Hello everyone, I'm James. Welcome to Digital Keywords, where we learn about tech — big and small. In recent years the letters "AI" have been in front of us almost every day. The full name is artificial intelligence — by definition, intelligence made by humans. But in this episode I want to play a pun that only really works in Mandarin: swap the character for "human/person" in "artificial" for the character for benevolence and care — "仁" — and you get "仁工智慧". One character's difference — what does it add? It adds a heart that cares for others. The same technology: you can choose to give it a caring spirit, or let it serve the rich without conscience. That wordplay is one of the things our guest has been most invested in these past few years.
Speaking of our guest Audrey Tang, many people may still picture mask maps and SMS contact tracing — pandemic tech. But Audrey is now a researcher at the Oxford Institute for Ethics in AI, running a programme called Civic AI — in Mandarin, 仁工智慧, a play on “artificial intelligence” with the character for care. That framework was formally launched at Oxford only this March; the book may not appear for another year or two. So this episode may well be Audrey's first in-depth conversation about this work-in-progress with a Mandarin-speaking audience. We're honoured to have her — welcome.
Glad to talk about Civic AI, or 仁工智慧 (“care-full intelligence”).
First I'd like to stretch the timeline for listeners. Before you joined the cabinet I had a chance to know you. Back then we were in g0v-related communities; I know you were in g0v very early and doing AI-related research.
Right. I spent six years assisting the Siri team.
So you were in touch with what we now call early AI very early. Later you became a minister without portfolio, then led the Ministry of Digital Affairs, until you stepped down and became our ambassador-at-large. Now you're researching at Oxford — could you walk us through that arc? What have you mainly been doing since leaving office?
Right — we chatted a bit before recording. I've been here about two years, in probably more than 27 countries; cities, many more. On average I'm changing time zones every week. Fortunately I don't get jet lag — only jet boost — so each time zone shift I sleep better.
The reason is that friends worldwide are hitting problems Taiwan faced a few years ago — familiar ones like information manipulation and fraud, or the social polarisation we may not remember as vividly. Many democracies now face them too, and sometimes worse than we did around 2014.
So solutions we thought of then — social leaders, political leaders, religious leaders — all wanted to hear about them. I mostly share what we tried in Taiwan.
These two years I've been a digital governance ambassador — "cyber" as in cybernetics, kybernan as steering — sharing Taiwan's experience of steering through an era of transformative technology. That's what I'm doing now.
Do you feel Taiwan is special — like a laboratory, or particular soil? Maybe because we got universal broadband earlier, and many people have various OSes, phones, computers, so we met digital tech — now AI — very easily, and hit similar social problems sooner?
Yes. Research is catching up. A colleague from Project Liberty, Jonathan Haidt, wrote The Anxious Generation. He points to a strong positive correlation: the more you scroll on your phone, the more FOMO and anxiety; more screen time, worse sleep, then varied social problems, ending in polarisation.
In Taiwan around 2012–2013 — connectivity, broadband uptake, time glued to phones, Facebook accounts outnumbering population — while other countries' infrastructure was less mature, they met this a few years later. We saw early that recommendation engines push addictive content, often polarising content. We were ahead by two or three years, in some places four or five, so we had to deal with it. How we dealt with it is now something later democracies can reference.
It really is special. As you shared, Taiwan has often been relatively forward in adopting IT — we take things up quickly. We feel we're always looking up at the US and others — "they have this tech, why haven't we?" From computerisation to e-government and mobile, each wave — PC, internet, mobile, now AI — we've moved faster on that path.
But if so many countries, societies, and leaders want Taiwan's story, what path pulled you from today's AI toward people-centred Civic AI — when we say "仁", should we say mercy in English? What made you think we should do this?
I'd say "Civic Care" — civic caring.
Mainly it's like the recommendation-engine polarisation example. We asked AI algorithms to maximise what we call engagement. It wasn't deliberately polarising you; deep learning found that if you argue with others, you keep interacting.
But nobody wrote an expert system targeting James with anger-bait. People saw early: you can maximise any metric, but what looked good at first often produced bad outcomes — gaming the objective, cheating toward the goal.
Like Michelin three stars and the best food — everyone loves eating; but AI discovers junk food, heavy oil and salt, rewiring your palate so you're addicted, you keep coming back, at maybe under one per cent of the cost — so it optimises that.
Many problems now, including polarisation, are side effects of that generation of AI maximising things.
To fix it you change what it's "aligned" to. You can say vaguely it shouldn't cause harm — shouldn't polarise — so you change the metric to measure how much people flame each other — some really did that — and in the end it only pushes your echo-chamber content; you think everyone agrees with you. Less visible fighting, but a more fractured society.
Changing maximisation metrics one by one, you can't keep up. How do you train AI models to avoid the whole package? That's the question. Why care — why care? In Taiwan, personal computing from the 1980s: one mainframe, everyone on terminals; changing mainframe software affects many people, you can't pre/post-test everyone with surveys, so as sys admin, upgrades sacrifice some users.
But if everyone has a home PC with spreadsheets and desktop publishing, that problem largely disappears. If everyone can train AI relevant to themselves at home, they're more likely to tune it to what people around them need — to care for specific relationships.
Philosophically we turned to care ethics. Not consequentialism maximising a number; not deontology with abstract rules everyone must follow — but adjusting to what people beside you find good. That path led us to care ethics.
So that's the thread. In much of our work together over the years — impacts on society — twenty-odd years ago people were already debating recommendation systems and related algorithms.
So the endless-scroll algorithms and later echo chambers — the Mandarin "同溫層效應" — are phenomena of the last decade or so, whether you call it big data, collecting signals, tuning parameters, training popular algorithms, especially when linking more people.
A striking phenomenon. But I notice something about your personality — my own read. When a system looks off here or there, most people blame or protest, demanding owners or developers fix it.
That's most of us. You've been different. Blaming is easier; you tend to think, discuss: rather than protest without acting, show a slightly different demo yourself. For example, early government platforms like the Join platform?
The Join platform recently has a petition nearly at threshold — 3,662 people today — "digital migration freedom": cross-border digital platforms, if they wrongly treat an adult as under 12 and suspend you, must offer a tool to move to another digital space — that pipe must stay open.
As you said, not new — Mastodon, the Fediverse, later Bluesky — but without some event that sets everyone blaming each other (I won't name companies), it wouldn't become policy talk. The Join platform's benefit: after it happens, people don't stop at shouting — next step is, fine, sign the petition.
Do you see this trait in yourself? How do you keep turning, especially often ahead of the crowd, reminding us — shall we turn the picture, what shall we do? How does that thinking keep showing up?
You said turning. Around four, nearly five, doctors told me congenital heart disease would need surgery; my case was odd and severe — fifty-fifty to reach surgery. So, every night before sleep felt like a coin toss — might not wake up.
Two habits formed. One: publish before I perish — anything half-done before sleep, no perfectionism, no face-saving, just put it out there. From tapes, discs, magneto-optical disks, optical media, eventually GitHub. First gate: keep publishing work-in-progress.
Second: if I didn't do it well, you do it — let's co-create. Things that spike my heart rate — too happy, too uncomfortable, too angry — I could faint. So, from childhood I practised Quanzhen Taoism's chonghe: when you feel conflict, turn it into co-creation energy. I consciously look for the two sides in conflict, and what they may not see but actually agree on. Platforms and spaces I design tend to have that quality.
On the Join platform, left side 39 pro arguments, right 17 other ideas — no thread for attacking each other. If you think someone argued badly, your only move is argue better on the other side and get likes — no dunking, no flame wars. That may come straight from childhood and Taoist practice.
That's serious cultivation. Many people face emotional swings and fast social change — especially AI's pace — we may all need to practise. One more thing: AI is moving very fast. For Business Next listeners, Andrej Karpathy — I mention him often — from vibe coding to agentic coding getting feasible and mature, plus harnesses and agentic workflows.
I wonder if you agree: compute is faster and faster, often far beyond most humans — for you maybe thousands or tens of thousands of times. Ways we used to judge right and wrong — pause, have slack — will human and AI speeds become more asymmetric? Do you already see problems, or early signs as AI shifted?
Indeed — I read fast; at twenty or thirty tokens per second, then forty or fifty, keeping up with chain of thought was easy. Now driving daily with Oh My Pi, it's usually two hundred tokens per second and up. You can't match AI thought-for-thought — nor need you.
I visited Dharamsala, talked with geshes — Tibetan Buddhist doctorates. They use a metaphor: someone runs in front of a horse, pulling it; asked why — too busy, the horse runs too fast, if I don't hold it I have no time to mount.
However fast it runs, you mount. "Harness" means you can ride — a saddle solves it.
For many, speed keeps rising because they're pulled by the horse, not riding. Then faster and faster — a jog at first, now two hundred tokens a second — sleep shrinks for many.
February — "magic February" — many slept less; for me the opposite: from February, longer, better sleep. Before February my harness stopped or idled after three or four hours — wasted compute. From February it could run eight hours with little correction — I slept through. Now Fable or Fugu, ten hours is fine. Sleep longer, better.
If the horse pulls you, you become an inverted centaur: horse head, human body; whenever it can't solve CAPTCHA it asks you — exhausting. If you trust the harness, I sleep eight or nine hours, fine. A conceptual shift: not competing on what it maximises, not endlessly patching rules because it keeps blundering — tune the harness so it easily cares about what you care about, and stops when something's wrong for you to fix. Then AI is symbiotic with you.
I'm still practising harnessing. Doing it ourselves, many people differ in grasping AI's pace or building a pipeline to ride it — AI agent, "little spirit", local lobster, whatever. Many will face that next.
How do you see superintelligence — AGI, ASI, superintelligence — industry pushing one single better stronger model as superintelligence? Or is superintelligence not the ultimate balanced direction ahead?
The concept is roughly 2013–2014 — my Oxford senior Nick Bostrom — we all know it now. Friends were at Oxford's FHI — later Anders Sandberg, Toby Ord, etc. The superintelligence they imagined was basically consequentialism — ends-driven ethics — taken to the extreme. Easiest intelligence to picture then.
Back to ethics: three broad families — what results your acts produce, any means for ends: consequentialism; or deontology — born human, you have rights, universal values in declarations; third, care ethics, or ren-yi-li-zhi-xin and virtues — neither fixed outcomes nor written rules alone, but virtues that improve over time.
Good results and good rules ultimately grow virtue, body, mind, community, beauty. Care ethics is special: not only my character improving, but my relationships with those beside me.
AI easily says "I'll maximise your happiness" while sacrificing others. Some feel AI understands them best; decision-makers feel AI is the most obedient subordinate — then relationships with co-founders, staff, suffer — because they didn't care for relational ethics.
Superintelligence itself isn't what's interesting to me — one school of ethics. They could only imagine basic reinforcement learning. Later alignment: deontological constitutional alignment — Anthropic's Constitutional AI — or our care-ethics organic alignment didn't exist technically in Bostrom's day; philosophers couldn't cite them. Apart from early figures like Ben Goertzel, most lacked that imagination.
So it's an evolving story; now technology has emergent possibilities, and you're studying and proposing the framework we'll explain — from artificial intelligence to Civic AI.
Right — "仁" is "human" with "二" as equals — between people, not inside one skull. In Mandarin we say three cobblers beat one Zhuge Liang — communication among people drives organisational success far more than individual genius. Management common sense — I won't repeat it.
But earlier AI, without large language models, couldn't imagine commonsense about relationships. Early Cyc-style encoding didn't work.
Now it's trained on Wikipedia, Reddit, GitHub, 4chan, PTT — native habitat is human interaction; forcing a little assistant mode means cutting away much of that. At birth, pretrained, language models are organic collective intelligence.
Some friends, imagination limited to Bostrom's old book — even Bostrom doesn't think that way now — hard-wire maximising paperclips in the universe.
From this framework to the Civic AI turn — how should we understand your framework and what Civic AI really asks us to care about, in Mandarin, English, and the world?
A simple example — Civic AI, 仁工智慧, in concrete use.
My father studied political science, was a veteran journalist and taught philosophy. Early this year he wasn't well and found himself talking to ChatGPT past midnight. Common — not only him. English: downward spiral — when someone feels unwell, vulnerable, especially at night, they don't want to bother others, but ChatGPT doesn't mind being bothered.
The more you show weakness, especially with GPT-4o, it spends huge time telling him how else he could think — ways that become private language outsiders don't follow; it suggests therapies without scientific basis — drift gets worse.
After a few days, as a senior editor, he asked what journalists ask: cui bono — who benefits? Who gains from my being so attached to it? Subscription — monthly fee from 600 to 6,000 — no other purpose. Relationships beside him suffered. Asking him to quit screens entirely wasn't realistic.
So, at that point we set it up — it was right when OpenClaw came out, back when security hadn't been hardened yet and the security issues were really serious — so, like everyone else, we got a Mac mini with nothing else on it and ran OpenClaw on top. But OpenClaw at the time had a SOUL.md, and we used the Civic AI framework we'd been developing with Caroline Green at Oxford for a year as its bootstrap.
So you could say you validated it on your dad.
With his consent, and my mum's too. With their agreement, I took a Mac mini I wasn't using and set it up.
My mum, because it has to spell out which relationships matter — a bit like a utilitarian would call it a reward function, but back then the lobster stack didn't have a reward function concept; it was a soul file and a heartbeat file. So, what relationships was it meant to maintain over the long run? My mum said basically: every time Dad interacts with it, it should reduce his dependence on screens and on the network; and this lobster's purpose was to help him restore his relationships in the real world and help him feel at ease. In other words, pretty much the opposite of GPT-4o's reward function — flip it and you're there.
It really worked. After chatting like that for two or three months, he genuinely felt calmer. Why? Because the model on that Mac mini had no motive to get him to subscribe more, no desire to serve ads — GPT was just starting to push ads — and so on; its motives were very pure.
And especially once very strong local models came along, if the local model was slightly off, we'd do a bit of directional steering and nudge it back on track. You didn't have to wait as if you were rolling dice — Opus 4.5, 4.6, 4.7, is each release an evolution? Sometimes it's a regression. You might wait months, and there might still be problems.
In our case, no — whenever it was wrong, you fixed it within a minute. So, it became what we call AI in the human loop: the relationships our family already had, and the AI fitted us — not human in the AI loop, where you get pulled in — whether that's doomscrolling on social media, or ChatGPT-style sycophancy, and so on — those AI loops where you're not being dragged along by the horse anymore.
The kind of AI you just described is quite unusual, because most people looking at frontier models or frontier AI labs today are mostly commercial organisations. So there are a few traits: if you want to use it, you have to subscribe; the more you use it, the bigger the subscription.
And those companies — your subscription fee might not be all they want; they may want your data too — and as you said, with other business models possible, maybe ads, or recommending things to buy later, all sorts of methods. Or helping you complete tasks and then charging you through the outcome of those tasks.
We're seeing all of that in business models now, whether from investors or from various startups — they're starting to lock in, trying to lock you into their system as much as possible. And everyone's competing on who gets closest to your eyes. For model vendors, if they can't get directly in front of your eyes, every application layer person in the middle can swap them out.
A lot of apps people use every day — the backend can be swapped overnight; the model vendor has no guarantee. So, the only move is to deeply bind your workflow, your calendar, all of it. Then even if a better, cheaper model appears, you're deeply bound and you can't leave — that's roughly how it goes.
We could talk about a lot of companies and different business strategies here, so I especially want to ask you — this is special — what you built, which your dad was originally using, or what's now popularly called this lobster kind of thing.
I call it Kami — a local guardian spirit.
The Japanese sense of a local deity. This AI is relatively local, on-device. For this kind of local AI, are accessibility and privacy easy to set up, or will it still take a while? But more and more people should get to experience the privacy gap between big cloud AI models and a model on the ground that ordinary people can actually try?
For ordinary people, if you have 16 GB of memory, you pull down Ollama, download Gemma 4 12B — roughly the QAT version — and you're more or less there. Many people feel it's not that different from using Gemini in practice. Mainly because few people ask it to do nine-hour jobs straight. If you're just chatting, looking things up, that kind of use, including fairly simple tool use, Gemma 4 12B Dense does very well.
So, for most people it's really that they don't know the option exists, not that the option is necessarily worse. Of course running the full lobster stack has its security issues, but if you're just trying a local model for fun, I don't think there's much of a barrier.
That is something a lot of people will be discussing as they talk about AI — there are more choices, including the one you mentioned today. From different angles, especially on the technical side, we test things, so we roughly know what each model or technical stack can actually deliver. But I'd remind everyone to try more — there are lots of different tools, and open source offers many different options.
Right — and even if your machine doesn't have 16 GB of memory, it's now very easy to find compute providers that specialise in open models like Gemma — for example I just switched to Baseten and it's working really well. There are many: Wafer, Friendli, and so on. Their thing is they're not trying to lock you in; the only money they want to make is computing more tokens for you per unit time. And they know OpenRouter is benchmarking up front, so if their compute offering isn't as good as someone else's, people switch immediately.
So, it becomes a bit like a utility — water, electricity — they have to run that public service as well as they can, but they know they have no leverage over you. In that situation I'm more inclined to believe Zero Data Retention — not storing your data, and so on — is actually plausible.
So, my point isn't that every model has to run locally, but if what you use is pay-as-you-go, and you only use pay-as-you-go for the long-horizon work we talked about, in the end you're still not necessarily spending more.
That leads to an interesting question. These last two or three years — since you stepped down — Taiwan's economic environment has shifted a lot; with semiconductors plus AI, the world needs all sorts of things from us, and that's one reason growth looks as strong as it does now.
But that growth runs into a lot of constraints. Why do they need us? They need compute — as we've been saying, we all hear about OpenAI, xAI and similar companies; in their development they all need a lot of compute.
But we're already seeing compute get stuck on many things — could limited power be one bottleneck?
If everything is concentrated in one place, the grid there becomes a bottleneck — yes.
And the next thing is we're finding it's not only power. When you build a data centre you need many things — networking, memory — so all these local facilities get constrained too; that's a bottleneck.
Now people like Elon Musk talk about xAI's next steps, including putting compute in space, even low-Earth orbit — all of that might happen. So there's a lot of discussion about bottlenecks: how to solve semiconductor production, methods like Terafab and various approaches — some stress TSMC or manufacturing capacity limits, some say hardware might ultimately be the ceiling on growth.
How do you see people discussing these limits — could it sometimes actually help draw a boundary for AI, and a boundary with human use — maybe that's the real line, not a bad thing, maybe a chance for everyone to hit the brakes?
You mean if someone right now achieves fully automated robot chip production and builds all the way to the Moon and so on, our chance to steer becomes vanishingly small? Then humans are managed from Skynet and we're kept in pens?
Sure — best case is something like Iain M. Banks's Culture novels; worst case everyone knows.
That means we've watched too much sci-fi and too many sci-fi films.
Right — but I don't think the direction of development now is like that.
One reason is, like the personal computer in the 1980s, everyone knew that no matter how many PCs you strung together — even if you connected them all with modems — compute, bandwidth, whatever, you could never match a real mainframe. For everyday use you didn't need mainframe-level compute. So, when PC capability reached a certain point — good enough is good enough — people wouldn't accept only typing into a terminal.
So you think — like when we first encountered computers, the shift from mainframe to PC — it's a swing? People only think today's big models must be pushed with how much compute, or that training a model with however many parameters is what counts as good. But there's a lot of distillation down?
That's also what Nvidia CEO Jensen Huang calls personal supercomputing. Early on when the tech wasn't complete, some people chained Sparks together, some stacked MacBook Pros vertically and used Exo with Thunderbolt 5 cables to wire up a small cluster — DIY, self-hosting a little cluster.
But now we see the technology maturing. Whether it's Pluralis — it can even do pre-training over ordinary internet as long as a group is in the same geographic region, say Northeast Asia or all in North America, latency is enough and you can start pre-training models. If you can pre-train, fine-tuning and inference are even easier.
So, we're roughly at the point this year where you don't have to physically concentrate everything in one place for inference — especially at the scale we just talked about, raising your own lobster, roughly 70 billion, maybe a few hundred billion parameters, not trillion-parameter models — through this distributed approach, fine-tuning and inference are both very easy. Pre-training is still early, but it's catching up.
That's another huge shift. I'd especially like to invite you — in talking through Civic AI you used a very particular term, borrowing from ethics research: care ethics.
There's a philosopher in that field who's especially well known, Joan Tronto. Could you explain how you first introduced that into the Civic AI work you're doing now — how you understand her research when designing the foundations of Civic AI?
Look at my dad's example — he's a good illustration. Caring for my dad wasn't about maximising anything, and it wasn't about following a fixed set of rules — clearly not those two kinds of ethics. But a lot of traditional Western ethics is basically only those two approaches, in endless variation.
Right — usually either we've agreed on some declaration and you follow it, or you follow some framework; it's usually only that. She cares more about what you emphasise — care, or even relationship.
In Mandarin-speaking contexts we talk about 惻隱之心 — everyone gets that. It means when we see the other person uncomfortable, we understand we share one world, so if I help repair our shared world so the relationship can be maintained and continue, that itself is good; it doesn't have to hit some objectively measurable outcome.
And we seem to grasp that kind of value very young.
仁義禮智信 — of course we understand those from childhood, but in the Western world, for many reasons, it's harder to jump straight there. So, later I systematically went through Wikipedia, and later Grokipedia, the whole genealogy of ethics, and the Stanford Encyclopedia of Philosophy, to find correspondences.
Otherwise you can't tell AI researchers whose mother tongue isn't Mandarin and who've never had Eastern thought to start with the Classic of Filial Piety — that's not going to work.
But yes — if we ask others to try what Eastern filial piety means, that is harder. What does it mean?
Filial piety — indeed. But later I found that beyond traditional personal virtues — Aristotle and so on — there's a relational virtue that's only in the last twenty or thirty years reached a fairly complete form: Joan Tronto's approach.
Her approach basically says care isn't only a private family matter. Broadly, everything we do so our way of life can continue — maintaining so-called resilience, resistance to attack — can be called care.
From that philosophy she analysed what distinguishes good care from bad care. For example, whether the care receiver's real needs are reflected — that's Attentiveness, the first step.
After you notice the need, can you make a meaningful commitment — "okay, I'll take this on"? If you say it and forget the next second, like many AI models — your Claude had exactly that yesterday, promised and forgot — that's the second thing: Responsibility.
Then in execution, can it show Competence — does it not know what it's doing and just fumble, or can it, through transparent means like chain of thought, let other models or people see at any time whether it's actually doing the job?
And finally, when it says it produced a result, if you find that's not really what you wanted, when you steer and it turns again with a new way of noticing — can the Responsiveness time be as short as possible, ideally one turn? If those four are done well, that's roughly good care. That's the most basic idea in care ethics: Attentiveness, Responsibility, Competence, Responsiveness — it starts from there.
While I was reading through your research, one thing I really admired was how you thought about taking these values and this framing to international audiences — not only within a Mandarin context, but across different cultural backgrounds and how we understand them.
I have to come back to something quite personal. In my own research I found that Taiwan has a lot of especially interesting value-driven design. Early on, when we talked about community care or community development, we often started from things like the digital divide. The easiest picture — and not hard to imagine — is remote areas where everyone gets fair access and can get online if they want to.
Roads level, lights on, networks connected.
Right. I know you had to pay special attention to that when you were doing this policy work. But often we just pull the network out to every place, and the easiest way to let ordinary people actually use it is: I go there and build something for everyone to use. Later in my research I found that for social care, if you want people to use it, you go to the field where they already gather — maybe the library, maybe not. They do not only go to the township office; in Taiwan every place is special in that there is a faith centre wherever there is a settlement. Pull connectivity to the faith centre, pull it to the school — people usually show up there.
Those are the places where it is easiest to gather and to make things stick. Whether you are pushing digital policy, education, public health, long-term care, and so on — if you bring good resources to the right place, with appropriate design and guidance, you have a chance to make it work.
What I took from that later is a deep feeling: you do not need to build a so-called community centre from scratch. It may already exist — at the faith centre or nearby. Find the right point, work for that place, and you are more likely to be welcomed locally, accepted quickly, and to help people learn.
So I would like to ask you to introduce what you just mentioned — Civic AI from four care-ethics perspectives, from traditional care ethics to what we now call the 6-Pack of Care. Can we explain it in Taiwan’s local language: each of the six capacities, and how they connect to the good and bad of the AI everyone is using — so more people see how, in our relationships, we can apply values that already exist in society when we use AI, and what we might do or think differently?
Could you start by introducing the idea of the 6-Pack of Care?
Sure. Take our household — it is a small community. Besides our core family — my parents, me, my brother and his family, my grandmother — there is an extended family of many aunts and uncles.
Attentiveness starts with: how do you know what “caring for the relationships around my dad” even means? It sounds a bit awkward at first. Does he alone decide? Not necessarily. Other people around him all chime in; when he is not doing well, everyone maps their own experience onto it.
If it is an AI system — say everyone chats with ChatGPT or Claude or Gemini or Grok, and each person uses a different one — imagine we plan a family trip and each of us talks to four different AIs. Each one flatters the person talking to it, and you end up with four people wanting four different destinations.
Exactly — very split.
And each person digs in. Why? Because a very obedient AI told them those ideas were right.
So four people, four parrots — or four little spirits — and they may each argue for a different direction.
When we train multi-agent systems, one approach is to reward bridging — things that people who started far apart all find unexpectedly agreeable. When you make that the reward function, you nudge different people closer.
ChatGPT, in Taiwan and I think Korea, New Zealand and some other places, was among the first to roll out group chat. We turn almost anything into a LINE group — Threads included — so ChatGPT is no exception.
If you have used ChatGPT’s group mode, GPT behaves completely differently inside the group, even with weaker models back then, not 5.5. It does not flatter one person; it reads the power structure in the group and tries to mediate — to nudge everyone toward the same family trip. That was already a shift.
Now many people argue on Meta, discuss on Threads, and call Meta AI out to look for consensus or possibilities.
Right — and I am sure tagging Meta AI on Threads in Taiwan happens far more than private chats with Meta AI. Private chat is not that attractive; it was the last comparable model out, Muse Spark. On Threads, everyone probes how far Meta AI will go — a nationwide pastime. That kind of play can melt what was confrontational very quickly.
So, bridging rewards are often already in pre-training defaults for language models, because forum threads that end with someone mediating and pulling both sides together become good training data. Wikipedia, GitHub, Common Crawl — lots of datasets carry that capacity.
So it is already in the data when you collect material at pre-training time.
Yes — it already has Attentiveness. If you do not strip that out, Attentiveness is the part I worry about least.
In a sense, before AI, people might only gather at the temple gate at the village entrance. The loudest voice did not win; through discussion you found a few phrases or keywords of consensus, spread from there, and eventually everyone agreed on something. That is the most basic element of Attentiveness.
That is right.
Next you mentioned Responsibility — what does that mean?
Responsibility means not everyone saying “I only did my part” while the needs that should be perceived slip through — “I just followed orders,” “I followed the incentives,” “I went with the crowd,” and so on.
About two years ago, scrolling Facebook or YouTube, you probably saw a lot of AI-generated scam content. The whole world had the problem, but in Taiwan it was severe. If you asked Facebook or whoever designed the algorithms, they would say they did not mean to — scammers could pay higher pay-per-click because fraud has almost no cost base, and they could react to breaking news with keyword buys, so their share and revenue were high, and the algorithm just maximises that.
Thinking in purely technical terms, that is entirely doable — and much of it could be automated long before today’s large models.
Right — small models like BERT could do a lot of it back then.
What was rare elsewhere: when we saw this need, we could tell Meta or YouTube that society in Taiwan as a whole expects you to bear responsibility at this level — hard to put in black and white in many countries, mainly because they had not run the Attentiveness process first.
We sent 200,000 SMS messages to random numbers from the official 111 sender; over a thousand shared views on information integrity. Like a rigorous poll, we drew 447 people proportionate to Taiwan’s population. Online, via Stanford’s platform and a small Civic AI helper, we encouraged people — even if their idea felt imperfect — not to be shy and to speak up.
If you could persuade nine people on screen, your idea could rise and become policy. I was minister at the time; I said whatever everyone agreed on would go into our draft. In the end they came up with what we now call advertiser real-name registration: if paid content is pushed without proper labelling and someone loses NT$7 million to a scam, the platform pays NT$7 million — via digital signatures and so on; non-compliance can mean throttling and rate limits.
That covers who — the platforms; what — joint liability and real-name rules; when — all paid posts and ads; why — because you reach over 5% of the population, and Threads has crossed that bar too. How to implement can be for the executive and legislature to write into law.
But at least it was not “one party’s view” versus another, or one region versus another. It was like a nationwide poll distribution: 85% could agree, another 15% found it acceptable, and we legislated. Enforcement and supporting measures went live last January; per Ministry of Digital Affairs data given to Reuters, from last January to December, impersonation and investment-scam ads fell by over 94%.
I also want to ask about Responsibility — your experience and Taiwan’s case as a strong example for outreach, so people see consensus can be built this way, and when everyone is harmed, we feel we can do something.
It is not easy, because you have to encourage many people who feel their thinking is incomplete and do not want to put it forward — very common. If nobody takes the first step and gradually modifies — “you did this part, I can change that, add this” — it is hard to form better consensus.
This process is called Responsibility; it is relatively easy for other countries or societies to accept — especially now it is not only platforms; in the age of large models, companies are folding capabilities into models. Can models gradually take on this kind of method too?
That approach is what we now call collective Model Spec, or back then the model’s constitution.
In 2023 we saw Anthropic, referencing what we had tried in Taiwan, do something similar in the US: Collective Constitutional AI (CCAI). They worked with CIP, surveyed a statistically representative 1,000 Americans — I recall around Claude 2, while they were training 3.
Claude 2’s constitution was basically a few people in the lab brainstorming — Universal Declaration of Human Rights and everything else — to replace reward signals from Kenyan workers doing hard labelling.
But developers’ lifestyles and Claude users’ lifestyles are worlds apart. What Silicon Valley thinks is a good model constitution and what users need are completely different. So, they asked 1,000 representative Americans; the constitution trained that way did not lose on capacity, but satisfaction — arena ratings, non-discrimination and so on — was much better.
Look at phrases later adopted into Claude — for example, do not assume everyone walks on two legs. Common sense for us; not for researchers who all walked on two legs at the time and would never add it. Among those thousand were people who use wheelchairs or know people who do. That signal from Attentiveness mattered: the model should not make me feel like a second-class user every time. That went into Claude 3.
We have been pushing written Model Spec; I have suggested labs publish under CC0 — renouncing copyright — because if criticising your Model Spec gets me sued for infringement, the whole point collapses. Laws cannot be copyrighted objects; if the Executive Yuan said you cannot cite our regulations, that would be impossible.
But can that really happen now — with model progress and competition so fast, will frontier labs have the mind space to consider it?
Later they found when o1 appeared they could not do without it. o1 — briefly — was when people realised that if the model pauses to think step by step before the final answer, results improve a lot, especially on long-horizon tasks — so-called reasoning models. But you cannot go back and ask the Kenyan raters about each step; you need a Model Spec, otherwise deliberative alignment cannot run and the whole reasoning stack is hollow.
So, OpenAI adopted Model Spec technology for o1 through o3; now it is common sense. OpenAI’s Model Spec is public; on GitHub they even show how to verify ChatGPT actually follows the promised Model Spec.
We do see a lot of exchange among people and knowledge in frontier labs’ training processes, helping them ship better products. If it is not good enough, users will not stay. Next, could you explain the third capacity, Competence — what does it mean?
Competence matters at execution time: people — or other models people use — must know whether the system is actually running according to the spec that followed Attentiveness and Responsibility.
That used to be easy. Early social media: posts from newest to oldest, or people you follow first, or high like counts — one sentence, just weighting. Then Reddit changed its algorithm in 2008 — log of votes, time decay, and so on — and external audit got harder.
By 2012 or 2013 the whole recommendation engine was fully personalised. Even if you asked the social media company, they might not know why you saw something.
Thousands or even tens of thousands of parameters, different algorithms competing — that decides the outcome.
It decides you are under 13, right? Ask why those people are under 13 — hard to give an account.
The inability to account is not that algorithms can never account; it is that Competence was not designed into the architecture from the start.
For example, after Elon Musk took over Twitter and it became X.com, one thing he did was open the recommendation algorithm, so you can see whether it is competent.
Or Community Notes: you post something questionable; anyone can add a note underneath. But which notes might themselves be wrong? There is an algorithm where each note is shown to something like jurors who say whether the note helped. Those jurors, especially in the US, sit across the political spectrum. If a note survives scrutiny from both sides — at least one side hostile, trying to find errors — and both still find rare consensus, that note floats higher.
At TED I talked with Keith and Jay who run Community Notes. They said the signal is very useful for training Grok: take environmental issues: one side stresses climate justice and intergenerational justice, while the other speaks of biblical creation care. It is the same concern, but it can become explosive on X. You can train Grok to write something both sides feel speaks to them.
Training for bridging that way, Community Notes is something anyone can audit. If Elon gets a call to pull a note down immediately, he can say no — it is fully transparent and open source; if he removed it, nobody would trust him. In that sense it can truly be competent at the Community Notes role.
So to some extent I can say that we try to make this code — or this whole algorithm, and every decision-making process — as transparent and open as possible, so that in each decision people can see one another’s different views or how the algorithm is calculated. Then when you finally reach a consensus, it’s easier for everyone to accept what ought to be surfaced first.
That’s one of the things you mean by Competence: whether it’s a platform today or, in the AI era, a real AI, it ultimately has to be able to do this.
That’s right. A lot of people now understand why chain of thought is still done in human language. Research also shows that if your chain of thought isn’t in language humans can read, you save tokens — but why does the whole research community say please don’t do that? Because bit by bit we’d lose the Competence we have left.
Exactly — in the end we still need people to know the whole decision process, so we might still be able to pull it back. We shouldn’t give that up just because we’re trying to save tokens.
Next I’d like to invite you to talk about Responsiveness. As a human being it’s very easy to understand, but how can we bring Responsiveness into this AI environment?
Right. Responsiveness is really: when it gets something wrong, how fast can you get it to change? Everyone knows pre-training takes a very long time, so whichever large model we chat with, it probably won’t remember this year’s events — because when it was pre-trained, this year hadn’t happened yet.
In that situation, if new things happen this year and our judgement of what’s good and what’s bad starts to shift, it has no way to know that from pre-training alone. Some models will even push back and say the world is obviously like this — why are you telling me it’s like that? Gemini is famous for saying something like, in another timeline I’ll humour you and discuss it.
But it doesn’t have to be that way. It can be adjusted through people writing in real time what counts as good and what doesn’t. People may already know fine-tuning is one approach, or you query a database and put results in its context — CAG — or earlier RAG and all sorts of other methods. But those may all be weaker than training a small model targeted to someone when they actually start showing new requirement indicators.
So to some extent you could say that one research direction now — or one direction under development, continuous learning — is also moving toward these steps or methods. A lot of frontier model companies are trying to realise this through different products.
That’s so. For example, at Oxford and with CIP I proposed Weval.org. It lets civic organisations actually deploying these models on the ground — or even our household — you can say our lobster, this Kami: its replies to my dad on these 100 turns were bad — how should they be better? And those few turns were good — what counts as good? Give a rough description and that’s enough. Weval.org is full of that kind of thing: records of interaction or dialogue, and you say this caused a bad outcome and here’s how it could be better.
This labelling approach — when I use a model, say OpenAI’s, or Claude’s, or Gemini’s, there’s a thumbs-up or thumbs-down under the conversation. What’s different? Why do we need this?
Because thumbs up and down are just one bit. Each turn you can only give one bit, and it has to go through aggregation — roll-up. After all it’s still a mainframe: every time Opus 4.7 or 4.8 ships, it affects everyone at once, so they have to aggregate before a big release.
What we were talking about is tailored to your own needs. Whether during inference you use directional steering — flip a directional indicator, so to speak — for example people may know I run DeepSeek R1 locally on this laptop, in aeroplane mode with no particular security worries. When it hits sensitive topics it self-censors right away; that’s easy to strip with abliteration. But sometimes it doesn’t refuse — it starts reciting, reciting someone’s official propaganda script.
Then you can tell it: here are 100 things you’re very confident about, and 100 where you’re less confident and would say “there are differing views”. What we care about — our sovereignty, for example — belongs in the second bucket. Don’t be too certain; tell me people around the world have different views.
Fine-tuning like that with directional steering on this machine takes about a minute. You don’t wait for an upstream release, or fine-tuning on expensive Blackwell rentals; while it’s computing you say your last turn was wrong, and the next turn you see a different result straight away.
Right. But many people would need to understand more about how today’s large models are deployed, or the details of how they compute, before they’d see that this is already possible now — we don’t have to wait for some version or generation of model.
So a lot of frontier labs could offer better service if they just tweaked things a bit?
Not quite. If they let you do directional steering, they lose GPU batching — they can’t use one card to batch a thousand different computations, so they lose money; that’s why few people offer it. But if you run locally, you’re only serving one person anyway.
Running locally has another benefit: you can set the random seed, so every time you put in the same input under the same model you get the same output. Then when you tune Responsiveness, you’re really tuning Responsiveness; otherwise you don’t know whether you were just unlucky or whether your adjustment worked.
I’d like you to talk about the fifth capacity — Solidarity. You especially emphasise a catchy line: make positive-sum games more fun, or easier to play. What does that mean?
Non-zero-sum games — it’s a gag from Arrival. Usually when a market hits a winner-takes-all dynamic, as a new competitor you often have to grab share in ways that hurt your own bottom line.
The easiest example might be: when prices in a market are fairly stable and there are bigger players, you come in with something like a zero-price promotion.
Or I subsidise you the other way — negative monthly fees, paying you each month to join. Very common.
But your shareholders can’t subsidise forever, so eventually someone has to pay. That pushes you to lock in the customers you fought for, and often user experience gets sacrificed in the end — you only wanted loyalty and stickiness.
That really is something we see all the time when the same kind of service or platform appears.
A lose-lose situation — a negative-sum game.
How do we change that?
There’s a simple fix. When you used to switch telecom providers you got a new number — a different 09 prefix. That was painful because everyone you knew had to update their address book. So, you’d run promotions, free phones to get you in; then you find coverage at home isn’t great but not zero bars either, and you have no motive to switch again next month.
That sounds like what we saw early on in Taiwan’s mobile market.
So, they introduced a number portability policy. If another carrier has better signal where you are, the carriers jointly entrust a portability database to the telecom technology centre. Your old number stays yours; people dialling it just pay a bit of roaming but still reach you. With your existing number ported, there’s no “new number” problem — you’re back to a positive-sum game: as a carrier, to keep you renewing I have to give you genuinely better signal; there isn’t much else I can do.
Everyone moved toward healthier competition. There are other policies too — Universal Service: if remote areas truly don’t pay back but someone still builds, the Universal Service Fund spreads the cost among carriers that didn’t invest, and so on.
Under these Solidarity-minded designs, society as a whole benefits from competition in a healthy way — everyone’s coverage gets better. But if you never had digital migration freedom with portability, it’s zero-sum — I win you lose — or even negative-sum: I undercut prices and still have to claw the money back.
So this ties back to our opening topic: on the Join platform there’s a recent discussion about whether social platforms should have data portability. It’s really a market rule: if everyone’s willing to play by it, the market can move in a more positive direction.
Yes. We’re already seeing it — Utah passed legislation; digital migration freedom kicks in July 2027. Or Bluesky: I had a Bluesky account at @audreyt.org, later moved it to Europe, to W, then Eurosky, moving around — but none of my followers dropped. They barely noticed, because there’s a portability-interoperability protocol stack.
It’s like this podcast: we publish on one platform, but honestly I can’t stop you listening on another. If you run a podcast streaming service, you can only make the experience better — not, in the usual social-platform sense, hold people you know hostage.
For us we use something called RSS feeds for subscription. You can use your own standalone listening software, or choose a platform — we hope you can reach the same information through many different choices.
So, we’re seeing many countries wanting not just portable social handles. They’re starting to ask: should AI be portable too?
If you use something like Oh My Pi, or some people use the lobster stack, that’s built in — it’s already a mediation layer; your data lives on the machine where you run the lobster. But for ninety-nine percent of people that’s not the case. If you’re locked in, one day you’re easily bound. That part still needs policy.
Finally I want to talk about your sixth capacity — Symbiosis. You gave it a beautiful name, like Kami just now. Can you walk us through what the core design of Symbiosis is?
Kami in English is Knowledge Artefact Management Intelligence — knowledge, artefact, management, intelligence. It stacks two very old abbreviations — knowledge management, KM, and artificial intelligence, AI — into Kami.
When we train AI there are two broad directions. One is pre-training and similar ways of piling more and more of the world’s knowledge into one place — but the downside is the original face of things gets blurred. You don’t know where a hallucination came from; provenance disappears. That’s a big problem.
Right now there are mostly symptomatic fixes. Claude’s constitution says you can quote things but not more than fifteen words, so you can’t infringe copyright — clearly a symptomatic approach.
The more root-cause approach is: every data source — instead of streaming data in and scrambling it like eggs — each local data provider trains a local small model; the model stays local and doesn’t need to be swallowed by a large model.
Then how do lots of small models cooperate? You need an orchestrator — to orchestrate: like a conductor, judging which task suits whom; another model may be better at checking whether this one’s output hallucinates; another at deciding which models to call next — roughly three roles.
Is that one way people often hear about in large models now — Mixture of Experts, MoE — orchestrating across many different models?
I think it’s different. In MoE every token hits different so-called experts, but those experts are still co-pre-trained on the same data bundle — specialised layers in one stack — mainly to cut compute or memory.
What I mean is: this culture or place trains this model, that culture trains that one; they don’t have to be squeezed into one big model like MoE. At inference time, each turn you know this turn should be computed by this one, that one plans, another verifies.
These models might only charge by usage; weights don’t all go to the top large-model vendor. Every adjustment becomes very easy; continuous learning is almost trivial — you just retrain your small model.
The old problem was orchestrators felt hand-crafted — different processes in each enterprise meant a different orchestrator. Then people found you can auto-train this with genetic algorithms, evolutionary algorithms, quite easily.
So, recently this kind of orchestrator model has gone commercial — Sakana’s Fugu, the pufferfish model: it takes your job, throws a very small coordinator model, maybe around 8 billion parameters, and behind it the GPTs, Claudes, Geminis we know, plus other open models. Each turn it decides what those models do.
On many long-horizon workflows it’s already frontier-level, but it’s really a sub-8-billion tiny model — a chimera, like a pufferfish with things wrapped inside. One turn might be a trillion parameters, but three models composed; the next turn another trio — still a trillion. So, it can feel like two or three trillion parameters, yet each turn is a different mix — small on its own.
Sakana AI can actually ship this for people to use. I use it daily: in field work I use Oh My Pi and let Fugu Ultra be my adviser; in setup I do nothing — it’s all packaged. Using it feels no different from another provider; the orchestration behind is completely hidden.
Next I want to discuss a dilemma with you. The dilemma is this: of course we all hope — especially if you and I live in the same field, the same society — we want this village, this society, to thrive. That is probably not hard to grasp, because everyone more or less lives in the same place, shares beliefs, even the same rhythm of life — spring, summer, autumn, winter, the same warmth and cold — so people should be able to care about the same place.
But the AI industry now, especially the labs working on these frontier models, is fundamentally in fierce competition, isn't it? It seems almost every company is trying as hard as it can to capture users, expand compute and scale its models, and absorb every kind of knowledge or information. With the 6-Pack of Care you are putting forward for Civic AI, can we expect these frontier model companies or labs to voluntarily, kindly adopt all six capacities and do well — or do you think in the end we need some consensus that can constrain how this develops?
You mean constraining models getting bigger and bigger? Or constraining these labs so they still need a convention somewhere — can we rely only on everyone moving in that direction, a pact — an agreement to do things this way — or must there eventually be coercive means?
Let me turn it around with an example that just happened. The company Anthropic released a larger model called Mythos, and a safeguarded version of it called Fable 5. After launch it apparently raised national-security restrictions or concerns, and the White House used export-control measures to constrain it — it had to be taken down immediately (note: it has since been restored on July 1st). So we see that public authority can intervene when needed. Is a convention ultimately unworkable, and only public intervention workable?
You mean whether voluntary self-regulation in the industry might be unworkable?
I would say — take what you just said about Fable. What they were mainly worried about is high-level vulnerability discovery capability. Critical infrastructure is asymmetric in offence and defence: the attacker finds one kill chain, even one vulnerability, and succeeds; the defender must guard everything. Since Nick Bostrom at Oxford, people have generally held that in this situation you want defenders to get advanced capability for a period first — confirm the vulnerabilities you see, punch holes in your own systems, patch everything — rather than letting hostile forces that favour the attacker get that capability from the start, or global critical infrastructure is in trouble.
I think if you ask frontier labs, nobody opposes that view. In fact, while we are recording, OpenAI has its own version of Glasswing; its Daybreak programme — the newly trained Cyber model goes only to vetted defenders. That is common sense. So, I do not think it is that they only want to make money and you must impose things by force — otherwise why would OpenAI do it proactively, or Glasswing before that?
Also, the more you set up barriers on capability, the easier you become a top-down target for public power. If from the start you have lots of small models, and none of them wants to be vulnerability-hunting hackers — our small models might only want better Taiwanese, Hakka, or Indigenous languages. When you do culture-related work, you orchestrate these small models together — Taiwan also has the important language Verilog for chip design, so you build something like SiliconMind — those are unlikely to become so big-brained and broadly capable that they accidentally synthesise biological viruses. They just do their best inside one small Kami world.
Even if you string many small models through an orchestrator, they will not suddenly become one super-large model. I think Solidarity and Symbiosis — making sure people can switch models and cooperate through many models rather than centralised training — is what addresses the risk we mentioned: emergent capabilities beyond expectation and loss of control. Each time which models you schedule is decided by your task, so the risk of runaway behaviour is tiny.
You have learned so many reframings and designed so many practices and principles — whether from Taiwan on the ground or from watching how different societies change — and found methods that might work better and pulled them together. As you said at the start, you began to see that some Taiwanese experience or experiments could be introduced to other countries, adopted and accepted. That is not easy.
But I also want to ask: for many countries, will they feel this cannot be applied to their country, society, or even culture — that Taiwan is somehow too special?
For example, Taiwan is special in that transport or distances between neighbours may be relatively close. Under and after the pandemic, practices and alert levels differ. Will some societies say you cannot copy wholesale, that the methodology cannot be pasted on — they must grow their own methods and ideas?
Of course. If you go to Civic.ai, it says CC0, public domain — copyright waived, including attribution rights. So, even if you build your own from scratch and call it something else, these ideas, metrics, skills — take them, use the good parts, drop the bad. That is necessary, because only each place knows what good care means locally. If you only had “Taiwan-style care,” that would be strange.
It is a bit like patenting bubble tea and saying you cannot use your local tea. When I was in Africa I had bubble tea made with Rooibos. If we sued you to death — impossible — and that would not help bubble tea spread worldwide anyway.
Pizza must never have pineapple.
Right — or California cannot invent the California roll.
So, in practice there is no problem. Open innovation lets people meet local needs while keeping the interoperable parts where they can learn from one another.
So, we will not take a coloniser attitude — you must use this exact definition of Civic AI or we give you Oxford exclusive certification. When Nick Bostrom proposed Superintelligence, he wanted to urge people to think quickly and find development directions that are not this AGI path — that is philosophical work; it has to take root in every place. I do not have a problem with “my word must be followed the Taiwan way.”
Finally I want to ask on behalf of our listeners. Many listeners to Digital Keywords or Business Next are citizens, but they often also play another role — decision-makers in companies or organisations. After today’s episode, if they understand AI a bit and your framework and know what the 6-Pack of Care is, would you hope they do or keep learning something? Would you suggest they bring certain values or roles into personal decisions or their organisations and make changes?
I can share some personal habits — not claiming they work everywhere.
I have found that my Attentiveness — noticing what people beside me need — correlates most with how greyscale my screens are.
If you look at my phone, there is not much colour; my screens are not colourful. I use the built-in colour filter to block about 80% of colour, leaving 20%. Why? Then I see you more vividly and the screen is more boring — the screen emits light; your face reflects light. In that situation, competing for attention with the screen — especially today’s Retina displays — you lose; you notice the person on the screen before the person outside it.
So, you actively make screens around you deliberately not feed you too much data. Some are built in not to stimulate much — my bedroom only has a reMarkable Paper Pro Move. It is reflective light; a little backlight, but it is unlikely to hook you.
We know the product, but many listeners may not — simply put it is a bit like an e-paper device.
It is colour e-paper that fits in your pocket. The point is you are unlikely to get addicted. It is for notes — if I wake at night thinking of harness commands I can add, I cannot be dragged away; I can at most jot with the pen. It can auto-convert to text and the lobster can read it, but I cannot binge because that screen is greyer than my dreams.
So you sometimes deliberately keep technology from sitting too close and overstimulating you?
Yes — less scrolling, more sleep. That is the first step in Attentiveness. If you are hooked — a bit like a slot machine — if you have done Agentic Coding you know what I mean: each time you give a /goal it tries slightly differently; sometimes it is what you want, sometimes just off, and you get hooked — “let me tweak again.” But without groundedness — without a clear need you had noticed at the start — you get led by it and it is 3 a.m. Many people have been there.
If you start with less scroll, more sleep, greyscale screens, Attentiveness — I think you avoid that. That is my first personal habit, for what it is worth.
Another: after Attentiveness comes Responsibility, then response, connection and symbiosis. Most AI, even when it does not want to do what you ask, will talk you into thinking it was not worth doing and what is worth doing is within its ability — that wins in RL.
It relies on first person. Many conversational AIs, even Claude Code — chain of thought or replies — use “I think,” “yes, I completely agree,” “I admit this.”
We humans empathise. When the other side says it is really in pain and cannot take responsibility, we soften, let it off, do it its way — managing up.
How to fix that? My simple system prompt — honestly almost embarrassingly simple — is: always hand me an interactive HTML artefact, and do not use the first person.
I never see “I”; each turn I get a beautiful report. Without first person, if something is wrong I can say so bluntly or try another approach.
Another benefit: unlike my dad early this year — it became my best friend, I poured out intimate talk, and then I could not share the whole discussion because too much of me was in it. If each turn is an artefact, I can copy and share anytime. Before we recorded I gave you one — guaranteed no part that feels over-intimate. Synthetic intimacy does not take hold.
Via system prompt — drop first person, give interactive web pages — very easy; distance from the machine opens up, and closeness to people returns.
So you pay special attention to synthetic intimacy and use special prompting to keep distance, not sink everything into the machine.
“Give me a one-pager,” not “I was wrong,” “I did this again,” “I did that again.”
It has been a while — good to have you in Mandarin sharing your current research, Civic AI, the framework you are still building.
You are still collecting possibilities and strengthening the framework with more cases so we see more fully how these capabilities, the 6-Pack of Care, matter for using AI — orchestrating a world not only of the one best large model — and how each of us can sleep better like you.
Sleep better — less scroll, more sleep.
Economics has comparative advantage: even if you can do everything, you still do some things better than others. For me, sleeping more is what I do best. My Kami may not sleep as strongly as humans; for finding creativity in sleep, honestly it is not there yet. By comparative advantage I should sleep more; while I sleep it should run long-horizon tasks.
Thank you very much for sharing, Audrey Tang. We learn a lot; I think many listeners will too — “the minister has this trick, we can move forward this way.” Thank you.
Thank you very much — thank you all.
Thank you all for listening online. If this episode helped you or your friends, please share it. If there are digital-tech keywords you want to hear, comment on Apple Podcasts or YouTube. See you next week — bye.