So, to your question, I think there’s a model around uncommon ground, around bridging, and there’s also another model around the loud extremes. It really depends on whether the space is a broadcasting medium or a broad listening medium. Just as we, in groups of six or ten, would not ...

A couple months ago, I had a hour-long conversation with Laura Loomer, a prominent influencer of the MAGA right. She and her 60,000 live-stream viewers on X and Rumble. We very quickly identified the bridging sentiment as countering authoritarianism without sacrificing free speech and finding this uncommon ground. It’s also ...

Yeah, the surprising validators. Yes. And I think that is really the point. With language models, one can now identify, in a huge amount of text—a rant that is a personal attack, very toxic or whatever—the few words in it that can be construed as constructive and bridge-making. I always ...

Closing the loop—not just the sense-making/listening part, but the closing the loop part—used to be so time-consuming. It was impossible to scale, even harder to scale than the listening part. The listening part, at least you could use some sense-making coding and human ‘small language models.’ But closing the loop, ...

The point here is that they also realized it’s now much easier to close the loop. If you, inasmuch as you had an online conversation or typed in a few words and suggestions, and then the Tokyo Metropolitan Government decides to implement some of them—previously, all they could do was ...

So he used sensemaking tools introduced in the book, like Talk to the City and Polis to crowdsource his campaign. He live-streamed on YouTube 24/7. You could even call into a number to talk to his voice clone and ask questions, or maybe hashtag #TokyoAI to post your suggestion to ...

Yeah, definitely. So I totally agree. And I think the risk-reducing part is now widely understood because if you keep backpedaling, you don’t go anywhere. The risk of vetocracy is well understood. Although the time-saving part, I think, is not well understood. I always stress this, and it certainly resonated ...

Yeah, because in Taiwan’s experience, like 90% of people are ‘maybes’.

Yes, MIMBYs (Maybe In My Back Yard). I love that.

Um, so one clarification though: Engaged California is a general-purpose platform that any ministry within the state government, as well as potentially city-level or maybe even district-level governments in the future, can reuse to introduce their own topics. There are also conversations about enabling, like in Taiwan, a citizen e-petition, ...

So it’s not that we manufactured <em> that </em> urgency, but rather we put people in the mood of figuring out what the short-term urgency is—the thing that if we don’t solve, we stand no chance of solving everything else after that. Kind of like a blocker in the path. ...

Yes. So, by highlighting an urgent aspect—for example, for AI governance, which, as we just learned, is all-encompassing—if you ask people, “What do you feel about AI governance?” you’re not going to get very useful results because it’s just too vague. Instead, last January or a few months before that, ...

It’s like chronic?

You could highlight the most urgent aspect.

However, on the day of the intended launch, the LA wildfire happened. I was in California, so we did not launch the initiative then. On the other hand, the wildfire provided an even more urgent meta-context, which was not just how to build back better when it comes to policy, ...

In California, we’ve been working on the second enabling condition, which was a law passed by their legislature demanding the state government prepare social media codes for young people under 18. This came from a widespread understanding that kids are kind of being trapped in the social media landscape. While ...

The second threshold is that there needs to be “air cover.” Either the governor, the mayor, the city council—somebody needs to pre-commit to close the loop. It needs to be a pre-commitment; that is to say, no matter what result the people want, leadership should at least be accountable in ...

The trade deal with Beijing in 2014 was certainly of this shape. The first generation of AI manipulation algorithms in terms of Uber, social media, and so on was certainly like that in 2015; the pandemic in 2020, of course, was also of this shape. The reason why is that ...

I think there essentially need to be two thresholds, one after another. The first is a widespread sense of urgency about something the British call a “wicked problem”—like there’s not a single actor that can solve it in a Pareto improvement kind of way, but rather, it requires everybody to ...

<hr> </hr>

What do you think about China’s “DeepSeek”? What are its problems?

DeepSeek, as many know, is an open model—like a LEGO brick that anyone can use to build their own towers. When DeepSeek first appeared, there was much publicity claiming it cost only ...

We still believe that in a freer society with a more common understanding base, there can be better checks and balances and less likelihood of harm to the entire society. Conversely, if most people lack the freedom to discuss or even understand the actual situation, with only a few people ...

As for your later question about where this leads when taken to the extreme—while I can’t speak for authoritarian regimes, theoretically, it means fewer and fewer people can make more and more decisions for the entire society. You no longer need to distribute decision-making power to those close to the ...

As journalists, you understand that journalism is a fundamental part of this common understanding. But press freedom in China before 2015 was quite different from the freedom journalists have now. So I don’t agree that high surveillance, maintaining stability, and creating a “clean and harmonious” internet is necessarily the best ...

If you deprive people of this ability, and even decision-makers can’t accurately know what’s happening because you’ve eliminated freedom of speech, you might achieve some safety objectives in certain aspects, but you sacrifice safety in all other aspects because the foundation for safety responses—our common understanding—has been taken away.

Safety has multiple dimensions. Many aspects of safety response rely on society having resilience. Resilience comes from everyone fully and transparently understanding the actual situation, with everyone able to find ways to respond to dangers—whether pandemics or other issues—from their perspective.

This reduced the number of people who could have been warned. From one perspective, perhaps this was “safe” in terms of not causing social panic. But regarding pandemic response, this was extremely unsafe because the situation spread until they had to lock down the entire city of Wuhan before taking ...

I also want to challenge the idea that “monitoring everything is safest.” As we mentioned, during the early 2020 pandemic, there were doctors like Li Wenliang who discovered in late 2019 that he was infected with something that didn’t feel like an ordinary cold but potentially a lethal disease. He ...

They took the opposite approach—making people transparent to the government. Whatever you do, whether expressing opinions or discussing online, the government knows what you’re doing. If your assembly or association freedoms pose a certain threat to the government, the government warns you not to continue such speech. So we can ...

As mentioned, around 2015 was a watershed moment. People could freely express opinions online, but this freedom potentially caused social division. In Taiwan, our approach was to make the government transparent to the people. In this situation, no matter how divided society became, at least people could understand some common ...

How do we defend against such attacks? We still need to use AI. These AI defense technologies are part of what we call the Trusted Technology Industry Chain in Taiwan. We have five Trusted Industry Sectors comprising various trustworthy technologies. These sectors are where Taiwan hopes to concentrate both private ...

Beyond CSAM and fraud, there are many areas where distributed defense helps security. Cybersecurity is another good example. Previously, attacking a computer’s vulnerabilities required professional hackers remotely controlling viruses or trojans. Now, viruses and trojans are becoming smarter and may not need remote control—they can move laterally using the target’s ...

Defensive technologies must be distributed because attack capabilities have become distributed. We call this concept “d/acc” (Defense-Decentralization-Democracy/Acceleration) because open dissemination in these areas doesn’t create weaponization effects—criminals already have these capabilities. So both security-focused and openness-focused individuals can invest in areas that can only be used defensively.

Instead, we need to open-source the ability to detect CSAM, training it into everyone’s phones or personal computers to immediately identify such material. We can also enable people to contribute data to analyze and train next-generation models.

Previously, it was easier to find the perpetrators because only criminal organizations could produce such material at scale. But now, synthetic deepfakes are easily created, making it harder to identify specific criminal groups. Any individual with a laptop or personal computer can mass-produce such photos or videos. This makes it ...

When I was in France, I worked with security-focused Eric Schmidt and openness-focused Yann LeCun to launch ROOST (Robust Open Online Safety Tools). This addresses issues like online exploitation of minors through images or videos.

As I mentioned, I’m most focused on ensuring that while AI spreads horizontally, we can address its harms through distributed methods. For example, the distributed ability for anyone to impersonate celebrities in social media ads to scam people can be countered with distributed electronic signatures verifying authentic identities, treating unauthorized ...

In the future, if some countries truly lack steering ability and rush ahead, causing harm, that harm might remind other countries and societies: “See, you can’t proceed without steering ability.”

The previous problem was that for issues caused by social media ten years ago, our steering wheel took a full decade to correct to a slightly better direction. That’s too slow. If steering could be faster, and our ability to see which aspects of the future are good or bad ...

My position is called “Cyber Ambassador”—“Cyber” meaning “Cybernetics.” Cybernetics is about steering a ship—the ability to steer. With cybernetics—the ability to steer—you can ask everyone: “If we continue in this direction, what are the consequences?” Through sense-making and broad listening, we collectively map out what’s ahead. After mapping it out, ...

As mentioned earlier, this all depends on your “steering” ability. If you have a car with a steering wheel that barely turns, you’ll drive a bit, suddenly hit a wall, or discover a cliff ahead and have to brake quickly, then turn slightly and try a new direction. But with ...

This capability for “pluralistic alignment,” which was previously expensive or time-consuming, is now fast and accessible. So starting this year, I think AI development will move beyond the vertical direction of increasingly larger models to a horizontal direction of increasingly distributed, open models. This doesn’t require people to adapt to ...

In this situation, individuals, companies, or small groups don’t need to wait for large capital or enterprises to train AI to their specifications. They can download many small models and train them themselves. Previously, the main issue was that only large mainframes computed fast enough—on small computers, you’d have to ...

But this year, we’ve seen many exciting inventions—small language models with reasoning capabilities approaching those of large models, perhaps at 90% of their level. In Japan, many companies specialize in these small and medium models, like Sakana AI. They help enterprises that need Japanese translation or spreadsheet software by taking ...

In the past two or three years, we’ve seen much of AI development concentrate large amounts of capital into data centers, training increasingly large models and powerful AI that seemingly can do anything. Most people are expected to just subscribe to these large models made by capitalists, so we’ve returned ...

But the personal computer concept was completely different. Software running on personal computers was “general-purpose.” Your operating system wouldn’t dictate what computations you had to perform. Software like spreadsheets didn’t restrict what you could calculate. Through the free software movement, where anyone could modify software and share it, people in ...

I don’t think people should adapt to AI development; rather, technology should adapt to everyone’s actual needs. When I was young in the 1980s, personal computers (PCs) were just emerging. This was a new concept because previously, only large enterprises or governments could afford mainframes. Everyone just shared a bit ...

Recently, we’ve seen initiatives like “Broad Listening” led by Takahiro Anno and friends in Japan, aiming to make the steering wheel more responsive. Or in California, Governor Gavin Newsom and I created “Engage California” two weeks ago to listen to people’s opinions on rebuilding after wildfires. These broad listening and ...

But you can see it took almost ten years for people to respond to the harm caused by AI. So my feeling is that if we want to contribute to this issue now, it’s better not to just develop AI faster, like pressing the accelerator, nor to stop all AI, ...

Of course, in recent years, people have realized that AI manipulation of human society cannot continue like this. Some places, like Australia, have said that children under 16 shouldn’t have their minds controlled by AI, so they’ve banned them from using social media. Some places require that if you place ...

Moreover, the content people saw—whether ads or network content with more clicks or shares—became increasingly extreme, as if arguments were particularly fierce. What started as finding like-minded people with common interests online ended up with arguments here and there until everyone fragmented. This was the actual situation we saw in ...