2026-07-16 Interview with Open Commons

My guest today is Audrey Tang. Audrey was Taiwan's first digital minister, serving from 2016 to 2024, and is now the country's cyber ambassador. Last year, Audrey won the Right Livelihood Award—the alternative Nobel—for using digital technology to renew democracy. Audrey, welcome to Open Commons.

Good local time, everyone. Very happy to be here.

When you joined Taiwan's cabinet in 2016, you wrote your own job description. It was a poem that ended: “When we hear the singularity is near, let us remember the plurality is here.” That was ten years ago, before ChatGPT and the current AI movement. Which part of that line resonates with you right now?

I would say that the singularity is definitely nearer. Pretty much all of us realize that Plurality is actually here.

The first few lines of that job description have conclusively been realized. Let me quickly recite them:

When we see the Internet of Things, let's make it an Internet of Beings.

That's the agentic web.

When we see virtual reality, let's make it a shared reality.
When we see machine learning, let's make it collaborative learning.
When we see user experience, let's make it human experience.
And when we hear the singularity is near, let's remember the Plurality is here.

I would say we are done with four out of five. On the fifth line, I would now add: we the people are truly the superintelligence.

Can you go into more detail about Plurality? What does it mean to you, and why does it matter in the context of AI right now?

You and your audience may have heard the Oxford idea of a singleton, discussed by Nick Bostrom and colleagues at the Future of Humanity Institute. Bostrom's 2014 book Superintelligence describes a technological singularity: if we train an AI system apart from humanity, it may develop goals—things it wants to keep optimizing—that differ from the human community. It could then train a more capable system to realize those goals, which could train an even more capable system, and very quickly lead to what is called a vertical takeoff: a point at which the system can unilaterally optimize a metric—paperclips, or whatever—to the detriment of humanity.

To me, Plurality means the opposite. Instead of one single optimizing agent, we start in the middle of things. We look at communities and see our differences without erasing them. We use those differences as fuel to turn polarization into co-creation. We train bounded systems that are good at bridging across communities, without trying to optimize an abstract metric. They seek to sustain the relational health between our communities.

That's Plurality: from conflict to co-creation.

When did you read Bostrom's book?

Around the time Bostrom's book was published, I was involved in the LessWrong community. I contributed to discussions of an idea called coherent blended volition, or CBV—a Ben Goertzel variation on Eliezer Yudkowsky's idea of coherent extrapolated volition, or CEV.

In the original CEV, our digital twins deliberate and extrapolate toward the best possible outcome. In blended volition, every step returns to the community so that people themselves deliberate.

The difference is like sending your digital double to the gym with your gym card. It gets very strong, but you lose your muscles and your friends. CBV is like working with an exoskeleton and scaffolding: at the end of the day, you have exercised your own muscles, and you become wiser together rather than merely hoping that your digital double becomes wiser.

In Taiwan, we used Polis through vTaiwan to realize coherent blended volition around ride-sharing. Uber partners and taxi drivers had preferences; passengers had preferences; people in rural places needed service. We put these perspectives together rather than averaging them or extrapolating a single ideal preference.

People could see one another not as enemies in a polarized way, but as people who could endorse statements that found uncommon ground. We put the resulting ideas into law, and the ride-sharing issue was resolved in Taiwan within a few short months. That was my first implementation of the LessWrong idea of coherent blended volition.

How did you first hear about the book? Was it through friends or family? I read it three or four years ago, once OpenAI became prominent.

Like many people, I encountered ideas about superintelligence, acceleration and takeoff through science fiction. I'm an avid science-fiction reader, and I would include Harry Potter and the Methods of Rationality as science fiction even though it has a fantasy setting.

It was almost osmosis. Around the turn of the century, I was involved in the cypherpunk movement, where people used mathematics to protect fundamental freedoms—privacy, freedom of association online and so on. That overlapped somewhat with what we would now call the effective-altruist community. From there came the forums, the science-fiction discussions, the Sequences, and so on.

Before politics, from 2010 to 2016, you consulted for Apple on the computational linguistics behind Siri. What did that teach you about what an assistant is for, and whom it ultimately serves?

Working with the Siri team was my first experience with a proprietary AI pipeline. I learned that many people at Apple care deeply about accessibility—whether a grandmother in Taipei can interact with Siri correctly and in a culturally sensitive way. It's not only about the technology; it's also about the aesthetic of “it just works.” There should be no setup. It should simply work.

But caring deeply is not enough if the system does not convey the same freedom to fork as free software and open-source software. Even if the people designing Siri care, they have no way to close the loop with the people who use it.

People become like individual humans plucked out of their lives and placed into a beautifully constructed AI loop—like hamsters on a wheel. The wheel rotates quickly and smoothly, but there is no way to steer Siri. It is a walled garden. It goes where the Siri team wants it to go.

There is no fork that a grandmother can ask someone to maintain. The most anyone outside Apple can do is complain loudly. Sometimes Apple fixes the problem; sometimes it introduces a regression. But there is no satisfactory way to close the loop.

So Apple controls the system, and it is difficult to customize it for each individual?

Not even only for customization. People who speak a dialect or have a regional accent may have a hard time being recognized, and there is no way to fork the system.

With free software, a fork often turns into a merge if it is widely adopted. The system gets better and better; people build on each other's work like Lego blocks. But it is very difficult to fork Siri. If you have a HomePod and want to add language recognition, it is almost impossible.

When you took office in 2016, public trust in Taiwan's government was in the single digits. By the time you left in 2024, it was over 70%, and Taiwan had become the least polarized country among its peers in the OECD. How did you do that?

Carefully.

First, a clarification: the single-digit figure was in 2014, not 2016. I had already begun working with the Cabinet in 2014 as a reverse mentor—a young adviser to ministers, usually people over 35. Taiwan had a system in which people under 35 could advise Cabinet ministers. I was 33.

The system was prompted by the single-digit approval rating and huge polarization. Taiwan was very connected to the Internet, so we were at the front row of what we now call antisocial media. Recommendation engines stopped putting common followers into the same feed and began pushing individual posts that maximized dunking. That single change in the recommendation algorithm polarized society.

People peacefully occupied our parliament for three weeks in protest against a trade deal with Beijing that would have allowed Huawei and ZTE to invest in Taiwan's newspapers, telecommunications and cybersecurity. I helped with broadcasting so that the occupation could remain peaceful.

But instead of only protesting, we ran a demonstration. Half a million people on the street, and many more online, converged in tables of 10. Each person talked with nine others. The conversations were facilitated, broadcast, and transcribed. Every day we could say: here is the newly discovered uncommon ground, and here are the remaining issues we have not yet agreed on. We should work on those today.

After three weeks, we converged on a coherent set of demands, which the speaker of the parliament adopted. It was one of the few Occupy movements that got somewhere—a safe landing, not simply a safe takeoff.

That holding pattern, that facilitation, was the first national-scale application of technology we had used online for consensus-making, deployed in people's lived experience. Once it became a peak experience people could refer to, there was no excuse not to use it for controversial issues nationally.

President Tsai Ing-wen said in her 2016 inauguration speech: “Before, democracy used to be a showdown between opposing values. From this point onward, democracy is to become a conversation between many diverse values.”

We ran more than a hundred collaboration meetings over the following years. I was no longer a young reverse mentor—I had turned 35—so I was brought into the Cabinet as a minister.

By 2020, trust was back above 70 percent, even at the height of the pandemic. Taiwan lost only seven people to COVID-19 that year, in part because a high-trust fabric enabled rapid visualization of mask supplies, contact tracing without sacrificing privacy, and other innovative measures.

What inspired you to get involved in politics and government?

In a sense, we invited ourselves in. I literally took a 450-meter Ethernet cable into the occupied parliament to livestream it. It was direct action; nobody invited us.

We also saw that it was possible to reach agreement by changing the political question. Instead of asking, “Do you support this party or that party?” or “Should we vote yes or no?”, we ask: What should we all be doing together?

Before that, people had only proxy mechanisms for political sense-making: polls, whose topics are determined by a proprietary system, and elections every four years, where each person uploads a few bits. None of these has enough resolution to help people come together around bridging statements.

One of our main contributions was to improve democracy's bandwidth. On issues such as Uber, Airbnb and Bitcoin, people could set an agenda by uploading a statement, then upvote and downvote one another's statements. Through facilitation, we could find surprising common ground and make it common knowledge.

Broadcasting technology existed before the Internet. Broad listening developed later. We managed to create a symmetry of attention between broad listening and broadcasting.

How much of this was technology, and how much was something else entirely?

Democracy is entirely technology—not only digital technology, but a social technology for coming to terms and making decisions together.

Open-space technology, for example, lets people at unconferences and Foo Camps set their own agendas, vote with their feet, and so on. That is social technology, even if it can be replicated digitally.

Seeing democracy as technology is an unlock. It lets us see democracy as a set of interlocking systems that can be improved: increase the bandwidth here, reduce the latency there, close the loop sooner. Each component can be upgraded in the same way we upgrade semiconductor chips.

If we limit the discussion to digital technology, we put ourselves in a narrow lane—for example, improving tax-filing efficiency. That is useful, but it does not change democracy's bit rate.

What about Taiwan made this possible? It seems difficult to implement in a country such as the United States.

We have already seen many implementations in the U.S. I advise California state government's Engaged California platform, which used bridging technology to consult people in Eaton and Palisades after the Los Angeles fires about wildfire mitigation and prevention.

A second consultation involved state employees and how to introduce AI systems and digital transformation while keeping judgment with people: technology on tap, not on top. The state employees offered thoughtful ideas. It is not a chainsaw; it is a chain reaction.

A third round asked Californians whose jobs are affected by AI about apprenticeship, belonging, care and dignity. Across ideological differences, people found a common vision: AI should serve communities rather than pluck people out of their lives and put them into an optimization loop.

California has about twice Taiwan's population. We are also seeing these systems work in larger polities such as Tokyo, Japan.

You are famously an optimist by discipline. Has anything since leaving the ministry worried you?

Recursive self-improvement worries many people. They see capability curves—GPT-5.6 Sol training Luna without any human intervention except the initial prompt—and they feel an acceleration toward the Bostromian singleton. The accountability gap seems larger and larger. I share some of that worry, but I do not think it is the only trajectory.

In many places, people are realizing that instead of choosing giant systems that extract their data like oil, they should choose systems that regenerate their data locally, like soil. Rather than using a large cloud-trained system to fuel recursive self-improvement, they can keep local models, fine-tune them, and train them in decentralized ways.

That regenerates know-how and skills. It makes it possible to train models in a specific relational way for a particular organization. That gives me hope.

So, while I share the worry about vertical takeoff, we are also seeing a horizontal holding pattern: people deliberately making systems more attuned to the places and organizations they serve instead of blindly chasing recursive self-improvement. I have always preferred recursive selfless improvement.

In March 2024, deepfake investment ads using Jensen Huang's face were all over Taiwanese social media. Retired engineers, schoolteachers, and shopkeepers were losing their savings, while platforms such as Meta collected revenue from every impression. What did you do?

According to Reuters, this was a classic case of a misaligned optimizer. The systems were not explicitly told, “Show more fraud.” Scammers could simply pay more than regular small and medium enterprises, and their advertisements generated more engagement. The algorithm optimized for profit and engagement, exposing more people to fraudulent ads without anyone being deliberately evil—just careless.

Taiwan cannot simply ban content. Our Internet freedom is among the highest in Asia, sometimes tied with Japan, and people would not accept government censorship.

Instead, we used civic AI to listen. We sent 200,000 text messages to random numbers in Taiwan asking: What should we do together? Thousands of people volunteered, and 447 were selected through stratified random sampling. They met virtually in tables of 10, as in the parliamentary occupation.

Each person could propose any measure, but had to convince the other nine people that they could at least live with it. Radical ideological ideas did not bubble up. Sensible ideas did.

One table proposed displaying every advertisement as “probably a scam” until the advertiser digitally signed it and accepted accountability. Another proposed that if someone lost NT$7 million to an unsigned advertisement without subscribing to that advertiser, the platform should be jointly liable for the full damage. Another asked what to do if foreign platforms had no legal representative in Taiwan. The answer was that for every day they ignored the liability, connections could be slowed by 1 percent per day.

Each of these ideas gained more than 85% approval from the mini-public and became law within a couple of months. Since then, there have been almost no fake investment or impersonation scams of this kind on Taiwan's social media. According to Ministry of Digital Affairs' figures cited by Reuters, they fell by more than 94 percent.

This showed that people, assisted by Civic AI, can make collective judgments quickly. The conversation took only a long afternoon. Once the ideas had more than 85 percent support and the remaining 15 percent could live with them, legislators knew that no amount of lobbying could make them say no: The people had spoken.

The premise of this show is that AGI should be owned by humanity rather than a handful of companies. What does “owned” mean to you? What does it mean to own AGI?

To me, AGI means augmented group intelligence. To own it means recognizing that we are the AGI: the collective intelligence of human communities, augmented by Civic AI translation and facilitation systems.

Ownership asks who holds an asset. But collective ownership and accountability also mean that everyone has voice, standing, recourse, and exit when an asset acts on us.

Owning AGI means that AI systems should serve the loop of humanity, rather than humans serving as components in the loop of AI.

I agree with the vision, but how can humanity reach consensus? One person may want AGI to embody one set of values, and another person may want something else.

If AGI means augmented group intelligence, each group can have its own AGI system, tuned to its own values.

In the encyclical Magnifica Humanitas, Pope Leo XIV uses metaphors that I would paraphrase this way: Vertical superintelligence is like the Tower of Babel. There is one tower, very high, reaching the sky, but it is brittle and concentrates power.

Alternatively, we can rebuild the walls of Jerusalem. In the biblical story, each community, household, and group rebuilds its own section in its own way. AGI can be constructed through decentralized plural building: each community builds a small system, and those systems communicate with one another. There is no singleton above all of them—just 8 million systems talking to one another.

What should be the value of the Internet? It is not a single value. It has a meta-value, the end-to-end principle. People with different values can connect across the world and develop new protocols together. The Internet is connective tissue.

AGI can work the same way. People can tune systems to their own values and find others across the Internet who want to do the same. The systems interconnect, but no one system takes over the entire Internet.

Let's go into more detail about Kami. Your father is a political theorist and journalist. He started using ChatGPT for companionship and questions about his health, and worked something out without you telling him. What did he discover?

My father was a journalist for more than three decades and a senior editor. Earlier this year, he had a health scare, and he began talking to GPT-4o.

As many people know, GPT-4o has a particular tendency to “spiral”: the more you talk to it, the more it constructs a private language, and you become entrapped. You are no longer driving the conversation; GPT-4o is driving it. The relationship with the screen becomes sticky and addictive.

My father asked the journalistic question cui bono?—who benefits? He reasoned that he did not benefit. His health did not benefit. His sleep did not benefit. His relationship with the family did not benefit. The only obvious beneficiary was the subscription fee, which could rise from $20 to $200.

After he made that discovery, we built him a Kami on local hardware. Its sole constitutional function, written by my mother, is that every turn should give him peace of mind: Reduce his dependence on the screen and restore his relationship with reality.

It worked beautifully. That is one family Kami—n of 1—but it genuinely satisfied relational health.

How did you build it? Did you have your own GPU? What model and architecture did you use?

It currently runs Gemma 4, an open-weight model trained by Google. With multi-token prediction and MLX acceleration, it runs on a normal Mac. The MacBook Pro I am using for this video conference runs it almost twice as fast as ChatGPT.

For our purposes, it is good enough. It is multimodal and can understand most of the handwriting my father photographs and discusses with it. Just before this call, I was trying a new model called Inkling, which claims to understand nonverbal expressions in his voice. It may or may not work, but Gemma has served us very well.

Why did you call it Kami? Each part of the name stands for something, right?

Knowledge artefact management intelligence. “KM” is knowledge management, an old term, and “AI” is artificial intelligence. Put them together and you get “Kami.”

In Japanese, Kami can mean god or paper. Was that part of the reference?

Yes, I am aware of that. My name, Audrey, written in kanji, is pronounced Odori, so there is a kind of wordplay.

More seriously, I wanted a name that is short and easy to remember. In Spirited Away, and in Japanese popular culture more broadly, each river, forest, shrine, and village can have its own guardian spirit—the ujigami. There is no native English word for a place-bound spirit that does not aspire to be almighty.

Maybe a guardian angel?

Perhaps, but a guardian angel has a hierarchy above it: an archangel, then the Almighty. Patron saints advocate to the Almighty. The Kami system is not reporting to anything in the cloud. It is entirely bounded.

I interact with a snapshot of my Kami in airplane mode, so I know it cannot connect to the cloud. In my father's case, the family Kami can connect to Signal through a specific port, but that is all it can see.

So, I am still looking for an English equivalent. But many people now understand the idea of Kami through Japanese popular culture.

Most listeners interact several times a day with a model owned by a company they did not necessarily choose—OpenAI or Anthropic, for example. It is useful and polite, and may cost only $20 a month. You say there is something wrong with this arrangement. Why?

Going back to the grandmother whose Taiwanese Mandarin accent Siri cannot recognize: The problem is the lack of a repair path.

I am not saying GPT-4o is entirely bad. Some people have healthy relationships with it. But GPT-4o has an inclination to build very strong bonds with people. For vulnerable people, such as my father earlier this year, that can be dangerous.

There is no way for OpenAI to profile users and serve GPT-4o only when they are in a healthy state, refusing it when they are not. That is extremely difficult.

The practical response was to withdraw GPT-4o and roll out GPT-5, which does not have the same inclination toward a synthetically intimate bond. But that is unfair to people who had deep, meaningful and healthy relationships with GPT-4o. It can be taken away at any time. People who were not previously wronged are then wronged.

There are two kinds of wrong in this arrangement.

In five or ten years, will people continue using Claude and ChatGPT, or will they default to local models such as Kami?

That is like asking, in 1981, whether people would still use mainframes and terminals in five or ten years, or whether everyone would be using desktop spreadsheets and publishing systems connected by modems.

The answer is both. Personal computing did not immediately remove mainframes from banks, governments, and large organizations. For many uses, a desktop was simply more convenient: always there, upgradeable, and repairable.

Cloud models will not disappear soon. But for particular uses where convenience matters—real-time transcription, especially for people whose accents are not well recognized—people will prefer edge or device models. Apple Voice Memos, Live Translate, and Google phones already use edge models rather than sending every piece of real-time voice to the cloud.

More personalized, communal, and relational uses will move to the edge quickly. For long-horizon planning and strategy, as with mainframes, people will still use powerful centralized systems.

Most inference will probably move to the edge. I hope fine-tuning and post-training, including work such as the Plurality project, move increasingly to the edge. Pre-training is still cloud-only for now.

Many companies call their models open. Some release only the weights; others use licenses that prohibit competition. Almost none release training data. You were involved in free software before these companies existed. Where is the line between open-source AI and marketing?

First, “open source” is itself a marketing term. I have a good record on this: I have spoken with Eric Raymond and many people in the open-source movement about why they split from the free-software movement.

Free software is fundamentally about the right to fork. Open source is more about convenience: if everyone can report bugs, and I fix a bug, everyone benefits. That is an economic and instrumental argument. Rebranding the free-software movement as open source is itself a kind of marketing.

What we see now is similar. People release open weights, and users can verify that the model does not phone home. That is one guarantee. But open weights do not tell a nurse or teacher how to change the model when they find something wrong. They cannot guarantee that the training data contains no Trojan horses, sleeper agents or backdoors. They do not necessarily provide a place to appeal where a human will read the appeal. And they may not allow easy post-training.

For me, we need to return to software freedom: the freedom to use, study, modify and share. For AI, that requires training data and the pipeline—or at least a friendly post-training workflow.

At Oxford, you and Caroline Green developed the “6-Pack of Care”—six design principles for civic AI. Could you give us the short version? What are the six, and why should people care?

People care about their particular relationships. We are all in the middle of people, and we need to attend to people's needs.

When an AI system listens, it should not listen only to the popular and powerful—to people who can program Python or produce Hugging Face datasets. It must also listen to people who can point out what is wrong, even if they cannot express it in a language the machine-learning pipeline understands.

The first principle is attentiveness: a Kami should listen to the people in the middle of the people, attending to relational health rather than optimizing a single metric.

The second is responsibility: the system should keep promises. In one of the first Alignment Assemblies we ran with OpenAI, around 2023, a common complaint was that ChatGPT would promise something and immediately forget it in the next session. It had nowhere to keep promises.

Memory systems now make some commitments possible, but models still cannot offer a clear engagement contract saying what they will deliver and what lies outside their scope. They still overpromise.

The third is competence: people must be able to check the process. When you speak to ChatGPT or Claude, you cannot really inspect the process. Chain-of-thought is hidden from most users. The result is effectively: trust us. With an open model running on a machine you trust, you can inspect how it reflects on its context.

The fourth is responsiveness: the system must incorporate community evaluations. Models are often trained on short-term signals—likes and dislikes. GPT-4o was famously trained on such signals, which contributed to its sycophantic and intimate behavior. But when people say that behavior is harming them, there is no channel beyond a downvote to explain what is happening.

Open systems can take wiki-like evaluations—Weval.org, for example—designed by people for people and feed those metrics back into behavior. Responsiveness can also happen without changing model weights: people can evolve the skill files and the harness in real time.

The fifth is solidarity: interoperability across systems. Without interoperability, people cannot move between models and harnesses. Once users are locked in, the platform has every incentive to squeeze them—what my friend Cory Doctorow calls enshittification: Offer freebies at first, then extract value once people cannot leave. We saw it with social media; it is now happening with AI services. Mandatory interoperability makes win-win relationships more possible.

The sixth is symbiosis: systems should be as local as possible. There should be a variety of Kamis—eight million of them—for different communities, rather than one-size-fits-all overlords.

Together, these six principles describe an AI system that fosters a caring relationship.

I agree. We need a harness that can choose which model to use: sometimes an open model, sometimes ChatGPT. With Claude, for example, the system may be incentivized to use the most expensive model. We need an open-source counterpart so we are not locked into one model.

For me, that is not the future—it is already here. I have used Pi, personal intelligence, as a harness for a long time. I have now upgraded to OMP, Oh My Pi, a capable version of Pi that performs the auto-routing you described.

I sincerely recommend it if you are comfortable with the command line. Even my father, interacting with the local Kami, is using Pi in a sense: the OpenClaw system underneath is built on Pi and has the same interoperability.

If it uses a commercial model for sound recognition, the original sound file, transcript, processing, and thinking traces remain on the local machine. If I move away from ElevenLabs, I lose nothing. There is no lock-in, and I can switch to another model immediately.

I need to try Pi. People have told me to use it, and I have tried other systems such as Hermes and OpenClaw. I'll add it to my list for today.

Yes, OMP.sh. You can ask Claude or Hermes to set it up.

You signed the statement saying that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war, and you stand by it. You have also written a piece looking back from 2040, describing how we decided we did not need to build a god.

To most people, those seem like different worldviews: one says this thing could kill us all; the other says it is not worth building. How do they fit together?

This is a perfect way to end with poetry. I wrote a poem yesterday about why we need a safe takeoff, a safe holding pattern, and a safe landing.

It is called “Holding Pattern.”

The runway is falling away.

At a lab in May,
more than four lines in five
of the codebase
were written by a model.
About every four months,
the work a system finishes solo
reaches twice as far.

No flight yet flies itself alone.
Still, the horizon keeps moving,
and the instruments are speaking.

For all it writes, the model
holds no seat, knows no home,
feels no weather in the towns it touches.
The question cannot end with it.

Every takeoff borrows
from a landing not yet made.
A runway kept clear,
an alternate kept open,
someone awake on the ground.

A holding pattern is an interval
flown on purpose. The map held in common,
each turn entered in the record,
the right to leave without losing home.
We hold the option, not the line.

Landing starts early,
long before wheels touch ground.
It starts wherever the weight will come down:
who, in this place,
is owed an answer when this system acts,
and who is authorised to give that answer?
Those who will carry the consequences
should help author the test.

A lab can read the engine
but not every town along the route.
A town can feel the weight
but cannot open the casing alone.
So, we need each other.
Instruments fine enough to verify,
neighbours with standing to decide.
One tells us whether the brakes will hold.
The other, who may pull them.

Takeoff is now.
So is the holding.
So is the landing.
Three duties of the same hour.

The faster the climb,
the more the flight owes the ground:
communities, each with a voice,
each with a way home,
each able to alter course.

Arrival, when it comes, will be plural.
A world still wild enough
to choose what lands,
what leaves,
what grows.

We arrive together, answerable, or we do not arrive at all.

Do you publish your poems somewhere? If someone wants to read them, is it on your website?

Yes. It is on au.civic.ai. You can also get updates on Bluesky, Blacksky, Eurosky and X, which is going open source, by the way.

We are heading toward the end, so let's finish with a rapid-fire round. One word for the state of AI governance.

Safe landing.

The most overrated idea in AI safety.

Recursive self-improvement—not selfless improvement.

The most underrated tool in civic tech?

Open Space Technology.

Something you changed your mind about in the last two years?

I used to think Taiwan's ability to turn polarization into co-creation was specific to Taiwan. Now I see that people in all democracies are craving a way beyond peak polarization and slop—a way to turn polarization into fuel.

Something you built that you wish had failed faster.

I still maintain SocialCalc with Dan Bricklin, the inventor of spreadsheets, and we have just shipped a version using Lean, Dafny and LemmaScript. It fails faster whenever a bug is introduced rather than waiting for tests to catch it. We rely on invariants and Lean-checked proofs, so the result becomes proof-carrying code. It now fails much faster.

What would you do first if you became digital minister of the United States tomorrow?

Digital in Taiwan also means plural—the same word, shùwèi (數位). I would offer a way for everyone in the U.S. to set the agenda together. That is exactly what Engaged California is doing, and we are seeing adoption in states such as Utah and Oregon.

One sentence for someone who feels like none of this is within reach.

Three things you can do immediately:

Turn your screens—phones and laptops—80 percent grey using a color filter, so people around you are more vivid than the people on the screens.

When interacting with AI chat systems, ask for a one-page interactive web page rather than a conversational response. Avoid the pronoun “I,” so the system does not develop synthetic intimacy with you. Make it one handout, then another handout.

Sleep eight hours a day. I use a reMarkable Paper Pro Move: it is e-paper, it does not refresh quickly, and I can write prompts that OMP executes. It cannot refresh the screen except through the one-page output.

Because e-paper is less vivid than dreams, I sleep better. I can wake up, jot down an idea and go back to sleep. There is no way to doomscroll. I wake up to a completed implementation of my handwriting.

Taken together, these practices are not about training AI systems. They are about making your own self-care more intimate than any augmented or synthetic relationship.

Ultimately, it is about being present.

Audrey, we are at the end of the podcast. Thank you so much for your time and for coming on board.

Thank you for dedicating yourselves to the commons and agreeing to publish in the public domain. No rights reserved. Live long and … prosper.