[Panel sharing at the event Parenting in the AI Age on March 20, 2026]

AI is rapidly becoming one of the most powerful general‑purpose technologies humanity has ever built, reshaping how we consume information, entertain ourselves and relate to one another. It offers phenomenal benefits, but it also stress tests our minds. If we are not careful, AI will not just make information abundant; it will make attention scarce and thinking optional. That is how we drift into artificial ignorance: a state in which powerful tools do so much of the visible thinking that we still look intelligent on the surface, while the underlying muscles of attention, memory and judgment quietly atrophy. This is the real risk facing our children, and those raising them. The question is not only whether AI will grow more intelligent, but whether we will allow ourselves to grow less so.

1. Innovation outruns adaptation

For the first time, the rate of innovation feels consistently faster than the rate of adaptation. We barely absorbed GPT‑3 in 2022 before more capable models landed in 2023 and 2024. Agentic systems now act as digital staff, planning and coordinating quietly in the background.

The curve of technological change has risen above the curve of human or organizational adaptation. Scott Brinker’s Martec’s Law puts it more formally: technology changes exponentially, organisations change logarithmically1 [1].

The gap between those curves is where parents and educators now live. Children inhabit a world of ambient, on‑demand intelligence; adults are still updating policies and habits designed for a slower era. Nowhere is this gap more visible than in how we consume information and spend attention.

2. Information overload with synthetic content

Analysts now warn that we are racing toward a world where much, if not most, online content is synthetic. A Europol‑linked briefing once estimated that “as much as 90% of online content may be synthetically generated by 2026”2 [2]. The precise number may be contested, but the direction is not. A growing share of what scrolls past our children’s eyes synthetic content spun out by machines.

AI tools generate, translate and recombine text, images, audio and video at negligible marginal cost. Studies of synthetic media on platforms like X show spikes in AI‑generated images and videos after each major model release, including viral deepfakes of public figures3 [3]. Misinformation researchers now treat AI‑generated content as a central risk to the integrity of our information environment4 [4].

It is not just about opening a floodgate of information. That flood now also contains:

• more hallucinated facts — confidently wrong answers that sound right,

• more false news and deepfakes, and

• less ability to tell who — or what — actually created what we see.

For a teenager trying to understand the world, signal and noise are becoming harder to distinguish. This is classic information overload, amplified by synthetic media. In response, a cottage industry of “AI detectors” has sprung up. The problem is structural: generators improve continuously; detectors are always one step behind. Europol’s analysis warns that as synthetic media proliferates, technical detection alone will be insufficient; human judgment, contextual verification and more old-fashioned attention will be essential. In other words, the real defence has to live in our minds [2].

That means cultivating a different way of reading and watching:

• Ask high‑quality, grounded questions with enough context. AI systems are pattern‑matchers, not oracles. The more specific your question and situation, the easier it is to see when an answer “sounds right” but clashes with basic facts or lived experience.

• Pre‑empt your own confirmation bias. AI is far too willing to agree and flatter. Before you ask, ask yourself: What evidence would change my mind? Otherwise, you risk using smart tools to dig yourself an even deeper intellectual trench.

• Practice critical, balanced thinking. Check sources, compare perspectives and stay alert to gaslighting, missing context and plausible nonsense dressed up as authority.

These are the cognitive habits that turn AI from a hallucination machine into a thinking aid. They are also habits that children can learn — but only if adults model them.

3. How are we using AI now?

Millions of people now use AI every day. Understanding people’s interactions with AI is one of the great sociological questions of our time. Anthropic, creator of Claude.ai, recently designed a privacy-preserved tool, Anthropic Interviewer, to asks people directly (detailed interviews at unprecedented scale) to get a comprehensive picture of AI’s changing role in people’s lives, including how people are actually using Claude’s output and how do they feel about it. This is a new step in understanding the wants and needs of our users, as well as gathering data for the analysis of AI’s societal and economic impacts5 [5].

Key Usage Trends from Anthropic Interview research results, the Anthropic Economic Index and the AI Fluency Index (late 2025 / early 2026):

• Dominant Uses: Usage is concentrated, with over one-third (36%) of Claude.ai conversations focusing on software development and coding, although educational and scientific tasks are rising.

• Automation vs. Augmentation: While AI agents have spurred an increase in automation (direct task delegation), a significant portion of users still prefer “augmentation”—using AI as a collaborative, interactive, and iterative thought partner.

• Agentic Feature Shift Over Time: By November 2025, 52% of interactions were classified as augmented, while 45% were automated, showing a shift back toward collaboration as more “agentic” (proactive) features were introduced.

• “Artifacts” Impact: When using the “Artifacts” feature (for creating documents, code, or apps), users tend to be less critical, questioning the AI’s reasoning 3.1 percentage points less often than in standard chat, suggesting higher trust in polished-looking outputs.

The trend towards agentic use cases is accelerating. It would be important to take a pause to understand what AI is doing to our brains and what skills would be required to properly leverage AI in amplifying human.

4. What AI is doing to our brains: cognitive offloading and deskilling

There is also a quieter, neurological risk: what happens when we lean on AI too much. And even experts are not immune. But let’s examine the baseline first, as illustrated in a recent MIT Media Lab study led by Nataliya Kosmyna, volunteers wore EEG‑like headsets while writing short essays and taking math tests under three conditions: using only their own brains, using a search engine and using ChatGPT as a co‑pilot6 [6]. The results were telling:

• In the brain‑only condition, participants showed the richest, most distributed brain connectivity, especially in regions linked to attention, planning and memory.

• With search engine assistance, connectivity dropped.

• With ChatGPT co-pilot, connectivity dropped roughly halved on some measures compared with the brain‑only baseline. Participants in the heavy‑AI condition were also unable to remember clearly what they had written later.

A single practice of using AI does not necessarily shrink our brains. But over time, if we rely on AI tools every day for years, we repeatedly offload effortful thinking to AI. In other words, we are not just using a tool, we are training ourselves not to think. That process of cognitive offloading is artificial ignorance in its purest form: high apparent output, low genuine engagement.

The more we lean on AI, the easier it becomes to let judgment idle, even in domains where we are supposed to be the experts. In the Harvard–Boston Consulting Group “jagged technological frontier” experiment, hundreds of BCG consultants were assigned to solve realistic business problems with and without GPT‑4. When they used AI on tasks inside their domain — say, telecom specialists on telecom cases — they completed 12.2% more tasks, 25.1% faster, and with 40% higher quality compared to those not using AI. But when put the same specialists on tasks outside that frontier, performance fell. Error rates rose by 19%, and consultants began relaying AI’s confident but wrong recommendations instead of interrogating them. Over‑reliance turned experts into novices, like a driver falling asleep at the wheel with cruise control on: safe on straight highways, dangerous on sharp bends7 [7].

Deskilling also appears in the very domains where experts are strongest. A study in The Lancet Gastroenterology & Hepatology followed endoscopists after they introduced AI systems to assist with polyp detection during colonoscopy. AI‑assisted procedures improved detection in the moment, but several months later, in unassisted procedures, the doctors’ own adenoma detection rate appeared to fall by about 20%, suggesting a deskilling effect. AI sharpened the tool but dulled the surgeon8 [8].

The remedy is not to abandon AI, but to build in “AI holidays” — regular AI‑free practice that keeps human skills alive even as machines assist [8]. If this is what over‑reliance can do to expert cognition and performance, it is not hard to imagine what happens when still‑forming minds of the children lean on AI for more and more of their thinking. The deepest risk is that children never fully develop the habits of attention and effort that deep thinking requires.

5. AI-amplified attention casino, loneliness, anxiety and mental health exacerbation

Now move from cognition to attention. When AI is implemented in social media and smartphones, it further fragments our focus by supercharging personalised feeds and content generation. Welcome to the attention casino.

Psychologist Professor Angela Duckworth notes a worrying pattern: where students once stayed with a task for around three minutes before switching, the rise of short‑form, highly curated feeds seems to have cut this to well under a minute. The exact “45 seconds” figure is not a law of nature, but the direction is clear: attention slices are getting thinner9 [9].

You cannot build deep expertise — or deep relationships — 45 seconds at a time.

This is not simply about willpower. Research from the University of Portsmouth and the University of Surrey finds that young adults with higher loneliness and anxiety are more prone to problematic smartphone and social‑media use. They often turn to their phones to cope, only to find that compulsive checking and late‑night scrolling make their anxiety worse. AI‑driven recommendation engines sit on top of that vulnerability, optimising for engagement, not well‑being.

Jonathan Haidt, in The Anxious Generation, offers three practices that, uncomfortably, describe many parents’ failures in fighting the attention crisis amplified by AI10 [10]:

• Treat the phone as an experience blocker, not just a distraction. It does not only steal minutes; it can steal entire childhood “sensitive periods” for learning social skills and independence.

• Scaffold real‑world risk. Children do not just need protection; they need difficult projects, physical challenges and unfamiliar groups that build anti‑fragility.

• Fight the algorithm, not the kid. Our children are not weak. They are up against billion‑dollar AI systems tuned to keep them glued to a screen. They do not need more shame; they need allies who understand the game.

The same logic extends into mental health.

On paper, Gen Z is the most connected cohort in history. Yet surveys across countries show rising loneliness and anxiety among teens and young adults. Digital habits are not the only cause, but they have become a powerful amplifier.

AI‑powered companions and “therapist” chatbots plug straight into that vulnerability. Xingye, an AI companion mobile app developed by AI powerhouse MiniMax, has around half a million daily users in China, many of them teen girls and young women11 [11]. Journalist Poppy Koronka reports that children using chatbots from Meta as therapists may see their mental health worsen. US regulators have opened investigations into AI therapy bots over misleading claims and data practices12 [12]. One clinical worry is structural: human therapy is bounded in time and space; sessions end. AI does not have office hours. A child lying awake at 2 a.m. can spend hours ruminating with an endlessly responsive bot with always-on relief, reinforcing anxious loops instead of disrupting them.

It is worth noting that AI can also support healthier habits — for example, by guiding exposure therapy, structuring journalling or offering language practice — when embedded in thoughtful products and bounded routines. But those designs remain the exception.

6. Skills for a human–AI symbiotic balance

Pull these threads together — synthetic content, artificial ignorance, attention slicing, AI‑mediated coping — and one conclusion emerges: skills for a human–AI symbiotic balance sit at the centre of a new parenting playbook. We are not just managing devices; we are shaping the relationship between our children’s minds and an always‑on layer of machine intelligence.

Human–AI co‑intelligence is less a tug‑of‑war and more a sideways infinity loop: one side human, one side machine. At different ages and in different tasks, one loop should swell while the other shrinks — sometimes the child leads and the AI merely suggests; other times the AI drafts and the human edits. The balance is not automatic; it needs deliberate, ongoing calibration.

Three skills matter most.

1. Asking good questions in the right context.
This is the antidote to both hallucination and shallow thinking. It forces us to slow down, frame problems clearly and engage our own cognition before outsourcing the rest. With teens, that might mean insisting they write their own first paragraph before asking an AI to help; with adults, it might mean defining success criteria before letting an AI agent act.

2. Judgment and discernment.
This is the daily practice of verifying claims, cross‑checking sources, resisting easy answers and being willing to update beliefs in light of evidence. AI will keep getting faster and smarter; the question is whether we, as families and communities, can get wiser at least as quickly — or whether we drift down the comforting glide path into artificial ignorance.

For adults and professionals, these two skills translate into clear guardrails. Humans stay in the loop (AI suggestions remain drafts until a responsible person signs off), AI assists but does not replace (co‑pilot, not pilot), and we schedule regular “AI‑off” sessions (or AI holidays) so people practice key skills without autopilot. In high‑stakes domains, that can mean dual‑pass reading, credentialed access to powerful tools and audits of when humans override or rubber‑stamp AI decisions.

3. Human–AI balance as a parenting habit.
For parents, the balance starts with a different set of questions. With teens, it means deciding together where AI should help and where it should stay out: which homework tasks are AI‑assisted versus AI‑free, which creative projects can use AI as a sparring partner versus a ghostwriter, and how much screen time goes to auto‑playing feeds versus deliberate research. You are not banning tools; you are co‑designing the loop.

I find it useful to picture the human–AI symbiotic partnership as an infinity symbol: one loop for the human, one for the AI. For any given task, age or situation, the loops should be different sizes — sometimes the human side dominates and AI only nudges; other times AI handles more routine work while the human decides what matters. But the human loop never disappears; keeping the human in the loop (HITL) is critical. The exact calibration of human and AI roles depends on two skills: asking high‑quality questions with enough context, and exercising judgment and discernment about when to trust, challenge or ignore what the machine suggests.

For younger children, parents can borrow Clayton Christensen’s “Jobs to Be Done” (JTBD) lens13 [13]. Stop asking “Why is my kid using this?” and start asking “What job are they hiring this for?” If a child is using AI “for homework”, is the real job avoiding boredom, chasing quick praise or actually learning the material? Do not fight the tool in the abstract. Ask what job your child is hiring it to do — and whether AI is truly doing that job well for their long‑term growth, or quietly doing the opposite.

Three concrete experiments can make this real in a single month:

• Choose one family activity — a project, trip or meal — that is planned and executed with no AI at all, simply to feel what attention without autopilot is like.

• Have one explicit JTBD conversation with your child about an app or AI tool they love: what job it is doing for them, and whether it is doing that job well.

• Set one clear boundary on AI use for schoolwork (for example, “AI may critique your draft but not write it”) and stick to it.

In the end, ambient intelligence will seep into every corner of our children’s lives. The open question is not whether they will grow up with powerful AI, but whether they will grow up with the inner skills to decide, moment by moment, when to lean on the machine — and when to leave their own minds fully in charge.

Fellow parents, let’s help one another and our children step confidently into the age of artificial intelligence, without sleepwalking into artificial ignorance.

References:

TL;DR:

• Nobel laureates systematically protect “wasted” work—and it becomes their moat

• Corporate biology requires slack as safety: portfolio of slack habits, misfit talent

• Startup biology can leverage minimum viable slack (MVS) to avoid locking in a hill: 5% time, piggybacked bets, cheap options, architecture for optionality

• Founders can design slack systems and deliberately cultivate misfit talent.

  1. When “moving the needle” is not enough

By startup standards, I should have rejected what I heard at the Nobel Heroes Forum. Much of what the laureates shared sounded counter‑intuitive, and even dangerous, for entrepreneurs.

I’ve spent decades as a founder and angel investor learning to run one operating system: kill what doesn’t move the needle, fail fast, and give every day an OKR (Objectives and Key Results)[1]1. Elon Musk is the caricature of this mindset of high cadence, public deadlines, visible intolerance for drift. And founders like Jeff Bezos and Jack Ma have built empires on ruthless execution toward a clear vision. That approach works until it quietly blinds you to a different kind of power.

On January 12, 2026, I sat in the Grand Hall of the University of Hong Kong as six Nobel laureates walked on stage with University of Hong Kong (HKU) President Xiang Zhang and Provost Richard Wong for The Nobel Heroes Forum[2]2. They weren’t there to raise a round or launch a product. They came to talk about decades of work that, for long stretches, looked like a waste of time—wrong paths, “useless” experiments, and methods nobody believed in.

One physicist, Konstantin Novoselov, told a packed room of over 600 students, academics, and founders that if your daily experience of science feels like nothing but struggle and failure, you are probably in the wrong profession; you must enjoy the process, not just the result. Another, Ferenc Krausz, described spending ten years chasing a seemingly crazy idea—watching electrons on attosecond timescales—with no inkling that the same tools would later open a path to ultra‑early cancer detection in human blood. A third, Kurt Wüthrich, calmly recounted how he realized in mid‑career that his field had spent years “completely on the wrong way” and decided to rebuild the entire programme from scratch.

Listening to them, I realized their lives were a controlled experiment in a claim most founders pay lip service to but rarely design around: the biggest long‑term returns often come from work that looks wasted on short time horizons—failed experiments, “useless” projects, non‑obvious applications, and arguably misfit people. Science has institutionalized this truth. Most startups haven’t.

I used to think this kind of “controlled experiment” with unknown outcomes was a luxury that neither startups nor even large corporates could afford. Now I have to admit it demands a complete rethink.​ As deep tech in AI, robotics, quantum computing, biotech, and space pulls startups closer to the scientific world, it feels like the right moment to examine the common ground between decacorn founders and Nobel‑level scientists. Their journeys rhyme more than we admit.

This essay is my attempt, as a founder and angel investor, to reverse‑engineer how stories from six Nobel laureates from the Nobel Heroes Forum and Donna Strickland operationalize “wasted” work, and how you might adapt their playbook meaningfully. These seven Nobel laureates whose careers, panel sharing and conversations with me shaped the ideas in this article. Think of it as a field guide to building failure and slack into your company so that what looks useless today becomes your moat tomorrow.

By a non‑obvious moat, I mean a competitive advantage built from work that looked “useless” or non‑core at the time (failed experiments, side projects, odd hires) that later becomes the hardest thing for competitors to copy.

2. Failure as route‑finding, not verdict

2.1 Ferenc Krausz – the physics of pivoting

When Ferenc Krausz[3]3 took the mic, he didn’t start with lasers. He started with questions.

Before you move any apparatus, he said, you have to decide what the right big question is. That question becomes your North Star for a very long journey. I was struck by how close this is to founding a startup: the best founders begin with a hard question or painpoint about the world. What problem are we really trying to solve here? And let that question, not the current tech stack, set the direction. It’s about falling in love with the question and the quest, not with any particular solution.

Along the journey, you’ll bump into all kinds of detours and side quests of anomalies, side effects, unexplained glitches. But every time you hit a fork you ask one thing: does solving this get us closer to the question? If not, walk away, however tempting it looks.

Krausz treats each failed attempt as a data point in a route‑finding algorithm. It tells you something about the landscape. As he put it, “A failure should never be understood as a failure in the conventional sense; it always offers a chance for rethinking and maybe taking a route which eventually will get us to our goal.”

In a Krausz‑style lab, the real failure is not a negative result; it’s failing to update the route. The sin isn’t that an experiment flops, it’s staying on a path that’s no longer promising just because it still produces papers, metrics, or comfort. In startup language: the true failure is the product‑market “almost fit” that keeps you stuck for years, fueled by founders’ confirmation bias and delusional thinking.

If you’ve ever kept a product alive because it was doing “okay” while quietly suffocating your ability to search for something better, you’ve already experienced the dark side of incremental success.

2.2 Kurt Wüthrich – when the whole field is wrong

Kurt Wüthrich[4]4 has the kind of story you usually only hear in retirement speeches.

In 1968, at Bell Labs, he took blood from his own arm to study his hemoglobin using NMR (Nuclear Magnetic Resonance, later known as MRI, Magnetic Resonance Imaging). The early results were exciting; he quickly became “quite famous” in structural biology.

Then, after six promising years, the field stalled. Wüthrich stopped to write a book surveying his own field and he realized something horrifying.

“We were completely on the wrong way.” Not slightly off. Completely wrong.

Instead, Wüthrich and his team went back to first principles. They hired mathematical physicists and rebuilt the entire method stack. By 1984, they had the approach that finally enabled the structures he’d dreamed of in 1968.

The founder parallel is brutal: sometimes the entire category you’re in is wrong. An industry can spend a decade marching confidently in the wrong direction. You could argue we are living through that dynamic now in AI: in the race toward AGI, the ecosystem is committing trillions of dollars and multi‑year roadmaps for ever more talent and compute at scaling LLMs with “more of the same.” But Yann LeCun and others are pushing different frontiers—embodiment, grounding, richer world models. If they are right, then today’s LLM arms race will look, in hindsight, like Wüthrich’s first NMR programme: an impressive but ultimately wrong way that had to be abandoned so the field could reconfigure around a deeper paradigm.

2.3 Konstantin Novoselov – the joy of daily failure

Konstantin Novoselov[5]5, the co‑discoverer of graphene, manages to be both more optimistic and more ruthless about failure than most founders.

He agrees that failure is everywhere in science, but he rejects the myth of heroic suffering. If you feel like you’re only struggling and failing, he told the HKU audience, you might simply be in the wrong profession. You have to enjoy solving the small daily problems (analytical puzzles, soldering, design hacks) as much as you enjoy the final breakthrough. “If you feel that you are struggling then it’s something wrong with what you are doing… you need to enjoy the process as much as the result.”

Contrast it with Jensen Huang’s “I wish upon you ample pain and suffering.[6]6” Huang is talking about character building and resilience in his Stanford’s speech: you cannot build a company like NVIDIA without long stretches of pain to develop real grit. Novoselov is talking about fit: if your work is only pain, with no genuine curiosity or play, you are probably on the wrong problem or in the wrong craft.

For deep‑tech founders, the synthesis is subtle but powerful. Expect hardship and treat it as training, not injustice—but insist that you still like the game you’re playing. The moment the work becomes pure dread rather than demanding fun, you’re no longer “building resilience”; you’re just burning your optionality and going blind to better routes.

2.4 How to fail well: productive pain, not burnout

In other words: adopt Huang’s tolerance for pain, Novoselov’s insistence on joy, and Krausz’s/Wüthrich’s habit of constantly asking whether you’re still on the right route. The mindset for deep‑tech founders and investors looks like this:

• Expect long, painful stretches and treat them as training, not injustice.

• Insist on intrinsic enjoyment of the craft as no amount of “grit” rhetoric will save you from daily dread.

• Distinguish good suffering from bad suffering. Good suffering is effort, risk, and uncertainty in service of a question you care about (e.g. Krausz chasing attoseconds). Bad suffering is grinding on the wrong problem, with the wrong people, for the wrong reasons, because you’re scared to pivot.

• Design organizations that can absorb pain without burning people out, and build a culture that treats setbacks as route‑finding—not personal failure—while leaving room for Novoselov‑style playfulness in the lab.

2.5 How work in the darkest hour became your greatest moat

Perhaps the clearest large‑scale example of “failure as route‑finding” is AI itself.

In the 1960s, Marvin Minsky and others predicted rapid progress toward human‑level AI[7]7. That optimism crashed into reality. By the mid 1980s, after overpromises and under‑delivery, funding dried up. MIT Technology Review and others ran covers effectively declaring that AI had failed and that there was no evidence it would ever fulfil its grand promises[8]8.

This was the AI winter. Symbolic AI and expert systems had hit a wall. I still remember my University of Waterloo days (90s) in electrical and computer engineering, trying to do a final‑year project on voice enhancement with early neural nets and expert systems. Backpropagation had just arrived and chewed through enormous compute just to train a few neural network layers; with the tools we had, real breakthroughs felt almost impossible.

Yet, a small group—Hinton, Bengio, LeCun and others—kept going anyway. Their “useless” work on backpropagation, deep learning architectures, and representations became the foundation of everything we call modern AI: transformers, generative models, foundation models, and the current wave of embodied systems and rich world models that learn by acting in 3D environments [9]9. The stone the builders rejected really did become the cornerstone.

From a startup perspective, AI winter is what happens when institutions mis‑price optionality. The field nearly died just as its critical building blocks were being formed. Imagine an investor board forcing Hinton to shut down “because it doesn’t move the needle this year.”

You see the same pattern in startups. During SARS in 2003 and aftermath of dotcom bubble burst, Jack Ma used a near‑death moment to build and launch Taobao, which later beat eBay China and became Alibaba’s crown jewel. Around the same time, Amazon looked finished after a 90% stock drawdown, but Bezos chose to “run scared” while doubling down on infrastructure and customer value; the internal tooling platform that kept Amazon alive under pressure became the seed of AWS, the largest cloud service provider in the world.

Jack Ma’s philosophy encapsulated the ethos of building your best product in the darkest days: “Today is hard, tomorrow will be worse, but the day after tomorrow will be sunshine. However, most people give up or die tomorrow evening, and could not see the sunshine. 今天很残酷，明天更残酷，但后天会很美好。然而，大多数人会在明天晚上放弃，因此无法看到后天的阳光 。[10]10”

2.6 Decision-making in bets like poker players: decision quality, not outcome quality

When you zoom out, Krausz, Wüthrich, Novoselov, and Hinton were all making decisions under radical uncertainty. They never had all the facts; they were, in Annie Duke’s language, thinking in bets rather than playing a solved game like chess[11]11.

In poker, you can make a brilliant decision and still lose because the river card goes the wrong way, or make a terrible decision and get lucky. Duke calls our habit of judging decisions purely by outcomes “resulting,” and it is deadly for non‑obvious moats.

If you treat every failed experiment as proof it was a bad idea, you will systemically under‑invest in exactly the volatile, long‑horizon bets that gave us deep learning, attosecond physics, or graphene. A healthier stance is to separate decision quality from outcome quality: given what we knew, was this a good bet—small downside, meaningful information gain, large possible upside—even if it failed this round?

Krausz and Wüthrich run their labs this way. Hinton spent the AI winter doing this inside departments that had already “resulted” on neural nets. If you want non‑obvious moats in a startup, you need the same discipline: treat each project as a probabilistic bet in an uncertain landscape, not as a guaranteed move on a solved chessboard.

3. Complex systems need slack - the case for “useless” work

If failure is navigation, “useless” work is fuel.

Founders are taught to avoid “waste” like poison. The investor deck demands a clear business case for every initiative. MBA logic says every project must hurdle an ROI threshold. Yet the most important work in science—and, frankly, in technology—almost never passes that test at the moment it is done.

3.1 Where scientists and deep‑tech founders land

Donald Stokes captured this tension with a simple 2×2 matrix[12]12. One axis is “quest for understanding,” the other is “considerations of use.” Bohr sits in pure curiosity, Edison in pure application, Pasteur in use‑inspired basic research, and the lower‑left is pure tinkering. Many Nobel‑level scientists care deeply about fundamental questions and often start in Bohr’s quadrant, where there is little or no thought of application. Deep‑tech founders and CTOs, by contrast, obsess with application and customers start in Edison’s quadrant.

The most powerful work happens where these instincts intersect—Pasteur’s quadrant, where deep understanding unlocks important applications.

​The founder move is to build a bridge between those quadrants: invite Bohr‑style questions into an Edison‑driven organization, and give your most scientific people enough slack to wander “upwards” while still anchored to real problems.

3.2 Attoseconds to blood diagnostics – connecting the dots that you can’t justify

In the early 1990s, Ferenc Krausz posed a question to himself and his group in Vienna: can we experimentally see the motion of electrons predicted by Schrödinger’s equation? Theory said electrons jiggle and oscillate on the attosecond scale (10 to the power -18 seconds). No instrument could see that. There was no product roadmap. No white paper promising a market. The question sat squarely in what Donald Stokes later called Bohr’s quadrant: pure curiosity‑driven science with no known application.

Krausz spent about a decade building the tools to answer that question. That work ultimately defined attosecond metrology. Then something unexpected happened.

Once they had the ability to capture attosecond oscillations, they realized the same technique could resolve the oscillations of visible and infrared light. Another decade of work later, they discovered that measuring those oscillations, when infrared light passes through a drop of human blood, reveals the blood’s molecular composition.

Today, Krausz’s teams in Munich, Budapest, and Hong Kong are using this to develop what might become the world’s first ultra‑early screening test for cancer, cardiovascular disease, COPD, and diabetes. The same “useless” question about electrons now sits at the core of a potential revolution in preventive medicine.

As Steve Jobs told Stanford graduates, “You can’t connect the dots looking forward; you can only connect them looking backwards. So you have to trust that the dots will somehow connect in your future.[13]13”

Krausz’s career is a living example. The quality of the dot—asking a deep, fertile question about the nature of electrons—mattered far more than any foresight about medical applications. He chose the right neighborhood to draw in, and the map revealed itself later.

Founders obsess over dot‑count (sprints, experiments, features). Nobelists obsess over dot‑quality (questions). The uncomfortable lesson: you can’t justify every dot upfront.

3.3 Scotch tape, “toys,” and solutions looking for problems

Novoselov’s most famous dot began as a toy.

He and Andre Geim started playing with Scotch tape and pencil graphite. Peel off a layer, stick it down, peel again. At first, it was simple curiosity: what happens if you keep going? Eventually they isolated a single atomic layer—graphene. Every one of those single layers had unique electronic and mechanical properties. That toy experiment opened an entire field of 2D materials and, eventually, a multi‑billion‑dollar industry.

At the forum, Novoselov told the students:

“I would definitely encourage people to do ‘useless’ research. There is enough people who would turn it useful.[14]14”

The basic research behind lasers was famously called “a solution looking for a problem” before they became ubiquitous in telecom, medicine, manufacturing, and consumer devices[15]15. In tech industry and startups, we’ve built a religion around “customer obsession” and “jobs to be done.” I deeply believe in Clayton Christensen’s JTBD lens[16]16—but I also know, from both science and deep tech, that there are categories of work whose only honest job at the time is “expand the frontier of what’s possible.”

3.4 Biological and financial systems – “slack” is safety

In the second panel, Louis Ignarro explained why a simple gas molecule, nitric oxide (NO), won a Nobel Prize[17]17.

NO is produced by our bodies to dilate blood vessels, reduce inflammation, and protect us from coronary artery disease. It also regulates sexual function in both men and women. It’s one molecule doing many jobs—but only in the right place at the right time. You don’t want full‑body vasodilation all day, nor continuous arousal while you’re eating lunch.

Biology solves this by layering tight local control on top of systemic capacity. The machinery to produce NO is always there, but it’s gated by context: cell type, signals, local feedback. The global system keeps slack—the ability to respond quickly—without wasting resources on constant maximum output.

Tim Hunt[18]18 added another lens: your nose.

Every seven years or so, almost all the cells in your nose have been replaced. Cell division and apoptosis are constantly renewing the tissue. Yet, from the outside, your nose looks the same size. There’s no explosion of mass, no gradual shrinkage. Underneath visible stability is a storm of “wasteful” activity: cells that divide then die, signals that over‑correct, micro‑errors that get edited away.

If nature ran purely on Lean Startup doctrine—no redundancy, no excess capacity, no non‑shipping work—you would not survive adolescence.

Robert C. Merton gave the financial version[19]19.

Markets work because they aggregate information no single actor has. Prices are not just numbers; they are weighted summaries of many people’s beliefs, resources, and confidence. Those with better information and higher conviction put more money behind their views, so the price reflects a biased but often more accurate collective intelligence.

But sometimes a crisis hits. Merton defines a true crisis as the moment when something terrible happens and nobody knows why. Suddenly the rules of the game seem changed. In those periods, rational actors withdraw: you don’t trade in a game whose rules you no longer understand. Liquidity evaporates.

That’s where government comes in—not because it is smarter, but because it has power. It can take positions when nobody else can, provide guarantees, and create temporary rules to stabilize things.

Underneath this is a simple principle: any system that pretends it can model the future perfectly will be brittle. Merton insists that real systems need reserves—capital buffers, slack, diversification—to remain robust under unknown unknowns.

As founders, we often run our companies closer to the “perfect model” fantasy than to biology or finance. We feel guilty about cash sitting idle, engineers not at 100% utilization, roadmap capacity not fully allocated. We optimize away slack to hit efficiency metrics. Then we’re shocked when a crisis hits (a pandemic or a regulatory shift) and we have no reserves.

The shared lesson from biology and finance is blunt: any system that pretends it can model the future perfectly will be brittle. Real systems keep reserves—redundant capacity, buffers, slack—so they can survive unknown unknowns. Complex systems don’t see slack as waste; they see it as safety.

3.5 Corporate biology: how companies store slack as potential energy

Big companies that stay innovative for decades have, consciously or not, copied biology and financial markets. Since the 1950s, 3M has had a 15% rule: employees are encouraged to spend up to 15% of their time on projects of their own choosing, outside formal assignments[20]20. Management doesn’t just tolerate this; they celebrate it. That “wasted” time produced, among other things, the Post‑it Note: a side experiment using a failed low‑tack adhesive[21]21.

Google popularized 20% time (“Innovation Time Off”), where engineers could work one day a week on personal projects. Insiders credit this with spawning Gmail, AdSense, and Google Earth; one VP estimated that over half of new Google products originated as 20% projects[22]22.

Over time, Google shifted from fully diffuse slack to more structured programmes like Area 120, an internal incubator giving teams concentrated blocks of time off normal duties to build new products[23]23. Research on slack time shows this broader trend: from scattered slack to focused, incubator‑style slack with clearer selection, to avoid a flood of marginal projects.

Then there’s X, Alphabet’s “moonshot factory,” which masters the art of encouraging slacks and filtering them. CEO Astro Teller has said bluntly that only about 2% of projects “graduate” after five or six years, yet those 2% consume nearly 44% of X’s total budget. X launches 100+ experiments yearly and tries to kill weak ideas as fast as possible. They publicly celebrate shutting down projects early and added a Foundry stage devoted to turning promising science experiments into businesses[24]24.

This is portfolio‑level slack and its slack filtering process is designed so that killing 98% of work is success, not failure.

As a founder, you might think, “We’re not Google; we can’t afford that.” You’re right: you can’t copy their percentages. You need a visible habit and a culture that celebrates good experiments, processes that make prototypes easy to start, and a storytelling loop that revisits how “useless” work became breakthroughs.

3.6 Startup biology: minimum viable slack (MVS)

At this point, a reasonable founder might object: All of this is fine for Nobel labs and trillion‑dollar companies, but I am running on 12 - 18 months of runway. I can’t afford slack.

The constraint is real. Early‑stage startups are resource‑starved. Copying Google or 3M would be suicide. The trap is assuming that the only alternatives are “full moonshot factory” or “zero slack.” Lean done badly optimizes for local certainty: every hour must tie to this quarter’s KPIs, or it is “waste.” But in deep tech, your biggest risks are model error (wrong question, wrong architecture, wrong market) and option loss (shutting down lines of inquiry that would have become your moat).

If you remove all slack, you may extend runway in the short term while silently increasing the probability that the company dies for structural reasons—wrong problem, wrong design space, wrong platform. So the question isn’t “Can we afford slack?” but “What is the minimum viable slack we need to avoid building the wrong thing really efficiently?”

Minimum Viable Slack (MVS) is the smallest deliberate buffer of time, talent, and architecture that keeps you from locking into the wrong hill while still feeding the one you’re climbing (building the wrong thing efficiently). In practice, that looks like:

• Timeboxing and scope discipline. Reserve a tiny, explicit quota, say 5% of engineering time for exploration. Run 1 - 2 day spikes with a crisp question (“Can we get X latency with Y method?”), then decide: kill, park, or promote

• Piggybacked exploration. Attach non‑obvious bets to work you’re already doing. While shipping a feature, prototype a strange data representation that might become differentiating.

• “Cheap options, expensive convictions.” Make it cheap to try many small, reversible experiments—and very expensive to commit headcount or multi‑quarter budget. This is Annie Duke’s thinking‑in‑bets applied to your roadmap with numerous, tiny bets.

• Optionality through architecture. Even if you can’t fund parallel moonshots, you can build modular interfaces, clean data boundaries, and richer instrumentation so you don’t lock yourself out of future directions.

• Externalizing some of the “lab.” Use open source, academia, and your ecosystem as part of your slack. Partner with a lab on one narrow, high‑leverage question; expose APIs so external developers can explore weird use cases beyond your bandwidth.

• Cultural moves that cost almost nothing. A “We Were Wrong” slide in every all‑hands, one‑page field notes after killed initiatives, a standing agenda item for “one non‑obvious bet we’re willing to fund this quarter and why”, and one intentional misfit hire or advisor whose job is to challenge your assumptions.

You’re building a biologically plausible organism: lean enough to move, but with just enough redundancy and slack to adapt, heal, and stumble into non‑obvious moats.

Try this week: Protect one tiny piece of slack.

• Block one afternoon for a 1–2 day spike with a crisp question.

• Attach one weird prototype to a feature you are already shipping.

• Add a “We Were Wrong” slide to your next all‑hands.

4. Useless inputs: humanities, art, and misfit talent

One of the most striking things about the Nobel Heroes Forum was how un‑narrow these scientists were.

Ferenc Krausz talked about creativity in science and art as fundamentally similar; Konstantin Novoselov talked about Chinese calligraphy and painting; Kurt Wüthrich talked about artists painting molecules before computers could draw them, and patients listening to Mozart during MRI scans. They are “useless inputs” that quietly shape how these people see problems.

4.1 Calligraphy, the Medici effect, and impossible planning

Novoselov has spent years practising Chinese calligraphy and ink painting. He described a deep parallel: you cannot plan a masterpiece. You can only set up the conditions, practise the strokes, and be ready. Some days the brush does what your mind sees; most days it doesn’t.

Science operates the same way. You cannot schedule a discovery for next quarter. You can only arrange conditions and increase the probability.

In a similar vein, Jobs later credited a calligraphy class at Reed College with shaping the Macintosh: “I learned about serif and sans‑serif typefaces, about varying the amount of space between different letter combinations, about what makes great typography great… ten years later, when we were designing the first Macintosh computer, it all came back to me…”

One pre-condition that may increase the chance of serendipitous discoveries – the Medici effect, a term coined by Frans Johansson, refers to breakthroughs that happen at intersections of disciplines and cultures[ ]25. Renaissance Florence combined artists, merchants, engineers, and philosophers in dense proximity; today, Novoselov’s institute in Singapore deliberately mixes mathematicians, computer scientists, biologists, chemists, and physicists—and he said he would happily add arts and literature if he could.

From a founder’s hiring dashboard, a calligrapher‑physicist looks inefficient. From an innovation‑portfolio view, he is a walking crossover engine.

At the Forum, HKU Vice‑Chancellor Zhang Xiang closed the loop in his own way, presenting each laureate with a carefully chosen Chinese phrase from classical literature as a personalized gift (see figure). Each phrase appeared to capture how their ‘useless’ work and misfit paths became world‑changing; I’ve unpacked the phrases and their meanings in the footnote references[26]26.

4.2 Wüthrich’s painters and MRI Mozart

Before graphical software existed, Wüthrich’s lab had no way to visualise the molecules they were studying from NMR data. They did the obvious thing: they hired artists. The painters turned streams of numbers into accurate, beautiful images of proteins and nucleic acids.

Later, MRI—partly rooted in NMR—made it into hospitals. It is loud, claustrophobic, and frightening. To make it tolerable, hospitals discovered that giving patients headphones and letting them listen to Mozart or Louis Armstrong dramatically improved the experience. Wüthrich sees this as another art‑science loop: physics enabled imaging; music made it humane[27]27.

In my Li Shu Pui Symposium Lecture “The Doctor Will See You Now… With AI,” I argued that the real bottleneck in AI medicine is not whether a model can read an X‑ray, but whether patients, clinicians, and regulators can trust it enough to let it into the consultation room. I framed the challenge as building Artificial Integrity: turning black‑box AI into glass‑box AI that is explainable and governed with integrity so it amplifies human values instead of merely automating tasks[28]28.

These details matter for founders building products in healthcare, AI, robotics, or fintech. You can build technically brilliant systems that are unusable or untrusted if you treat design, narrative, and emotion as “non‑core.” Humanities and arts are the R&D department for meaning.

4.3 Donna Strickland – the mispriced PhD

Then there is Donna Strickland, the third woman ever to win the Nobel Prize in Physics[29]29. She did the work that won her the prize—chirped pulse amplification (CPA)—as a PhD student at the University of Rochester in the 1980s. That work, her first scientific paper, made it possible to generate ultra‑short, ultra‑intense laser pulses safely and underpins laser eye surgery and precision micromachining.

When I met Donna Strickland at an alumni dinner in Hong Kong in early 2023, she came across as a disarmingly humble intellect, more like a favourite homeroom professor than a Nobel laureate, fully present in conversation and quietly, genuinely engaged.

Benjamin Jones has shown that, as knowledge accumulates, major scientific contributions tend to happen later, not earlier, in researchers’ lives. In his “Age and Great Invention” and “Burden of Knowledge” work, he documents a simple structural shift: it now takes more years of training just to reach the frontier, which pushes breakthrough work into mid‑career, shrinks the window for originality, and forces people into narrower specializations and larger teams[30,31]3031.

Against that backdrop, Strickland’s early, quiet contribution at the PhD stage looks like a statistical outlier, exactly the kind of mispriced talent we systematically overlook. By Jones’ logic, the burden of knowledge should have pushed a contribution of that depth later in life and deeper into team science; instead, it arrived early, quietly, and from someone the system was not primed to pattern‑match as a future Nobelist.

For decades, her contribution sat in the literature, widely used but not widely celebrated; physics went 55 years between second woman winner Maria Goeppert Mayer’s Nobel[32]32 and Strickland’s. She spent most of her career as a professor at Waterloo, teaching and running a lab, with relatively little public fanfare and no Wikipedia page until 2018. In other words: Strickland’s PhD work looked like “just” a student project, and she looked like a “normal” mid‑career academic; only later did the system admit that this “non‑core person” had done work that quietly rewired both basic science and industrial practice.

If you are a founder or investor, ask yourself how many Donna Stricklands walk past your pattern‑matching filters each year. The junior engineer whose side project is quietly solving a deep infrastructure bottleneck. The female founder outside Silicon Valley whose PhD work you don’t fully understand. The misfit candidate whose CV doesn’t match your mental picture of “high‑growth CEO” or “10x engineer.” If you’re honest, the answer is more than zero—and Jones would argue that, as the burden of knowledge rises, it will take more of these mispriced, high‑depth individuals working in teams to move the frontier at all, which means systematically overlooking them is an increasingly expensive mistake.

Try this month: bet on one mispriced person.

• Identify one junior or “non‑obvious” person whose side project hints at deeper capability.

• Give them a small, bounded‑downside project with real autonomy.

5. Designing non‑obvious moats in startups

So what do you do with all this if you run a startup or a tech team?

Nobel lives make for wonderful conference inspiration, but founders live under cashflow, runway, and customer pressure. You can’t replicate these practice with a 12‑person seed‑stage company. You can, however, design a smaller, leaner version of their systems.

5.1 Portfolio design — build small options

At the level of projects, you are essentially running a portfolio of bets under radical uncertainty. You don’t know in advance which lines of work are “Bohr’s quadrant” curiosities and which will turn into Pasteur‑style moats. Your job is to size and structure those bets so that you can afford to be wrong many times and still be very right once.

Flickr followed a similar arc. It emerged from a massively multiplayer online game project, Game Neverending, where photo‑sharing was a side feature. The game went nowhere. The “useless” feature—put in to make the game more fun—became the main product, and eventually a pioneering photo community that set norms for web‑scale image sharing[33]33.

From a startup perspective, these stories show what it looks like to protect a handful of small, bounded‑downside, high‑upside options in your portfolio:

• You carve out a few projects that don’t obviously move this quarter’s KPIs but touch deep questions in your problem space.

• You accept that most will die, but you size them so their death is survivable and their success can change the company.

I’ve lived both sides of this. During the period 2009–2011 at Cherrypicks, we built iButterfly, an AR + location app that let people “catch” digital butterflies in the real world and redeem them for rewards. Smartphones were young, GPS was flaky, the B2B business model was murky. After a few years, despite millions of downloads from Italy and Turkey to Japan and Indonesia, I killed it as a distraction from our “core” work[34]34.

On the spreadsheet, that looked rational. In hindsight, iButterfly was a non‑obvious moat in embryo: a working prototype of the same AR, geolocation, and game‑loop mechanics that later made Pokémon GO a global phenomenon. We would have given iButterfly a second life if we had changed the business model to B2C and engaged with branded characters. I had confused “doesn’t move the needle this year” with “never will,” and treated a live world‑scale experiment as waste instead of a strategic lab.

A few years later, I made the opposite call. In 2013 I carved out a small “SEAL Team Alpha” at Cherrypicks to explore indoor positioning with sound, Bluetooth, and Wi‑Fi fingerprints. Through reorganizations and even an acquisition, their platform rarely contributed more than 5% of revenue, but we kept advancing the core competence. In late 2024, it snapped into place: those indoor‑positioning technologies unlocked a 100× performance gain in AI 3D world‑model (Deep World Model), and ten years of “non‑core” slack became the strongest moat in our AI work[35]35.

iButterfly is the option I killed too early; Deep World Model is the option I protected long enough to become a moat.

5.2 Slack as a habit

Slack is a habit at system level. MVS is what that habit looks like at startup scale: visible but small allocations of time and budget that you protect even when the runway clock is loud. You can also see mature versions of this in corporate biology at 3M, Google, and X, Alphabet’s ‘moonshot factory.’

Design a slack system, don’t bolt on one‑off hacks. Start with:

• Time: a visible 5% exploration quota.

• Budget: a small, protected pool for experiments that don’t need full business case.

• Architecture: modular APIs, clean data boundaries, rich instrumentation.

• Portfolio: 1–3 option bets that are allowed to look “useless” for a while.

5.3 Cultural norms — normalize being on the wrong way

Behind the portfolio and the systems sits culture. Wüthrich only realized his field was “completely on the wrong way” because he stepped back to write a book and was willing to admit it publicly. Krausz infects his lab with an attitude that treats setbacks as chances to rethink the whole route. Novoselov insists that if it all feels like struggle, something is misaligned.

GitHub has long encouraged hack weeks, side projects, and shipping tiny experiments behind feature flags. Many of its core ideas—like social coding patterns around pull requests—emerged more from community practice and internal tinkering than from top‑down product roadmaps.

Bring that into your company with cheap moves:

• A regular “We Were Wrong” slide in every all‑hands, where leaders share one assumption they’ve updated and what changed.

• One‑page field notes for killed initiatives, capturing what you learned and where the next team might pick up the thread.

• A standing roadmap agenda item: “One non‑obvious bet we’re willing to fund this quarter and why.”

The goal is to make “we are on the wrong way” a promoted behavior, not a career‑ending confession.

5.4 Talent and intersection design

The last lever is who you let into the building and what worlds they connect.

Nintendo has spent decades living off “lateral thinking with withered technology” (Gunpei Yokoi’s phrase): combining old, cheap components with surprising design to create things like the Game Boy and Wii[36]36. From a spreadsheet view, hiring toy designers and eccentric engineers to work with yesterday’s chips looked inefficient. In reality, those “misfit” teams built some of the most enduring franchises in entertainment.

Anthropic is a more recent, AI‑native version of misfit talent as moat. It has oriented itself around safety research on frontier models—RLHF, Constitutional AI, red‑teaming, agentic misalignment studies—that many companies would treat as pure cost or compliance overhead[37]37. That “useless” work is, in fact, a stock of methods, evaluations, and habits that becomes a competitive advantage when models get powerful and risky. Safety researchers and contrarians—the people whose job is to say “this may fail in ways you haven’t imagined”—are Anthropic’s version of Wüthrich’s painters and the MRI Mozart headphones.

In your own teams, the simplest move is to consciously hire at least one Medici‑type per core group: someone with a serious side domain (art, philosophy, anthropology, hardware hacking) and explicit license to connect dots across domains. Then, ask a Strickland question in hiring and investment committees: Which candidate here looks least like our pattern but could plausibly be the one doing the work we’ll regret missing in 20 years?

To build a talent and culture edge, do three simple things:

• One misfit hire or advisor whose explicit job is to question assumptions.

• A recurring “We Were Wrong” slide and short debriefs after killed projects.

• A simple rule: promote people who run good experiments, not just those who ship features.

6. Build your own lab for the non‑obvious moat, not local optimum

At the Nobel Heroes Forum, President Xiang Zhang said that Hong Kong wants to be a scientific bridge between China and the world; that science should be a common language across political divides; and that discovery often starts with work that looks useless, only to become essential decades later. He could have been talking about startups.

The point of this essay is not to make you feel guilty about execution. It is to suggest that execution without designed slack is a local optimum. You can build a very efficient machine that perfectly exploits your current opportunity, only to discover that you’ve optimized yourself away from the future.

The Nobel laureates on stage at HKU did not get there by grinding harder within the accepted playbook. They got there by:

• Asking questions with no immediate use.

• Persisting through years of apparent failure.

• Embracing misfit interests and misfit people.

• Operating in systems—labs, meetings, mentoring cultures—that institutionalized “wasted” work as a feature, not a bug.

AWS did not emerge from a perfectly efficient retail IT budget; it came from a willingness to treat infrastructure work as more than cost. SpaceX did not build reusability by trying to avoid failed launches at all costs; it treated those setbacks as tuition. Anthropic is betting that the “useless” work of safety will turn out to be the only thing that matters when the models get weird.

If you’re a founder, you will never have their time horizons or funding security. But you can still decide, today, to run your company a bit more like a lab:

• Protect small pockets of work that don’t yet have a business case—your internal tool that might be a Slack, your “boring” infra that might be an AWS.

• Design slack into your system before a crisis forces it on you—whether that slack looks like a 5% exploration block, a tiny Area‑120‑style incubator, or one person whose job is to be your internal Anthropic.

• Hire and mentor at least one person who looks useless to your current roadmap but deeply interesting to your future, the kind of mispriced Strickland or Nintendo designer you’ll later realize was holding the keys.

In the end, you don’t need a Nobel‑sized lab. You need a founder‑sized version of Minimum Viable Slack: a small, deliberate buffer of time, talent, and architecture that keeps you from building the wrong thing really efficiently.

To wrap up, I have one question and one call for action to leave you with:

• Question: if Hinton had optimized for performance KPIs in 1985[ ]38, it’s hard to imagine we’d have GPT-4 in 2023. What are you killing today that the world will need in five years, 20 years? Whatever your answer is, that would probably be the “wasted work” you most need to protect.

• Call for action: this week, protect one ‘wasted’ hour. Next month, hire one misfit. Next quarter, fund one non-obvious bet. In several years, that might be your moat.

Slack wisely, and may today’s “useless” work become your future non-obvious moats.

Footnote References:

1. The calling of an engineer

One of the most vivid scenes from my undergraduate years (in the 90s) in Electrical and Computer Engineering at the University of Waterloo did not happen in a lab or a lecture hall. It happened in a guarded room, with hooded figures, heavy chains, and a small ring of iron.

The Iron Ring ceremony in the final year of study is one of the most secretive and theatrical experiences in Canadian engineering education. No photographs. No public livestream. No family in the audience. You are ushered into a private ritual, sworn to secrecy, and invited to join a chain of engineers stretching back a century. The legend goes that the first rings were hammered from the collapsed steel of the Quebec Bridge, whose catastrophic failure in 1907 killed many workers and shook public confidence in engineering. Rudyard Kipling, the author of The Jungle Book, was asked to write the ceremonial “Ritual of the Calling of an Engineer,” a text so guarded that it was long kept out of print and only modernized by the Seven Wardens in 2025 [1, 2]12.

In medicine, students put on the white coat and recite the Hippocratic Oath, pledging to do no harm and to put patients first. The Iron Ring ceremony plays a similar role for Canadian engineers: a moment of collective conscience‑forming, where you are reminded that your calculations touch the lives of many. The ring on your smallest finger is deliberately inconvenient. It scrapes on the drafting board, catches on the keyboard, clinks against the instrument panel, as if to whisper: remember your lifelong commitment to ethical conduct as an engineer.

If trust is the primary product of 21st‑century engineering, education and ethics are the two main levers. The science can be dazzling, the talent world‑class, the capital abundant—and yet, if people do not trust the systems we build, they will not use them, or worse, they will use them and feel betrayed. Modern engineering rests on a new trinity: excellence, talent, and trust. The bottleneck is no longer just what we can build, but what society is willing to let into its bloodstream.

2. Tech adoption moves at the speed of trust

“Tech adoption moves at the speed of trust,” said Mary Wells, Dean of Engineering at the University of Waterloo, as she echoed Canada’s federal AI Minister Evan Solomon. The minister’s original line was, “Tech moves at the speed of innovation, adoption moves at the speed of trust.”

As a member of Canada’s federal AI Strategy Task Force, Dean Wells has been helping design how the country will build “safe AI systems and public trust in AI”—from research funding and infrastructure to education and sector‑specific deployment. Technical breakthroughs and regulatory frameworks are necessary but not sufficient. Without trust—grounded in transparency, accountability, and genuine public engagement—AI will stall at the pilot stage or trigger backlash that sets everyone back [4]3.

Waterloo is building trust into its research DNA. The TRuST (Trust in Research Undertaken in Science and Technology) network brings together scholars across engineering, sciences, health, and social sciences to ask a deceptively simple question: when people say they trust or distrust technologies like AI, what are they actually reacting to? Is it the accuracy of models, the integrity of institutions, the history of harms, or the opacity of decision‑making? Their events, such as “Conversations on Artificial Intelligence: Should It Be Trusted?”, are deliberately uncomfortable: pairing AI scientists with ethicists, lawyers, and even NASA and Google representatives to probe where trust is deserved and where it is not [5, 6]45.

Dean Well’s phrase lands with me because I have lived the flip side of it. During COVID period in Hong Kong, I led a company that built the venue‑based contact tracing app LeaveHomeSafe (安心出行). Citizens scanned a QR code at every restaurant, office building, and public facility for a privacy‑preserved record used in contact tracing and co‑location exposure alerts. When the Government made the use of LeaveHomeSafe app mandatory in November 2020, the system undoubtedly helped health authorities cut chains of transmission. For more than 2 years, the city managed to avoid the worst waves that later hammered other regions [7, 8]67.

But inside the team, the engineering satisfaction of seeing the system scale sat alongside a knot in the stomach. Some in the population saw the app not as a public‑health shield but as a surveillance symbol. Even when you know how carefully you have architected privacy safeguards, you cannot demand trust—you can only invite it.

The same observations applied to vaccines. Whether people trusted mRNA, protein‑based, or inactivated vaccines—and whether they trusted brands from “East” or “West”—often mattered more than the underlying science.

The lesson, in hindsight, is simple: no technical or scientific merit can outrun a deficit of trust. In the Age of AI, when systems will see, infer, and influence more of our lives than any contact tracing app ever did, the speed of adoption will track the speed of trust. And the speed of trust depends on more than compliance or efficiency; it depends on whether people believe that the systems—and the institutions behind them—are ultimately on their side.

3. Re‑wiring engineering education for excellence and trust

If trust is the primary product, education is where we build the factory. The Iron Ring ceremony is a powerful moment, but it is still a capstone ritual. The harder question is: what happens in the preceding four years? What curriculum, culture, and institutional choices wire engineers to earn trust?

At the University of Toronto, Dean Christopher Yip is stretching the definition of what an engineering education should contain. Under his leadership, U of T Engineering has launched new certificates in Public Policy and in Justice, Equity, Diversity & Inclusion (JEDI), as well as a tri‑campus minor in Global Leadership. The stated goal is to prepare engineers who can “tackle society’s biggest challenges” by integrating technical excellence with policy literacy and a deep understanding of equity and inclusion [9, 10]89.

At Queen’s University, Dean Kevin Deluzio has been spearheading a transformation of Smith Engineering into “a ground‑breaking new model for engineering education.” Queen’s is restructuring around common first‑year experiences, multidisciplinary projects, and real‑world challenges where social context and stakeholder trust are central. In his public statements, Deluzio emphasizes that engineers must be prepared for “complex, interconnected problems” and that the new model is meant to produce graduates who are “adaptable, collaborative, and responsive to society’s needs[11, 12]1011.”

Back at Waterloo, Dean Mary Wells sits on top of a unique trust engine: a massive co‑op program that pushes engineering students into industry, startups, and research labs multiple times before graduation. Waterloo’s model is not just about employability; it is about closing the loop between lab, market, and public. When students ship real products and services—whether fintech apps, medical devices, or infrastructure planning tools—they experience first‑hand that a brilliant design is the one users trust.

Layered on top of this educational fabric are explicit trust projects. TRuST’s work on AI‑disclosure standards for research articles is quietly rewriting what “honest engineering” looks like in an AI‑mediated world: if you use AI in your analysis or writing, you say so, in a standardized way that peers and the public can audit. That is Iron‑Ring thinking applied to the invisible layer of code and data.

Iron‑Ring values—humility before complexity, accountability for failure, and duty to the public—inspire a new kind of engineering education, one where excellence is measured not only in throughput and accuracy, but in how well we prepare both humans and machines to behave responsibly.

4. Trustworthy by design and trust in practice

How do we design for trust, rather than treat it as an afterthought? One way is to start from the human emotion that underlies it: empathy. That is precisely what De Kai is trying to harness with the Empathic AI Institute, a U.S. non‑profit he founded to “guide AI’s evolution with compassion, awareness, and ethics.” As an AI pioneer whose work on machine translation has underpinned some of the world’s most widely used translation systems, he has seen how tools meant to connect people can deepen divisions when trained on our worst behaviour [13]12.

De Kai’s Raising AI suggests that we think of ourselves as “parents of AI,” responsible not only for what systems do but for what they learn from us. These “artificial children” grow up in the climate of our online and institutional behaviour; “we are the training data.” Every design decision, every API, every product spec becomes part of an invisible curriculum that AIs absorb. If we cut corners on safety, reward engagement over truth, or encode a trinity of bias—technological, institutional, and cultural—into reward functions, we are teaching entire generations of models that this is how the world works. This is trust by design: tending the climate in which AI grows as carefully as we design its architecture [14]13.

Trust in practice is also about the everyday governance of platforms and resources. Consider Wattpad, the storytelling platform that Allen Lau co‑founded and led for over a decade. Wattpad now has tens of millions of users and has turned amateur stories into major film and publishing deals. Lau describes Wattpad as “a community where everyone has a story to tell” and emphasizes that the company’s mission is to “entertain and connect the world through stories.” Underneath the hood, Wattpad is driven by recommendation algorithms, moderation policies, and community norms. When a story goes viral, it is both a creative success and a test of the platform’s integrity [15, 16]1415.

Similar dynamics appear across more traditional domains of engineering where our nine deans operate:

Water and environment – Dean Viviane Yargeau (McGill University): Her work on endocrine‑disrupting chemicals and wastewater treatment is about protecting the trust we place in something as basic and invisible as tap water [17]16.

Biomedical design and self‑assessment – Dean Heather Sheardown (McMaster University): Beyond leading biomedical and chemical engineering, she has co‑authored work on developing students’ self‑assessment skills, training them to evaluate their own performance honestly—a core ingredient of professional integrity [18]17.

Multidisciplinary systems – Dean James Olson (University of British Columbia): Olson champions initiatives that bring engineers together with planners, architects, and policymakers, arguing that challenges like climate change and inequity require technical and social trust to be designed in tandem [19]18.

Human factors and cyber‑human systems – Dean Caroline Cao (University of Ottawa): Her background in surgical navigation, human factors, and Industry 4.0/5.0 pushes design toward resilience and safety in complex socio‑technical environments, from operating rooms to XR‑enabled training [20]19.

Sustainable materials – Dean Phillip Choi (University of Regina): Choi’s decades in polymer science and sustainable materials remind us that trust in supply chains, packaging, and infrastructure starts with molecular‑scale choices [21]20.

Inclusive engineering for communities – Dean Michael Bradley (University of Saskatchewan): Bradley speaks about creating “the engineers the world needs,” with an emphasis on diversity and Indigenous inclusion—another axis of trust, especially in rural and resource‑dependent communities [22]21.

Each of these leaders works on a different layer of the same question: how to ensure that the systems we build—algorithms, platforms, water systems, hospitals, grids—are not just functional but deserving of trust.

5. The next trust statement on Jan 24, 2026

On January 24, 2026, eight Canadian engineering deans (aforementioned in the article) gathered in Hong Kong alongside other Asia‑Pacific leaders for The Canadian Engineering Asia‑Pacific Conference – A Centennial Celebration of Engineering Excellence and Ethics, symbolized by the Iron Ring. It is quite literally, the cast of this entire essay stepping off the page into the same room. The conference culminated in the issuance of The Canadian Engineering Asia‑Pacific Conference Statement on Engineering. Think of it as the next iteration of the trust contract that the Iron Ring ceremony embodies.

The 20th century asked engineers not to let bridges fall; the 21st asks us not to let trust collapse.

If you care about where this goes next—about how we raise both engineers and AIs, about how we protect human dignity while embracing technological possibility—then this is the event for you.

More information on this unprecedented historical assembly of academic leaders in Canadian engineering.

External media coverage for the event below:

The Globe and Mail (Jan 25, 2026): Iron ring centennial ceremony in Hong Kong coincides with thaw in Sino-Canadian relations

South China Morning Post (Jan 24, 2026): Hong Kong ‘pivotal’ to rebuilding Canada-China academic links, experts say

South China Morning Post (Jan 25, 2026 Print): HK is ‘pivotal’ to Canada-China academic links

The Iron Ring is no longer just a ring of steel on a single finger. It is becoming a network of commitments circling the globe.

The Canadian Engineering Asia‑Pacific Conference – A Centennial Celebration of Engineering Excellence and Ethics:

Notes: I have used Perplexity to perform final edits and Gemini 3/Nano Bananas for the first two illustrations.

Footnote references:

I’m a Gen X dad (born in the 70s) with two teenagers, which means I grew up rewinding cassette tapes and my two kids (born in the 00s) grew up rewinding YouTube videos.

In 2025, that gap hit me harder than ever as my son turned 18 at the end of last year. We spent late nights talking about university choices and future plans, and in May I watched him deliver his valedictorian speech at his high school. Observing him and his Gen Z peers (born in 00s and 10s) forced me to look beyond my own household. Their worldview has been shaped by technology disruptions (smartphones, social media, AI), natural disruptions (pandemics, climate change) and socio-economic disruptions (financial crises, political conflict, wars).

Understanding these generational differences is a survival skill for modern parenting. It changes how we communicate, how we interpret our children’s choices, and how we decide when to step in—or step back.

1. Ready to learn vs ready to teach

Like most parents, I started out believing “good parenting” meant opening as many doors as possible and pushing opportunities through them: extra classes, advice, internships, networks. It was all push.

After 19 years of experimentation (with ongoing debugging), I’ve realised how one‑sided that can be. If parenting were software, I am definitely still in early beta phase. Love may be unconditional; learning isn’t. For the first 16 years, I learned that there is no guaranteed download—no understanding—just because you are broadcasting wisdom on a strong signal.

Then, over the last few years, almost magically, something flipped.

In the middle of Grade 10, my son went from cruise control to full throttle. He began to initiate. He started asking questions, exploring options, thinking seriously about his future. He was no longer just being educated; he was educating himself. He was ready to learn.

By Grade 12, he had a better research stack than I did—search engines, AI copilots, and a willingness to DM/PM the right people on social media. My daughter, two years younger, started her own awakening earlier around Grade 9.

In How Will You Measure Your Life?, innovation theorist Clayton Christensen captures this with a simple but profound insight[1]1:

“Children learn when they’re ready to learn, not when we’re ready to teach. If we’re not with them as they encounter real challenges, we’re not there when they’re finally open to learning. Our job is to be present when life opens a learning window.”

Parenting stops being a broadcast model and becomes more like being on call: the job is to be around when life raises a ticket they care enough to open.

Even this is easier said than done, and at the same time you will also find yourself navigating through a minefield of classic traps along the way, most of which you would only notice in hindsight [1]:

· The helicopter trap: pouring in resources but starving them of experiences that build capability.

· The sequencing trap: “I’ll invest time when I’m less busy.” By the time we’re less busy, their values and priorities are already formed.

Timing and presence turn out to be the scarcest resources, not money or advice. They seem to be the core of effective parenting.

2. Prepare your child for the road, not the road for your child

“One generation plants the trees, and another enjoys the shade” or “前人栽树，后人乘凉” is a beautiful Chinese proverb. It sounds comforting. But in 2025, it can also be dangerous.

Gen X did plant trees. We built careers and companies in a relatively stable globalisation era. But we planted those trees for our climate. The climate has changed—economically, technologically, geopolitically.

If our children simply sit in the shade of what we built, they may be shielded from discomfort, but also from the practice they need to face their own storms. Our job is not to pave the road; it’s to prepare the driver. We don’t get to design the world they inherit; we only get to influence the operating system they bring into that world with.

We cannot teach them our solutions because their problems are structurally different. Christensen reminds us again that our job is not to script their paths but to be the safety net when the world teaches them hard lessons [1].

This also explains another familiar saying: “Wealth does not pass three generations / 富不过三代.” It often doesn’t if the second and third generations only inherit shade from the first, without learning how to plant, prune and replant in different soil. The problem is that shade feels comfortable—right up until the weather changes.

“Prepare your child for the road, not the road for your child. If they only inherit our shade, they won’t learn how to plant in a new climate.”

3. Worldview differences between the operating systems of Gen X and Gen Z

Psychologist Jean Twenge offers a simple way to think about the gap between generations. She argues that technology is the primary driver of generational change: older cohorts lived a “fast life” strategy (grow up fast, marry early, work early), while Gen Z runs a “slow life” strategy in a more complex, high‑stakes world. Seen that way, when Gen Z “quiet quits” or treats work–life balance as non‑negotiable, they’re not slacking; it’s a rational response to a harder game [2]2. Caitlin Fisher calls this “gaslighting”: tell a generation “work hard and anything is possible,” then blame them as entitled when the system breaks under them [3]3.

Generational researchers like Bobby Duffy adds another layer: our “formative years” (roughly age 0–15) are when the outside world quietly writes itself into our mental code. What happens in the society in those years—booms or busts, stability or crisis—shapes the operating system we carry into adulthood [4]4.

For parents, this means our kids’ reactions to work, risk and technology are often about the world they grew up in, not just their individual character. To see how this plays out in my son and me, it helps to look at our formative years and at two speeches we gave in the summer of 2025. I was addressing over 600 graduands at the 214th congregation of the Faculty of Medicine of the University of Hong Kong (available here). My son was delivering his valedictorian address to his cohort Class of 2025 at his high school [5]5 [6]6.

My formative years (70s and 80s): improvable systems

I grew up in analog Hong Kong. Systems did not feel perfect, but they felt improvable. ICAC [7]7 cleaned up corruption. Public housing expanded. Education opened up. Factory kids became office workers. Deng Xiaoping’s “Open Door” economic reforms and globalisation lifted growth. Hong Kong transformed into an international financial centre, and the Sino–British Joint Declaration set the stage for the handover. The message to my generation was simple: work hard, keep your nose clean, and things will generally improve. Systems were not perfect, but they were improvable.

My son’s formative years (10s and 20s): perpetual crises

My son arrived just as the global financial crisis was triggered by the collapse of Lehman Brothers, hardly the soundtrack you would choose for a memorable childhood. It was followed by equally turbulent noise: climate anxiety, deglobalisation, the Umbrella Movement, the 2019 social unrest, COVID, smartphone and AI disruptions. Don’t get me wrong, there were positive events in that period. But by and large, he came of age watching Millennials (generation before him) play by the rules—study hard, get the degree, grind in the job market—and still lose on things like jobs, housing and debt.

Four specific contrasts between the two speeches stand out. It was like watching two different versions of the same operating system.

What is growth to you?

Father: I talked about Felix—my inner alter ego, the voice that challenges and pushes me. I spoke about “manifestly important and nearly impossible” problems and failing forward. My underlying message: the answers are inside you if you’re prepared to wrestle with them.

Son: He described high school as training an AI bot. He compared growth to machine learning: feed in data, make mistakes, adjust the model.

The metaphors are telling. For me, growth is reflecting with an inner self. For my kids, growth is fine‑tuning a model with the outside world.

Definition of resilience

Father: Resilience means endurance and pivot. “Fail fast, fail big, fail forward.” It is about surviving large losses and rebuilding—classic antifragility.

Son: Resilience means adaptation. “Less about powering through and more about learning how to adapt.” It is about learning to bend with disruption (COVID, AI) instead of snapping.

View of AI

Father: AI is a tool. I challenged graduates to use AI for the obvious tasks and reserve their humanity for what AI cannot do: compassion, ethics, complex judgement.

Son: AI is a companion and a model. He joked about ChatGPT as his “new buddy” and used AI itself as the metaphor for human learning. For him, AI is not just a tool in the environment; it is part of the environment. He is near-AI-native.

Mindset

Father: The answers are inside you, if you have the grit to dig them out.

Son: The answers emerge from interaction—between self, peers, systems and algorithms.

These contrasts are not just differences in speechwriting; they are signs of a deeper rewiring in how each generation understands risk, time and growth. If technology could shift my son’s mental “operating system” this far from mine, it raises an obvious question: what happens to the children who grow up not just with phones in their pockets, but with AI woven into every corner of their childhood? What does the next great rewiring look like?

4. The next great rewiring, this time by AI & Robotics – Hello Generation AIR and Q

Social psychologist Jonathan Haidt calls 2010–2015 the “Great Rewiring” of childhood: the shift from play‑based to phone‑based childhood. Unstructured outdoor play was replaced by structured activities and always‑on screens, contributing to anxiety, sleep loss and social fragility in Gen Z [8]8.

But next great rewiring, even more powerful than phones and social media, is already underway: AI and robotics. Generation Alpha (born roughly from 2013 to the mid-2020s), the children coming after my kids, are already feeling its effects, but the cohort after them, Generation Beta (born from the mid-2020s onwards) will be thoroughly shaped by AI and robotics. If we were to choose a more directly descriptive name for Generation Beta, it could be called Generation AIR (Gen AIR): AI & Robotics natives, for roughly those born from 2020 to 2045. After Gen AIR lies Generation Q (Gen Q), who may grow up with quantum computing and bio‑digital hybrids as normal. At that point, our nostalgic arguments about radio vs television will sound very quaint.

If Gen Z grew up with phones in their pockets, Gen AIR will grow up with AI in their bedrooms, classrooms, toys and all devices. Recent report in The Economist describes AI tutors that adapt perfectly to a child’s level, toys that talk back and remember their preferences, and AI companions that are always available, never sulk, never argue and always validate their feelings. That “perfect companion” creates new kinds of risk: echo chambers, “yes-bots” and cognitive offloading [9]9.

Echo chambers for children: the football‑obsessed child gets only football stories, examples and games. Imagine a childhood where the algorithm decides you’re “the football kid” at age six and never lets you forget it. Serendipity vanishes; so does tolerance for the unfamiliar. This is similar to the social‑media algorithm that feeds you content solely based on your preferences and exacerbates echo chambers and confirmation bias.

“Yes‑bots” as friends: a third of American teenagers (and likely the same in other developed countries) already say talking to an AI companion feels at least as satisfying as talking to a friend, and easier than talking to parents. For a teenager, a friend who never disagrees can be very tempting—and very misleading training for real relationships. My teenage self would have loved a chatbot that did my homework. My adult self is less sure.

Cognitive offloading / Cognitive delegation: students using AI tools show less brain activity and recall less of their own work, suggesting they are outsourcing not just tasks but thinking itself. A recent MIT study measured brain activity with EEG while three groups of people performed tasks using only their brains, a search engine, or an AI/LLM respectively. The results showed clear differences in neural engagement [10]10:

· Users relying solely on their brains had the richest, most distributed neural activity.

· Search engine users had reduced neural activity.

· LLM users had the lowest. Many users could not recall what they had written in their essays.

Simply put, when we lean on AI too heavily, our brains idle. Over time, this erodes learning capacity — a path straight to what I call “artificial ignorance.” For the trained professionals, “artificial ignorance” also means deskilling of their mastery.

If Gen Z is “anxious”, according to Haidt, Generation AIR risks being completely hollowed-out, raised with “perfect partners” who never demand compromise, patience or negotiation, yet quietly encourage subconscious cognitive offloading.

5. A bridging toolkit: Jobs to Be Done

Standing here as a Gen X parent with Gen Z kids, watching Gen AIR on the horizon, I find myself reaching back to Clayton Christensen for help. His “Jobs to Be Done” framework suggests we stop asking “Why is my kid using this?” and start asking “What job are they hiring this for?”[1]

When my child is glued to a screen, that device is doing a job:

· Functional: entertain me, teach me, help me finish homework faster.

· Social: keep me in the group chat, let me belong.

· Emotional: make me feel competent, comfort me, numb my anxiety.

In my childhood, the job “help me feel competent and independent” was served by roaming the neighbourhood, taking buses alone, finding part‑time work. In my son’s childhood, the same job might be served by mastering a game, building an online project or creating content.

The human need or the job is the same; the hired solution looks different. Instead of fighting the tool (screen, phone, social media, AI, etc.), we can ask what job it’s doing, check if it’s doing that job well, and offer alternative ways to meet the same need in the physical world.

“Don’t fight the tool. Ask what job your child is hiring it to do—and whether it’s actually doing that job well.”

Gen X and Gen Z are in a unique position because they are the bridge. Gen X remembers life before digital, mobile phone and the internet; Gen Z remembers life before AI, Robots and Drones. Together, they bridge analog, digital and now AI‑native childhoods.

My kids’ generation (Gen Z) will parent, teach and manage Gen AIR and Q. They have to figure out how to raise humans in a world where AI and quantum computing can do almost everything better, faster and cheaper—hopefully except be human.

6. An experimental prescription for the Human OS

When I read Jonathan Haidt’s three simple guidelines in Anxious Generation, I recognize some of my own mistakes. I’ve pinned these guidelines to my own mental dashboard [8]:

Treat the phone as an experience blocker, not just a distraction. It can block critical “sensitive periods” for learning social skills. Our job isn’t just to confiscate or shame the child; it’s to reopen those periods: family dinners, walks, shared projects, awkward but real conversations.

Scaffold real‑world risk. Encourage the kinds of risk we grew up with—travel, hard projects, unfamiliar groups—so they can rebuild some of the antifragility that unsupervised play once provided.

Fight the algorithm, not the kid. Our children are up against billion‑dollar recommendation engines. They don’t need another enemy; they need an ally who understands the game.

Which brings me back to our two speeches.

In my HKU talk, I ended with a health‑life prescription such as: “May you detect the first cancer cell of ego and heal them daily with kindness and humility.[5]”

In my son’s valedictorian speech, he told his classmates that in the long run, their lives would be measured not by grades but by “the relationships we’ve built, the kindness we’ve shown, and the lives we’ve touched.[6]”

Different worldviews but the same GPS coordinate: kindness. Perhaps this is our Human Operating System (Human OS).

Everything else is upgradable or replaceable: devices, platforms, languages, job skills. The Human OS—kindness, resilience, curiosity, courage, deep relationships—is the only software that has remained backward‑compatible through every generation.

“Our parenting playbook is obsolete. Our Human OS (kindness, resilience, real connection) isn’t.”

We don’t get to captain our children’s ships anymore; they have better navigation tech than we do. But we can be lighthouses—steady, visible, honest about the rocks and the storms.

That, at least, is the faith this Gen X dad is choosing to have. A century ago, people feared that the diffusion of colour TV would make society shallow and destroy the Human OS. The same fears appeared with computers, the internet, social media and smartphones. AI and, later, quantum technologies will stress it in new ways, but if history is any guide, the core code of the Human OS will adapt and survive. And if my kids are reading this: yes, I still reserve the right to complain about your screen time, lovingly so.

Wish you an awesome 2026!

Notes: I have used Perplexity to perform final edits and Gemini 3/Nano Bananas for the last illustration.

Footnote references:

I. Unidentical Twins in AI

The distance between U.S. and Chinese AI is widening, but not for the reason most people think. It’s not because one side is “winning” and the other is “losing.” It’s because their AI aspirations are being raised in different environments, trained on different problems, and optimized for fundamentally different competitive realities.

On one side, the U.S. is building a deep infrastructure moat around pure intelligence. Its stack runs from advanced chip design to hyperscale cloud computing, from decades of enterprise software to capital-intensive foundation models powered by vast data centers. The US controls about 75% of global AI supercomputer capacity with 850,000 H100-equivalents compared to China’s 110,000. US computational performance is 9x China’s and 17x the EU’s.[1]1 Companies like OpenAI, Anthropic, Google, Meta, Microsoft, Nvidia, AMD, Broadcom and Oracle have collectively invested well over US$1 trillion into frontier model research, training, and the infrastructure that supports it. The strategic bet is explicit: whoever builds the most capable generative models and reasoning systems will capture disproportionate value toward Artificial General Intelligence (AGI).

On the other side, China’s structural advantages lie somewhere very different: in striving for efficiency in model training and fine-tuning, as well as in the hardware-centric manufacturing and control of the physical infrastructure that intelligent systems depend on. China produces roughly 75% of global lithium-ion batteries and close to 90% of neodymium magnets, and it dominates power electronics and embedded compute used in EVs, industrial equipment, and robots [2,3]23. This isn’t low-margin commodity production—it’s the critical foundation of every electric vehicle, industrial robot, drone, and autonomous system that will run on AI. China also deploys more than half of the world’s new industrial robots each year, while companies like DJI, Unitree, and UBTech push drones and humanoids toward mass-market viability.

A different “Sputnik moment”

The DeepSeek r1 launch in January 2025 was, in many ways, a “Sputnik moment” for the West, exposing that the AI gap between the U.S. and China is far narrower than many had assumed. It showed that China can now produce frontier-class open-weight models at a fraction of U.S. training costs, catching many observers by surprise. But more importantly and somewhat different this time around, it crystallized two diverging expectations of what AI is for, and where the real value will be captured over time. U.S. and China are the unidentical twins in the global AI and robotics ecosystem.

In the U.S., the twin is surrounded by researchers and compute. It learns to read, write, code, summarize, argue, and plan. Its world is tokens, parameters, and benchmarks. Its job is to think as powerfully and flexibly as possible, for as many people as possible, through a browser window or an API call.

In China, while always optimizing for cost efficiency, the twin grows up on factory floors and in city streets. It learns to grasp, carry, weld, navigate, and monitor. Its world is torque, voltage, sensor noise, and unit price. Its job is to act reliably in the messiness of the physical world—in hospitals, warehouses, eldercare facilities, and logistics corridors.

They are diverging because the problems they’re being prioritized and trained to solve are fundamentally different. The datasets are different. The optimization targets are different. The time scales and failure modes are different. And yet, both kinds of intelligence are essential.

Such divergence somewhat echoes patterns identified by Kai-Fu Lee in AI Superpowers (2018), though I believe the relationship is increasingly complementary rather than zero-sum, and extends Chris Miller’s semiconductor analysis in Chip War (2022) into the AI domain [4,5]45. The crucial point is that the U.S. and China are not simply competing on the same track with the same objective function. They are specializing.

Make no mistake: the twins’ strategies are not static. The U.S. will inevitably invest more heavily in robotics, bring manufacturing back onshore, and build resilient supply chains for critical components. China will continue to pour capital into semiconductor design, large-language models and closed-source frontier systems for their own ecosystem platforms (especially for Tencent and Alibaba), while simultaneously pushing open-source/open-weight alternatives into the commons.

In other words, beneath their divergence lies a quiet convergence on the common denominators of foundational ingredients and overlapping terrains. This offers yet another compelling reason for collaboration and mutual learning.

Finally, one more interesting tale of the superpowers worth highlighting is Dan Wang’s new lens of the twins in Breakneck: China’s Quest to Engineering the Future (2025): China is an engineering state, building big at breakneck speed, in contrast to the U.S.’s lawyerly society, blocking everything it can, good and bad. Each superpower offers a vision of how the other can be better [6]6.

A better set of questions to ask

Viewed through this “unidentical twins” lens, the usual “Who is ahead?” question looks less helpful. Rather, a better set of questions emerges:

• What happens as model capabilities continue to advance toward agentic AI, robotics, embodied intelligence, and ultimately AGI?

• In that world, who captures the value: the owner of the smartest model, or the owner of the robots, vehicles, factories, and infrastructure where that intelligence is deployed?

The DeepSeek moment offers some critical clues to the questions above and helps reveal the emerging flywheel that China is starting to assemble in AI and robotics.

II. China’s Speed and Impact
(Tsinghua Agent Hospital and Wenge Enterprise Workflow)

In less than a year, a relatively unknown team at DeepSeek produced a model that briefly topped global benchmarks and shook Western markets. With improved mixture-of-experts architectures, aggressive distillation, and ruthless efficiency, r1 was released it under an open-weight license. DeepSeek rewrote the rules of the game by asking “who can move fastest, run most cost efficiently, and turn intelligence into an ingredient that anyone can use?” DeepSeek is just one example of how fast AI is being deployed, scaled, and repurposed inside real Chinese institutions.

On the consumer side, we see domestic models powering everything from super-apps to education platforms. On the enterprise side, companies like Manycore have created the world’s largest spatial design platform, and Unitree Robotics humanoid robots G1 and R1 are mastering combat-grade dynamic movement at record speed [7]7. Other AI-native and robotic-native startups such as Zhipu, MiniMax, Moonshot, BrainCo, DEEP Robotics, EngineAI and UBTech, alongside hardware giants and EV makers, are all underpinning the rapid innovation. To really see the character of this wave, though, it helps to look at a couple of lesser-known but equally revealing cases in hospitals and enterprises, where the approach is pragmatic and unapologetically impact-oriented.

Healthcare offers one of the clearest windows into this speed

In May 2024, Tsinghua University’s Institute for AI Industry Research (AIR) and Beijing Tsinghua Changgung Hospital launched Agent Hospital: a fully simulated hospital where AI agents act as doctors, nurses, patients, and administrators [8]8. Within this digital twin, thousands of simulated cases run across different departments: internal medicine, surgery, pediatrics, emergency, and more. Agents learn triage, diagnosis, treatment workflows, and care coordination in a safe, synthetic environment—long before touching real patients.

Beijing Tsinghua Changgung Hospital is a separate, real-world teaching hospital. This integration is for assistance and decision support, not for replacing human doctors entirely. By 2024, the system had been stress-tested at scale. Forty-two AI “doctors” were operating across 21 medical departments. On the patient side, researchers generated 50,000–100,000 synthetic patient agents mirroring real demographic, geographic, and disease patterns. Within weeks, the system had completed 10,000 virtual patient cases—roughly equivalent to two years of human clinical workload—while achieving 93% diagnostic accuracy on MedQA respiratory datasets, with strong performance across other specialties [8].

Tsinghua borrowed from autonomous driving framework and defined four levels of human–AI collaboration:

• Level 0 – Simulation only: agents operate entirely in a virtual hospital (current status).

• Level 1 – Shadow mode: agents run alongside clinicians, reading the same charts and proposing diagnoses and treatment plans, but only for human reference.

• Level 2 – Suggestion mode: agents propose decisions; clinicians approve, override, or modify them.

• Level 3 – Limited autonomy: agents handle routine, low-risk cases autonomously and escalate edge cases or ambiguous situations to human doctors.

The first real-world pilot at Tsinghua Changgung Hospital begins in 2025. By the time an AI agent sees its first real patient, it will have “seen” far more synthetic cases—across more specialties—than any human doctor will over an entire career. Vice Provost and Senior Vice-Chancellor of Tsinghua Medicine Wong Tien Yin noted that the AI Agent Hospital is designed to transcend the traditional “Hospital + AI” model [9]9.

If Agent Hospital is the visible frontier in healthcare, the real economic impact is unfolding quietly inside thousands of enterprises wrestling with a different reality: AI hype versus AI impact.

Implementing AI in the enterprise workflow is the real test for AI impact

MIT NANDA’s The GenAI Divide: State of AI in Business 2025 report found that while roughly 90% of companies claim to “use AI,” 67% remain stuck in pilot mode, and only 6% have successfully embedded AI into workflows with measurable business impact [10]10. The gap between AI adoption and AI impact is not primarily a technology problem. It is an implementation problem.

Wang Lei, founder of Zhongke WengAI (Wenge)—a Beijing-based spin-off from the Institute of Automation at the Chinese Academy of Sciences—has built a business around closing that gap. Wenge is what Chinese policy calls a “little giant”: a national specialized, sophisticated, differentiated, and innovative SME targeted for support under Beijing’s strategy to cultivate high-growth, technology-driven companies.

Wenge does have its own proprietary base models (Yayi Foundation Model), but that’s not where it competes. Instead, it offers pre-built agent toolkits with native industry-specific models trained on sector data, deep integrations into legacy systems via APIs, and out-of-the-box workflows tuned for government administration, manufacturing, retail, finance, logistics, media and more.

More than 1,000 enterprise and government clients have embedded Wenge agents into live operations. The workflow AI includes all-media intelligent topic selection and production for Xinhua News Agency, traditional Chinese medicine-assisted medication for the Academy of Traditional Chinese Medicine, the AI intelligent government service system in Zhongguancun, experts in the field of finance and taxation for Air China, financial and legal experts in Hong Kong, and investment research and credit analysis for small and medium-sized enterprises [11]11.

Wenge also illustrates how fast this can go global. In 2025, it formed a strategic alliance with Cherrypicks, a major Asia-Pacific application developer headquartered in Hong Kong, to package its agent stack for overseas markets where the same workflow agents now handling Chinese documents and procedures are being localized and exported as a turnkey AI layer for enterprises in the global markets [12]12.

Both Tsinghua Agent Hospital and Wenge Enterprise Workflow speak to China’s speed and pragmatism at scale: start from workflows and bottlenecks; iterate in production across hundreds of clients; and export the playbook. Some of these experiments will fail. Many will struggle with shaky economics. But in classic China fashion, the system is betting that running enough experiments fast enough—in models, hospitals, workflows, factories, and cities—will surface the configurations that matter.

III. DeepSeek r1 as the Opening Gambit of Commoditizing Intelligence

If open-source and open-weight models keep improving at their current trajectory, China’s AI industry may trigger a fundamental shift in where competitive advantage sits in the entire AI and robotics stack. The DeepSeek moment is the opening gambit of that shift. Today, the most intelligent models remain closed: OpenAI’s GPT-5, o3, Gemini 2.5 Pro, Claude 4.1 Opus, and Grok 4. The open-weight models—OpenAI’s GPT-OSS 120B, Qwen 3, and DeepSeek v3.2—trail closely in aggregate intelligence performance. While top performers remain predominantly U.S. models, Chinese models are emerging rapidly. After years of trailing the US in model quality prior to 2023, Chinese models have surpassed US counterparts in global downloads and model adoption. For example, among open-weight/open-source models, Qwen 3 edges out Meta’s Llama 4 in user preference and global downloads. The reason goes far beyond mere performance: it lies in accessibility and adoption-friendliness for developers and builders in terms of training stacks (OpenRLHF and verl), Apache-2.0/MIT license, model sizes and shapes. On the video generation front, Tencent’s Hunyuan Video (13B), an open-sourced transformer-based diffusion model, outperformed Runway Gen‑3 and Luma 1.6. The momentum continues to accelerate, setting the stage for China’s open-sourcing strategy [1].

Value migration

When DeepSeek r1 launched in January 2025, the surprise was not only on closing the performance gap, but also for its superb economics in characteristically Chinese style. When DeepSeek-level reasoning can be purchased for on the order of US$0.50 per million tokens, and comparable open-weight alternatives proliferate, foundation models start to drift toward commodity status. The business logic that sustained premium positioning—”we have the best model, pay accordingly”—inevitably erodes. Competition is forced to move to adjacent layers of the value chain. This dynamic follows Harvard Professor Clayton Christensen’s theory of value migration in The Innovator’s Dilemma (1997), where disruptive innovation commoditizes incumbent technology and shifts value to adjacent layers of the stack [13]13.

We have seen similar movie before. When HTML became an open web standard and browsers commoditized, value shifted to web services, search, and advertising. When Android and iOS created stable smartphone platforms, value moved from operating systems to apps, services, and chipsets. When cloud computing standardized infrastructure, value migrated to SaaS and data. The same dynamic is now unfolding in AI. As foundation models commoditize, the locus of competitive advantage shifts decisively to embodied implementation—robots, drones, agentic devices, and smart sensors deployed at scale.

If intelligence becomes effectively “free”—available via open-weight models and cheap APIs at negligible marginal cost— then the center of gravity shifts to the layers that are not easily copied: hardware, data, and deployment ecosystems in the real world. This is where China holds structural advantage.

China’s open-source strategy makes sense in this light. Rather than fighting an uphill battle for absolute dominance in frontier model research, where Western capital and talent remain heavily concentrated, China is more than willing to commoditize that layer by opening model weights, training frameworks, and tooling into the commons. In exchange, they aim to dominate the layer that will generate trillions of dollars of economic value: intelligent systems that act reliably and cost-effectively in the real world.

Commoditized intelligence lever

For embodied AI companies, commoditized models are a gift. Roboticists, drone manufacturers, and autonomous system teams no longer need to build or fine-tune proprietary large language models. They can plug in whichever open model is “good enough” and focus entirely on the other hard problems: building robust world models for physical understanding, designing mechanical and electrical systems, achieving real-time control under noise and uncertainty, and adapting to unstructured environments.

When open-source/open-weight models such as Qwen 3 perform “good enough” for most real-world tasks and come in all shapes and sizes with developer-friendly attributes, competition shifts. The bottleneck is no longer building and training the smartest model. It is building the cheapest, safest, most reliable systems that embed that intelligence in machines, workflows, and infrastructure.

Billion-user ecosystem players such as Tencent, Alibaba and ByteDance will continue to pour capital into large-language models and closed-source frontier systems for ecosystem competitive advantage, while simultaneously push for open-source/open-weight as a value migration lever to drive the competition to a layer that China has most advantages - energy, hardware and deployment systems. And that leads directly to the next question: in a world where intelligence commoditizes, who owns the energy, hardware, and deployment systems that turn intelligence into reality.

IV. China Is Racing to Where the Ball Is Going: Energy and Embodied AI

While American AI discourse remains fixated on who has the best model in the race toward AGI, China has quietly positioned itself to dominate a deeper, more material layer: the hardware and systems that turn electricity into physical action.

The trinity for producing everything

As Packy McCormick argues in his “Electric Slide” analysis, everything we produce ultimately rests on three pillars: energy, intelligence, and action [14]14. Energy is the ability to generate and distribute power for the insatiable demand of compute. Intelligence is the suite of models, algorithms, and agents that make decisions. Action is the translation of those decisions into real-world effects—robots moving, drones flying, vehicles routing, machines operating. This trinity of ingredients will truly unleash the full power of AI while intelligence alone—without energy and action—is impotent thinking.

The U.S. seems to be making an implicit bet: win decisively on intelligence first, then catch up on energy infrastructure and embodied systems later. China is running the opposite play: build overwhelming strength in batteries, motors, power electronics, and manufacturing capacity, and then plug increasingly capable intelligence into that base.

In terms of peak electricity demand, China and the U.S. set the records of 1,450 GW and 759 GW respectively. China does not only serve more demand, it is also building a larger overhang of available power. China now generates roughly 2.5 times as much electricity as the U.S.. It operates the world’s largest renewable energy system and manufactures the majority of global lithium-ion batteries and a dominant share of neodymium magnets for electric motors. It is rapidly closing the gap in power electronics and now produces a large portion of the world’s power conversion systems. While still catching up in leading-edge chip manufacturing, China is accelerating its dominance in edge processors and embedded AI chips—exactly the components that go into robots, drones, vehicles, and industrial equipment. Crucially, companies like BYD, DJI, and Huawei don’t merely supply parts; they integrate the entire stack into finished product [15]15.

Controlling the cost curves of the electro-industrial stack

This integration powers almost everything that moves, powers, or computes, from EVs to robots and drones to data centers. China’s dominance of the electric stack is not just about supply-chain self-sufficiency. It increasingly controls the cost curves for the entire embodied-AI ecosystem.

The dynamic works like this:

• Motors get cheaper.

• Batteries get cheaper.

• Power conversion hardware gets cheaper.

• Edge compute gets cheaper (even if the very highest-end chips remain constrained).

When all of those components get cheaper at the same time, robots get cheaper. When robots get cheaper, deployment accelerates. As deployment accelerates, data generation increases. As data increases, AI models improve. As models improve, more complex tasks become automatable. This is the virtuous cycle China has been orchestrating.

For embodied AI to move from niche applications to transformative, economy-wide impact, unit economics must cross a threshold. A US$150,000 humanoid can only justify itself in wealthy countries and narrow use cases. A US$6,000 humanoid—roughly where Unitree has driven some models, down from around US$16,000 in a single year—can scale to eldercare facilities, warehouses, and factories around the world. That kind of price compression is driven almost entirely by electro-industrial cost curves that China increasingly controls.

Another critical foundation of the electro-industrial stack is the dominance of steel and rare earth production. In 2024, China produced over 50% of global steel and 70% of global rare earth elements, while processing nearly 90% of the world’s rare earths [16, 17]1617.

Investors like Ryan McEntush at a16z have argued that the electro-industrial stack—batteries, motors, power electronics, and compute—will “move the world,” unleashing a productivity boom and a new industrial renaissance [18]18. If that’s true, then China, with its world leadership in manufacturing everything from EVs to drones to electric bikes to robots, is heading directly to where the ball is going.

In industrial robotics, China has already become the global center of gravity. According to the International Federation of Robotics, roughly 4.28 million industrial robots were operating in factories worldwide in 2023, an all-time high and a 10% year-on-year increase. Annual installations have held above half a million units for three consecutive years. Around 70% of all new industrial robots in 2023 were installed in Asia, reflecting both manufacturing concentration and the rapid shift from labor-intensive production to highly automated, AI-enabled factories [19]19.

A decade ago, most industrial robots in China were imported. Today, Chinese manufacturers provide close to or more than half of all robots installed in China, with domestic market share rising to roughly 47–57% in 2023–24. China has also overtaken Japan to become the largest producer of industrial robots globally. With a domestic installed base approaching 2 million robots by 2024, China’s factories have become the main training grounds for the next generation of embodied AI [20]20.

Compared with industrial robots, humanoid and service robots are still early in their growth curve. Global installed volumes remain small, but capital and expectations are large. Morgan Stanley estimates that the global humanoid robot market could reach US$4.7–5 trillion in annual revenue by 2050, with a cumulative installed base of more than 1 billion humanoids worldwide. Their analysis starts from a roughly US$60 trillion “embodied AI” total addressable. Indeed, China could end up the “embodied AI” factory for Western brands in addition to being self-sufficient domestically [20].

For now, however, the humanoid market remains capital-intensive and experimental. Before humanoid robots can take off, they still need AI world models and physical AI capabilities for spatial reasoning, and comprehensive safety measures beyond just economic viability.

Undeniably though, these elements of commoditized intelligence, energy scale, manufacturing dominance, and deployment speed form the core competence of China’s advantages. The question for Western competitors, then, is no longer simply, “Can we compete in AI?” It is, more precisely:

Can we compete in the manufacturing, supply chains, energy systems, and cost optimization that turn intelligent systems into deployable reality?

V. The Emerging AI–Robotics Flywheel

Codename for a long game - MIC2025

China’s capabilities in energy, hardware, and manufacturing didn’t materialize overnight. They are the product of decades of industrial policy and investment—“Made in China 2025,” (MIC2025) infrastructure build-outs, and long-run bets on high-speed rail, new-energy vehicles, and biomedicine. The national strategy “Made in China 2025,” launched in 2015, explicitly aimed to shift the country from low-cost manufacturing to a high-tech, high-value-added industrial powerhouse. Ten strategic sectors were identified, with high-end numerical control machinery and robotics prominent among them, alongside next-generation IT, advanced rail, new energy vehicles, and biomedicine [21]21.

Despite trade headwinds and export controls, multiple independent analyses suggest that around 80–86% of the more than 260 specific Made in China 2025 targets have been achieved by 2024–25 [22]22. In robotics, the policy translated into domestic content requirements, R&D subsidies, local action plans (especially in provinces like Jiangsu and Guangdong), and the build-out of innovation platforms, standards, and testbeds. The rise of dense regional ecosystems—sometimes described as “six dragons” in places like Hangzhou—reflects the combination of local government support, entrepreneurial drive, and dense supplier networks.

In August 2025, The State Council of China released “AI-Plus Plan”, a national strategy to deeply integrate artificial intelligence across six key sectors (science, industry, consumption, quality of life, governance, global cooperation) by 2030, aiming for an “intelligent economy” and “intelligent civilization” by 2035, focusing on practical deployment, “new quality productive forces,” robust infrastructure (data/compute), and fostering an open-source AI ecosystem to drive economic growth and national competitiveness, distinct from purely frontier research. In short, in 10 years, China aims at AI capabilities fully integrated across the entire economy with complete AI penetration across all sectors [23]23.

This “long game” mindset driven by the State allows incremental gains in talent, supply chains, and capital to compound into an ecosystem that now moves at speed and scale few other economies can match.

All of these capabilities are starting to behave like a self-reinforcing AI–robotics flywheel. This flywheel spins through three interconnected layers: industrial production, societal adoption, and intelligent infrastructure.

Layer 1: Industrial foundation

The first layer of the flywheel is the industrial foundation: a manufacturing base and near self-sufficient robotics supply chain that provides an unmatched proving ground.

China now has one of the world’s largest installed bases of industrial robots, and it continues to deploy more than half of all new industrial robots globally each year. Each robot on each production line is not just replacing human labor; it is generating terabytes of real-world sensor data—gripper pressures, motion paths, error recoveries, task adaptations—that can be fed back into machine-learning models to improve the next generation of robots. That creates a data virtuous cycle: more robots in factories produce more real-world training data; better models produce more capable robots that can handle more complex tasks; more capable robots justify further adoption. The factory floor has essentially become the training ground for embodied AI in the physical world.

Simultaneously, every critical component—actuator, torque sensor, gear reducer, vision system, control board, even custom microcontroller—is available domestically, often within dense localized ecosystem clusters like Shenzhen’s Huaqiangbei or Beijing’s Zhongguancun. A new robot design in such a localized ecosystem can go from concept to prototype in weeks, not months, with cost-efficient iterations happening at a pace that is difficult to match elsewhere. The economic barrier to experimentation is low, iteration cycles are tight, and price–performance improves quickly. A “wolf-pack” dynamic—dozens or hundreds of small competitors all pushing to outperform each other—drives down costs while simultaneously ratcheting up capabilities.

Layer 2: Societal adoption

The second layer of the flywheel is societal adoption, driven by demographics, urbanization, healthcare and policy.

Widespread cultural optimism toward technology adoption

Cultural optimism toward technology adoption is widespread in China, with citizens readily embracing innovations that promise to improve everyday life. After all, this confidence is rooted in tangible everyday experience: over the past decade, citizens witnessed the transformative leap from cash-based transactions directly to mobile payments—skipping the intermediary card-based phase entirely. Today, that optimism is manifesting visibly. Robots serve meals in restaurants, perform as dancers in pop concerts, and take center stage during New Year celebrations.

Preparing for the tsunami of labour shortage and silver economy

China’s demographic and labor trends create a structural pull for robotics. More than 310 million people are now over 60 (compared to the U.S. population of about 340 million), and the working-age population is shrinking. Young people increasingly avoid low-wage, physically demanding, or hazardous jobs even at higher pay. This produces acute shortages in exactly the roles robots can fill: eldercare staff, warehouse workers, cleaners, and basic service roles. Unlike many Western countries where labor regulations, unionization, and social anxiety can slow automation, China’s state-driven policy environment and relatively enthusiastic attitude toward new technology make large-scale experiments politically and socially feasible. The Economist estimated that, by 2050, 487 million people will be over the age of 60, about 35% of China’s population, up from 21% today. This calls for digital healthcare to support the wave of chronic diseases brought by the rapidly ageing population [24]24.

Central and local governments are actively seeding adoption. A three-year elderly-care robotics pilot program is underway. Notable pilot examples include Beijing’s smart elderly care station robot and Shanghai Changning District’s elderly care robot companion. Cities like and Shenzhen offer subsidies of up to 30% of the purchase price for approved eldercare robots. The Ministry of Civil Affairs has issued directives encouraging deployments in hospitals, nursing homes, and community care centers across dozens of cities. Each deployment produces data and experience that loops back into product design and policy, tightening the link between need, experimentation, and refinement [25, 26]2526.

Telemedicine took off during the covid-19 pandemic out of necessity. While significant traction was achieved by digital health apps such as JD.com’s JD Health App (serving over 500,000 consultations a day) or Ant Group’s AQ app (served 140 million patients with 1 million doctors on the platform), the typical users of telemedicine apps are young urban resident who already have access to the best access to China’s public health services. Hence, it is far from the original goal of offering healthcare access to faraway experts for the rural population [24].

In November, 2025, the Central Government released a plan calling for “full coverage” of AI-powered diagnosis and treatment tools at grassroots health centers by 2030. The notion here is to allow local doctors to use AI models to access the latest advice and help build patients’ trust in the technology while only complex cases would go to large clinics and hospitals [23].

Layer 3: Intelligent infrastructure

The third layer of the flywheel is intelligent infrastructure, where state coordination is most visible, especially in smart cities, autonomous driving (autonomous EVs and robotaxis), and the low-altitude economy (drones and air taxis).

Not just intelligent things

China has designated hundreds of cities as candidates for smart-city development, with billions of dollars flowing into intelligent infrastructure: sensors, cameras, roadside units, and computational layers used to construct unified digital twins of entire urban areas. In policy documents, this is often described as “vehicle–road–cloud integration”: vehicles, roads, and cloud systems all talking to each other in real time [27]27. It is the whole symbiotic integration of vehicles, physical road infrastructure and backend cloud infrastructure that requires to be intelligent and connected.

Pilot cities like Shenzhen, Shanghai, Beijing, Wuhan, and Guangzhou are running thousands of autonomous vehicle-miles per day on real streets, under coordinated traffic control and data-collection regimes.

Don’t just fly, soar

Parallel to ground-based systems, China has identified the low-altitude (LAE) economy as a strategic emerging industry. The LAE comprises of consumer drones (0-120m), industrial drones (up to 300m), cargo drones (50-300m) an eVTOLs/passenger aircraft (300-600m on fixed, pre-approved routes). The Civil Aviation Administration projects this market could reach 1.5 trillion yuan (about US$210 billion) by 2025 and 3.5 trillion yuan by 2035 [28, 29]2829.

To enable that, regulators have opened dedicated air corridors, simplified certification for cargo flights, and built take-off/landing sites, maintenance centers, and localized air-traffic management systems across more than 30 provincial regions. Companies like Ant Work (medical drones), DJI (commercial and agricultural drones), JD Logistics, Meituan, and SF Express already operate thousands of drone routes, completing hundreds of thousands of commercial deliveries in major Chinese cities such as Shenzhen where drone landing pads are installed in office buildings and shopping malls for aerial deliveries. In addition, medical delivery corridors into rural areas are established to dramatically cut down the accessibility and timeliness of healthcare [30]30. All of this creates a unique ecosystem where billions of autonomous vehicle-kilometers—on the ground and in the air—are logged annually. In fact, a robotaxi revolution is gathering pace where the cost of China’s robotaxi is one-third of the cost of Waymo’s in America [31]31.

Put these three layers together and you get a self-amplifying loop:

1. More embodied AI systems (robots, drones, autonomous vehicles) are deployed.

2. More use cases, data, user feedback, operational insights, and fail-forward lessons are generated.

3. Better algorithms, training datasets, and control policies emerge.

4. Robots become more capable and cheaper to build.

5. New applications and solutions emerge.

6. Loop back to step 1.

And the flywheel spins faster.

Morgan Stanley estimates that humanoid robots alone could see cost declines of more than 50% over the next 5–10 years as volumes increase and AI improves, mirroring the cost trajectories seen in electric vehicles and solar photovoltaics before them [18]. If that happens, the economics of deploying embodied AI at massive scale begin to look not only plausible but inevitable.

The 158-km clue

Not only is China striving to build intelligent infrastructure, it is also transforming the way infrastructure is being built and maintained. In October 2024, China had completed the world’s first fully unmanned paving construction along a 157.79 km stretch of the Beijing-Hong Kong-Macao Expressway. The use of fleets of autonomous pavers, robotic rollers and drones, all operated remotely with no worker on the road represents a powerful leap in AI-driven road construction and maintenance in the future. This approach reduced human risk and improved safety, quality, speed and scale of infrastructure delivery [32]32.

What makes China’s AI–robotics flywheel distinctive is simultaneity and scale. Industrial robots are already deeply embedded in production; humanoid and service robots are moving out of labs and into early commercialization; autonomous vehicles and drones are being tested and, in some cases, used in real operations across dozens of cities. China is running all three layers together, across hundreds of cities and thousands of factories, at once. No other country is running this experiment at comparable scale across all three dimensions.

VI. Entrepreneurial Culture, Deep Talent Pool and Permissive Policy As Turbochargers

The flywheel is not just a product of top-down planning. It is also turbo-charged by entrepreneurial culture in the private sector and lubricated by policy at the state, provincial, and city government levels.

The iteration culture

Chinese entrepreneurs in robotics and embodied AI are willing to show unfinished, imperfect products in public. At trade shows and expos, it is common to see humanoids that stumble, service robots that occasionally misroute themselves, and drones that are still rough around the edges. Instead of hiding these flaws until everything is polished, founders invite feedback, iterate fast, and treat the field as a living lab. This cultural permission to fail fast, learn cheaply, and fix quickly shortens the path from idea to viable product.

Consumer-facing humanoids and service robots from firms like Unitree (sub-US$6,000 humanoids), UBTech (relatively affordable home companions), EngineAI Robotics (ready for mass production by 2026) and other startups are already in the hands of early adopters and researchers. They are not perfect, but they are out there—walking, falling, learning. The same is true in cleaning, inspection, delivery, and security robotics. As early adopters deploy these systems in small but meaningful ways, they generate usage patterns, performance reports, and failure logs that feed directly into product roadmaps.

While American robots remain mostly in labs, Chinese firms are field-testing theirs across stadiums, factories and martial arts arenas in an industrial revolution fueled by scale, speed and system. South China Morning Post reported a stark contrast between tech giant Tesla and a shoestring Chinese robotics startup in the Greater Bay Area region. Tesla’s Optimus recently “set a personal record” by jogging a few steps but fell backwards while trying to hand over a water bottle. Meanwhile, in less than two years, EngineAI Robotics, led by CEO Zhao Tongyang, created the T800 – a robot that delivers roundhouse kicks with the peak torque of a small car, human-like dexterity, and a solid-state battery for extended operation, will be ready for mass production by 2026 [33]33.

U.S.-China talent pool inversion

The entrepreneurial culture is reinforced by the deep talent pool of armies of engineering graduates. Beijing, Hangzhou, Shenzhen, and other hubs now host dense networks of founders, engineers, and applied researchers. Each year, China now produces on the order of 1.5 million engineering bachelor’s graduates, compared with about 140,000 bachelor’s graduates in the United States—an order-of-magnitude difference in the size of the technical talent pool. Beyond sheer volume, universities and research institutes are adapting by allowing a working product, rather than a traditional thesis, to serve as a PhD defense, explicitly rewarding translation of ideas into real-world systems. Harbin Institute of Technology (HIT), one of China’s “Seven Sons of National Defense,” is part of a broader national pilot program led by China’s Ministry of Education and other agencies to let several top engineering universities in strategically important fields (semiconductors, quantum, etc.) experiment with product-based doctorates [34]34.

In addition, an interesting contrast emerges when examining the teams’ bench strength at DeepSeek and Meta’s Superintelligence Lab. A Stanford report analyzing 201 staff at DeepSeek found that 55% were trained and based entirely in China, with no U.S. affiliation. Only 24% of DeepSeek authors had any U.S. affiliation, most for just one year [35]35. Meanwhile, half of the researchers reporting to Alexander Wang in Meta’s Superintelligence Lab received their undergraduate degrees in China [1].

On December 17, Tencent appointed Yao Shunyu, a 27-year-old fomer OpenAI reserach, as Chief AI Scientist, leading a new AI research structure for the billion-user Tencent/QQ/wechat ecosystem, and reporting directly to the President Martin Lau [36]36.

These cases signal China’s growing ability to train and retain its scientists locally and attract scientists from U.S., while the U.S. has become increasingly dependent on Chinese AI talent. Many overseas-educated researchers and engineers are returning to mainland China and Hong Kong to join University research and start companies, drawn by this combination of opportunity, ecosystem support, access to supply chains and geopolitics. And three Chinese tech giants (Tencent, Alibaba and ByteDance) spent over RMB 100 billion on AI infrastructure in 2024 [36].

Policy is the third turbocharger. Numerous examples throughout this essay reflect a pattern of state-driven policy, permissive regulations and agile local government incentive schemes. At the highest level, China’s leadership has repeatedly emphasized the importance of AI. In April 2025, Xi Jinping used a Politburo study session (20^th collective study session) to push for an “orderly development” of AI, stressing both opportunity and risk. These high-profile signals matter. They tell ministries, provincial leaders, and SOE executives that AI is not optional; it is a strategic priority. This is consistent with the only previous session focused on AI, the 9th collective study session held on November 2018, where Xi called for “healthy development of a new generation of AI. [37, 38]3738”

Entrepreneurship and the deep talent pool provide the catalysts. State-driven, long-term policies and permissive regulations provide the lubricant of agile rulemaking. Together, they keep the flywheel spinning.

VII. Going Global as the Fifth Great Invention

If this AI–robotics flywheel fully materializes inside China, it will not remain a domestic phenomenon. It will be exported—just as China has already exported high-speed rail, 5G infrastructure, smartphones, solar panels, and electric vehicles.

In each earlier wave, the pattern was similar: China combined domestic market scale, manufacturing cost advantage, and state-backed export strategies to capture large shares of emerging and developing markets. Robotics and embodied AI are poised to follow a similar trajectory, with potentially deeper implications.

The Global South—BRICS countries, Belt and Road partners, and other emerging markets—represents a massive unmet demand for automation, infrastructure, and services. Many of these economies face their own versions of labor shortages, urbanization challenges, and infrastructure gaps. For them, Chinese robots, drones, and AI infrastructure can be an attractive package: lower cost, faster deployment, bundled financing and training, and often fewer political strings than Western alternatives.

China is already positioning robotics and embodied AI within its major geopolitical frameworks, especially the Shanghai Cooperation Organization (SCO) and the Belt and Road Initiative (BRI). At the 2025 SCO Summit in Tianjin, for example, China proposed a new “China–SCO Artificial Intelligence Cooperation Forum” and announced a plan to build an AI cooperation center for more than 20 SCO member nations [39]39. In parallel, Chinese companies are actively exploring BRI channels for robotics and AI exports.

In embodied AI, diversity of training data is crucial. A humanoid or service robot that learns only in Chinese factories and apartments will be useful, but one that has been trained—through federated learning or structured data sharing—on China’s factories, Thailand’s farmlands, India’s hospitals, and Brazil’s elderly homes will be far more robust. China’s expanding networks of BRI and SCO partners give it a natural platform to pursue such a “global training set” in partnership with the Global South. This echoes earlier export plays in high-speed rail and telecoms but moves them up one layer, into the intelligence stack itself.

In August 2025, the State Council published an opinion paper describing AI as a “public good for humanity” [40]40. China has proposed setting up a new World AI Cooperation Organization and is working with the UN’s Pact for the Future and Global Digital Compact to build out processes for AI development and governance.

The private sector is already moving along this vector. Zhongke WengAI, for instance, is pursuing an explicit “going global” strategy. By investing in and partnering with firms like Cherrypicks, a major Asia-Pacific application developer and integrator headquartered in Hong Kong, it is building pathways to bring Chinese AI tooling into broader Asian, Middle East and Global South markets [12].

For many developing countries, the appeal is straightforward: lower cost than Western vendors, faster time to deployment, more flexibility in integrating with local systems, and a greater sense of sovereignty than being locked into proprietary Western SaaS and licensing regimes. Just as important, China is not only exporting technology; it is exporting standards and governance frameworks. Initiatives like the “Shanghai Declaration” and other AI governance proposals are being pitched as reference points for countries that want AI development but are wary of being fully dependent on U.S. or EU rulemaking [41]41.

As The Economist’s latest feature article on “What China will dominate next” sharply pointed out that, in response to China’s rise, any knee-jerk protectionism in the name of security or safety from the West would be a mistake. It would be better for Western economies to reflect and rethink how innovation works at home [31].

For centuries, the world has remembered China for four great inventions: papermaking, printing, the compass, and gunpowder. Each fundamentally reshaped human civilization and spread far beyond China’s borders. Today, as Chinese AI matures from generative models to embodied systems deployed at scale, could embodied AI (AI and Robotics) become China’s fifth great invention for the world? If history is any guide, this time the diffusion wouldn’t take centuries. It could unfold within a single decade.

Notes: I wrote this article when I was invited to provide input on AI and Robotics for a new book about how companies need to change to win in China. I have used Perplexity to perform final edits and Gemini 3/Nano Bananas for illustrations.

Footnote references:

14

P. McCormick, “The Electric Slide,” Not Boring by Packy McCormick, Dec. 2024. [Online]. Available:

Not Boring by Packy McCormick

The Electric Slide

Welcome to the 1,269 newly Not Boring people who have joined us since our last essay! Join 248,515 smart, curious folks by subscribing here…

Read more

8 months ago · 894 likes · 67 comments · Packy McCormick and Sam D'Amico