So, as you can see, this group, this just started, so not a lot of people. 10 people actually occupy a higher area than 37 people. This is dimensional reduction, right? And so, if these people mobilize and vote exactly the same, the number may increase, but the area does ...

But because Polis measures what we call group-informed consensus, which means if you have two groups, and one… let me just show you the Polis conversation. It’s easier that way. We’re actually having a Polis conversation right now around public money, public code.

It’s a marketing email. So, it’s open to everybody, but yeah. It’s open to everybody, it’s self-selection. So, it’s actually local Taiwanese. It’s not a foreign interference, but it’s an interference by a stakeholder saying to everyone.

Right, exactly. Yeah, it’s a sortition, right? A random selection. So, in both self-selections, but guarded by SMS and FidO, or sortition, which is based on a known customer list for statistical representativeness, these two are quite robust. In the Airbnb conversation, what happened was that Airbnb sent an email ...

In the Ideathon which I just showed you, you have to either have a local SMS number, or you can use FidO if you’re abroad, and you’re enrolling, for example. In a week or so, we will send the government SMS, starting from our ministry and six other ministries, from ...

Exactly.

That’s right, that’s right. So, that’s grassroots, that’s not my ministry.

Oh, you did?

Exactly. And the original vTaiwan community, which is not sponsored by the government in any way, recently just partnered with the Turing Institute at Chatham House. And they together formed this interesting democratic alliance that also worked with OpenAI, which is distinct from the Alignment Assemblies that I just mentioned.

Not really.

Exactly. Start such a conversation.

No, it’s grassroots, right? So, vTaiwan is just a bunch of people applying this methodology, like Polis and so on. And it’s not a government project.

Yeah, and part of those edge models can run this kind of crowd deliberation by itself. You don’t even need an outside connection. They can run it in a community, in a workplace.

Yeah.

20 gigabytes will do the job.

But lately, we’ve also worked on Edge AI, which is a language model that compress this kind of collective volition in it, so that it fits in a USB stick. It’s just 10 gigabytes, 13 gigabytes. And then, if you can send that in, it’s like the, I would say, ...

Yeah, I think… so, I thought about this a lot. This is kind of why we have a Department of Democracy network instead of Department of International Cooperation, like every other ministry. Because, first of all, we only work with democracies, but also, we also work with democratic elements in ...

Yeah, so this is my blog. You can just search for Digital Democracy and the Age of AI.

Yeah, of course.

Yeah, sure, sure.

And I have an interactive demo of this in my blog.

Yeah, of course. This is called Talk to the City. You can just take a picture.

Yeah, of course.

I think it’s very tuned now, so it’s really useful.

And so, the AI basically synthesized every argument it can find from the cluster in the latent space. And then, they synthesize a persona to answer this question based on the collective blended volition. So, this is very useful for me as a decision maker.

Yeah, all this is public. And more interestingly, instead of just summarizing, you can actually challenge a cluster. So that you can go into a cluster, and you can say, “can you elaborate?”, “I disagree.”, “Why do you believe in this?”, “What about, I don’t know, mass unemployment?”, and so ...

Yeah, yes. So, this and Polis are two different front ends, and they both are admitted to this conversation. So, we actually run our own as well with our Ideathon, which is part of the Ministry of Digital Affairs. So, this is what people felt, and you can easily summarize ...

Exactly.

Right, exactly. So, once you start voting, you can at any time add your own idea.

Yeah, it’s free text.

This is from some researchers in Iceland. So, when it comes to making an essay for the public, I want to make sure. And then there’s two choices written by your fellow citizens. And what’s important to you, like they’re designed to stay in your box. Or I’m not saying ...

Right.

So, um… but the idea is really quite simple. Each person gets these choices between two equally important… Like this is a wiki survey.

So, one is using all our ideas, which is like Polis, except it’s paired, ranked. Basically, for a person… I can just show it to you right now, why not?

I think Nick is quite interested in this. We had two long discussions last time in Kyoto just a week ago.

I know, I know. It would legitimize a lot of people.

Any community can just have this brainstorming session about what is better speech. And then, they can tune their local Llama models so as to fit people’s expectations.

And it’s been working really well. Now we’re probably going to work with Meta on the third iteration. The Meta one is quite unique, because their Llama model has openly available weights, it means that we don’t have to wait for Meta to retune, like we did for OpenAI or ...

So, yeah. I work with an ex-DeepMind researcher, Saffron Huang, and also an ex-Microsoft researcher, Divya Siddarth. But the three of us kicked this off this March. Feels like a lifetime ago.

Yeah, this is my 20% project.

Yours truly.

Yeah, so basically, we run two experiments. One was OpenAI, one was Anthropic. With open AI, the prompt was in terms of generative AI, what are…

So, we’re very interested in using language model technology to shape the conversation, not just in a moderation kind of way, but rather as an active companion kind of way.

Well, it’s actually our term.

Yeah.

And this AI performs equally well in capability. But in terms of care, like less bias and so on, performs much better. The people’s AI is much better. So…

…language model. That’s how they distinguish themselves from the OpenAI way. So… but now, the constitution is co-created by around 1,000 statistically representative Americans, who’ve done this wiki survey using Polis technology, exactly as you mentioned, a constrained setting. And so, they invite the best bridge making statements to be ...

Anthropic is basically using a language model to supervise.

Yeah, that’s the Kenyan workers I mentioned for OpenAI.