Well, it’s on an as-needed basis.

So, these, like Class A agencies that hold the private information for all 23 million people, is going to switch to the security and communication, the T-Road and the Zero Trust architecture, and so on, by the end of next year also. So, the timelines are the same for all ...

So, we need to work, as I mentioned, with all the ministries and agencies that hold similar-sized services that has the personal data of everyone, 23 million people.

Yeah, so we’ve done quite a lot, at least for our ministry and the two administrations. We’re pretty ready for such a scenario. On the other hand, to coordinate is not just our ministry.

The third one is the data portability and backup system to be agnostic of the data centers of our core public services.

We currently use Google for our internal communication because it’s the first to satisfy such requirements. But it’s also on public transcript that we’re working with multiple providers, including Signal and so on, who are all quite willing to consider working with local resilience requirements.

The local resilience?

They need to be portable, because if they’re tied to one specific data center, one specific set of computers, and the location is known, then we can safely assume it’s going to be destroyed.

Like Google, but also Microsoft, Amazon and so on, right? Cloudflare and so on. So, this is the local part. This is the international communication part. This is the local communication part. And also, our software systems, services, and platforms need to also work, regardless of which data centers they ...

By working with the public cloud providers that are domestic in Taiwan and physically operated by people without citizenship.

Right, the communication resilience. And another is this local infrastructure resilience.

So, all three, the data resilience, the local infrastructure resilience and the communication resilience are going to be needed. And without any of the three, we’re not going to coordinate very effectively in terms of response to large earthquakes.

That’s the beginning. And it only solves the international correspondence part. There’s at least two parts as important as that. The part that I keep talking about is the local resilience of the coordination for domestic users. And I will also add to that a good cloud-based backup and high ...

We are going to see a lot of live reports from Taiwan that are completely synthetic. So, for the international correspondence to still communicate to the world and for the domestic coordination to still function, even when all the physical known places are destroyed. That is what we’re preparing ourselves ...

And so, for the communication internally to nevertheless still work and for the international correspondence to nevertheless still have the way to reach to the global audience. Because otherwise, the deepfaked FIMI are going to win the day.

Because all these locations are not secret and we can safely assume Shannon’s maxim: “the enemy knows the system.” Basically, all of these is public information, so we can assume that they’re going to be disrupted, jammed or destroyed in a huge earthquake.

The worst-case scenario is that any physical points of communication that’s fixed in space and known will be destroyed. That’s actually the assumption we’re working with. So, not just the landing sites of the cables, but also the core network of the three major telecommunication providers, also the power supply ...

That’s just the beginning.

Yeah, we’re part of the World Wide Web Consortium, W3C, and the FIDO Alliance, Fast Identity Online. And because this is a global scale problem, it’s not going to be Taiwan only, so, we’re working on standard making to make decentralized identifiers, verifiable credentials, all those zero knowledge, new ways ...

But then that, if implemented incorrectly, will take away pseudonymity on the Internet and that will create a lot of repercussions on civil liberty. So, we’re working on a set of research called Zero Knowledge, meaning that you’re going to be able to prove you’re a citizen or you’re 18 ...

Whereas previously in online forums, we assumed the people we interact with are human until they exhibit bot-like behavior or content. Now, we probably have to shift to a different norm in that it’s assumed to be a bot unless it’s signed in a convincing way.

So, they evolved a set of norms of a strong reliance on digital signatures, on mutually verifiable credentials. Essentially the norm is if you don’t see the entire message, behavior and content digitally signed and notarized in a public way, then you assume it’s fake. So, it flips the default.

But I think we can learn from the actor model from the Web3 community because even before we had all this generative AI stuff, the Web3 community is already saturated in scams. It’s a maximally scamming place. If you lose your wallet, some money, you’re not going to get it ...

And then, the behavior part may have some clues. For example, Twitter recently basically says that if you pay up and then use the financial institutions to do KYC, then this behavior is going to be more trustworthy than people who are not going to pay to get a blue ...

In terms of ABC, Actor Behavior Content, it means that it’s no longer feasible to tell a fake from real by content alone, because the content is even more real than real. That’s what virtual means, actually.

We already see in the terms of scams; the voice cloning scams are very rampant. And what used to take minutes or hours to pre-process now only takes a fraction of a second. And it’s only as a sample of you picking up and saying, hello, and then, hi, why ...

I think it’s going to influence the election.

But these two, of course, may be working together, right? So, the email invites you on a video call with me and the synthetic me started to convince them to do some other things.

And the second, shift in behavior from a broadcast-oriented viral FIMI attack to highly individualized messages. So, these may not be interactive deepfakes, these may just be emails and so on. But the point is that it’s going to be very inexpensive to operate an individualized email campaign to tens ...

So, the language model is going to empower two things. One is interactive deepfakes. So, voice and video cloning that are very convincing and that can interact with you instead of being canned or pre-cooked. That’s one.

So, I think if the society can improve coordination, to produce very quickly this kind of principles and apply it to AI, then we have a real possibility of doing the defense in a democratic way. And I think that’s the way forward.

Except the human for open AI used to be people from Kenya and so on, that helps aligning it. But now there are new techniques like constitutional AI that if you have a clear set of constitutional principles, it automatically aligns the AI to that principle.

There’s a lot of research in this regard, like reinforcement learning by human feedback.

What we need is a way for the society to agree on how those language models should behave, and for those agreements to be directly applied to the language models so that they behave in such a way.

But the even most advanced language models fit on a USB stick, so anyone with just a couple dozen gigabytes of hard disk can run a full-size language model. The whole internet is compressed is very, very small. So, that means that anyone with a personal computer with some GPU ...

So, in my opinion, it’s high time to democratize the safety research. And the nature of this model does have one thing to work in our favor, which is if you want to run an internet-scale search engine that rivals Google at home, it’s not possible. It requires a lot ...

So, I noticed in the hearing just yesterday, many people pointed out this fact. And so, the question now is whether we empower each and every one with this kind of laptop to participate in safety research, to collaboratively develop the way to align such language models. Or do we ...

It’s dwarfed by comparison, by the capacity investment. Ideally, in new technology, you want these two to be balanced for transformative technology at least. But at the moment, because there is an arms race competitively in the capitalist dynamic, this part receives a huge amount of investment, whereas the safety ...

And I think Dr. Hinton made a point in that there’s a lot of investment in even increasing the capability of synthesizing human speech and human interaction. But an investment toward the safety part of the equation of how to steer it so that it fits the societal norms and ...

It’s not great for that. However, if you’re going to use it for scamming, for the con art, and so on, it’s perfect because it creates fake intimacy, fake trustworthiness, and so on. So, by the very nature of the newer generation of language models, it’s going to empower the ...

The answers to that collectively is going to determine a lot of the courses of the century. But just to set the ground, the current generation of language models, they basically create plausible sounding continuations of text, so it’s not great if you want to use it for fact checking.

Yeah, that’s a great question. It’s in fact a question of the century.

Yes.

Previously, FIMI relies on viral spread but it’s also self-limiting in a sense because it’s the same message, just slightly mutated, and people develop antibody to it. But this time around, it’s the same scale, however, completely individualized, and that’s going to be a very different attack vector.

What’s used to take creativity and what is cybersecurity, we will go social engineering, phishing attacks, spare phishing attacks, and so on, cannot be done on a mass scale, yet completely individualized.

But this year, the algorithm is powerful enough so that any laptop that’s good enough can run it. So, interactive deepfakes is going to be a thing.

And it also can run voice cloning and face cloning in real time with non-perceptible latency. That’s to say less than half a second. So, if you video conference with my digital devil, it could be entirely automated. And you will not be able to tell the difference. And that ...

First, any good enough laptop can run what used to take a server render farm to run. So, this laptop, I can run entirely offline a language model that’s as good, even better, than GPT 3.5. And so, I don’t need an open AI internet connection to fully simulate interaction ...

Yes. Well, Gollum did suffer from an earthquake in the movie and in the books. A volcano actually. Anyways, so, my point being, we know the capability is there, but this year is different in two ways.

Conversation.