Could you clarify what you mean by “a lot of people don’t?” People certainly engage with each other on instant messengers, social media, and so on. So more or less, everyone looks for information…

Well, I mean, there’s millions of attempts every day, so that’s like the background.

Well, they tried, of course, and saw that we switched to a different defense posture, and it became asymmetrical this time in our favor, right? So, they did not attempt the same amount of denial of service, because they found that, instead of focusing on one server, one computer cluster, ...

So basically, we engage the public crowd, the democracy network to jointly defend against this kind of disruptions, and we’ve been successfully transferring this mode of thinking, this decentralized thinking, to other ministries and so on, so that their website’s data flow too.

After I said that publicly, of course, we got a lot of free testing. [laughter] But still, we remain resilient thanks to our engagement with the web3 community, the decentralized web community. Basically, it means a copy of our website, free of copyrights, can be replicated automatically on everyone who ...

Well, first of all, the stock market didn’t plummet, instead it increased that day. [laughter] So obviously, the society already has some inoculations, antibodies of mind already. And I said publicly to the press that our ministry’s website went online the same hour the drill started, and have not even ...

And when journalists go and check the original website for clarifications, they find it’s busy and difficult to get in. So, there is a kind of strong coordination for the first time between the cyberattack tampering, denial of service and botnets, and disinformation.

But in addition to denying people’s access to those websites by having botnets trying to disrupt the capacity, there were also tampering. The Taiwan rail stations, the advertisement billboards, and so on, were hijacked to display hateful messages against Speaker Pelosi. And also, convenience stores like 7-Eleven that had all ...

So last August, right before our ministry started, former US House Speaker Nancy Pelosi visited Taiwan. And around the same time, the PLA, their military, from the PRC, started military drills with missiles and all that. And at the same time, we witnessed in a single day more than 23 ...

So, I would argue that radical transparency improves societal resilience against the interference of polarization.

And so, publishing the context of policymaking is a great depolarizing tool, because people, for example, civic journalists, all the different stripes and parties, can clearly see what’s actually going on. So instead of having to wait for four years to replace the administration in order to actually get to ...

Yeah, so we talk about societal resilience, which is the society’s ability to defend against polarization, as well as the interference of information that gets people on the extremes, hate each other more, right? So that was the main angle of interference from autocracies to democracies. It’s not pro something ...

There was not a single year where…

Reunify? There’s no reunification.

Re? When was it unified?

And all these things you can read on our website, not just our monthly ministry meetings, but also all the interviews, including this one, and lobbies, visits, and so on, to give a context of the how and why of policymaking, not just the what or the products of policies.

It’s going really well. Now that we have a dedicated Ministry of Digital Affairs in charge of not just open data and co-creation on data-related applications but actually also spectrum allocation, universal services to the rural and remote areas, and so on, platform economy, and even cybersecurity.

Thank you. Thank you so much.

So, my conversation with the OpenAI co-founder, Wojciech, is on public record online, and we talked about pretty much the same topics. So, if on your Google DeepMind, Gemini team, there are experts that are interested in exploring this together, I’m always available for a video conference on the record.

可以，就是隨時都可以聊。

這個是比較接近的風險，因為我有跟三百多個人一起簽 extinction statement，說 AI 就跟 pandemic 或者 nuclear proliferation 一樣，搞不好會變成大概可能大家都只剩幾年可以活的那種狀況。那我們要化解那個狀況，我覺得就要先看我們在哪些領域上最靠近 extinction risk，那目前看起來這種自走的 bot，可能是現有的幾個大的 vector，包括 Biohazard、Infohazard 等等裡面，Cyber 好像是比較接近 extinction 的那一個。

就是 fully stealth。那這一套我有看一些 paper，目前看起來 open source 的 model 已經快要接近可以做這件事情的程度，那其實蠻令人憂心。因為在這樣的情況下，如果是人在執行 kill-chain，你還是可以做一個大範圍的 attribution，你打掉它就好了。但是如果現在是有點像是勒索病毒，它對面其實是機器人，你可能沒有辦法找得到它。

所以等於是一個攻擊手的腦袋，跑在你的電腦裡，那這種型態是最難防禦的，因為從外面看起來是沒有 traffic，幾乎沒有任何的徵兆。

就是說進入你的內網、潛伏取得你的機器之後，什麼都不做，但是就按照實際的狀態開始寫 Zero Day Exploit，但是寫的都不用，就一直存在那台電腦上，一直到整個 Kill-chain 我覺得做好了，再一次全部執行。

對，因為我剛剛講的 Threat Profile 比較是類似 PaLM 的邪惡版本。

這我們了解，包括 SBOM、SPDX 等等。

之後 NICS 資安院也會有一個驗測機制，我們可以在這個機制裡面做一些調整。我想之前很多立委提出來的是說，好像沒有>一個很系統性的方式來解決這個問題。

但是這個可以自動測試，就像我剛剛講的 OpenCC 其實有一張表。所以你們自己拿那張表進去就知道，到底是不是有 localize，理論上不應該只由我們一個一個來反映才對。

Sure.

So specifically, low resource or lower resource languages is something I care deeply about. We’re currently exploring open source as a solution. But if you’re interested in working with us on that, we can also work on that together.

When you green light PaLM. in the training material, PRC materials could vastly outnumber Taiwanese materials. And it’s almost inevitable that it will carry some of the PRC bias on the Mandarin parts, not to mention the Taigi or Hakka parts.

I don’t know whether you have similar ambitions. I understand that in the speech recognition model, there’s a pretty good spoken language support, but I don’t know whether you have similar ambitions in the language model world. We’ve heard many MPs saying that the Google Translate is polluted with PRC ...

So, I think it’s to this end that OpenAI decided to work with Icelandic language communities specifically to preserve that language. And they have reached out to us as well to working on Taigi, Hakka and the 16 indigenous Pacific Islander languages in Taiwan.

And finally, I think is about bias. In Taiwan, we’ve got 20 national languages. PaLM currently doesn’t speak any of the 20. And GPT only speaks Mandarin and not very well at that. So, there is… because I, before entering the cabinet, worked with the Siri team for six years. ...

I don’t have a very good solution to that, and maybe you have thought about that already- a persistent advanced self-writing Red Teaming bot. So, that is something I’m interested to explore together with the Nice and Google partnership.

The other risk I think is more short-term, which is like Red Team AI bots. So, if there is a ransomware that can write itself using the system resources on the target system, then to most of the threat indication mechanisms, they simply don’t exist until they finally decide to ...

And if people do get access to those language models that they can run locally or within the community, it makes innovations on safety easier because then everybody who cares about safety gets to innovate instead of having to wait for the retraining for it to pass your guardrails and ...

And so, the other thing is, here in Taiwan, we do believe in open source, the current generation of language models because they cannot do much harm anyway. If they can do harm, like synthesizing biohazards or things like that, we’re already gone. And the fact that we’re still alive ...

I think there were some solutions being bandied about, like watermarks and so on, which as soon as Facebook releases the llama, it’s gone, right? Because paraphrasing attacks is something everybody can execute locally. So, I don’t think watermark is the way to go. It’s actually counterproductive.

And soon, I think we will be at a stage where all the content that is not verified in this way is assumed to be fake online. We’ll probably get there within a year or so. I think this agrees with the extinction risk paper published by DeepMind recently. So, ...

And so that requires a widespread adoption of digital signature, and by extension, provenance technologies. We need to ensure that any content that’s posted online can be verified as posting, for example, from a resident or a citizen, but that carries its own civil liberty risks. And so, we’ve been ...

Yeah. As I mentioned, anything that can be transmitted over a phone, including passwords, is probably gone. We assume phishing attacks, scam calls, will always succeed. This is just like our local resilience. We assume any communication hardware infrastructure that is public in its location will be destroyed. This is ...

但 Workspace Duet 可以先開沒關係。謝謝。

像我現在所有真正機密的 Workload，我都是在這台 MacBook 上面算，因為它有 96G RAM，所以算起來還蠻快的。

了解，就是一個 Redundancy 的概念。

我了解了。

好，那像我開了 Duet，在 Workspace 裡面用，這個 workload 是在 Workspace 台灣的 region 算的還是不一定？

所以現在彰濱是有一些算力的？

我了解，不然 Logistics 太複雜了。但它是在台灣 Fine Tune 嗎？就是它的 TPU Workload 是送到別的地方去算嗎?

了解，所以我不可能跟你買一塊這個，然後我自己來 fine tune 這樣子。