ones.
• The layer definitions — Generative AI (language and vision-language models) → Agentic AI (plus digital tool use: search, solvers, code execution, external APIs) → Physical AI (plus physical tool use: cameras, LiDAR and microphones for perception, arms and grippers for action).
• Robot learning formalised — a high-dimensional mapping from perception (sensory input) to action (per-joint motor torque; a robotic hand can have 24 degrees of freedom), usually modelled with a neural network optimised by gradient descent.
• Defect one, generalisation — a deep-RL agent trained on 8×8 mazes solves any 8×8 layout perfectly and looks like it has learned the right-hand rule (follow the wall; not the shortest path but guaranteed to exit any connected maze). Put the same model in a larger maze and it fails completely, proving it memorised every training layout with high capacity.
• Defect two, interpretability — a black box cannot explain its decision, so users cannot trust it: when the feeding robot fails you cannot find out why or debug it, and the same applies to language models.
• The pipeline has three modules — Task Interpretation / Program Inference (synthesise a readable program in a domain-specific language from a video demonstration or reward function before acting), Primitive Skills Acquisition (a pre-learned set of building blocks: move, grip, assemble), and Task Execution (run the program to produce a high-level plan, then hand off to low-level motor control).
• Demo-to-Program — neural networks cut corners on rare if-else branches in the data because ignoring them keeps loss low; requiring a program that covers all demonstration logic catches the rare else branch correctly.
• LIPS — learn a continuous embedding space over a random program dataset (nearby latents behave similarly), then sample in that space against the task reward, decode to a program, run it, and re-sample by return. On the maze it produced exactly the right-hand rule, generalising to any size with no performance drop.
• The interpretability experiment — USC undergraduates paid US$10/hour debugged a not-yet-perfect generated program, and needed to change only 3 to 5 tokens to improve performance substantially. Neural-network weights are essentially impossible to hand-tune by comparison.
• LLM-guided search — search and RL find good programs but slowly, with no commonsense assumptions. An LLM has commonsense and coding ability (he calls it a text-version clone of human civilisation) but is hard to optimise directly for a task, so they combine LLM-generated initial programs with search-based refinement. A further experiment tried letting the LLM iterate alone from environment feedback and found it lacks credit assignment — it cannot tell which part of the program did well or badly — so search or RL is still needed.

If you genuinely figure out how to solve the maze task, the right-hand-rule program generalises to a maze of any size whatsoever. It is a general solution.

10 · Chih Ming-yang — Chihan Quantum Security

Open standards and development platforms for drone systems

Buying a drone is not the same as running a successful drone programme. Illustrated by America's mature police drone model, Ukrainian mine clearance, and the lesson of Taiwan's storm season, when not a single local drone got airborne — plus the association's standards platform, disaster-resilience drills, and a post-quantum cryptography framework.

• The first success of the open-standard platform — mine clearance in Ukraine: a drone with sensors photographs the terrain and the back end computes the best route for laying a limited length of clearance cord. Planning that manually took a full day; with drones it takes about an hour.
• The association, founded in 2023, runs four programmes — de-risking the red supply chain (auditing key technologies such as thermal cameras, batteries and autonomous flight every six months since March 2025, already drawing listed companies in), a drone standards platform, robotics and robot dogs from 2026, and a disaster-resilience participation programme.
• The US police case — since 2014, nearly every police unit has drones, and the drone often reaches the scene before officers do: an average of 71 seconds. Phoenix PD flies over 10,000 missions a year. Drone as first responder is now a mature commercial model.
• The Taiwanese reality check — during this year's storms in Tainan and Chiayi, every local drone including those at the Chiayi drone park failed to get airborne, because over 300 utility poles came down (power and telecom share the poles) and 4G/5G went with them.
• Three lessons from international disaster response — communications must be standardised; training is inadequate (overseas practice is 40 hours of cross-agency drill per quarter across mountain rescue, typhoon road-clearing and chemical decontamination); and capability and equipment tiering is missing (3D disaster-zone modelling, GPS-free flight, night flight, live video relay — without getting data back to the command centre in real time the drone team is useless in a rescue).
• The 2025 resilience exercise programme covers — aerial 3D modelling relayed to the command centre to judge road passability; sweep-pattern autonomous search (thermal plus optical to find survivors, cheaper and more efficient than a helicopter); broadcast to reassure survivors (about three hours without communications and morale closes down); supply drops (prescription medicine, power banks); communications relay (about 100 km on current drones, roughly 300 km on military and commercial systems, 2–100 Mbps, covering an emergency service area about 5 km out from the safe zone); stored power (a hydrogen-producing storage system delivering about 5 kWh, because after poles come down the water has to recede before they can be raised and reconnected — a gap that can run past five days); and post-disaster disinfection (a golf course takes 8 hours by hand and one night by drone, without closing).
• Post-quantum cryptography — quantum computers will break current encryption; the US government has set 2030 as the deadline. Google's quantum computer announced earlier in 2025 solved in five minutes what previously took a year. NIST provides a risk-management framework for asset assessment and algorithm migration, and the association has built an encrypted transmission, storage and authentication platform for drone systems and sovereign AI infrastructure.
• Regulatory friction — high-speed rail corridors are no-fly zones. The FAA precedent took 3–4 years of negotiation to bring police drone route approval from 6–12 months down to as little as 6 hours (at most 6 days).

You bought a drone and brought it home. Getting to the point where you can actually apply it is a completely different thing.

Losing communications in a disaster zone is more dangerous than the disaster.

11 · Lightning · Ian Wu — Accu Crazy

Content monetisation 2.0 in the AI era

Marketing has moved from precision targeting (audience base) to content is king (content base): mass-producing localised word-of-mouth copy with a fine-tuned model, replacing studio photography with Nano Banana image editing, and running a multi-agent system that generates topical marketing posts automatically.

• The shift — mass buying used to solve reach. Now generating content is trivially easy, so the real difficulty becomes producing the right content.
• Fine-tuning — training a client's word-of-mouth articles into a model produced invented internet slang the speaker verified had never appeared online before, which he calls emergence.
• The control experiment — training a model on all of Dcard produced something full of hatred and harassment. Garbage in, garbage model out: curating the data matters more than the model.
• A Turing-test-ish field trial — an AI-written article posted to a forum got 1,200 likes and 1,300 comments and was reposted by the platform's own editors into the featured section (normally a paid placement), with no reader spotting it as sponsored or AI-generated.
• The multi-agent system — Moana produces word-of-mouth copy by persona and brand angle, Pandora monitors sentiment, Eve tracks breaking news, Johnny plans article angles and turns sentiment data into a knowledge graph.
• Cost collapse — a bridal client generated a Bali wedding concept image live in a meeting, replacing a NT$500,000 shoot at an AI cost of about NT$10. A designer's ad asset goes for about NT$3,000 each; AI produces a dozen at once.

12 · Lightning · Eric Chang — Taiwan AI Labs

The open-model wave: how enterprises keep control of AI

Enterprises should stop agonising over open versus closed and on-prem versus cloud, and focus instead on how the organisation designs its agentic workflow — including the decision points where a human must intervene.

• The common myth — grab an open model and fine-tune it yourself, while forgetting you are not NVIDIA's biggest customer and you do not have Google's or Microsoft's thousands of engineers. Work out what problem you are actually solving first.
• Three success factors from MIT research with his former employer — partner with a specialist ISV rather than doing everything yourself; do not start from the pain point where everything hurts, pick one seed case that can succeed bottom-up and let it pull the rest of the organisation; and choose an extensible platform you can grow more vertical applications on.
• Why agentic flow matters — a real procurement process (request → manager approval → chairman approval above a threshold → purchasing → cashier → accounting → shipping) has many decision points, each needing human-in-the-loop. That is what AI governance actually means, rather than a slogan.
• Deployment essentials — tunable back-end parameters, tiered permissions (chairman, GM, admin, ordinary user), a user feedback mechanism to flag wrong answers, and industry-specific pre-trained models. A small, well-made brain suits an enterprise better than one that knows everything.

13 · Lightning · Jason Lin — iCook Intelligence

How brands use AI to build consumer communication

Marketing's decisive factor has moved from ad-buying technique to the company's own data assets, because technical equity means everyone has the same AI tools — so differentiation comes from industry know-how and first-party data.

• Three drivers behind the data industry — digital transformation created a flood of data sources, processing and AI analysis tools matured, and personalisation demand pushed toward a different face for every customer.
• Marketing's own shift — early advertising was shallow contact; the last 5–7 years of martech focused on putting 80% of revenue on the 20% of loyal customers (lifetime value); the last two years opened channels end to end, with QR-code membership binding and 360-degree cross-channel journey profiles.
• Schumpeter's point — genuine disruptive innovation comes from new combinations rather than wholly new inventions, and data × AI is a textbook new combination.
• Citing a Lenny's Podcast interview with a Microsoft AI platform CVP: Microsoft's two biggest worries are enterprises over-relying on AI, and security holes. He reads both as a business opening.

14 · Lightning · Nick Chuang — Chengyu Intelligence

Unlocking LLM inference with vLLM: a quick benchmark guide

A purely hands-on session: how to benchmark inference hardware with vLLM, covering Docker deployment gotchas, NCCL version compatibility landmines, how to read the key metrics, and how to use MLPerf for comparison.

• Three inference frameworks — vLLM suits multi-GPU and multi-node deployment and is the enterprise default; Ollama works like a Docker CLI and suits fast local testing with a ready model repository; LM Studio is the third option.
• Deployment gotchas — use the official vLLM Docker image (there is a separate one packaged for GPT-OSS), and always download the model locally first: loading online usually fails. Override the entry point and run two command lines, one serving the model and one running the benchmark (which needs pandas and datasets installed).
• The landmine — NVIDIA's NCCL version has to match the card tier. A lower-tier new card such as the RTX 6000 series needs a manual upgrade to a newer NCCL (2.26.5), or the official image (built for B200 and H100 class cards) simply will not run.
• Key parameters — tensor_parallel_size (four cards in his test) and GPU memory utilisation pushed as high as it goes without OOM. The approach is platform-agnostic; any container environment works.
• Reading the report (add --profile for the concise version) — Time to First Token, in his example about 1,644 ms, which is on the fast side; average latency after the first token; and inter-token latency. The trade-off is fixed: shortening TTFT lengthens the gap between subsequent output tokens, tunable via max_model_len and max_num_seqs.
• MLPerf — public scores from different hardware vendors are comparable on the MLPerf site and useful as procurement leverage; his own test results matched the official numbers.

15 · Lightning · Shiny Lai — Sinotech Intelligence

AI-driven AIoT × GeoAI for disaster prevention

Combining IoT sensing (rainfall, water level, seismographs), a GIS spatial platform and a fine-tuned llama 3 into a disaster-prevention digital twin that runs from live monitoring through risk simulation to a fast decision-recommendation report — already used by New Taipei and Taipei emergency operations centres and in the Hualien earthquake response.

• Differentiation — not just integrating Taiwan's existing IoT device data (rain gauges, water-level gauges, seismographs) but bringing domain knowledge in disaster prevention, so it can do further AI and big-data analysis and integrate UAV aerial and satellite imagery.
• The LLM piece — llama 3 fine-tuned on fire-agency and district-office disaster SOPs and historical earthquake response experience, producing a draft decision-recommendation report that staff edit on the platform for the commander, replacing the time cost of compiling reports by hand.
• Deployed cases — New Taipei's smart monitoring platform (the commander can judge where to enter a rescue), Taipei's EOC platform (decision-makers need not be physically present to hear a briefing), Kaohsiung earthquake and flood simulation, and the Hualien earthquake, where live UAV imagery was turned into 3D models so the commander could judge access routes from a bird's-eye view.
• Where it extends — financial and insurance use: banks assessing asset risk for lending and insurers assessing high-risk areas for claims, using public soil, geology and meteorological history.
• Digital twin extension — AI image recognition strips sensitive objects before rebuilding a site model, for factory training, industrial-safety incident simulation and AR/VR integration.

16 · Lightning · Ken Yeh — Deep Wave

Audio AI: what it can and cannot do

The most valuable talk of the conference about a failure: why factory anomalous-sound detection — machine maintenance and speaker end-of-line sweep testing — never became a product, and why technology-product-market fit matters more than recognition rate.

• The technology worked — four consecutive months of one motor's audio, indistinguishable to the human ear, were correctly analysed by similarity to predict that the machine needed servicing, which the repair confirmed. On soundbar sweep testing, normal, air-leaking and mis-assembled units sound alike to most people, and basic spectrum features with a simple model such as a GMM already beat the human ear in the lab. The traditional method is an experienced worker sitting on the line for eight hours a day listening.
• The scaling barrier — a real production floor is far messier than a lab: multiple lines running at once interfere with the recording; the hardware fails (a microphone's wiring burnt through because acid mist concentration on the line was too high); choosing recording equipment is hard. The human ear is about 80% accurate and the machine reaches 90–95% in a controlled environment, but the line floor cannot reproduce that even with noise reduction — and each site may need its own retrained model. Scalability never arrived, so the team dropped productisation.
• The five design principles behind their current AI meeting-notes product — strong model performance, broad applicability (online and offline meetings), scalability (PC, app and web, with the LLM swappable on the back end), focus on being genuinely usable rather than piling every AI feature in, and a cost that lets them make money.
• The conclusion — what decides an AI product is not recognition rate, but how closely it fits the user's experience.

An AI product needs more than product-market fit. It needs technology-product-market fit. A good model doesn't mean it maps to a suitable scenario; mapping to a scenario doesn't mean it scales; scaling doesn't mean users find it usable.

17 · Lightning · Jerry Wu — APMIC

The key ideas in enterprise LLM privatisation

The decisive question for enterprise LLM adoption is not which open model but fine-tuning — because public benchmarks, mostly multiple choice, cannot reflect the open-ended questions enterprises actually ask.

• The three-layer fine-tuning architecture — sovereign AI model (national and cultural knowledge) → industry knowledge model (finance, semiconductors) → enterprise knowledge model (yours alone, which only you can build).
• Searching → conversation as a paradigm shift — traditional search costs the same whatever the query is worth; in conversational AI a trivial question and a complex commercial one burn the same tokens for very different value, which argues for a hybrid architecture running ordinary traffic on-prem and complex traffic in the cloud.
• Fine-tuning creates valuable tokens — an un-tuned model facing a question outside its domain keeps thinking, burning tokens and compute for an imprecise answer.
• His six recommendations — open models cannot meet enterprise market needs directly; fine-tuning should be layered sovereign/industry/enterprise; establish measurement criteria before an agent goes live rather than going on feel; build your own value baseline and knowledge-training software; buy tools rather than build; and note that not every model is worth fine-tuning, because malleability varies.

You cannot weigh the depth of water on a scale.on judging a model's commercial fitness with multiple-choice benchmarks

18 · Lightning · Huang Liang-hsun — Twinkle AI

The art of GRPO training

Twinkle AI is the only community in Taiwan focused solely on open Traditional Chinese models. He explains GRPO against PPO with a study analogy — GRPO drops the separate value model for a hand-written reward function, saving substantial GPU — and backs it with their F1 model going from 50 to 64 on MMLU.

• The community — not a company. Founded 2 January 2025, members from companies and institutions across Taiwan participating out of interest, aiming to build Taiwan's next strongest Traditional Chinese model. Earlier in the year he personally gathered about 30B tokens of Traditional Chinese corpus and single-handedly trained and open-sourced a llama 3.2 Traditional Chinese model.
• The reasoning-model project — with APMIC and a defence centre, prepared from February and released in April: a non-reasoning Traditional Chinese model, a reasoning version showing its working, quantised versions, and open reasoning datasets (mathematical, everyday, tool-calling), all under MIT or CC licences on Hugging Face.
• Benchmark results for the 3B model — clearly ahead on TMMLU+ (general knowledge including Taiwan's civil-service exam bank, nearly 10,000 questions); Breeze 2 scored higher on the legal subset (he volunteered his competitor's advantage); function calling including MCP scored above 90, higher than GPT-4o; and after GRPO training, MATH-500 and GPQA Diamond both improved clearly over the base model.
• The small-model experiment — a Traditional Chinese version of Gemma 3 270M. The stock version mixed Simplified characters into Traditional answers and held only basic, sometimes wrong knowledge about Taiwan; after training it used Traditional Chinese correctly and answered with richer local knowledge. On a maths question the stock model got stuck repeating one token, while the trained version produced the correct formula. He expects home-GPU-trainable small models to act as routers and to excel in specific domains.
• The limitation — if the base model never saw the knowledge in pre-training, GRPO cannot train it in either.
• Community output — TwinkleLM Lab (free online training sessions every week or two, nicknamed the late-night diner because they run after 10:30pm), in-house evaluation tooling, self-collected training datasets, a Traditional Chinese coder model, an on-device LLM phone app, and an LLM red-team tool.

19 · Huang Neng-fu — Memorial Lecture

CHEN SHENG-WEI MEMORIAL LECTURE — Industrial AI deployment and hands-on AI talent

Ten years of visiting farms across Taiwan — open-field and greenhouse, soil and hydroponic — showing how AI vision and hearing actually solve labour shortage, climate change and an ageing workforce, plus the no-code AI Maker platform he built for hands-on talent development.

• Architecture — sensors in the field (soil, water, air, micro weather stations, cameras, drone sampling) → transmission (WiFi and Bluetooth are too short-range for a farm, so LoRa and NB-IoT carry data and 4G/5G carries video; solar power for remote sites without grid) → cloud modelling and decisions → back to farm automation. LoRa reaches about 20 km on the ground, NB-IoT about 3 km; they are working with the National Space Organization on low-earth-orbit satellites (500–600 km) for the Central Mountain Range's communication blind spots.
• Why edge AI — a cloud round trip takes about 2–3 seconds, too slow for many real-time uses, so models move onto edge hosts (NVIDIA Jetson Nano/Orin) and then onto the device itself (smart glasses) for one-sided inference with no comms at all.
• AI glasses versus AR glasses — AI glasses can only capture, recognise and answer by sound; AR glasses have optical projection and can overlay results for visual interaction, such as drawing a box around ripe coffee cherries. Prices: the demo AR unit about 100 grams at NT$200,000 (not mass-produced), a mass-produced version about NT$30,000, Xiaomi's AI glasses about RMB 1,999. He thinks AI/AR glasses may be the best tool for landing AI.
• Cucumber-harvesting robot — driven by farmer demand: cucumbers grow so fast they need harvesting twice a day and become seconds if overripe. Japan already has a suspended harvesting robot at about NT$600,000.
• Banana grading (Pingtung, Taiwan's largest banana cooperative, supplying 7-11) — AI vision detects bruising and the bee-sting black spots the eye struggles with, beating human sharpness especially once presbyopia and fatigue are factored in.
• Shrimp-larva counting — manual visual counting is off by up to 20%, which is the whole profit margin; AI vision counts precisely within a second and can also track feeding time and body-length changes.
• Dengue mosquito egg monitoring (Tainan) — 37 districts and over 3,000 ovitraps, previously opened and counted by hand, now recognised, counted and transmitted live by AR smart glasses.
• Soft IV bag foreign-object inspection — AI detects fibres around 500 micron and plastic fragments around 400 micron; the customer wants 100 micron, entirely invisible to the eye. The current method needs 20 people with 1.0 vision.
• AI tapping stick for pineapples (Pingtung Kaoshu, 80 hectares, exported to Japan) — too much water content means the fruit leaks in transit and a whole container is rejected. Traditionally a worker taps with a hot-glue stick and judges by ear. His stick has an AI chip that can load different crop models and shows red/amber/green within one second of the tap, because farmers said they cannot wait three seconds. A fully automated conveyor-plus-robot-arm version has been patented.
• AI clam grader — clams drop onto a sensor plate on a high-speed conveyor and the sound tells live from dead. Currently 10 clams a second (about 600 a minute) per line against 2,800 a minute by hand, so the plan is ten lines for 6,000 a minute.
• Solar AI drying — the panel produces about 100°C hot air rather than electricity, feeding a drying cabinet for fruit, vegetables, herbs, coffee and tea, with AI vision watching the drying to protect colour, texture and nutrients. Zero-carbon and ESG-friendly against open-air drying (animal urine, sudden rain) or diesel dryers. The Indonesian and Thai student team behind it won the 2025 Hult Prize Taiwan championship.
• AI Maker — a no-code platform for students and engineers to build vision and hearing models: collect data → label → train → download and deploy (phone, glasses) → run → evaluate → iterate, with no programming at any step. Alongside it, a distributed compute-sharing alliance across universities (NTHU, STUST, National University of Kaohsiung and others, now extending to India, Malaysia, the Philippines and Japan) so schools without compute can take part.

Doing AI means being fast and accurate at the same time.

If your machine isn't better than me, don't bother coming.the clam farmer, to his team

20 · Tsai Ming-shun — Closing

Ten million dreams, one Taiwan: putting AI competitiveness into practice

A photograph from Manzhou, Pingtung — a tractor with egrets following behind — stands in for eight years of AIA's work, closing the two days and setting out the next eight: public-private collaboration, co-creating an ecosystem, and connecting industry, government, academia and research.

1. Keep collaborating across public and private and connecting communities — bring in more people so Taiwan's gears keep turning.
2. Co-create the ecosystem — he judges Taiwan unlikely to become an AI research power, but with a strong semiconductor and communications supply chain and dense industrial clusters it can absolutely become an AI application power, even an artificial intelligence island.
3. Bring in industry, government, academic and research resources — integrate what exists toward deployment, and bring the international in while taking Taiwan's own work out.

• Proof of action — the agreements signed the previous day with three universities will, beyond certification, build characterised regional industrial clusters in the north, centre and south, integrating companies, universities and research institutes. (The universities were not named here; from Chen Ling-jyh's Day 1 announcement they are National Chengchi University, Feng Chia University and Southern Taiwan University of Science and Technology.) AIA will also help take the AI Maker platform to more sites nationally, explicitly including Hualien, Taitung and the offshore islands.
• The Taiwan Inside formula — AI + D (domain) × T (Taiwan way, and also technology, training, talent, tool) as a frame for companies and individuals driving AI.
• Next chapter from October — a hundred trades and professions angle, starting with the industries most exposed to large language models: advertising, sales and marketing, news media, and management consulting.
• Announced for 14 October — a news-media AI forum plus workshop with international speakers including Hugging Face and Singapore's AI Center, aimed at journalists and advertising and marketing workers. Application-reviewed but free.

Taiwan probably does not have a particularly high chance of becoming a so-called AI research power. But we can absolutely become an application power.