After that, I’m going to Barcelona for Mozfest and then to Berlin for the Freedom Forum.

So that’s literally what’s next. I’m really thankful for the Ethics and AI Institute here in Oxford for giving me this senior fellowship. I’m actually an “Accelerator Fellow,” which I would nuance by saying I’m accelerating steerability instead of accelerating blind maximization.

Well, literally what’s next is a seminar in the Schwarzman Center of Humanities a couple of days from now, on training machines in the ethics of care. I believe this is very promising because if you just train machines to maximize some number in the utilitarian sense and try to ...

One can be culturally and traditionally conservative without top-down coercion. This is something Elinor Ostrom worked out long ago, and this is just an implementation of some of her ideas.

This is fundamentally conservative, right? It’s about preserving the tradition, the lineage, what matters, what is meaning-making to a certain community. Then we add to that the ability to translate across epistemic norms, across communities, so the wider society becomes more plural. It’s conservative, but it’s not stopping new norms ...

It’s not really a paradox. The anarchism, to me, just means I don’t believe in coercion. I don’t believe in top-down, takedown, shutdown, lockdown-style policy levers. I believe only in voluntary cooperation. But what the voluntary cooperation is doing is not sacrificing the existing community’s norms in the service of ...

That is to say, it attracts the kind of people who otherwise would not engage in policy conversation because they don’t have the vocabulary for it. Instead, they just post some rants and automatically get an invitation into this online conversation. It turns out that this language model speaks their ...

Not at all. Part of the point of working at a Lagrange point, between governments on one side and movements on the other, is not to be pulled into one particular orbit. So I don’t think this is distracting; I think this is actually <em> attracting </em> .

Of course, this is emotional labor if you have to do it by yourself. But because of this exoskeletal language model that I wield, I manage to hug trolls on a kind of industrial scale.

If they have a lot of grievances and gripes, chances are they’re actually suffering somehow from existing policy and just didn’t have the right words for it, or they’re systemically ignored by the political apparatus. So they just use gender or whatever as a way to get my attention. But ...

All the language models, small and large, are pre-trained on my thoughts, basically, and on my A-game. Whenever I receive a trollish attack, my language model helps find the five words within that 1,000-word rant that can be construed as constructive, and then I engage them meaningfully based on these ...

Well, like anything, right? I “hug the trolls.” There are a lot of trolls on the internet, and many of them actually have legitimate grievances. Many of them just make personal attacks, 1,000-word rants, or whatever. I’ve published all my meetings, journalistic visits, and so on for the past ten ...

Each community can talk in whatever words they find comfortable, but they can also reach “uncommon ground” with other communities using different epistemic standards. I think the pronoun “whatever” is one example of this generalized principle: if you work with the community—tech <em> for </em> a community—instead of making some ...

This is why I believe this kind of bridge-making algorithm and bridge-making dictionary is so important. We’re working with what’s called the Green Earth project, which prototypes on Bluesky, but also BitChute, I believe, and hopefully Truth Social, so that each community can have its own conversations, but the tech ...

Well, my pronouns are “whatever,” so this also works in English. I think it’s interesting that many people see online conversations as so polarized because they get caught up on the specific ways words are used. I used the example earlier of “climate justice” versus “creation care.” If you actually ...

After five hours of deliberation, a language model wove together the 45 resonating ideas from those 45 rooms into a core package. We voted on it and showed the parliamentarians that, no matter their ideological position, this core package left more than 85% of people happy and nobody very unhappy. ...

In groups of ten, they deliberated. For example, one room said, “Let’s make sure all ads on social media are displayed as ‘probably scam’ unless digitally signed by someone using a digital signature KYC (Know Your Customer) we can trust.” Another room said, “If somebody loses seven million dollars to ...

In fact, last year when we faced a new wave of deepfake scams in March, we sent 200,000 text messages to random numbers around Taiwan asking one simple question: “How do you feel about this new challenge to information integrity? What should we do about it?” People gave us their ...

Yeah, we believe in actor- and behavior-level liabilities instead of content-level takedowns. We don’t do administrative takedowns; there’s just no censorship. We’re top of Asia, I believe, and maybe one of the top in the world when it comes to internet freedom. But our constitution never protected foreign robots, you ...

Every time we face something new—the pandemic, the infodemic, deepfakes, fraud, whatever—we need to figure out a way that does not backslide on our internet freedom and democracy, but still overcomes that challenge. Every time we do that, people share a peak experience and become even more resilient against the ...

We were on two tracks that were directly opposite. There’s no winner, I guess; you win by default once you turn in that direction. The main challenge Taiwan faced during the past decade—in addition to the polarization attacks from Beijing (for the past 12 years, Taiwan was top of the ...

Well, around ten years ago, both sides of the Taiwan Strait looked at the same phenomena: recommendation engines, parasitic AI that drives engagement through enragement. But we came to very different conclusions. Taiwan, as I mentioned, focused on making the state transparent to the people so we can build common ...

Because of that, it’s just like number portability. If you switch from one telecom to another and cannot keep your number, there’s a lot of misaligned incentive for the existing large player to squeeze you and not offer good service. But as soon as you have portability, you can take ...

One big part is the way people in Taiwan use social media. For example, we’re probably one of the only jurisdictions where there’s not a lot of growth for TikTok or other recommendation engines that amplify outrage and division. In fact, I think we’re the largest segment on the Fediverse ...

Well, I don’t think people are <em> actually </em> polarized. I think we have a high “polarization per minute” (PPM) context, thanks to social media and the 24/7 media cycle. To me, this is more about whether people feel they can still go to some civic space where there’s a ...

In fact, in Finland, before they had the national conversation using this platform, many of their city councils had this online agenda-setting. They commit, every week when they’re having the city council meeting, to reading the consensus. That’s it. As a literally agenda-setting-only power, that is very achievable.

I don’t think this is hard to get at a hyperlocal scale. Of course, it might be harder if you’re talking about hundreds of millions of people, but in a town of 10,000 people, it’s trivial.

Well, I think the only thing one needs for this kind of broad listening is to have a committed listening partner that agrees to respond point-by-point to the results summarized by a language model, in any particular polity. In Bowling Green, Kentucky, they got local developers, museum curators, and so ...

So I think the switch from the individual to the community is entirely feasible. I don’t think this is delusional at all. People are generally waking up to the fact that the only way to get past this polarization—this <em> illusion </em> of polarization—is to put people in the context ...

This kind of forum software is basically saying that it’s much easier to get generative policy ideas if you poll communities instead of individuals—if you aggregate the preferences of people when they’re around other people. And it’s just that. We’re now working with pollsters, such as the Napolitan Institute and ...

Well, I mean, these are just ways for people to make better meeting summaries. These are not some magical tech platform that takes eons and a state to set up. They’re literally just polling mechanisms, right? One insight is that people are more sociable when they’re around other people. If ...

It also enables translation so that those communities, when they’re actually talking about something, can enable this kind of cross-community communication and translation without relying on a single large model. I think we’re now gradually seeing the largest AI companies see that most communities will prefer their locally tunable steering ...

So maybe one community can say, “We care about climate justice,” and only uplift that sort of conversation. Another community can say, “We care about creation care,” and only have full biblical quotes, and so on. It’s not opinionated; you can just have your policy as a text file. The ...

Well, first of all, not all AI companies are proprietary. There are AI companies built upon the idea of open access; for example, Mistral is famously doing that in France. Even OpenAI, which has “open” in its name, started doing open models this year. Just a week ago, we worked ...

It’s putting AI in the loop of humans, instead of putting humans in the loop of AI, which feels like a hamster in a hamster wheel. It just keeps going faster and faster, but the hamster has no control, no steering over the hamster wheel. It is possible to develop ...

I would also say that this is not the only way to train AI systems. In Taiwan, our national development fund, instead of funding one national champion that trains a huge foundational model, rewards more than 150 different smaller models, each knowing what they’re doing for a particular industry or ...

Well, I mean if you mean communism as absolute centralized state control and redistribution, then yes, I think the current development of large transformer models, which cost a lot of electricity to train, is naturally gravitating toward this over-centralization of power.

But we never do humor for humor’s sake. All the humorous communications are linked with a bridging conversation. We know that people on both sides, however polarized, do agree with this bridging statement. The humor was more like a payload to make sure it travels faster than rumors.

Yeah, well, because it turns out that humor travels faster than outrage online. During the pandemic, especially on social media, people live in a very high “PPM” (polarization per minute) environment, and the fog makes it almost impossible to agree on something. But humor just cuts right through it. People, ...

So, if I want to wear a mask and you don’t, I’m just reminding you to wash your hands. Nobody could really push back against that idea, because the science wasn’t very clear which side was winning. We found this one thing that could bridge and translate across the two. ...

So instead, we worked with the “participation officers” (POs) embedded in each ministry and agency. They did a very quick, crowdsourced survey and we found the “uncommon ground”—the one thing that both sides <em> could </em> agree on. The very next day at 2 p.m., we released this meme where ...

Another example: Early in 2020, as we were rationing masks, there were two very dangerous strains of memes making the rounds. One said that because of our SARS experience, only N95 masks are useful; everything else is just placebo. Another strain said any kind of mask actually hurts you, and ...

I’ll use a few examples. There was a 10-year-old boy who called 1922 saying, “You’re rationing masks, which is great, but all I got were pink masks, which is not so great because I’m a boy. I don’t want to wear pink; people will laugh at me.” Instead of the ...

Yeah, in 2020, the presidential approval rating was 71%, and the Central Epidemic Command Center (CECC) was over 90%. The way the CECC earned so much trust is because we gave people a lot of trust. Anybody could call the toll-free number 1922 and say whatever they want about the ...

That’s right.

Or, if they claim they’re over 18 or under 18, they can produce a zero-knowledge proof that proves this fact as an attestation or credential, but they do not need to reveal their ID. There’s no way to dox that person, which leads to much better-quality conversations. It also made ...

Well, I mean, we’ve been doing “meronymity,” or partial real-name, partial anonymity, for quite a while now. On the national platform of participation, join.gov.tw, anybody can collect 5,000 signatures and force a national-level policy discussion, deliberation, or participatory budgets, and so on. But on those forums, if somebody claims they’re ...

I can go on. I think a thousand or so projects from g0v during the past ten years really changed the norm, so that people always ask how you can fork the government so it can merge you back in. This creative, parallel sense of experimentation makes democracy and policy ...

For example, when people criticized the government for not rationing out masks well during the pandemic, they co-created their own mask visibility map system, which was merged swiftly in just 24 hours into the national system. The same goes for contact tracing. People were saying that the old contact tracing ...

It created this notion of, instead of protesting <em> against </em> something, you need to demonstrate <em> for </em> something. So instead of saying, “This doesn’t work,” people are invited to make something that works better—which may or may not always work, but some of them did work.