Episodes
Real conversations with the people building Hong Kong's AI future.
I-Ching, Taste & Transparent Systems
Aug steps in front of the camera for this one. We talk about his journey with the I-Ching, the challenges in teaching AI 'taste', and why transparent systems matter more than clever ones.
How AI Is Changing Who Gets to Build Software
We dig into how AI is changing who gets to build software, how rabbit holes are a good thing, and what it's like to ship product in Hong Kong. Alexander shares how he went from oat milk to building conversational interfaces that replace forms for wholesale F&B accounts.
AI, IP & the Dark Secret Nobody Talks About
AI and intellectual property — it's the question everyone runs into eventually. We dig into copyright vs. training data, trade secrets as a practical alternative to patents, music IP disputes, and what Hong Kong's regulatory landscape means for AI builders.
Spatial Computing, Robotics & the Great Reversal
Nils dives into multi-robot coordination and collaborative perception — how phones, glasses, and robots build a shared understanding of physical space. From Warhammer AR overlays during COVID to a patented phone-to-phone calibration technique and retail robots deploying this year. Plus why Hong Kong is the perfect testbed for spatial computing, and the "great reversal" of computers coming out to us.
Latent Space, World Models & What AI Actually Knows
William has been doing AI since high school — 12 years before GPT made it mainstream. We talk about what's really inside these models, the difference between knowledge and intelligence, latent space research, and why the labeled data problem shaped everything.
Legal AI, IP & the Engineer-Lawyer (Cantonese)
Our first fully-Cantonese episode. Kenneth Yip is the founder and general counsel of MakeBell — an AI legal tech startup building domain-specific AI assistants for Hong Kong legal professionals, with a focus on human-in-the-loop design, accuracy, and attorney-client privilege protection. We dig into why lawyers can't just use ChatGPT, IP and patent law in the AI era, and what a dual-track engineer-to-lawyer career looks like in Hong Kong.
Why Search Is Broken & What Comes After RAG
Michelangiolo quit quantitative finance during COVID to teach himself computer science from his home in Italy — now he builds some of the most interesting search infrastructure in Hong Kong. We dig into why vanilla RAG is broken, his "covariate search" approach that applies variable weights across multiple semantic dimensions at once, and an algorithmic tagging method he's publishing that aims to outperform zero-shot LLM labelling at scale. Then the fun part: Michelangiolo wrote and illustrated a 460-page Lovecraftian comic called "The Crimson Duke" using Midjourney — discovering a probabilistic consistency trick for character generation along the way.
Legal LLMs, Pre-Training & Custom AI Hardware
Jeremy is a private M&A lawyer who runs a legal AI startup — and runs his own LLM inference hardware out of his living room because he's been building custom PCs since age 14 (for Call of Duty), then mined Dogecoin on graphics cards, and discovered all that knowledge transfers perfectly to training and serving models. We go deep on why lawyers can't use ChatGPT, Claude, or Gemini with client data, why the entire private equity domain is essentially absent from LLM training corpora, why Qwen punches way above its weight, and why the pre-training ceiling is real.
Expert Systems, Agent Risk & the Fixer Economy
Ronald's tech career predates AI being a boardroom imperative by decades — IBM (on the same GM manufacturing team as pre-Craigslist Craig Newmark), Wang Labs, HP, FedEx Singapore (where the client-server system later used for Apple/Foxconn supply chain was first architected), then computer forensics, then law. He co-founded MakeBell with Kenneth and teaches AI Law at CityU and CUHK. We dig into the symbolic-to-generative-to-neurosymbolic arc, why vibe-coded agents will break at enterprise scale (and the coming "fixer economy" for cleanup specialists), and one oral-exam trick for assessing students in the GPT era.
Shader Art, 3D Worlds & Vibe Coding
A self-taught shader artist from Diamond Hill who beat Google projects to reach the Webby Awards top 5 — Lok Lok Wong builds 3D worlds that run on an iPad using nothing but math, Three.js, and raw curiosity. From particle simulations and shader alchemy to building his own 'Next.js for 3D' framework, Lok Lok shows how creativity and experimentation can replace expensive tools. Plus: using vector embeddings to search the Bible for life wisdom, running local AI on a MacBook Air, and why the metaverse is really about social connection.
Cantonese Lyrics, Vibe Coding & 0243 (Cantonese)
Jenny Wong is a corporate marketer with a creative writing background and zero IT experience. Six months ago she set out to build an AI that could write tonally-correct Cantonese pop lyrics. Every major LLM failed at Cantonese tone-matching. So she taught herself through Dify, n8n, Manus, and Genspark, found the local HK AI community on Telegram and Instagram, met 黃志華 — the godfather of the 0243 tonal notation system — and shipped Cantolyrics.ai, which now generates Hong Kong-quality lyrics in around four minutes. Our second fully-Cantonese episode.
Taxi Driver by Night, AI Startup Founder by Day (Cantonese)
Jacky Man (文子軒) is a 22-year-old HKU double-degree student who co-founded Mastermock — an AI-powered exam question generator that helps teachers save four hours a week on paper-setting. He drives a taxi at night to fund his zero-revenue startup, pitched his way through dozens of rejections to land a government grant, scaled from ten employees down to two (and got better results), and still finds time to hunt for a girlfriend. A raw, honest look at what it's like to be young, broke, and building in Hong Kong's AI scene.
Brain Mapping, Refugee Privacy & Cantonese Tones
A neuroscience PhD who once mapped neonatal mouse brain wiring under a microscope, Marcus Leiwe now runs an AI consultancy in Hong Kong applying the same academic rigor to business problems. From building a private, on-device AI pipeline for Branches of Hope (so refugee records never touch the cloud) to a passion-project Cantonese tone visualizer that runs entirely in the browser, Marcus shows what AI looks like when 'move fast and break things' isn't an option. Plus: Air Canada's chatbot lawsuit, the Vedic 17 levels of consciousness, and why HK is misread as just fintech.
Floppy Disks, Language Models & the I-Ching (Cantonese)
Dr. Bruce Cheung (張維) has been doing AI in Hong Kong since before most people had a computer. One of the first Computer Science graduates from HKU in 1985, he built statistical language models on floppy disks using LISP and Prolog — decades before anyone said 'LLM.' Now Head of the College of Life Sciences and Technology at HKU SPACE, he bridges AI with law and education, arguing that Hong Kong's universities have been quietly world-class in AI research since the 1980s while the public only noticed when ChatGPT arrived. Plus: why Manus costs too much, whether Gemini will ever serve Hong Kong, ADHD in the AI age, and Aug's theory that the I-Ching is the 'bones' for structured AI thinking — which Bruce agrees with.
Apple Park, AI as Clay & Saying No
From Yahoo Mail in the 90s to building EA's online games business out of Shanghai, to leading a prototyping team on Apple Vision Pro — Hanley Leung has worked through every major platform shift of the last 30 years. Now in Hong Kong running an AI startup and the Lunatechs community, he argues this AI cycle rhymes with the 90s internet but moves faster than mobile. Why he calls AI 'clay, not Lego', why ex-Apple founders frustrate VCs, why HTML now beats PowerPoint, and why focus — saying no to good ideas — is the real currency of the AI era.
Compliance AI, Improv, and Running Your Own Models
Andy Green has spent thirty years building systems that cannot be allowed to fail — GSM anti-fraud for mobile networks, the early terminals at the Hong Kong Exchange, and now Three Points Compliance, which builds AI tools for SFC-licensed firms. He argues compliance is the one place where a plausible answer is worse than no answer, and explains how you build AI that knows when to stop. Also: why he runs Qwen at home on modified hardware, why 'never' is a dangerous word, and what improv comedy taught him about engineering.
The Recovering AI Skeptic
Most of our guests arrive already sold on AI. Daniel Sarosi did not. A JVM specialist with twenty years in Hong Kong, he watched people generate code that broke every practice he had spent a career defending — and said so. This is the conversation about what changed his mind and what did not. Along the way: what bad code actually is in plain language, why he still will not trust non-engineers to vibe code anything that matters, why data security is the part nobody talks about, and what foreign companies get wrong about entering China. Then the second half: what Great Wall Connect does day to day, what living in Taiwan and travelling China taught him about rules and flexibility, and where he actually lands on AI — a sophisticated pattern matcher he would not trust to build a secure system alone.
Google, Alibaba, and the New Production Relations
Jason Xu has built product on both sides of the divide — years at Google, then years at Alibaba — and he is unusually precise about how differently the two think. Asked to name two good things about each, he does, and the answers say more than a critique would. Then the conversation turns philosophical: if AI collapses the cost of reorganising a team, what happens to the relationship between a company and the people in it? Jason reorganises his own workforce every three months. He explains what that means when most of the workforce is a model.
Google Analytics for the Real World
Online, every impression is measured. Step outside and the whole industry goes blind. Steve Walsh is CTO of Smart Retail, an edge computer vision platform that brings web-style analytics to physical space — and he is careful to explain how it counts people without identifying them. The engineering is the interesting part: a 30-to-40-million-parameter model small enough to run on a standalone device with no connection, uploading only aggregates. Also: what four years of AI coding tools actually changed about how his team ships.