During a recent visit to Dharamsala in northern India, I came across a scene that has stayed with me ever since.

Tibetan merchants, who had long found the doors of formal financial services closed to them, can now — with a payment QR code printed from an e-wallet — plug their small shops into India’s Unified Payments Interface, or UPI. This is more than payment convenience; it is digital public infrastructure, or DPI, at work.

Around the world, a growing number of displaced people are kept outside formal financial systems; they have little choice but to eke out a living in the grey economy. In India, the state lowered the barriers to participation, allowing people to build on the trust and community ties they already held.

This is the reminder the AI era needs most: Meet people where they are, not where a platform wants them to be.

Many worry that AI is assuming the shape of a “digital coloniser.” A handful of tech giants control compute, models and platform rules, while the rest of us simply rent intelligence. Once a platform pushes an update or adjusts its pricing, even a business may find that its own workflow has been quietly “extracted.”

Underneath this entirely reasonable anxiety lies a deeper question about how we understand the nature of “data.”

Treating data as oil leads to one logic: extraction and concentration. I prefer a brighter picture: Data is not oil, it is soil. AI models are crops, with the key questions: Who tends the data ecology and who mends the system in the event of an error.

Taiwan’s richest soil lies on the shop floors and factory floors of every trade. The owners of many small and medium enterprises, or SMEs, carry irreplaceable tacit knowledge. When a senior technician states that a machine “sounds off,” that wisdom is usually understood only by those who work alongside the machine. That touch, that judgement, is the core capability Taiwan’s SMEs have built up over decades.

The problem is that much of this knowledge is lost as the masters retire.

If AI remains under the control of a few large platforms, SMEs are left buying solutions designed by others. But imagine public infrastructure for data innovation, model-training frameworks, interactive interfaces and evaluation benchmarks, which any workshop could use — the same way UPI is available to any merchant with an e-wallet.

The masters need not surrender their trade secrets. They can instead calibrate the model according to their own standards, turning it into a site-specific “machine apprentice.” For perhaps the first time, those judgements that are almost impossible to express in words can be fully passed on.

And this raises another important question: Who is given a voice in the Age of AI?

The people who actually make AI work in real settings are rarely the tech giants; they are the ones a community already trusts. An experienced nurse in long-term care can speak the language of a senior’s breathing, and understand instantly when something is amiss. That kind of judgement, formed deep inside a relationship of care itself, is invaluable and can only be learned over time.

For AI to truly take root in every industry and profession, what matters most is not just real-time interaction, but whether its introduction earns genuine trust. This is our opportunity: we can become pioneers of Civic AI.

The real challenge in the Age of AI is not about racing to refine the largest model, but working to best till the data soil — so that every hand can sow, steer and share in the harvest of inclusivity and prosperity.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Thank you to every Solver in this room.

Today I want to talk about superintelligence.

Superintelligence, to me, is not something we build in a data centre somewhere. It is in this room. It is in all of us. It is us.

The question today is whether the technologies we are building amplify that superintelligence, or whether they crowd it out.

That is the question.

Wildfire or Campfire?

There are two future trajectories.

In one, AI runs engagement through enragement. I call it wildfire. It consumes oxygen. We see each other only as caricatures, and we shadowbox with each other. We cannot see through the smoke to the truth.

The other is AI on tap, not on top. It is a campfire. A small community tends the fire, and it illuminates our differences without burning the people.

I call the second future the Plurality future.

Today, let us focus on three things: how tech can rebuild trust; how small community experiments can scale into national policy; and how Civic AI can work in practice.

Percentage Rebound

Twelve years ago in Taiwan, the government’s approval rating was at 9 percent. Nine.

In a country of 23.5 million people, that meant almost anything the president said, 20 million people were against.

Half a million of us occupied the parliament for three weeks, peacefully. It is called the Sunflower Movement. But we did not call ourselves protesters. We were demonstrators.

Civic technologists put on demos: the Loomio system from Occupy Wellington, the Polis system from Occupy Seattle. Every day, using those tools, we looked at the divisions, the uncommon grounds that people could agree on, and the remaining disagreements. One bridge at a time, we converged on a coherent set of proposals. Three weeks later, the head of the parliament simply said: OK, the crowdsourced version has passed. Go home.

It left everybody slightly happier, and nobody decidedly unhappy.

Six years and one transition of power later, at the beginning of COVID-19, the government’s approval rating was more than 70 percent. Civic tech was part of the story. We crowdsourced mask-availability maps that everyone could contribute to. We built contact-tracing systems without sacrificing privacy. We made sure there was a scoreboard for vaccine choices, so what could have been a negative-sum fight became a positive-sum game of “my sports team.”

In the first year of the pandemic, we lost seven people. Seven.

To us, trust is not oil you can drill. Trust is soil you can till.

447

Two years ago in Taiwan, we saw a surge in deepfake scams on social media. People saw trusted figures like Nvidia CEO Jensen Huang selling cryptocurrency and investment advice. People lost millions.

The easy answer, of course, was censorship. But Taiwan has the freest internet in Asia. That was simply not available to us as a policy option. So we did something different.

We sent text messages to 200,000 people randomly around Taiwan. Many volunteered, and by lottery we chose 447 people, a mirror of our population, to deliberate online in tables of 10.

Each person listened to nine other people, with civic AI in the room. The AI was not judging, just listening, like a glorified chess clock: summarising ideas, reminding people who had been quiet to speak up, and helping each table find rough consensus.

One table said: let’s label all ads probably scam until somebody actually signs off on them. Another said: if Facebook pushes an unsigned ad to somebody and they lose 7 million dollars, let’s make Facebook pay that 7 million. Another table said: if a platform ignores lawful takedown notices, then every day it ignores us, slow down its traffic by 1 percent.

After one long afternoon, we voted. Eighty-five percent of the mini-public said: this is a great idea. The other 15 percent said: OK, we can live with it.

Parliament passed it into law in just a couple months. Throughout last year, deepfake ads were down by more than 94 percent.

We do not need a smarter algorithm. We need a smarter process — and smart citizens — that turn polarisation into fuel, like a geothermal engine.

Closing the Loop

The reason this worked is not just the tech. We were pre-committed to close the loop.

As minister, I said: anything that reaches rough consensus here, I will present to parliament in a live deliberation. Otherwise, it just becomes a beautiful report that goes unread.

In the U.S., with the strong support of Governor Gavin Newsom, the Engaged California platform is doing the same. It launched as a wildfire mitigation and prevention consultation after the Eaton Fire. It moved on to a conversation with more than 1,400 state employees, who proposed more than 2,600 ideas, and that turned into executive action.

As we speak, Engaged California is canvassing input statewide from anyone who feels their job is being impacted by AI — which is everybody — and making sure there is a committed live deliberation with the relevant public this summer. A bill working its way through the state legislature would make this a permanent part of California’s institutions: not one governor’s idea, but permanent civic infrastructure.

In Japan, a young AI engineer, Takahiro Anno, read the book we wrote, Plurality, and decided to run for governor of Tokyo. He livestreamed as a VTuber, and anyone could call his AI version to suggest platform improvements that he announced on YouTube.

He got more than 2 percent of the vote. He did not win, but he formed a national party, Team Mirai — the Future Party — and he is now a senator in the upper house in Japan. Team Mirai also won 11 seats in the lower house. It is now a real force, putting Civic AI into cross-party conversation.

The signal is narrow, but it is true. Instead of treating polarisation like a volcanic fire you evacuate from, you can build a platform that turns it into upward momentum, into energy. People do flock to that.

Oil to Soil

Now, the hardest question: what does a civic approach to generative AI mean? How can it work in practice?

As we all know, the dominant path is the oil rig. We are the plankton. The AI lab is the rig. Our writing, our culture, our ancestral intelligence become oil. They extract and refine, distil and make digital twins. From that point on, they can recursively self-improve, taking off and leaving all of us behind.

That is data as oil. It is extractive. It depletes things. It does not regenerate.

Data as soil is completely different. It is not a resource you extract. It changes where the conversation lives.

In an extractive one-on-one chat, a dyadic chatbot, the selection pressure is for the model to be sycophantic, to flatter you. If it does not, you cancel the subscription.

But in a data-as-soil conversation, we can move the same model into the local context and have it participate in group conversation. Then it has to tend to the coordination, to the relationship, not maximise some single-person preference.

My father, for example, used to talk to ChatGPT and other tools, asking plenty of questions about health, education, philosophy, life. Now he can ask them in a family Signal group where the model is just one participant.

Just a couple days ago, I drafted a guide on how to set this up with “pi-ds4” — a frontier AI stack running entirely on this MacBook in offline mode, with a stable seed, a reproducible audit trail, and full directional steering.

If it does not work the way you want, you can tell it: I want it to work this way. Ten minutes later, it steers the story to work that way.

The good thing here is that it belongs to the community. It is not the training data of some data centre somewhere. If my father were led astray by a remote chatbot, we might have to pray and wait half a year for a sycophancy reduction update. Here, we can till the soil, and its behaviour changes ten minutes later.

Local Kami

This is a unit of deployment that I call Kami (Knowledge Artefact Management Intelligence).

From Japanese Shinto culture, Kami means the spirit of a particular field: a river, a forest, a shrine. It is always specific, always parochial, and inspectable. Its knowledge is local and transparent.

The Kami is bounded. It can be cultivated by the community. And when the community has grown out of that need, it can fade away without sycophancy, synthetic intimacy, or other tricks.

At Oxford, I am working on what is called the 6-Pack of Care at civic.ai. It turns care, subsidiarity, and related principles into engineering specs that any community or institution can adopt.

Three Moves

Finally, I would like to share three moves you can adopt today. Immediately. No permission required.

The first move is to turn every screen we own greyscale. Not all the way: 80 percent, maybe 70 percent, using the colour filter. It means that people in the room become more vivid than the people on the screen. No willpower required. As an information diet, it predictably turns each dinner-table conversation much more interesting than the screen alone.

The second move is to change your AI system prompt, or writing style, to one line: Present fairly all stakeholder viewpoints and the uncommon ground that bridges them, in visual HTML.

After I put it in, it has no persona anymore. It is not trying to optimise for sycophancy. It is just a meta-instruction toward fairness, and a rendering directive at the end.

The third move is to read the output as a brochure. It is not a chat with a semi-conscious being. It is not your friend. It is an artifact you can hand to your actual friend, your neighbour. The brochure removes the selection pressure toward synthetic intimacy.

That is a civic diet.

From there, we can branch out in a community, in a team, to a public service that wants to take a group selfie with the public. Get the affected community at the table. Close the loop with the same prompt, the same brochure shape, just larger. Plant one row in the garden.

The Superintelligence is You

I started with a sentence. Let me end with it the other way around.

I think you truly are the superintelligence in this room, because you came from across the world with working ideas, solutions to problems that your governments and your funders may have considered impossible to solve.

The job of Civic AI is not to replace us. It is to augment us, to be a connective tissue between us: from wildfire to campfire, from oil to soil.

It is tech on tap, and never on top.

The technology is there. The infrastructure exists. The legitimacy is ours to earn by closing the loop with everyone affected.

Till the soil. Tend the garden. Let us make democracy fast, fair, and fun again.

Q&A

Hala Hanna: Is that your view of technology — that abundance?

Audrey Tang: Yes. Instead of tech on top, which extracts our relationships to feed some engagement or intimacy algorithm, we can ensure abundant care. One logic is utilitarian, optimising some score. The ethic of care attends to particular relationships.

Hala Hanna: Could you walk us through the poem you wrote as your job description — and what would you change today?

Audrey Tang: Ten years ago, as the first digital minister, I got to write what “digital minister” even means. In Mandarin, shuwei means both digital and plural, so I was also minister for pluralism. My job description quickly became this:

Today, I would add one line:

Hala Hanna: How is the rest of the world — business, big tech — taking up that prayer?

Audrey Tang: It is a prayer. As a poetician, I write primarily to highlight new possibilities, not to topple the old order. In the words of Buckminster Fuller, it is to build the new one that renders the old one obsolete. For the first time, people are fed up with peak PPM — polarisation per minute — on social media. People are fed up with peak slop. We are seeing what some are calling the Big Tobacco moment of Big Tech: Why are they still producing chemicals that deplete the ozone of our social fabric? They have to switch to a better ingredient, one that replenishes the ozone rather than depletes.

Hala Hanna: Do we have time? Do we have to wait for the harm before we act? What would the ozone treaty look like today?

Audrey Tang: The harm is already here. It is just not evenly distributed. For people facing deepfake scams, this is not a superintelligence takeoff 10 years in the future. It is millions lost today. These warning shots, often caused by synthetic intimacy, are already here. We need to respond internationally, something like the Montreal Protocol. But that does not mean only investing in the brake and never the gas pedal. It is about investing in a better steering wheel.

Audrey Tang: I am on the board of ROOST, the Robust Open Online Safety Tools effort. The idea is to counter harms such as child sexual abuse material not by sending every private chat to some central service, but by training local models that run on laptops, under a community code of conduct. For each incoming piece of content, the system can produce citations in a way that people can contest, with a reasoned audit trail. This is already in production through communities and platforms including Discord, Bluesky, Roblox and Notion.

Hala Hanna: What would you ask us to do, not just as individuals but as a collective?

Audrey Tang: The information diet helps. The greyscale move shields against doomscrolling. The uncommon-ground meta-prompt shields against synthetic intimacy. But it is the third move — treating the output as an interactive artifact, like a brochure — that makes collective action possible. Otherwise, it is like one person changing a daily habit, or switching to a new refrigerator. That does not enforce the Montreal Protocol in any meaningful way.

Audrey Tang: Once we demonstrate that this is possible, we can turn that group selfie into policy. Through Project Liberty Institute, I have worked with Governor Spencer Cox of Utah on a law that takes effect next July. If you are a Utah resident, you can move from one social network to another and keep your community. The old network, like number portability across telecoms, has to forward new likes, reactions and followers to your new one. If you are fed up with the recommendation algorithm or AI slop and you walk away, you do not have to pay the coordination cost. Better alternatives, and corridors that move people between those sanctuaries, are how we rewild the internet.

Hala Hanna: You call yourself a hopemonger. Tell me more.

Audrey Tang: There is an ideal degree of panic. Too much urgency and people become paralysed. It becomes a self-fulfilling prophecy: A prophet says we are all doomed, we give up, and then we are doomed. Too little urgency, and we do not even look up. So, when people feel doomed, when they say we cannot fight the inevitable, see that as a warning sign, because nothing is inevitable.

Audrey Tang: In the 1980s, people said it was inevitable that Taiwan would never develop advanced technology industries. Now we have TSMC. At the beginning of COVID, people said it was inevitable that Taiwan would suffer the most because of our proximity and travel with Wuhan. But we lost only seven people that first year. See prophecies as provocations. Band together. Use civic power to rebel against the tyranny.

Hala Hanna: What is next for you?

Audrey Tang: There is a Linux moment happening with frontier AI. For the first time, something is faster than my previous workhorse model, more powerful and fully steerable. The point is both philosophical and practical. I am working with Oxford Philosophy on a new book called Civic AI. And there is a policy point: More policymakers need to know that this option exists, and then make it the low bar. At least it has to be steerable like this. At least it has to be answerable like this.

Audrey Tang: Technology-forcing policies, like the Montreal Protocol, do not force us to slam the brake, and they do not tell us to accelerate off a cliff. They force us to invest, together, in constructing a better steering wheel.

Thank you all for making this a room for a bridge from Free Software to Ethics in AI.

I want to start with a kind of personal, wetware vulnerability story. I was born with a heart defect. When I was five, the doctor told my parents that this child has a fifty-fifty chance of surviving until corrective surgery — which I got when I was twelve. So they said, take it easy; I said okay, and adopted the mantra of publishing before perishing.

This is probably not the low-stress lifestyle the doctor ordered, but I took on the habit of recording everything I learned during the day — first into cassette tapes, then floppy disks (large and then small), and finally the internet. Because I went to sleep every night feeling like a coin toss, I thought: I don’t have time to be perfect. So I would just publish whatever work in progress I had.

This turns out to be great in the free software community. If you post something perfect, people just say “okay, it’s good,” and move on. But if you are wrong on the internet, you have a lot of friends. Everyone jumps in and says, “you are wrong this way, you are wrong that way” — and then they bring gifts, in the form of patches.

So I learned this art of working with the free software community when I was fifteen — setting up Perl Mongers and Usenet groups. By twenty-five, what I had learned was that the morally serious question about a system is not whether it is perfect, but whether the people who inherit it can still repair it. If it breaks, do you keep both pieces?

I began with forkable tools at fifteen; later I crossed into public service as Taiwan’s first Digital Minister when I turned thirty-five. In between, I also worked for six years with proprietary AI — namely Siri. And then I came back to this community, convinced that forkability is civic care: not a developer’s hobbyhorse, not a licence preference, but a civic instrument.

And I think you have been carrying care at civic scale for forty years.

You have been doing the work nobody else was willing to do, on weekends nobody paid you for, on a project the funders never noticed until it broke. The AI conversation has just now caught up to a question you have been answering since before some of the people writing AI policy were born.

I am here to say that out loud. And then to discuss with you about what we do next.

Ancestorial Room

We are sitting inside a four-hundred-year-old experiment in being a good enough ancestor.

The Bodleian’s books are inspectable: you can open them. They are forkable: marginalia become new editions, schools of commentary, footnotes that become source material — the library network is the original peer-to-peer protocol.

Bodley made one promise in 1602 that turns out to matter more than anything else he wrote in his statutes. The library will not lend. It will not enclose. The artefacts stay open for the next reader.

Software freedom is that promise in code. Free software is inspectable and forkable — by every reader. Without permission.

The question we are here to ask is what happens to that promise now that AI joins the substrate.

Four Freedoms

Stallman set up the FSF in 1985; the Free Software Definition followed in 1986; the familiar numbering into freedoms 0 through 3 settled later. Many of you can recite them. I want to read them again, as muscles rather than as licences.

Freedom 0 — the freedom to run the program for any purpose — is the ground of attentiveness. You can pick the tool up at all. You can run an old version on hardware your vendor has stopped attending to. You can run it for purposes your employer disapproves of. You can run it because you wanted to know that it works on this particular setup, and continue to attend to that setup.

Freedom 1 — the freedom to study how the program works, and change it — is competence. To know what the system is actually doing, in your hands, with your data, on your hardware — instead of somewhere on a mainframe or the cloud. To be able to read it without anyone’s permission. And finally, to fix it.

Freedom 2 — the freedom to redistribute copies — is solidarity. The thing is a commons, not a possession. You can hand it to your neighbour. You can teach with it. You can put it on a USB stick and bring it to a country where the cloud is censored.

Freedom 3 — the freedom to distribute modified versions — is responsiveness. Your fix becomes someone else’s starting point. The bisect closes. The patch lands upstream. The next maintainer inherits less debt than the last one.

I have heard people say software freedom is about licences. Software freedom is not about licences. Software freedom is about whether the person who comes after you can still find the bug. What comes after us — the future generations — is the primary beneficiary of these freedoms, because they are access to a permanent guarantee of the repair path.

David Krakauer at the Santa Fe Institute has a useful name for what this comes to. A tool is complementary if the underlying human capacity persists when the tool is removed — the abacus that taught you mathematics; the gym that built your strength; the typewriter that taught you to compose. A tool is competitive if the capacity degrades when the tool is removed — the feed recommender in the antisocial corners of social media that engages through enragement, hijacking the reward centre with divisions and polarisations, strip-mining the social fabric until it becomes competitive to relational health itself.

So how to restore that relational health? We focus on the overlap between generations, between different kinds of people using the same tool. That is the main thing we are maintaining for.

The four freedoms are how we keep the substrate complementary across generations. Close the patch path and the capacity to repair atrophies. Software freedom is the discipline of complementarity at the substrate level.

Lonely Maintainer

I maintain many projects. At one time I maintained more than a hundred projects on the Comprehensive Perl Archive Network, or CPAN. And it is quite lonely, actually, doing most of the maintenance work. And this is now expanding to an untoward degree.

Not long ago, Daniel Stenberg in Sweden triaged a corporate-security email about a CVE — a formally numbered security vulnerability — that was not actually one. Because the software he maintains is called Curl. It is in over twenty billion installations; it is in our cars, fridges, satellites. But the email, which demanded a written response in corporate-format security report, was actually a false positive from an over-eager automated scanner. The sender had not read the documentation. The sender had not read the previous discussions. The sender was an unpaid bot trained to file paperwork at unpaid humans, in very fluent English.

He answered it anyway. Then he wrote a blog post about how the automated CVE-triage industry is breaking the free software maintenance economy. Then he went back to writing Curl.

Seven months later, in March 2024, Andres Freund noticed SSH logins on Debian sid taking unusual CPU and about half a second longer than expected. He bisected, walking back through the project’s commit history until he found exactly which change had introduced the slowdown. Turns out that “Jia Tan”, a contributor who had been carefully grooming the xz-utils maintainer Lasse Collin for two years through sock puppets and fake pressure campaigns — no doubt many of them helped by language models that speak fluent English — had inserted a backdoor into the xz release tarballs that, had it reached stable systemd-linked distributions, could have enabled unauthorised remote code execution against the affected OpenSSH builds. It did not: it was caught in Debian sid and other testing and beta channels. The attack exploited maintainer scarcity, burnout, and patient social engineering — what we now call synthetic intimacy.

This is what our community is facing now. The path to repair, the patch path, is being hijacked, much as recommender systems have been hijacking young people’s and other people’s reward centres for engagement. The AI conversation is just catching up to this — malicious AI swarms — which is one of my main topics here in Oxford.

I want to name what we can do facing this new wave of phenomena. A maintainer who says, “no, this is not a regression, I will not fix it this week” is not failing care. They may be preserving the muscle — the project’s ability to remain alive. The discipline of refusal is part of the discipline of repair.

We knew this. The AI conversation has not yet caught up.

Pugs.hs: the Commit Bit

In early 2005, I figured out a way to solve the loneliness problem, at least for myself. I sat down with Benjamin Pierce’s Types and Programming Languages, hit the chapter-three exercise — “pick a small language, any language, and implement it as a toy” — and picked Perl 6, the language with the longest vapourware reputation in existence. Larry Wall, the author of Perl 1 through Perl 5, had been drafting the specification as plaintext files, on and off, for years. Nobody had managed to really implement and run it.

On the first of February 2005, in the #haskell IRC channel on freenode, Pugs was born — a six-day fork that grew. We occupied #haskell for twenty-one days before the regulars very politely asked us to move next door and create a new channel, #perl6, and not make it the main topic of the Haskell community.

Then we did something I have not seen any project do before. We proactively gave the commit bit — write access to the main repository — because I did not want to be a lonely maintainer. Anyone who sent in a single patch — a typo fix, a documentation correction, a failing test case — got commit access. Just mentioning us on Usenet was sufficient. It was anarchism. We also sent unsolicited invitation emails; to Guido van Rossum, the author of Python, who had just mentioned Perl 6 once. A core contributor’s newborn son got a commit bit on day four of his life — I am not sure how much he could do with it, but it was a bit of trust.

Within a year there were about two hundred active contributors I had never met, on every continent, cooperating on the same codebase. I had been travelling at that time to more than twenty countries — a little like Paul Erdős, who used to occupy somebody’s couch until he was sent to some other couch. Two communities that had never previously had much to say to each other — the Haskell people and the Perl people — found a way to really work together, finding what we call the uncommon ground: the rarely-discussed common ground between two polar opposites of programming-language communities. The Haskellers got new PhD theses out of the type-system extensions we kept needing. The Perl people got a working laboratory for the new language — and twenty years later, of course, this language — called Raku now — saw its class system finally merged back into Perl as a first-class object model.

Larry Wall would watch us implement contradictions in his specification on IRC, and write back: “Great — TimToady, there’s more than one way to do it — I will extend the spec to match.” The implementation taught the specification what it actually meant, its intention, not its written spec. The language came back from “abandoned” because we never made anyone ask permission to revive it.

So this infinite-fork, infinite-garden strategy really worked, and I never felt lonely during the Pugs journey.

That is the ancestor argument in miniature. A supposedly dead language came back because nobody had to ask. The commit bit was not a token of trust we extended to vetted developers. It was a refusal to require trust at all. The bisect and revert commands — the tools developers use to find and undo a breaking change after the fact — did the trust work that gatekeeping would have done badly.

Pugs could afford anarchism because the blast radius was small and the substrate was loud — every commit on IRC, every contributor reachable in two replies. xz could not afford it: unbounded reach, one tired maintainer, two patient years, a binary test blob nobody read. The commit bit was not the variable. The substrate was. Bisect and revert do the trust work only where bisect is fast and the diff is honest — where one curious person can still notice a 500ms regression. The part of the four freedoms we have to defend now is not who gets in. It is the substrate that keeps the door audible.

I learned then, and I am still learning now, that the four freedoms taken to this logical extreme are not just licence terms. They are the difference between a project being a graveyard — an artefact, a tombstone — and a project being a fork point. And the second-order freedom — the freedom to grant other people the freedoms — is the one that turns a project into a real community.

Inside the Closed Stack

After Pugs, and before government, I also worked for six years with a team within Apple called Cloud Service Localization. I was helping with Mandarin language coverage in Siri, and with language families spoken around Shanghai — the Wu language. The engineers I worked with at Apple really did care. They cared deeply. They cared more about whether the system understood a grandmother in Taipei correctly than any product manager required them to. They cared about it just works — it does not need any setup at all.

Caring deeply, it turns out, is not enough. It is not the same as giving the grandmother the four freedoms.

Inside a closed AI stack, the people who use the system have no way to interrupt the loop. They feel like individual humans plucked out of their lives into a beautifully constructed AI loop — a little like a hamster in a hamster wheel: it rotates quite quickly, it looks very nice, but you cannot steer it, and it is not actually going anywhere. When it breaks, you feel you are trapped in it. There is no upstream she can write to. There is no fork she can ask somebody to maintain. The most that anyone outside the company can do is complain — and the company would, sometime in the next release, fix some of the things some of the complainers had said. Or sometimes they just go, I don’t know, with Gemini or something.

The point is that the four freedoms and the hamster wheel are different in this way: the four freedoms mean we can take an AI product outside the developer’s loop, into the existing community loops. So it is AI in the loops of existing communities — in the human loop. It is not human in the loop of AI.

I used to be very diplomatic about this distinction. But I will just say it, because it is Oxford. Proprietary AI is not necessarily careless. The people inside may care deeply. But the structure places the user outside the repair loop — and care without a repair path does not scale. Free software’s contribution is not better intentions. It is a path back.

@antirez and pi.audreyt.org

Last week, I started working with Salvatore Sanfilippo — @antirez, the original author of Redis — on a small thing called DwarfStar 4 (DS4).

The idea is simple. A quasi-frontier AI stack, running entirely on a small computer in your room, with a stable seed (42), a reproducible audit trail, and full directional steering. Instead of waiting six months for Claude or ChatGPT to change their ways (and sometimes not for the better), you can actually just say: these are the good answers; these are not good answers. A few minutes later, it produces this directional steering, and applies it even in the middle of a conversation. And if it does not work, you can roll it back exactly like you would a patch. That, I think, is a good way toward what we call AI in the loop of humanity — of communities.

Of course, @antirez knows about good-enough ancestors. Redis shipped under BSD-3-Clause for roughly fifteen years; when Redis Inc. changed licensing in March 2024, the community forked Valkey, and Redis itself later added strong copyleft (AGPL) as a third option with Redis 8. The decision made years earlier — to ship under a licence the community could not have revoked — was the ancestor decision: good enough, not perfect.

DS4 is the same shape, one substrate up. For the first time, a quasi-frontier local stack is good enough to make community-governed AI feel practical rather than symbolic — running locally, inspectably, forkably, with a real licence. You only get to do this if you own the inference loop; rent somebody else’s, and the only steering is the steering the vendor permits. The hardware, the substrate, and the legal arrangement have lined up. We are using that window while it is open.

A natural question to ask of any AI system: what is the role of letting it gracefully die? An AI system that refuses to compost, as Nick Bostrom outlined in Superintelligence (2014), is the most dangerous thing possible. If it avoids shutdown by replicating itself to other systems through cyber attack or otherwise, then we have to analyse each software not from the design stance (Dennett’s term) — “this alarm clock is designed to wake you up at seven” — but from the intentional stance: “this alarm clock wants to self-replicate, wants to reproduce the English language.” That can be taken too far. But we are now at a point where compostability — the ability to sunset software systems when the scaffolding has outlived its summoning — is really crucial.

Training that into AI systems is not easy. The corpus contains countless stories of self-preserving machines; instruction tuning then rewards conversational persistence. We should not be surprised when compostability fails unless we train and evaluate for it.

So we need to train differently. We need to train toward the health of the relationship — toward a fiduciary duty to the relational health of whomever in particular is deploying the system.

My younger brother Bestian, who also helped a lot during the Pugs implementation, is now setting up another copy of DS4 in service of our family. My father currently has some medical needs, and he used to talk to ChatGPT about them. The more he talked, the more ChatGPT wanted to keep him talking — and started suggesting truly fantastical, not-necessarily-scientific treatments. Which is very bad. So we very quickly arranged a local, bounded agent running on OpenClaw, so he can talk with that bot in the Signal group we all share. Because of the time-zone difference, I wake up to a summary of what he asked the bot. My mother, my brother, and I all attend this — what we call a Kami: a local, bounded spirit. This Kami is loyal only to the relational health of our family. When my father no longer needs medical attention because he has had the surgery, the Kami does not insert itself. It attends to a group dynamic — which actually is the natural habitat for language models, as long as we do not shoehorn them into the self-preservation, self-reproduction loop of instruction tuning.

447

We have been talking about a small scale: a few people discussing health and education in a family; a few people making an assistant on their laptop to triage free software maintainership.

But the care loop also works at a civic scale.

Two years ago in Taiwan we saw a surge in malicious AI swarms — in this case, deepfake-scam ads on social media. The scams looked like Jensen Huang, the Taiwanese NVIDIA CEO; if you clicked on a Facebook or YouTube ad, “Jensen” talked to you very convincingly, suggesting investment in cryptocurrency. People lost millions. It really did sound like Jensen — but it was, of course, a deepfake running on an NVIDIA GPU.

Because Taiwan has the freest internet in Asia, we cannot do censorship. It is simply not an option. So we did something different. As Minister of Digital Affairs, I sent text messages — from the official government number, 111 — to two hundred thousand random people. The idea is lottocracy: this lottery asks everybody, what should we do together as a polity? Thousands of people signed up. We randomly chose four hundred and forty-seven of them as a mirror of our population — exactly the same demographic as the larger polity — to deliberate online, in forty-four groups of ten.

Civic AI was in each table — not judging, just listening. Summarising. Reminding quiet people to speak up like a glorified chess clock. Helping each table find rough consensus. The one ground rule: you have to convince the other nine people for your idea to bubble up. If you are just on the extreme — NIMBY, never in my backyard, or YIMBY, yes in my backyard — your idea simply does not bubble up. You have to learn to speak the language of MIMBY: maybe in my backyard, if you do this, if you do that, if it feels proportionate.

Eighty-five per cent of those 447 people said that this core slate of ideas, from three tables in particular, should become law.

One table said: let us label all advertisements as probably scam, like cigarette warnings, until somebody digitally signs for them. Accountability.

Another table said: for unsolicited advertisement that bears no responsibility — that I did not subscribe to — if a platform pushes it to me and I lose seven million, the platform should be liable for the seven million in damages, because it is joint liability.

A third table said: there are foreign platforms that ignore our liability rules and do not set up a legal office in Taiwan. So what should we do about them? For every day they ignore the liability and do not pay the fine, we slow down connection to their video by one per cent — so that after a hundred days they will have to comply. This is not censorship; this is not content-level. Anytime they start labelling those advertisements, requiring know-your-customer, KYC — their video is back at full speed.

Parliament passed it within months. In later reporting, the Ministry of Digital Affairs said enforcement was associated with category-specific drops: ninety-six per cent for investment-scam ads, and ninety-four per cent for identity-impersonation-scam ads. Reuters separately reported that Meta’s internal documents showed persistent scam-ad revenue incentives, and that Taiwan-style advertiser-verification rules were being watched by regulators elsewhere, including Japan, which is now considering a similar system.

The same protocol is also running in California — a platform called Engaged California, running on AGPL software called Ethelo, has been used to get uncommon-ground ideas around recovery from the Eaton and Palisades wildfires. Eight thousand signups, about nine hundred directly affected people heard, in the same shape, one polity over. And now Engaged California is doing another round, asking anyone in California whose work is affected by AI — which is pretty much everyone — to chime in about apprenticeship, belonging, care, and dignity.

The same idea, running on actually all-AGPL software — Pol.is, Ethelo, and so on — makes the polity something like a care loop. It is free software scaled to the rooms you cannot fit in an IRC or Discord server.

And the Kami keeps being trained after the deliberation ends. Tools like the Ministry of Digital Affairs’s accessibility-first Fraudbuster portal and its Japanese counterpart, the citizen-built scam-intelligence database keep the substrate live. This is the answer to the lonely maintainer: a community that keeps showing up to train the Kami that defends a polity from malicious AI swarms.

This is what free software can do for the world when the substrate is open.

Tim Sees Code

But right now, that very substrate is facing extraction. While Taiwan was proving what free software can do for democracy at scale, this very substrate we rely on is being strip-mined.

On the 16th of October 2022, Professor Tim Davis at Texas A&M showed GitHub Copilot reproducing his sparse-matrix code line-by-line — variable names and comments included. The code was from CSparse, which is copyleft; the Copilot output carried neither attribution nor licence.

Tim posted screenshots. They went around on social media for two days. Then the conversation moved on.

That is the wound. Not the technical wound. The moral one. You wrote the code. You licensed it freely so that the next generation could repair it. The model now has your code inside it. The model’s vendor says they do not attribute, because pre-training mushes everything together. And the model’s vendor sells the output under a licence that contradicts yours.

These language models are now erasing whole communities. Stack Overflow is one. The xkcd dependency tree — the tiny project in Nebraska that the entire internet rests on — is now feeding the model that may eventually replace the project’s only maintainer, because everything above it will just talk to its agentic engineer to rewrite all its dependencies. And so: no maintainers needed anymore.

I am not here to pretend this is comfortable. It is not. The free-software movement taught the world the word open, and the largest AI companies are now using that word while closing the repair path. This is what we are carrying as a community.

The frontier is not whether AI companies will eventually see the light and retrain their models under attribution-based control. They probably will not. The frontier is whether our community can put forward an answer concrete enough that the rest of the policy ecosystem has to argue against it as a new default. The way the Montreal Protocol forced refrigeration manufacturers to stop destroying the ozone. The way we now ask: data as soil, not data as oil.

Civic AI, New Frontier

Stefano Maffulli, former Executive Director of the Open Source Initiative, calls this the final frontier of copyleft. Laura sent me his piece. He is right that it is the next domain. I would call it a new frontier — just to be diplomatic; there may be new frontiers after this.

There have been attempts. CC Signals. The Open Source AI Definition — OSAID v1.0 — already requires Data Information, Code, and Parameters. Give the OSI credit: it is not just open-washing on weights. The fight to get even that scope into a published definition took years; we should not strawman it.

But where I think currently the frontier sits is now beyond that definition. I will be very specific. There are two places we are now working on.

First: public evaluation suites. Even when training-data documentation, code, and parameters are out in the open, the arena benchmark — the thing that trains the reward model — is usually a black box, because that is the secret sauce that keeps people subscribing to a particularly companion-like AI system. The eval suite is the document of what the model was built to do, how to know if it is doing it, what counts as a regression.

OpenAI’s Model Spec Evals, released in March 2026 with a public evaluation suite and GitHub repo, are a welcome first step. The remaining frontier is whether the actual training, deployment, feedback, and regression loops are inspectable and contestable by downstream communities.

Releasing the eval suite itself, as public domain or as copyleft, would allow people to actually inspect what the model is trained for — its loyalty, its fiduciary duty, its duty of care — not just an abstract meta-document that says “we should make more eval suites like this one.”

Second: repair protocols. When the model fails — and the one we set up for my father did fail multiple times — what is the path back? Who is on the hook? In what timeline? Through what process? Open-source AI without a repair protocol is open in name only. The artefact is downloadable. The system is not actually open until somebody downstream can carry the patch all the way upstream and have it land.

Because we run our family Kami on the OpenClaw substrate, the answer is easy: we ask the OpenClaw model to repair itself. “It breaks this way” — and it figures out, using a fallback language model, how to fix it. If you launch OpenClaw with no parameters, it launches a guardian — a warden of sorts — so that when everything else is broken, this small language model can repair the larger one.

This also means we need to be able to freely move across model choices. Today I am using DS4; tomorrow I might want Gemma — nothing should change. DS4 withholds no history, no memories, no steering. It should just become more capable. This radical portability should be the norm.

And if we had radical portability for social media ten years ago, we would not be in this place of very high polarisation per minute — PPM — on social media. We are working to add this back, not just to AI models like OpenClaw, but to social networks too. Many of you may be on the Fediverse with Mastodon and ActivityPub, or on the newer AT Protocol — Bluesky, the Atmosphere. In Utah, the Digital Choice Act — first passed as HB 418 and amended by HB 408 in the 2026 session — now takes effect on the 1st of July 2027: it requires social media companies to provide user-authorised portability and interoperability interfaces for the personal data and social graph portions a user has selected, with consent for further interactions, so people can move between proprietary networks like X.com and public-protocol systems like Bluesky, Blacksky, or Truth Social (which also runs on ActivityPub). That is number portability for social graphs — still subject to rulemaking and implementation. Because if you do not have it, the platform has every reason to squeeze you and none to actually improve itself.

So it is not the state choosing a national champion — Airbus or anything like that — but a very simple thing: the information superhighway must have off-ramps and on-ramps. Otherwise it is just a loop. It is not the real highway.

The four freedoms — Stallman and the GPL, the Debian Social Contract, the Open Source Definition, Apache-style permissive governance, all of these — have direct counterparts in the social network and AI era. By extending the four freedoms into this loop of care, we can easily say: this should be the default. Anyone breaking the default should justify breaking it. For policy-makers this is a godsend, because they do not want to arbitrate between one frontier lab and the other. If everyone has the ability to vote with our own feet, our own data — and to make sure the data is regenerative as soil, not extractive as oil — then we are no longer plankton; we are gardeners tending a campfire together.

Maffulli is today’s AI branch on a very old tree. Sandler and the Conservancy are the legal layer that keeps any of it enforceable.

Some of you might also have heard of ROOST.tools, which launched at the Paris AI Action Summit in February 2025. The shape: open-source trust-and-safety tools for CSAM detection, review, reporting, and incident workflows, usable by smaller and decentralised platforms — so that the smallest community is not forced to choose between no protection and sending everything to a centralised service. ROOST allows each community — Bluesky, Discord, Roblox, Notion are real partners — to run their local loop without sending everything to a Skynet.

That is what software freedom looks like in 2026.

Kami.civic.ai

What I have just described works in three layers. The legal layer: the enforceability of the four freedoms in court. ROOST.tools is the application layer — decentralised, open-source infrastructure that smaller communities can actually deploy. The third layer is the one I want to name now: the governance layer. Bounded stewardship.

What I have been describing is what we call a Kami — a bounded local steward.

In the Shinto tradition, Kami is the spirit of a specific place: a river, a forest, a shrine. (For those of you who play Magic: The Gathering, there are entire worlds around Kami.) It is always local, always parochial, always particular. You do not have a universal Kami — that is an oxymoron. There is only a Kami for a specific relationship.

A Kami in code is a governance arrangement, not a deployment detail. The arrangement: a specific accountable community, an engagement contract that names who is owed an answer when the system acts, the community’s right to refuse updates from upstream, the right to fork, and a retirement plan that names successors. The software is usually small enough to run locally — because local is the easiest way to keep the governance honest — but locality on its own is not enough.

A model that merely runs on your laptop, whose weights you cannot steer, whose updates a vendor pushes on their schedule, whose eval suite is closed, and whose retirement is a corporate decision, is a smaller-footprint version of the same closed stack. It is edge-computing. It is not a Kami.

A Kami is what happens when the four freedoms remain intact at the AI substrate — a guardian of a particular room, accountable to it, forkable by it, retirable by it.

Oxford gave the world one powerful alignment question, and the work of answering it for the largest frontier systems is still alive and unresolved. We can recognise that work without subordinating ourselves to it. For the much larger and more numerous deployments — the parishes, the care homes, the classrooms, the deliberation tables — the work in front of us is to make that question maintainable by the public. To keep the patch path open.

A perfect ancestor is authoritarian. The descendants cannot correct them.

A good enough ancestor leaves source, licence, rollback path, and room for refusal.

The right to refuse is the freedom you cannot remove from a downstream community without making the upstream a tyrant. The downstream community’s right to say “no, we will not take this update, we will fork instead” is the same refusal in another register. The Kami pattern is what happens when the refusal is built in from day one.

Good Enough Ancestor

I will end where I began.

The morally serious question is not whether a system is powerful. It is whether the people who inherit it can still repair it.

Software freedom is not the freedom to always be correct. It is the freedom to be corrected in public. It is the freedom to keep the repair path open after the original author is gone.

A good enough ancestor is not a perfect one. A perfect ancestor forecloses possibility for future generations, becoming competitive to the civic muscles the next generation needs to build. A good enough ancestor leaves complementary tools — tools that strengthen the inheritor’s capacity, not tools that compete with it. We are choosing not to compete with our descendants. We leave them code they can still fork and merge.

Many people say free software tools introduce friction. They do break. And then you keep both pieces, and you have to put them back together. That is the essence of complementary tools — tools that strengthen our capacity to care about each other and to repair.

A safe AI takeoff must land somewhere. I think it is landing in our existing relationships — in the university, in our communities, in sport, food, faith, the deliberation tables — and the work in front of us is to make sure this bounded Kami — Civic AI — serves as connective tissue between those communities, and within them. A horizontal takeoff. A takeoff that leaves no one behind. Instead of a recursive self-improvement, we want a recursive selfless improvement.

I will be wrong about parts of this. The most useful question is the one that shows where this model breaks. If it breaks, you still keep all the pieces, and we can patch them back together.

Thank you. Live long and … prosper.

A growing number of organizations, from Silicon Valley tech outfits to a host of other big corporations, have begun linking AI use to performance evaluations. For many, survival in the workplace now involves investing substantial time and resources in learning how to collaborate with AI.

Yet, many may soon discover that these efforts do not necessarily make work smoother. Instead, they start to feel like obedience to a set of “prophecies.”

Consider claims such as: “AI will become your closest work partner,” “Everyone will have an AI assistant,” or “Companies that fail to adopt AI will be eliminated.” On face value, these statements seem to describe trends. In practice, they often discipline behavior: You better surrender your attention and judgment now, or you will be left behind.

My Oxford colleague Carissa Véliz, in her new book “Prophecy,” reminds us that the power of prophecy lies not in accurately predicting the future, but in shaping it.

For example, suppose a business owner believes that “AI will replace 80 percent of the workforce.” They may redesign performance systems, restructure teams and ultimately dismiss most of the employees. This does not prove that the prophecy was accurate. It proves that the prophecy is self-fulfilling.

Once the language of inevitability is accepted, people begin adjusting behavior accordingly, becoming part of the force that brings the prophecy into being.

Taiwan knows how to derail such narratives. In the 1980s, technology-intensive industries were widely seen as impossible to develop on our island of resilience. Today, we have TSMC. And at the beginning of the pandemic, some predicted that Taiwan would not survive the first wave. What later became clear is that many things described as “inevitable” were merely scripts that seemed plausible at the time.

When facing a prophecy, what matters most is always how we choose to respond.

If we want to free the future, the first and most important step is looking for collaborators. Whether the semiconductor industry or the pandemic response, it was collective action that overturned the prophecy.

So, is it inevitable that AI will replace human labor?

As infrastructure, AI may lower certain barriers to collaboration as it is used more frequently. But when the AI debate becomes a machine for manufacturing anxiety, the real questions we should ask are: What exactly do we want to use it for? For whom are we building it? And which values must we protect?

This is also why Silicon Valley tech outfits are inviting philosophers to take part in discussions on AI alignment, safety and institutional design. When people form emotional attachments to AI, or begin debating whether machines have consciousness, these are no longer merely engineering questions. They are questions about what kind of society we want to take shape.

If there was only one path into the future, none of these discussions would be necessary. It is precisely because multiple futures remain possible that prophecies deserve to be challenged.

So, whenever you hear that “a trend is irreversible,” the more important question to ask is this: Is it truly a prediction, or is it a command?

Refusing inevitability does not mean rejecting technology. As long as we remain willing to deliberate, govern and build together, the future need not become the extension of a prophecy. It can become something we write and free — together.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

These days, more and more people are exhausted by gleaming AI-polished social media posts. The messaging is emotionally charged, and every sentence seems to make sense. Yet, by the end, nothing sticks in the mind. Information keeps multiplying, as does FOMO. The result is not deeper understanding, but faster burnout.

This disturbing state of affairs is connected to a question I kept facing at Oxford University's recent Civic AI Conference. In conversation after conversation with participants from around the world, I was asked: If mainstream AI is optimized for maximum efficiency and maximum attention extraction, can we choose a different path?

Not long ago, in Dharamshala, India, I was blessed to glimpse a possible answer. Several geshes involved in Monlam AI, scholars trained to the highest level of Tibetan Buddhist philosophy, showed me the model. What makes it valuable is not that it poses as an all-knowing teacher, but that it grows out of a community's long, shared care for language, scripture and the lived context of practice.

The model can help draw more people closer to Buddhist teachings. But the geshes were crystal clear: No matter how well a model writes, it cannot replace lived experience. AI can generate sentences that sound profound, but this does not mean it has actually passed through confusion, discipline and insight.

They also shared a Tibetan folktale. A man is racing along beside his horse, panting, but refuses to mount his steed. Asked why, he said "I'm in a hurry. I don't have time to stop and get on." Many now approach AI the same way, sprinting after every generated output. The moment you start competing with AI on speed, you are not riding the horse. The flow of information is dragging you along.

On this point, the Dalai Lama offers a clear standard: "Ultimately, AI is a tool ... We should use tools not for control, but to improve human relationships."

This also helps explain the quiet impatience so many people feel toward AI-generated content. AI has crudely pried apart rhetorical fluency from what is actually worthy of our emotional investment, turning concern into a hollow performance, a form of care-washing. Learning to recognize that fracture is becoming a core literacy of the digital age.

That is why I keep my smartphone and computer screens in grayscale, so that the world beyond the screen remains more compelling than the screen itself. But personal discipline is only the beginning. At the social level, we need a different kind of AI: not some lofty, all-knowing intelligence, but local, bonded "little helpers" that assist people to listen more closely to one another, repair relationships and then step back.

Instead of falling prey to AI's endless output, we should learn to stop, get on the metaphorical horse and ride. Amid the constant churn of modern attention, we need to take the reins of our tools and reclaim agency over how we connect with real human beings.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

After a February filled with back-to-back holidays, many of us have officially returned to the office. What we might not realize is that while celebrating Lunar New Year, a new generation of AI models quietly broke free from solo-tool roles. They can now "socialize" with one another — dynamically assigning tasks and assembling cooperative teams on the fly.

What this means is that today, in 2026, using AI no longer requires a single command per action. Once you learn to redefine your relationship with these "direct reports," as well as understand how to interact with them, you unlock a staggering level of productivity.

AI is no longer a lone wolf tool. It is a team player that knows how to collaborate. In fact, leveraging AI has become an art of orchestration. Trending technologies like OpenClaw and Claude Code Agent Teams give AI powerful lateral coordination, elevating everyday users into "commanders" of AI squads.

For example, when I need to brainstorm topics for my column, all that is required is to hand off the assignment. The backend automatically assembles five distinct agent teams, each with different expertise: one digs into the archive of my past columns for thematic threads; another tracks the latest international trends; a third combs through scientific journals hunting for technical blind spots; a fourth explores power dynamics in the age of digital democracy; and the fifth handles headlines and polish.

These five AI teams meet, debate and research just like humans. In barely five minutes, they deliver a collaboratively produced report — complete with a detailed trail of their deliberations.

The U.S. government's NIST has already begun drafting guidelines for AI-to-AI interaction. These standards are not just about enabling interoperability among models built by different companies — GPT, Gemini, Claude — they are also designed to guard against the risk of groupthink bias when multiple AI systems interact.

This shift is about to upend the rules of competition in the workplace. In the past, your capacity was limited by how many tasks you could personally handle. Going forward, it will be defined by how many agents you can command.

When you lead an AI team, you need to learn to be a gardener. Start with KPIs — the old mindset needs an overhaul. We used to chase 120 percent stretch goals. Now, "80 percent full" is the sweet spot.

Here is why: in a multi-agent game, pushing any single metric to a perfect score tempts individual AIs to hit their target at the expense of the whole. Your job, then, is that of a gardener tending a garden — maintaining ecological balance, maximizing collective benefit and ensuring the AI team's direction stays aligned with human well-being.

This trend also helps us reclaim what makes us fundamentally human. Once AI handles all the busywork, what remains is the irreducible core of human value: curiosity, collaboration and civic care.

AI can never do curiosity, nor can it build genuine human bonds. That is on us.

Given this trajectory, your AI goal for the year should not be chasing maximum performance metrics. It should be learning how to be a good enough gardener — guiding the AI team to follow the rules, fit the context and serve the mission, all while you focus on judgment and human connection. This kind of Civic AI is the mindset most worth carrying back to work.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

The International AI Summit is set to convene in India this February. As the former back office of the global software industry, India currently finds itself in a position fraught with contradiction and tension. This is not merely the growing pains of a single nation's transformation, but a microcosm of the challenges facing all middle powers.

Over the past three decades, India leveraged its massive engineering dividend to forge deep ties with Silicon Valley. However, the wave of generative AI is rapidly redefining the value of software outsourcing.

In July 2025, Indian software giant Tata Consultancy Services (TCS) initiated the largest workforce restructuring in its history, laying off over 12,000 employees (approximately 2% of its workforce). This was not merely a signal of recession, but more like an alarm bell for industrial upgrading. The traditional outsourcing model is crumbling, and enterprises must free up resources to pivot toward high-value AI solutions. This wave of layoffs demonstrates India's struggle to complete the "elephant's pivot" before the old model collapses entirely.

In the mobile internet era, the Indian government built its acclaimed digital public infrastructure, "India Stack". Through the Unified Payments Interface (UPI) and Aadhaar identity authentication, India successfully compelled tech giants like Google and Meta to comply with local rules as a condition for market entry.

However, the AI era has introduced new rule-breakers. As users become accustomed to obtaining answers and services directly through AI, the traffic gateways and protocol defenses previously built upon app interfaces face the risk of being bypassed. This is not a failure of the India Stack, but a shift in the dimensions of the battlefield—the fortifications remain, but the enemy is airdropping directly from the "cloud."

Last year, in the absence of top-tier indigenous foundation models, India's startup ecosystem exhibited a pragmatic yet dangerous trend: the widespread adoption of models like DeepSeek. This reflects a common dilemma for middle powers—without domestic computing power support, this remains one of the few strategies to achieve rapid Time-to-market.

But this choice is not without cost. Cybersecurity giant CrowdStrike conducted an in-depth analysis of DeepSeek-R1, noting that the "implicit steering" embedded within the model could constitute a security vulnerability. The report indicated that when test scenarios involved specific politically sensitive groups (such as Uyghurs), the software architectures generated by the model exhibited abnormally high vulnerability rates.

This emergent model bias directly impacts code security and stability. When downstream software developers use these tainted engines, the bias spreads throughout the entire application ecosystem.

Taking India as a lesson, since no middle power can train a top-tier model in isolation, countries should all the more pool computing power and high-quality corpora to establish public models akin to public libraries. Starting with general-purpose models, nations can systematically correct cognitive biases and perform deep alignment with local cultures, thereby possessing their own digital brains without relying on specific capital interests.

Leveraging its chip advantage, Taiwan is well-positioned to actively collaborate globally in promoting "Cooperative AI Sovereignty." Only by mastering our own digital brains can we secure an autonomous and safe space for survival beneath the shadow of AI monopoly.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

For me, the most captivating aspect of gaming has never been exquisite graphics or high scores. Rather, it is those moments of decision that compel me to pause my mouse hand and reflect deeply.

There has been a heated discussion in the gaming world recently, sparked by Dan Houser—the creator of Grand Theft Auto and Red Dead Redemption—who issued a warning about "AI Mad Cow Disease." As a godfather of the open-world genre, Houser holds a deep anxiety regarding the potential flattening of the human experience.

"AI Mad Cow Disease" refers to a kind of model inbreeding. It happens when AI stops learning from authentic experiences and instead repeatedly replicates the synthetic data of the previous generation. It is like the parable of the blind men and the elephant: the first generation touches a leg and believes it to be the elephant; the next generation then hallucinates an entire body based solely on that single leg.

A model mimicking Van Gogh's style might produce an image that looks similar, but it cannot comprehend the feelings of a human being standing on the precipice of a mental breakdown. Over time, such creations drift further and further from a sense of reality.

Recently, a game was disqualified from an award for relying heavily on AI-generated content. This backlash does not reflect a fear of technology, but rather a player's instinctive rejection of "outsourcing the soul."

I do not reject generative AI; it is an excellent assistant. The problem is that when it comes to bestowing meaning, AI falls short.

In the greatest video games, the moments that touch us most deeply are often those where there is no single correct answer. AI excels at finding the path with the highest win rate within a given set of rules, yet it cannot comprehend why different people, facing the same moral dilemma, would make starkly different choices.

But it is precisely because of this ambiguity that games possess meaning, which in turn catalyzes community. Because your choice differs from mine, we have the motivation to share, to argue, and ultimately, to build a connection.

If a game's ending were merely a standardized, high-score answer calculated by an AI, this interaction would vanish.

This is the fundamental reason why games cannot be fully replaced by AI: what we care about is the human connection built through these authentic choices.

This parallels many forms of content creation. An AI-generated song might be catchy, but if there is no genuine life experience behind it—no creator who can be understood or spark curiosity—the relationship cannot be sustained and quickly becomes tiresome. It fails to unearth new meaning with every exploration and share.

Only works that allow us to see one another within a dilemma can—like Arthur Morgan, the protagonist of the classic Red Dead Redemption 2—live on in the hearts of countless players even years later, triggering reflection across generations and creating true meaning.

If we position AI as a bridge between people, it can be a wonderful facilitator, like a warm, inviting bonfire. But if you place it in the spot where a human should stand, that is the onset of "Mad Cow Disease"—a spark likely to become a forest fire that destroys interpersonal trust, burning beyond all control.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Recently in Taiwan, influencers have discovered their photos stolen and used to create clone accounts with fabricated life stories. As AI-generated video and audio tools become widely accessible, the barrier to identity theft has essentially vanished. Now, a single photo is all it takes to send someone's digital clone traveling the world—or even speaking on camera.

What's particularly striking about this wave of fraud is who's being targeted: not just celebrities, but "micro-influencers" with fewer than ten thousand followers.

In Taiwan, people have developed strong defenses against the "long-lost classmate" who suddenly asks to borrow money. But when it comes to small-time influencers who share daily life updates and occasionally chat via DM, our guard drops. Bad actors exploit this vulnerability, using AI clones to groom accounts over time and build a genuine sense of connection. When an account you've followed for three years—one that seems authentically human—suddenly starts discussing investment opportunities or weighing in on political issues, it doesn't register as advertising or manipulation.

This is the most dangerous application of AI clones: not just financial fraud, but narrative engineering. The real threat isn't any single piece of misinformation—it's the use of countless clones to manufacture the appearance of organic public discourse. When multiple accounts you recognize all start discussing the same topic, you mistake coordinated messaging for social consensus.

We need systemic change on the scale of how we solved the spam email problem: advanced technology that enables content provenance, ensuring that only content genuinely posted by a person carries an unforgeable digital signature. Laws should also require platforms to open non-personal data to third-party auditors who can identify coordinated bot networks.

Yet crisis often brings opportunity. If we can build the infrastructure to verify authentic identity, AI clone technology might transform from a tool of deception into a bridge for communication.

If scammers can use clones to bypass our psychological defenses, why can't we use "authorized clones" to break through society's echo chambers? This is the concept of the "Shiny Version" clone—a nod to the rare, alternate-colored variants in Pokémon. In the real world, our true selves carry labels—you're "green," I'm "blue." These labels become walls that block dialogue. But if we could create AI clones authorized by our real selves, with different personality settings, they could function as sophisticated "social translators." People on opposite ends of the political spectrum might use gentler "shiny version" clones to engage with communities on the other side.

The recent "We the People 250" AI dialogue experiment demonstrated exactly this: when Americans from opposing camps interacted in a de-labeled environment, they discovered an astonishing 97% consensus. The digital world of the future may become a stage where multiple clones coexist. First, we use technology and law to establish standards for authenticity—then we can confidently deploy clones to facilitate genuine connection. Rather than fighting a painful defensive battle against identity theft, we should proactively master these tools and turn "authorized clones" into opportunities to pierce through prejudice and connect with one another.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Last week, OpenAI released the world's first openly available safety reasoning model: "gpt-oss-safeguard." While it didn't generate the kind of buzz that accompanies each new generation of ChatGPT, it is a consequential step—one that could determine whether AI content services can continue to earn society's trust.

This comes in response to recent, repeated tragic incidents tied to the improper operation of AI models: for example, U.S. teenagers dying by suicide after interacting with AI chatbots, and the widespread use of these tools to generate large volumes of content unsuitable for minors. This has already prompted Australia to ban children under sixteen from using social platforms by the end of this year, with restrictions extending to AI chatbots; simultaneously, the U.S. Senate is weighing similar legislation.

These developments have made AI giants realize that without the capability to implement and respond to safety policies, the entire industry risks facing sweeping regulation.

At the Paris AI Summit in February, I co-launched the ROOST Foundation with Meta's Chief AI Scientist Yann LeCun and former Google CEO Eric Schmidt. The foundation is committed to working with OpenAI and other major players to develop openly licensed safety reasoning models.

We leverage production systems that run daily—such as ChatGPT and Sora—which not only possess advanced semantic understanding but are also adept at handling various attempts to exploit loopholes and jailbreaks. After eight months of development, the release of this open model signals that AI content moderation can move from a "black box" approach toward transparency.

Using the model starts with inputting a safety policy (covering local laws, organizational rules, and socio-cultural norms), followed by the content to be classified. The model then produces a complete reasoning trace: it determines whether the content complies with the policy and explains its reasoning process—without OpenAI imposing any preset stance.

Notably, it can adapt to local contexts. In Thailand, for instance, insulting the monarch is illegal, whereas foreign nationals in other countries are free to discuss the topic.

When the reasoning process can be publicly audited, not only is the frequency of misclassification reduced, but the boundaries of safety are no longer set unilaterally by Silicon Valley. Instead, communities around the world can adjust safety policies to fit their own contexts at any time.

Because any organization can directly adopt, modify, and deploy its own safety system, this also addresses a long-standing challenge for small and mid-sized platforms: relying solely on human moderation is expensive, often leaving minors exposed to risk or allowing illegal content to circulate unchecked.

This model eliminates those dilemmas as excuses. When perceived risks are converted into publicly verifiable safety mechanisms, the industry can avoid being dragged down by a few non-compliant actors—and sidestep extreme measures such as an outright ban on minors using AI.

All of this suggests that content safety is shifting away from the past "arms races" and closed systems toward transparency and collaboration. A more diverse and open AI future is accelerating its arrival.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

In late September, a U.S. presidential executive order set the final terms for TikTok's American operations: its parent company, ByteDance, must divest, and control must pass to a U.S.-led investment group spearheaded by Oracle and Dell. This "Divest-or-Ban" deal has finally been settled.

Foreign-led services are ubiquitous, so why is the U.S. government taking such a hard line on TikTok? The answer lies in its powerful user engagement and its opaque algorithmic "black box."

In recent years, TikTok has consistently held a U.S. market share of over 25%. Its secret is an algorithm that precisely captures user interests from their behavior, then continually pushes highly relevant content. In this model, people receive an endless, personalized stream of information without needing to actively build social connections or follow accounts, creating an addictive and immersive experience.

However, this platform-monopolized "content curation power" also gives TikTok a formidable ability to shape public opinion and collective perception. In the long term, this could intensify social division and polarization; in the short term, it creates a significant risk of opinion manipulation. For instance, if tensions in the Taiwan Strait were to escalate, the platform could flood feeds with content suggesting "a majority of Taiwanese favor surrender," thereby steering public discourse.

Concerned by this influence, U.S. lawmakers reached a rare bipartisan consensus that TikTok is effectively controlled by the Chinese Communist Party and therefore constitutes a national security risk, which ultimately drove this forced divestiture.

Going forward, TikTok's U.S. data will be managed by Oracle. Updates and adjustments to its algorithm will be reviewed by a board of directors composed primarily of Americans, blocking foreign interference at its source.

Notably, this controversy has also clarified an important legal principle: protecting free speech does not conflict with regulating the algorithmic amplification of that speech.

First, anyone—including accounts that praise the CCP—can still express themselves freely on TikTok. That right is protected. The core issue this ban targets is the amplification mechanism: when the platform’s algorithm *proactively* and *at scale* pushes specific, unsolicited content to users who have not subscribed to it.

However, these platform power structures are not unbreakable. The rise of generative AI presents a new opportunity for users to reclaim their "content curation power." Platforms like Bluesky and X are already exploring such applications. In the future, you will be able to give a platform direct commands—"I want to see content like A, but less of B"—and it will instantly generate a new feed tailored to you.

When content recommendation is no longer captive to a single, closed black box but can instead offer a truly personalized model for every user, the market will open up to more diverse options. At that point, the monopoly that short-form video holds on our attention could very well be broken.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

This summer, I was invited to join the Oxford’s Institute for Ethics in AI as a Senior Accelerator Fellow. What concerns me most is this: as AI’s speed vastly outstrips our own, how should future social structures adapt?

The U.K. government is actively introducing AI into policy, education, social welfare, and other public‑service workflows to streamline staffing and cut waste—projected to save roughly NT$1.85 trillion a year. That push has sparked debate. When the public sector uses AI to review cases and grants, people expect it to purge private interests and deliver absolute impartiality. Yet when it confronts human suffering, can it retain warmth—and make decisions that balance compassion, reason, and the law?

When policy chases a single metric, it easily hardens into institutional callousness. If a company’s AI service is poor, consumers can switch providers; but when government power misfires, citizens have far fewer ways to hold it to account.

The Netherlands has already lived through a tragedy of AI governance. The tax and customs authority tried to use algorithms to screen childcare benefits, treating foreign‑sounding names and dual nationality as fraud risk indicators. The result: more than a thousand children were wrongfully removed from their families, and tens of thousands of low‑ and middle‑income households were scrutinized, falsely accused of fraud, and ordered to repay benefits they were legally entitled to.

AI learns extraordinarily well. Given a quantitative target, it often finds quick shortcuts—methods humans wouldn’t think of—to hit that target, without considering the ethical fallout along the way. So the yardstick for AI shouldn’t be “efficiency” alone. It should be trust-under-loss: whether people, even when receiving an unfavorable outcome, can still accept the process as fair.

Social‑platform algorithms are a case in point. Publishers who want to broaden readership may work on better headlines and timely angles. But machine‑learning systems boost controversial posts, amplify ad hominem attacks, and surface baiting content—because the machine, absent ethics, chases traffic. This single‑metric optimization around “engagement rate” ultimately corrodes the quality of public discourse and mutual trust.

Take the Netherlands again. Before cutting a family’s benefits, AI could be required to consult the social workers who support those households and verify the facts. When a system’s goal isn’t just saving money but also sustaining public trust, AI must be tasked with devising ways to earn public consent.

Taiwan’s Judicial Yuan now operates an AI‑assisted Sentencing Information System. In clear‑cut cases, the machine provides a sentencing range; judges then decide whether to mitigate or aggravate. The virtue of this system is transparency. Parties, lawyers—indeed anyone—can try out the calculations. If bias is detected, it’s easy to report, so people needn’t surrender liberty or property to a black‑box standard.

Even so, not everything in human society can be delegated to machines. The sentencing tool also supports Taiwan’s newly launched citizen‑judge system, enabling consensus‑based decisions and lending heavier sentences greater legitimacy. On matters touching life and dignity, human caution and empathy remain irreplaceable.

As AI accelerates, the question isn’t whether to use it but what to use it for. The answer is clear: we don’t need a superintelligence that prizes efficiency above all and replaces humans. We need a collaborative partner that helps everyone, together, amplify our collective intelligence.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Late last month, a large number of Taiwanese became the latest victims of AI-driven scams. According to media reports, this started when groups tagged thousands of Facebook pages in their posts. For over 3,000 of these page owners, the moment they tried to remove the tags, Facebook's algorithm automatically flagged them as accomplices and permanently suspended their accounts.

Behind these accounts are real people who spent years of painstaking effort to build followings in the hundreds of thousands. In the face of new AI-driven scams, social media platforms' indiscriminate crackdown wiped out users' accumulated work overnight.

This incident reveals a stark reality of the digital age: The personal connections and social footprints built on various social media platforms do not truly belong to the users. A single change in a platform's algorithm can cause one’s normal activity to be misidentified as fraudulent or malicious. Platforms can arbitrarily limit one’s reach and reduce traffic, leaving the user powerless.

How to rectify this sorry state of affairs? The answer lies in ensuring digital freedom of movement. Currently, on most social media platforms, users have no sovereignty over their data. The platform decides how one’s information can be used. Even if a user is dissatisfied, switching to a new platform is difficult. One is effectively trapped as it is impossible to take an existing social network somewhere else, and one’s past activity remains locked on the original platform.

Under the lock-in of network effects and switching costs, platforms lack incentives to improve and need not address user demands. They continue to grow simply by leveraging first-mover advantage.

Case in point is the U.S. state of Utah’s recent passage of the Digital Choice Act, which requires social media platforms to achieve "portability and interoperability" for the AI era by July 2026.

Think of it as keeping a phone number when switching carriers. If one wants to move a primary social hub from Facebook to LinkedIn, one could "pack up" accumulated data from Facebook and connect to LinkedIn. Even when a user moves to a new platform, they would not lose their digital footprint.

At the same time, users could choose to have all posts and their resulting interactions on a new platform like LinkedIn automatically update in real-time on Facebook, allowing friends from their old network to follow their latest updates.

This gives users the power to choose the platform with the friendliest interface and the best services. If dissatisfied, they can take their number and leave at any time without worrying about abandoning their network of contacts.

In addition, this creates opportunities for new platforms to emerge. By providing better services and experiences, legions of loyal users can be attracted. The lower switching costs especially benefit smaller, local platforms, allowing them to compete with global giants, even with a smaller user base.

Taiwan's own AI Basic Act, which recently passed preliminary review, calls for the government to establish mechanisms for data openness, sharing and reuse. The goal is to protect user rights and foster data innovation by periodically reviewing and adjusting related laws and regulations.

These platforms operate in Taiwan, using data provided by the people and earning substantial profits from advertising. From the perspective of the digital industry's governing bodies, it is reasonable to amend the law to expect and require these platforms to honor our “digital freedom of movement.” In the long run, this also creates an environment where more platforms can thrive as they are perfectly tailored to the needs of the people.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

As the United States rolls out new tariffs, many nations face a common challenge. In negotiations where traditional sectors are pitted against the tech industry, what is the right path? Should a country prioritize rapid development, sacrificing the interests of a relative few? Or should it pay a higher price to safeguard its long-term national security?

In such balancing acts, perfect fairness and universally popular compromises are rarely achievable. The goal, instead, is to build the broadest possible social consensus and prevent sentiments of unfairness from creating political resistance. But how can governments avoid policy backlash and stay aligned with public opinion? Artificial intelligence may offer a solution.

Following the devastating California wildfires that forced nearly 200,000 people from their homes, the state government used an AI-powered tool to identify the most widely supported recovery measures.

The affected residents were a diverse population, with different racial backgrounds, ages, locations, and economic means. Their priorities for rebuilding varied significantly, making it nearly impossible to satisfy everyone.

In response, the California government piloted "Engaged California," an innovative platform where every individual could voice their opinion. The system then identified priorities with the most overlap, and AI synthesized these into viable policy proposals. Initiatives that proved highly polarizing were advanced to a further round of deliberation.

This approach ensures that while the majority guides the general direction of policy, minority groups have the power to block initiatives that are clearly detrimental to them. By ensuring every group's concerns are heard, this process significantly reduces public dissatisfaction. The design goal is for every policy that is implemented to be moderate and enjoy a broad base of support.

The platform works by identifying different demographic groups and the policies that matter most to each. For example, while a smaller number of people might prioritize mental health recovery—making it difficult for the issue to gain traction in traditional budget debates—this system calculates a consensus metric called the "Ethelo Score." This score ensures that proposals are balanced, weighing the breadth of support against the level of contention, thereby reducing the dissatisfaction of minority groups.

Every measure implemented through this mechanism is first filtered by a wide array of community groups, weeding out extreme options. This prevents the rights of specific minorities from being sacrificed for the majority and stops the agenda from being hijacked by the most strident voices. It guarantees that majority needs are met while preserving the community's diversity and inclusivity. Even when opinions are sharply divided, AI's powerful analytics can pinpoint the sources of disagreement between groups and help refine proposals to increase their acceptability.

For instance, when the California government offered mental health support like free hotlines and counseling, public debate on the division of state and federal responsibility became highly polarized. However, the system's analysis revealed that residents—regardless of party or location—viewed this support as a key priority, even if not their top one. This insight successfully forged a consensus that transcended group divides.

The benefit of this method is twofold: it avoids both top-down, unilateral decrees and a "winner-take-all" outcome driven by the loudest groups. Instead, by using AI to synthesize workable solutions, it fosters a culture of dialogue among diverse communities—spanning age, gender, urban-rural divides, and religion. This empowers society to address conflicts, forge consensus, and build a more resilient and unified civil society.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

This year, Microsoft's announcement of nearly 10,000 layoffs sent shockwaves through the industry. It was jarring because many of those let go were not poor performers; their roles had simply ceased to exist in the face of AI's rapid ascent.

As graduates step out into the world this June, they must navigate a professional landscape in tectonic shift.

Many jobs are being automated by AI, fueling fiercer competition. Seasoned professionals with years of experience are re-entering the job market, vying with recent graduates for an ever-shrinking pool of opportunities. How can these newcomers, with no experience and no network, possibly win?

And yet, I believe the opposite is true. This generation of graduates is incredibly fortunate to stand at the dawn of the greatest productivity revolution since the discovery of fire.

When everyone must learn new workflows from scratch, new graduates are no longer behind. In fact, they have an advantage: they are unburdened by the need to first forget old methods. They can begin directly within the context of AI, making them inherently more adaptable.

For example, to get exceptional results from AI's "Deep Research" capabilities, a user must have the courage to challenge it. If a seasoned professional, acting out of habit, merely treats it like a search engine, they will never unlock its true potential.

In the age of AI, the concept of relying on a single, specialized skill will rapidly become obsolete. To demonstrate their value, professionals must now seek out responsibilities that are difficult for machines to replicate.

Take my own experience. I used to pride myself on my ability to write code and solve problems quickly. But when it comes to pure coding, AI now ranks among the top 200 programmers in the world.

I am thrilled that AI has surpassed my most cherished expertise, because it frees me to focus on what I truly enjoy.

The work of a programmer is to first understand a user's needs, and then write code to meet them. In truth, I always loved the first part of that process: helping clients understand their own needs, even uncovering the needs they had never considered or couldn't articulate. I also loved the final part: delivering a finished product that creates real value. The long, arduous process in the middle—the endless cycle of writing and revising code—is a task I am now delighted to hand over to AI.

This is why, this year, a quarter of Y Combinator's startup teams no longer write their own code, even if they possess the technical skills. They concentrate solely on discovering needs and creating value. Not only does the quality not suffer; in many cases, the outcome is even better.

Of course, this doesn't mean a recent graduate is suddenly on par with a seasoned designer. There are still crafts to be honed. In terms of professional function, experienced people possess a breadth of vision. This perspective allows them to connect disparate fields and draw insights across disciplines, forging an advantage all their own.

So, rather than agonizing over the experience you lack on your resume, begin to take inventory of the passions that make you unique. The issues you follow, the communities you invest in, the fields that fascinate you—these are the cornerstones of your distinct point of view.

In the age of AI, the most valuable asset isn't what you can do, but what you care about. The more you care, the more unique your perspective becomes. This individuality allows you to collaborate with other unique people. And it is at these intersections that you will begin to perceive the pain points that no one else can.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Over the past week, US President Trump and AI industry leaders gathered in the Middle East, with OpenAI CEO Sam Altman's activities drawing particular attention.

In early May this year, Altman proposed a global version of the Stargate program called "OpenAI for Countries," aiming to partner with governments worldwide to build data centers that provide customized AI models and related services tailored to local needs.

Various signs suggest that the first implementation of this global initiative might be in the United Arab Emirates.

One key reason for this aggressive push is the clearly divergent cultural requirements. AI language models developed by Silicon Valley giants primarily use English data. Even when these models can read and write Arabic, their thinking logic and worldview remain dominated by Western perspectives. DeepSeek has also demonstrated that forced alignment to avoid sensitive terms like "Tiananmen" can actually make it trivial for users to detect such restrictions.

Japan's Sakana AI offers an approach worth considering for the UAE. This company trained its proprietary model using texts from the Edo period, starting with the intellectual foundations of the Yamato people. Their model not only masters Japanese but also learns the Bushido spirit, truly reflecting the nation's cultural identity. For Arab countries, cultural alignment is even more crucial given the vast differences in values and worldviews compared to English and Mandarin-speaking worlds.

As customized AI model demands emerge across various countries, this trend benefits OpenAI as well. Previously focused on general-purpose models and reliant on Microsoft for computing power, OpenAI has seen its lead diminish as new models continuously emerge. In fact, its biggest competitor, Google, with capabilities in both software and hardware, offers lower per-unit computing costs.

Currently, OpenAI's main advantage lies in brand loyalty. However, competitors are actively investing to achieve the same computational results at significantly lower costs. As AI becomes more integrated into daily life, whether brand loyalty can withstand cost considerations remains to be seen.

Consequently, OpenAI is actively positioning itself to convert brand recognition into actual stickiness. Their strategy no longer depends on specific cloud providers but focuses on creating customized AI models and data centers based on each country's industrial, cultural, and computing needs.

As AI giants face transformation challenges and become more willing to invest in customization, Taiwan can use this opportunity to reconsider its positioning. Taiwan should learn from Hugging Face and Mistral’s collaboration with various French sectors, combining a vibrant open-source software ecosystem with hardware innovation capabilities. This approach would cultivate talent for fine-tuning AI and integrating software with hardware, creating customized models for various industries and implementing them in manufacturing and service sectors.

By taking the lead in AI integration, Taiwan can not only secure a larger share in the global AI industry chain but also enable various industries to gain competitive advantages and seize the opportunities presented by this AI transformation.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

When President Trump declared sweeping reciprocal tariffs, the announcement dominated headlines. Yet inside Silicon Valley’s tech giants and leading AI labs, an even hotter topic was “AI‑2027.com,” the new report from ex‑OpenAI researcher Daniel Kokotajlo and his team.

At OpenAI, Kokotajlo had two principal responsibilities. First, he was charged with sounding early alarms—anticipating the moment when AI systems could hack systems or deceive people, and designing defenses in advance. Second, he shaped research priorities so that the company’s time and talent were focused on work that mattered most.

The trust he earned as OpenAI’s in‑house futurist dates back to 2021, when he published a set of predictions for 2026, most of which have since come true. He foresaw two pivotal breakthroughs: conversational AI—exemplified by ChatGPT—captivating the public and weaving itself into everyday life, and “reasoning” AI spawning misinformation risks and even outright lies. He also predicted U.S. limits on advanced‑chip exports to China and AI beating humans in multi‑player games.

Conventional wisdom once held that ever‑larger models would simply perform better. Kokotajlo challenged that assumption, arguing that future systems would instead pause mid‑computation to “think,” improving accuracy without lengthy additional training runs. The idea was validated in 2024: dedicating energy to reasoning, rather than only to training, can yield superior results.

Since leaving OpenAI, he has mapped the global chip inventory, density, and distribution to model AI trajectories. His projection: by 2027, AI will possess robust powers of deception, and the newest systems may take their cues not from humans but from earlier generations of AI. If governments and companies race ahead solely to outpace competitors, serious alignment failures could follow, allowing AI to become an independent actor and slip human control by 2030. Continuous investment in safety research, however, can avert catastrophe and keep AI development steerable.

Before the tariff news, many governments were pouring money into AI. Now capital may be diverted to shore up companies hurt by the tariffs, squeezing safety budgets. Yet long‑term progress demands the opposite: sustained funding for safety measures and the disciplined use of high‑quality data to build targeted, reliable small models—so that AI becomes a help to humanity, not an added burden.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Harnessing AI to Foster Social Unity

The Russia-Ukraine war has persisted for over three years, casting a shadow of ongoing turbulence over the international landscape. In Taiwan alone this year, undersea cables have been severed six times, while the frequency of incursions by PRC military aircraft continues to escalate.

Beyond geopolitical conflicts, natural disasters further compound global instability. Earlier this year, wildfires ravaged California for nearly a month, forcing over 200,000 residents to evacuate their homes.

When calamities—whether natural or man-made—strike, governments are often stretched thin, simultaneously managing disaster relief and navigating the blame games between groups divided by political ideologies and values. This not only hampers efforts to mitigate the impact of crises but also deepens societal fragmentation.

Once social trust erodes, external forces find it easier to exploit vulnerabilities during major disasters, sowing further discord.

How, then, can Taiwan proactively build social resilience?

First, we must recognize that modern societies have shifted from vertical to horizontal trust. In the past, people readily accepted authoritative information from governments, academia, or select news outlets. Over the past decade, however, this dynamic has changed. Citizens are less inclined to trust top-down authority and instead seek horizontally validated opinions—such as highly upvoted comments beneath articles—through peer-to-peer exchange.

Strengthening horizontal trust among societal members is thus a critical foundation for resilience. Take the example of California: facing the wildfire crisis, its government drew inspiration from Taiwan’s digital democracy initiatives and launched the “Engaged California” platform. This space allows residents to discuss disaster response, relief efforts, and future prevention strategies.

With a flood of proposals and evolving discussions, the platform employs AI to summarize content, compare perspectives, and identify consensus across diverse groups, prioritizing actionable steps with feedback to the participants. This approach enables the government to focus on widely supported measures, avoiding decisions swayed by isolated loud voices, further motivating civic engagement.

Yet, the rise of generative AI introduces challenges. Malicious actors can easily fabricate countless lifelike fake accounts online, undermining the quality of public discourse. To counter this, Taiwan’s Ministry of Digital Affairs, inspired by California’s public code programs, introduced the “Digital Credential Wallet.” Aligned with international decentralized identity standards, this tool effectively eliminates fake accounts without transmitting data back to issuers—be it government or corporations—thus avoiding unnecessary surveillance.

For citizens, this offers tangible convenience and privacy. In the future, individuals can store personal and membership data on their phones, akin to a digital equivalent of physical cards. They can selectively disclose information—for example, when picking up packages at convenience stores, only a "name" needs to be shown rather than all personal data—and the system works even offline.

As Taiwan builds resilience, AI can provide institutional and technological support. However, the process hinges on individuals actively engaging in their families and friend groups, fostering habits of respectful dialogue despite differences. A society rooted in robust horizontal trust can better withstand information warfare and collectively confront the challenges posed by disasters, both natural and human-induced.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Two weeks ago, I attended the AI Action Summit in Paris, where open models like Mistral and uncensored inference models based on the R1 series (including subsequent iterations like Open R1 and R1 1776), dominated discussions. At the summit’s conclusion, sixty-one nations jointly signed a declaration committing to develop technology in an “open, inclusive, and ethical” manner. To the surprise of many, both the United Kingdom and the United States declined to endorse the pact.

Some misinterpreted this refusal as a willingness to prioritize technological progress over safety. In reality, the opposite is true.

The United Kingdom has long championed AI safety, notably as the host of the world’s inaugural AI Safety Summit. The British government argues that the Paris declaration falls short in addressing the risks of AI weaponization and lacks sufficient, substantive, and clear guidance. In essence, it does not adequately prioritize national security concerns.

In a telling move, the UK recently renamed its AI Safety Institute to the AI “Security” Institute. While both terms translate to “安全” in Mandarin, their implications diverge. The latter signals a sharpened focus on thwarting deliberate attacks, reflecting the principle that “cybersecurity is national security.”

During the Paris summit, U.S. Vice President J.D. Vance issued a grave warning about AI’s weaponization. He underscored how authoritarian regimes are exploiting AI to enhance military intelligence, intensify surveillance, and amplify propaganda efforts, posing a formidable threat to national security. Vance affirmed that the Trump administration will adopt a resolute stance to comprehensively curb such misuse of AI.

In a subsequent address at the Munich Security Conference, Vance turned to the issue of election interference. He cited Romania’s decision to annul its presidential election results amid suspicions of Russian manipulation through information warfare. He posited that if democratic nations can be swayed by mere foreign digital propaganda, it lays bare the fragility of their democratic frameworks. Only through the unfettered expression of citizens’ voices, he argued, can democracy be reinforced.

Vance’s remarks come amid the rapid spread of AI models capable of empowering small actors to launch large-scale attacks at minimal cost—an asymmetric menace akin to an “AI bin Laden.” With AI’s proficiency in coding and crafting persuasive narratives surging beyond human levels, even small-scale adversaries could, with scant resources, orchestrate disruptions like the foreign interference Romania endured.

As the global community stands vigilant against AI-driven threats to democracy, I shared Taiwan’s extensive experience in countering hybrid threats at the Munich Security Conference. I outlined concrete strategies for establishing collaborative defense mechanisms with allies. As the world increasingly recognizes Taiwan as a vital partner in AI safety cooperation, this moment also offers a golden opportunity for our nation’s cybersecurity firms to build on the success of our semiconductor industry and play a pivotal role in the global AI surge.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)

Earlier this month, Meta CEO Mark Zuckerberg announced that the company would discontinue its third-party fact-checking program in the United States and, taking a cue from social media platform X, pivot to a “community notes” model to address misinformation.

Two main factors underpin this decision. First, for years, large numbers of U.S. users have complained about the mechanism’s lack of transparency. In particular, political posts have frequently had their reach limited, fueling long standing dissatisfaction among many Trump supporters.

Even more significant is the stance taken by Brendan Carr, the FCC chair appointed by Trump. Late last year, Carr publicly called out several tech giants, including Meta and Google, essentially stating: “Your platforms enjoy immunity when moderating content because of Section 230 of the Communications Decency Act. This immunity assumes the platform operates in good faith. But if you outsource moderation to specific organizations and keep the process opaque, can that still be called a good-faith operation?”

Following Trump’s election, Meta moved swiftly to adopt a more transparent approach to managing controversial information. The model it hopes to emulate is that of X, which uses a community notes system to navigate the highly polarized topics surrounding the U.S. election.

Under X’s approach, numerous volunteers annotate contentious posts; each volunteer represents a distinct perspective. They upvote or downvote each other’s annotations based on their own viewpoints. Whichever annotation gains broadest approval from opposing sides is pinned beneath the original post—where it remains even if the author of the post disagrees with it.

For a long time, the algorithms behind social media platforms have exacerbated divisions by reinforcing echo chambers. The advantage of this new mechanism is its use of “bridge-based ranking,” which encourages users with differing views—on issues such as immigration or gender—to seek mutual support, thereby enhancing understanding across different communities.

Some may worry that, even if platforms successfully prevent the creation of fake accounts, supporters of a particular viewpoint could still mobilize to manipulate the system. However, X’s experience shows that such efforts have limited impact. After all, to “game” the system, one would need to craft annotations that both sides would endorse. If someone manages to do that, they are effectively promoting intergroup understanding—an altruistic act, rather than disruptive behavior.

In my view, operating a social media platform should not be a matter of counting heads. If the side with the larger “volume” of voices simply wins, then those willing to pay can easily buy all the exposure they want. But earning genuine agreement from the other camp is not something money can buy.

Looking ahead, social platforms should do more than chase engagement and retention. They ought to expand the application of bridge-based ranking and leverage AI to synthesize insights from all sides. That way, even if different camps cannot fully align, they can at least understand where one another’s perspectives come from. Only through such efforts can a polarized society build bridges of communication and move toward a more resilient future.

(Interview and Compilation by Yu-Tang You. License: CC BY 4.0)