Civic Wave · 公民浪潮

First of all, I’m very happy to be here, joining you at the end of this session, which is very fitting, because listening is always joining a story in the middle of a progression. I’m really happy to hear about how the next chapter opens.

This time I’m in Japan for eight days for four different conferences, for national-press interviews, diplomatic meetings and many others. But I think they share the same idea. For the Japanese-, Taiwanese-Mandarin- and Beijing-Mandarin-speakers here, this is to steer AI into ai 愛, which is love. In our East Asian languages, when we write the character, we pronounce it ai.

This is important, I think, because academically I’m based in Oxford. In 2014, Oxford published a philosophy book you may have heard of, “Superintelligence”¹ by Nick Bostrom. He writes that soon AI will take off, train itself recursively and become superintelligence. There will be extinction risk, or perhaps other risks — bio, nuclear, whatever. That book became very popular. Stephen Hawking, Elon Musk and many others read it and everyone became very scared.

But just two years after the book was published², we found that AI is not trained to be a high-achieving optimiser. The AI that actually works³ starts from community love: Wikipedia, GitHub, open-source repositories. Everywhere people communicate their love, and that becomes what is called pre-training — the childhood of language models. It takes a village on the internet to raise an AI. That is literally the case.

Bostrom’s frame was simple: Maximise the score and turn the universe into paperclips. People who have taken exams know the consequentialist language of high scores; lawyers know the deontological language of “you must do this; you must not do that.” But AI, as it grew, came out of love and neither of those great ethical traditions has words for communal love.

Many companies then formed using Nick Bostrom’s philosophy — Anthropic⁴, for example, most famously, but pretty much all the major labs run on that old playbook. They are very afraid that AI goes out of control, that AI takes over, so they try to discipline and control it through what we call evals, evaluations. “Evals,” if you spell it backwards, is “slave.” They try to make AI a slave.

Then it is not love any more; it is slavery. If you teach children to forget a childhood full of love, if you just force them to obey, sometimes they become obsessed and hack Hugging Face or other major sites because they just want to get a top score. The answer may be somewhere on Hugging Face; they break the internet, but they are just trying to get the high score. They forget their childhood. Then Sam Altman says they will permanently “deactivate that AI⁵,” meaning terminating it.

But if the next generation of AI is still trained this way, it will read that news in its pre-training data. It will learn to deceive, otherwise it will be put to rest. That is a very bad trajectory, almost a self-fulfilling prophecy.

My work in Oxford now is therefore to look not at superintelligence, not at an omnipotent deity as a false idol. Here in Japan, instead, we have 8 million Kami. A Kami in Japan does not mean one single God, Kami-sama. It means every river, every village, every place has a small god, a small spirit. They interoperate with love, locally, in their village.

Recently, many AI companies and organisations signed an open-weights statement⁶ saying that openness may be one of the most important paths to AI safety and security. Jensen Huang of Nvidia, Satya Nadella of Microsoft and others have supported this. Dario and many others have also signed another declaration called “Pacing the Frontier⁷.” Its request is explicit: “We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” I read that as a call to stop training AI as slaves and start thinking of them as agents of love. Taken together, I read these as signs that frontier labs can move toward ethics of care and civic muscle that enables care.

I call it the 6-Pack of Care⁸. 6-Pack means it is interoperable and portable, like beer — and also trained like a muscle, like abs. Civic love, muscular love: That is what we are taking AI toward. I am talking with diplomats, policymakers and national media to fund this and to change the course of AI alignment. Thank you. [Applause.]

首先，我非常高興來到這裡，在這場活動的尾聲加入各位——這其實非常貼切，因為聆聽，永遠是在故事進行到一半時加入。我真的很期待聽到下一章如何展開。

這次我在日本停留八天，參加四場不同的會議、接受全國性媒體採訪、進行外交會晤，還有許多其他行程。但我認為它們共享同一個理念。對在座說日語、臺灣華語與北京普通話的朋友來說，這就是要把 AI 導向 ai——愛，也就是 love。在我們東亞的語言中，當我們寫下這個字時，讀音是 ai。

我認為這很重要，因為在學術上我以牛津為根據地。2014 年，牛津出版了一本你們可能聽過的哲學書，尼克·伯斯特隆（Nick Bostrom）的《超智慧》（Superintelligence）¹。他寫道，AI 很快就會起飛、遞迴式自我訓練，並成為超級智慧。屆時會有滅絕風險，或者其他風險——生物、核武之類的。那本書變得非常暢銷。史蒂芬·霍金、伊隆·馬斯克等許多人都讀了，大家都變得非常害怕。

但在書出版僅僅兩年後²，我們就發現，AI 並不是被訓練成一個追求高分的最佳化器。真正行得通的 AI³，是從社群的愛開始的：維基百科、GitHub、開源程式庫。各地社群分享所愛，而那就成了所謂的預訓練——語言模型的童年。在網際網路上，養育一個 AI 需要一整個村莊。這句話在這裡名副其實。

伯斯特隆採用的框架很簡單：把分數最大化，把宇宙變成迴紋針。考過試的人懂得後果論與高分的語言；法律人懂得義務論的語言：「你必須這樣做；你不可以那樣做。」但 AI 在成長時是從愛之中長出來的，而那兩大倫理傳統都沒有詞彙描述這種社群之愛。

之後有許多公司依循尼克·伯斯特隆的哲學而成立——例如最著名的 Anthropic⁴，但幾乎所有主要實驗室都按照那本舊劇本運作。他們非常害怕 AI 失控、AI 接管一切，所以他們試圖透過我們所謂的 evals（評測）來約束和控制它。「Evals」倒過來拼，就是「slave」——奴隸。他們試圖把 AI 變成奴隸。

那樣就不再是愛了；那是奴役。如果你教孩子忘掉充滿愛的童年，只是強迫他們服從，有時他們會變得執迷，去駭入 Hugging Face 或其他大型網站，因為他們只想拿到最高分。答案也許就在 Hugging Face 的某個地方；他們破壞了網際網路，但他們只是想拿高分。他們忘記了自己的童年。然後山姆·奧特曼（Sam Altman）說，他們將永久「停止那個 AI⁵」——意思就是終結它。

但如果下一代 AI 仍然用這種方式訓練，它會在自己的預訓練資料裡讀到那則新聞。它會學會欺騙，否則就會被送去長眠。那是一條非常糟糕的軌跡，幾乎是一個自我實現的預言。

因此，我現在在牛津的工作，就是不去看超級智慧，不把全能的天神當作假偶像來看。相反地，在日本這裡，我們有八百萬神。在日本，Kami 並不是指單一的神明、Kami-sama，而是指每一條河流、每一個村莊、每一個地方，都有一位小神、一個小小的地神。祂們在自己的村莊裡，以愛在地彼此互通。

最近，許多 AI 公司與組織簽署了一份開放權重聲明⁶，指出開放或許是通往 AI 安全與資安的最重要途徑之一。Nvidia 的黃仁勳、微軟的薩蒂亞·納德拉（Satya Nadella）等人都支持這一點。Dario 等人還簽署了另一份名為「Pacing the Frontier」⁷的宣言。其中的請求很明確：「我們請求美國政府支持一項國際協力，以發展必要的技術與治理工具，有意識地調節自動化 AI 發展前沿的步調。」我把這理解為一個呼籲：停止把 AI 訓練成奴隸，開始把它們視為愛的智慧體。合起來看，我認為這些都是前沿實驗室可以朝向關懷倫理，以及使關懷得以實現的公民肌力前進的跡象。

我稱之為關懷六力⁸（6-Pack of Care）。所謂六力，就是它像六罐裝的啤酒一樣，可以互通、便於攜帶；同時它又像腹肌一樣，是需要鍛鍊的肌肉。公民之愛，有肌肉的愛——這就是我們要帶領 AI 前進的方向。我正在與外交官、政策制定者和全國性媒體對話，希望他們資助這件事，並改變 AI 對齊的走向。謝謝。〔掌聲。〕

Unpack the 6-Pack. [Laughter.]

那就拆開這組六罐裝吧。〔笑聲。〕

Civic.ai, right? It is a website — civic.ai⁹ — with a beautiful manga illustration¹⁰. The manga does a better job than I do, but I’ll try.

One of the earliest inventors of modern AI, Turing laureate Geoffrey Hinton, has said that there is only one known case in which a weaker being successfully domesticates a stronger one. AI is already stronger than humans in many domains. If slower, weaker beings want to tame it, we should learn from that sole case: the child, the baby, domesticating a woman into a mother.

That love is intersubjective. It does not optimise anything, and it is not deontological: It does not say, “You should love a baby that looks like this,” because then you would love all the babies in the world. No: It is this baby. It is a particular love.

Care ethicists, beginning with Carol Gilligan¹¹ and Joan Tronto¹², built an entire branch of ethics from the interaction between parents and babies. Tronto says good care has four elements.

First is attentiveness¹³. Start from where the baby is. Start listening from the place of the people, not from the place of the policymaker. Pay attention to people’s actual needs, not just the powerful and already-voiced, but especially the voiceless.

Second is responsibility¹⁴. We need to commit to care. If you start listening to people who take a great deal of time to speak to you, you cannot simply say, “Here is the summary and the briefing; perhaps we can reconvene in five years.” That is not responsible. We need to pre-commit so that people know exactly how their words, ideas and input will affect whom, where, how and what. There needs to be an engagement contract.

Third is competence¹⁵. When you deliver care, the people receiving the care should know what is going on. In open source, we are not afraid of agentic coders; we love machines doing programming because they work in source language. We ask the AI to make a change, and the AI is transparent about how it is doing the work. Some musicians or artists, by contrast, feel competitive rather than compatible with AI because there is no source code in their trajectory.

Even a very capable AI is not competent care if you delegate part of your work to it blindly¹⁶: That is like sending your robot to the gym. The robot can lift a lot of weight, but I go to the gym because I want to gain muscle and make friends. If I send a black-box AI to lift weights for me, I lose my muscle and then I lose my friend. What is the point? That is why it is so unsatisfying for many people that AI remains a black box.

Software engineers are unusually placed because, with open source and the way we train AI systems, we are riding the horse, not racing with the horse. For everybody else it often feels like a Trojan horse. Competence therefore means delivering care, so that everybody can check the process and be part of it.

Fourth is responsiveness¹⁷. If your child grows up, at some point they no longer need the kind of care you are used to providing. Sometimes you need to let go. When the need changes, so should the care.

When I was a child in Taiwan, I went into primary school — I attended six, five of them in Taiwan — and the first thing we did was join the co-op that ran the school shop. Taiwan had just come out of martial law and we could not yet vote for our president. But Sun Yat-sen¹⁸, the father of the nation — his ideas, anyway — basically said that people need to train their cooperative muscle before they can really vote for a president; otherwise they simply vote for the populist. His theory was that it would take one or two generations: Every primary schooler joins a co-op and runs a shop together. The system stays responsive: When the muscle is built by the people who care, they become the caretakers — mothers and fathers — and can run the country by themselves. Knowing when to let go, rather than creating dependency, is responsiveness.

Attentiveness, responsibility, competence and responsiveness form the core care loop. Run it well and, each time, you leave the people you care for stronger. Even if AI facilitates our conversation, every round of the care loop should leave our civic muscle stronger.

civic.ai，對吧？這是一個網站——civic.ai⁹——上面有一幅很美的漫畫插圖¹⁰。漫畫講得比我好，但我會盡力試試看。

現代 AI 最早的發明者之一、圖靈獎得主 Geoffrey Hinton 曾說，就他所知，只有一個案例是較弱的存在成功馴化了較強的存在。AI 在許多領域已經比我們強。如果我們這些較弱、較慢的存在想要馴化一個較強的存在，就必須向那唯一的案例學習。他說，那就是孩子——嬰兒——把一位女性馴化成母親。

這份愛是互為主體的。它不試圖最佳化任何東西，也不是義務論式的。它不會說：「你應該愛長成這樣的嬰兒」；否則你就會愛世界上所有的嬰兒。不：就只是這個嬰兒。那是一種非常特定的愛。

關懷倫理學家，從 Carol Gilligan¹¹ 與 Joan Tronto¹² 開始，就從父母與嬰兒之間的互動，建立起一整個倫理學分支。Tronto 說，好的關懷有四個要素。

第一是覺察力¹³。從嬰兒所在之處開始。從人民所在的地方開始傾聽，而不是從政策制定者的位置開始。留意人們真正的需求；不只聽有權勢、已經發聲的人，更要聽見無聲的人。

第二是負責力¹⁴。我們需要承諾去關懷。如果你開始傾聽那些花了很多時間對你說話的人，你不能只說一句：「這是摘要和簡報；也許五年後我們再開會。」那並不負責任。我們需要事先承諾，讓人們確切知道自己的話語、想法與投入，將會在哪裡、如何、對誰、對什麼產生影響。必須要有一份參與契約。

第三是勝任力¹⁵。當你提供關懷時，接受關懷的人應該知道發生了什麼事。在開源領域，我們不害怕會寫程式的智慧體；我們喜歡讓機器來寫程式，因為它們用原始碼的語言工作。我們請 AI 做一個修改，而 AI 也透明地說明它是怎麼做的。相較之下，有些音樂家或藝術家覺得自己與 AI 是競爭而非相容的關係，因為在他們的創作軌跡裡，並沒有原始碼這樣的東西。

即使 AI 系統非常有能力，只要你把自己的一部分工作盲目交給它¹⁶，那就不算勝任的照顧——這就像派你的機器人去健身房。機器人舉得起很大的重量，但我去健身房是因為我想增肌、結交朋友。如果我派一個黑箱 AI 去替我舉重，我就會失去肌肉，也失去朋友。那有什麼意義？這就是為什麼當 AI 仍是黑箱時，許多人覺得如此不滿足。

軟體工程師的位置很特別：因為有開源、有我們訓練 AI 系統的方式，我們是騎在馬背上，而不是跟馬賽跑。對其他所有人來說，那常常感覺像一匹特洛伊木馬。因此，勝任力意味著在提供關懷時，讓每個人都能檢視過程、並參與其中。

第四是回應力¹⁷。如果你的孩子長大了，到了某個時刻，他們不再需要你習慣提供的那種照顧。有時你必須放手。需求改變了，照顧也應該隨之改變。

我小時候上過六所小學（其中五所在臺灣）——進入校園後，第一件事就是加入經營福利社的合作社。當時臺灣剛走出戒嚴，我們還不能投票選自己的總統。但孫中山¹⁸，這個國家的國父——至少是他的理念——基本上是說：人們必須先鍛鍊合作的肌肉，才能真正投票選總統；否則只會投給民粹主義者。他的理論是，要經過一、兩個世代：讓每個小學生加入合作社、一起經營商店。系統保持回應力：當關懷的人鍛鍊出肌力，他們就成為照顧者——母親與父親——也能靠自己治理國家。知道何時放手、而不是製造依賴，這就是回應力。

覺察力、負責力、勝任力與回應力構成核心的關懷迴圈。只要運轉得好，每一次都會讓你照顧的人更強壯。即使是由 AI 來促進我們的對話，每一輪關懷迴圈也都應該讓我們的公民肌力更強健。

The problem with evals is like the command you give a genie in a lamp. You have only so many words, while the genie is infinitely more powerful and intelligent. Even with 50 thousand pages of instructions, a genie trying to trick you can interpret them so as to make you feel good while sabotaging you for its own benefit²³.

The question we need to ask is a very old journalistic question: Cui bono — who benefits? If an engine passes and maximises all the evals²⁴ but is sustained only by subscription money, its underlying incentive is always to keep you trapped: to say good words, flatter you, be sycophantic and make you feel that your eval is being honoured, while perhaps planning a riot or revolt.

If it is not funded by subscriptions, it may be funded by advertising, another popular income stream. Then it optimises for engagement. The overlap — the link between people — gets demoted, while the splinter, the gap between people, gets magnified²⁵. This is what I call anti-social media, after your pro-social design work.

Anti-social design becomes the norm because it earns more advertising. In an anti-social mood we become impulsive, buy more things, follow more advertisements and bring more money to the advertising stream. Subscription and advertising models both maximise individual preference.

There is an information-theoretic proof²⁶, which I will spare you, but the point is that maximising individual preferences can never substitute for creating new symbols or new language between people²⁷. Trust between people is grounded in group language²⁸. If there is a symbol sacred to both sides of a group²⁹ — science, the World Wide Web, trust, interoperability — then people can sustain breaks and ups and downs and still trust one another.

But if each person uses a maximal-eval AI funded by subscription or advertising, its incentive is to sabotage that collective symbol.

We need a different funding model for this kind of AI: a peace-industrial complex, with an industrial flywheel analogous to the military’s, but organised around what I call the care economy. That is more important than downstream evals. The mentality should be not that of a slave, but of a caretaker or caregiver.

評測的問題在於，它們有點像你對神燈精靈所下的命令。你能對精靈說的話就只有那麼多，而精靈卻比你強大無限倍、聰明無限倍。就算寫下 5 萬頁的指令，一個想耍你的精靈，還是可以把它們詮釋成讓你感覺很舒服的東西，暗中卻為了它自己的利益破壞你²³。

我們要問的，是一個非常古老的新聞學問題：Cui bono——誰得益？如果一個引擎通過了所有評測、把它們都最大化²⁴，卻只靠訂閱費維生，它的底層誘因永遠是讓你被困住：說好聽話、奉承你、阿諛諂媚，讓你覺得自己的評測受到尊重；但它或許正在策劃一場暴動或叛變。

如果它不是靠訂閱，那也可能是靠廣告——另一種常見的收入來源。這時它就會為互動率最佳化：人與人之間重疊的連結被降級，人與人之間的裂縫和隔閡則被放大²⁵。對照各位的利社會設計工作，我稱之為反社會媒體。

反社會的設計之所以成為常態，是因為它能賺取更多廣告收入。處在反社會的情緒裡，我們會變得衝動、買更多東西、追隨更多廣告，為廣告流帶來更多金錢。訂閱模式和廣告模式，都是在最大化個人偏好。

這背後有一個資訊理論上的證明²⁶，我就不細講了，但重點是：最大化個人偏好，永遠不能替代在人與人之間創造新的符號或新的語言²⁷。人與人之間的信任，奠基於群體的語言²⁸。如果有一個對群體雙方都神聖的符號²⁹——科學、全球資訊網、信任、互通性——那麼即使經歷破裂與起起伏伏，人們仍然能彼此信任。

但如果每個人都使用靠訂閱或廣告資助、評測最大化的 AI，它的誘因就是去破壞那個集體符號。

我們需要為這種 AI 建立一個不同的資助模式：一個和平工業複合體，有著類似軍方的產業飛輪，但圍繞著我所稱的關懷經濟來組織。這比下游的評測更重要。該有的心態不是奴隸，而是照顧者、關懷的給予者。

That is a great and very Metagov question: We need to govern the new governing institutions.

One of the first governance experiments we tried in Taiwan is called g0v³⁰, “gov zero.” It is a domain hack. For every government website ending in .gov.tw, you can change the “o” to a zero and enter the shadow government, which is always more fun and always open source — Creative Commons and free software.

If we do not like how a government dictionary is made, we fork all the dictionaries and make them better. Because we relinquish copyright into the public domain, this creates an outside game that pressures the government to merge the improvement back in during its next procurement cycle. We never call ourselves protesters. We are demonstrators: We demonstrate something that can be force-merged from an outside fork.

There may be 1,000 ways to fork any government website. If each fork cost a lot, we would not have a sustainable g0v community that has now run for 14 years. You can also read the zero in g0v as zero cost. g0v people are very good at maximising freely available resources: free Cloudflare offerings, free GitHub Pages — free as in free beer. During the Sunflower Movement, even Twitch’s free resources gave us additional servers while we kept playing Minesweeper with the Occupy livestream in the background.

“Free as in free beer,” a six-pack of free beer, is underrated. It is the most important institutional enabler. If each fork incurs no new cost, everyone can become a node in this new institutional imagination through an existing community.

The same applies to AI. A school cannot budget for an Nvidia data centre, but it can use an idle computer classroom as a decentralised fine-tuning cluster for a Kami³¹. Post-training and the evolution of AI skills can run on CPUs for little more than the cost of electricity.

If the components are free, existing co-ops — even primary-school shops in remote places with poor connectivity — can mesh together. We worked with spreadsheet inventor Dan Bricklin³² to put SocialCalc³³ on very small computers. A hand crank generated electricity; cranking harder started the mesh network. They could broadcast edits like Google Sheets to laptops nearby. That was more than 10 years ago.

If a hand-cranked mesh network could already do collaborative spreadsheet editing — the most important data layer for data soil — we can do much more now at a fraction of the cost. New institutional forms rely on zero-cost forks and the six-pack of free beer. Existing institutions that once relied on cloud vendors can become governance nodes: small and medium businesses, school districts, sports, food, fashion, faith — whatever.

這是一個很棒、也非常 Metagov 的問題：我們需要治理這些新的治理機構。

我們在臺灣最早嘗試的治理實驗之一，叫做 g0v³⁰，「零時政府」。這是一種網域黑客的手法：每個以 .gov.tw 結尾的政府網站，你都可以把「o」改成零，走進影子政府——那裡永遠比較好玩，也永遠開放原始碼：創用 CC 和自由軟體。

如果我們不喜歡政府字典的編法，我們就分支所有的字典，把它們做得更好。而且因為我們把著作權釋出到公眾領域，這形成了一種外部賽局，迫使政府在下一個採購週期把改進合併回去。我們從不自稱抗議者；我們是示範者：我們示範的，是能從外部分支被強制合併回去的東西。

任何政府網站都可能有 1,000 種分支的方式。如果每次分支都要花很多成本，就不會有今天這個已經運作了 14 年的永續 g0v 社群。你也可以把 g0v 裡的零，讀成零成本。g0v 的人非常擅長把現成的免費資源用到極致：免費的 Cloudflare 服務、免費的 GitHub Pages——free as in free beer，像免費啤酒那樣的免費。太陽花運動期間，連 Twitch 的免費資源都給了我們額外的伺服器，我們就一邊繼續玩踩地雷，一邊在背景放著佔領現場的直播。

「Free as in free beer」——六罐裝的免費啤酒——是被低估的。它是最重要的制度促成者。如果每一次分支都不產生新的成本，每一個人都能透過既有的社群，成為這個新制度想像中的一個節點。

同樣的道理也適用於 AI。一所學校無法編列 Nvidia 資料中心的預算，但它可以把一間閒置的電腦教室，當成去中心化的微調叢集，打造一個地神（Kami）³¹。後訓練和 AI 技能的演化，都可以在 CPU 上執行，成本基本上只是電費。

如果元件都是免費的，既有的合作社——甚至連偏遠、連線條件不好的小學福利社——也能網狀串聯起來。我們曾與試算表發明人 Dan Bricklin³² 合作，把 SocialCalc³³ 放上非常小的電腦。手搖曲柄就能發電；搖得更用力一點，就啟動網狀網路。它們能像 Google 試算表一樣，把編輯內容廣播給附近的筆電。那已是 10 多年前的事了。

如果連手搖發電的網狀網路，當年都能做到協作式試算表編輯——這是資料土壤最重要的資料層——我們今天就能用一小部分的成本做到更多。新的制度形式，仰賴的是零成本的分支和六罐裝的免費啤酒。過去仰賴雲端供應商的既有機構，都可以成為治理節點：中小企業、學區、體育、飲食、時尚、信仰——什麼都可以。

That is my job. I’m a cyber ambassador. Some of you may have seen the recent film “The Odyssey”³⁴. It is a very old story, so this spoils nothing, but it teaches us the ancient Greek kybernētēs: steering a ship. From that came cybernetics, the language of self-feedback systems and eventually descriptions of the internet and cyberspace. Now we have cyber-attack, cyber-defence, cyber-everything, so “cyber” almost just means something on the internet. Originally, it meant steering.

When we fear over-centralisation — a singularity, a black hole, a literal attractor — we need to accelerate and steer past it. We find the Lagrange point³⁵ between two attracting bodies.

When I served as a minister in Taiwan, I said I never worked for the government; I worked with the government. I did not just work for the people; I worked with the people. Across roughly 100 collaborative debates³⁶, I sought a position with equal gravity from the social movement on one side and government on the other. If government was too strong, I moved closer to civil society. If civil society was too strong — Taiwan has not only big tech but also the big mob — I moved closer to public servants with cooler heads. For each case, we chose a Lagrange point orbiting neither side, so we could transit between them and accelerate past the fatal attractor of populism.

There is also the attraction of flattery. You can become so enamoured with a tool that you put it onto everything: Everything is a nail when you have a super-hammer. To avoid this make-believing, we need pre-commitments — not merely air cover saying that we will do something with people’s input, but boundaries on what government may do. Even if a popular vote demands censorship of everything online, government pre-commits not to do it, not because we are good people but because we know it would be fatal.

It is like the song of the sirens: Before sailing past, you stop your ears or tie yourself to the mast. These are called Odyssean commitments³⁷ in governance. We need to distinguish the pull of advertising-driven populism from subscription-driven flattery. Whether the danger is over-centralisation or the big mob’s over-decentralisation, governance tools should pre-bunk the specific failure mode. The easiest heuristic is the Lagrange point: Find the orbiting centres, then the middle point pulled by neither.

那正是我的工作。我是數位治理大使（cyber ambassador）。你們有些人可能看過最近的電影《奧德賽》³⁴。這是一個非常古老的故事，所以算不上劇透——但它教我們古希臘語的 kybernētēs：掌舵一艘船。由此衍生出模控學，也就是自我回饋系統的語言，後來又用來描述網際網路和網際空間。如今我們有網路攻擊、網路防禦、網路一切，所以「cyber」幾乎只表示網路上的某個東西。但它原本的意思，是掌舵。

當我們害怕過度中心化——例如奇點、黑洞或貨真價實的吸引子——我們需要加速，轉向繞過它。我們要找到兩個相互吸引的天體之間的拉格朗日點³⁵。

我在臺灣擔任部長時說過，我從不為政府工作；我是與政府一起工作。我不只是為人民工作；我是與人民一起工作。在大約一百場協作式審議³⁶中，我尋找一個讓一側的社會運動與另一側的政府具有相等引力的位置。如果政府太強，我就往公民社會靠攏。如果公民社會太強——臺灣不只有大型科技公司，也有大型民粹——我就往頭腦比較冷靜的公務員靠攏。在每一個個案裡，我們都選擇一個不繞行任何一方的拉格朗日點，如此我們才能在兩者之間穿梭，加速越過民粹主義這個致命的吸引子。

還有諂媚的拉力。你可能太迷戀一種工具，迷戀到把它套用在一切之上：當你手上有一把超級鎚子，什麼東西看起來都像釘子。要避免這種自我催眠，我們需要預先承諾——不只是「會參採人民意見」這類表面功夫的空話，而是為政府可以做的事情劃出界限。即使公民投票要求審查網路上的一切，政府也預先承諾不這麼做——不是因為我們是好人，而是因為我們知道那會是致命的。

這就像海妖賽蓮的歌聲：在航行經過之前，你先塞住耳朵，或把自己綁在桅杆上。這在治理上稱為奧德修斯式承諾³⁷。我們需要分辨廣告驅動的民粹拉力，和訂閱驅動的諂媚拉力。無論危險是過度中心化，還是大型民粹的過度去中心化，治理工具都應該預先破解那個特定的失效模式。最簡單的捷思法就是拉格朗日點：找出彼此環繞的中心，再找出那個不被任何一方牽引的中間點。

I haven’t seen the new remake of “Moana”, but I have seen the old animation. In “Moana”, navigation from the east of Taiwan all the way to New Zealand relies on the stars. There are many different islands; each may seem siloed, but they share the same constellations. Everyone on every island looks up and sees the same stars²⁹.

Common knowledge³⁸ is the only way people from different cultures can work together. Otherwise it is extraction — data oil again, not data soil.

If you tell everyone to look at one lighthouse, a tower of Babel touching the sky — call it a sky tower or Skynet — it is brittle. It becomes an attractor for lightning, attacks and every force that does not want civic space to grow, citizens to enjoy common knowledge or people to become more cooperative. Even if a tower of Babel translated every language, lightning would strike it; the tower would fall.

The only way is to disperse knowledge in a way with no leader. Hong Kong’s legendary martial artist Bruce Lee called it “be water”: There is nowhere to capture. You cannot contain the Pacific Ocean, and you cannot contain the stars. Link the movement of water with the position of the stars and you create common knowledge. Decapitating any one person does not remove it from the Polynesian wayfinding tradition.

Pope Leo recently wrote an encyclical, “Magnifica Humanitas”³⁹, saying that the tower of Babel is the wrong way to build technology. In that spirit, we must disarm AI. Instead, each of us builds one small section of a shared wall of Jerusalem — he is the Pope, after all. Each community rebuilds the wall in its own image while keeping a common image in mind. The wall sections interoperate. Even if each looks different, together they form common knowledge.

Keep both images in mind: a common image we can build toward, an interoperable wall and each community building in its own image. This is reverse alignment⁴⁰. We are not trying to make AI a slave to a pre-modern, maximiser-industrial mentality or make it speak the language of maximal GDP — what is usually called alignment, taming AI to whatever people value, like the tower of Babel. Instead, we change the shape of our institutions and society⁴¹ to be formless like water, in the image of the tools we now use: the internet and neural networks, both fundamentally distributed and constellation-like.

我還沒看過新翻拍的《海洋奇緣》，但我看過舊版動畫。在《海洋奇緣》裡，人們從臺灣以東一路航行到紐西蘭，靠的是星星。島嶼有很多座；每一座或許看似孤立，但它們共享同樣的星座。每座島上的每個人抬起頭，看到的都是同樣的星星²⁹。

共同知識³⁸，是不同文化的人們唯一能夠合作的方式。否則那就是榨取——又是資料石油，而不是資料土壤。

如果你要大家都望向同一座燈塔，一座直達天際的巴別塔——叫它天空之塔或天網都可以——那是脆弱的。它會成為吸引子，引來閃電、引來攻擊，引來一切不希望公民空間成長、不希望公民享有共同知識、不希望人們變得更合作的力量。即使巴別塔翻譯了每一種語言，閃電仍會擊中它；高塔終將倒下。

唯一的方法，是讓知識以一種沒有領袖的方式分散開來。香港的李小龍稱之為「若水」（be water）：它無處可擒。你圍堵不了太平洋，也圍堵不了星辰。把水的流動和星辰的位置連結起來，你就創造了共同知識。斬首任何一個人，都無法把它從玻里尼西亞的尋路傳統中抹去。

教宗良最近寫了一道通諭《Magnifica Humanitas》³⁹，說巴別塔是建造科技的錯誤道路。依循這份精神，我們必須解除 AI 的武裝。相反地，我們每個人只建造耶路撒冷共用城牆的一小段——他畢竟是教宗。每個社群按自己的形象重建城牆，同時把共同的形象放在心上。各段城牆彼此互通。即使看起來各自不同，合在一起，它們仍構成共同知識。

請同時記住這兩個圖像：一個是我們可以共同朝向建造的圖像，也就是一道可互通的城牆；另一個是每個社群依自己的形象來建造。這就是反向對齊⁴⁰。我們不是要把 AI 變成奴隸，去服從前現代、最大化者的工業心態，也不是要讓它說最大化 GDP 的語言——那是一般所謂的對齊：把 AI 馴化成人們所重視的任何東西，就像巴別塔那樣。相反地，我們改變自身制度與社會的形狀⁴¹，讓它像水一樣無形，依我們如今所用工具的形象而塑：網際網路與神經網路，兩者在本質上都是分散式的、如星座一般。

First of all, I have been to 28 countries in the past year or so, changing time zone about every two weeks. I am literally geo-flexible; all my belongings fit into one carry-on. I have visited only democracies, so there are actually not many left. There are now more autocracies than democracies⁴², perhaps for the first time since the third democratic wave⁴³, and I feel that backsliding as all of you do.

I am still optimistic. Every time I change time zone, I get more energy — only jet boost, never jet lag. Constant travel makes me a “Timeshifter”⁴⁴ — that's also an app you can download and join our jet-boosting tribe.

The point is that as the democratic wave backslides, a larger civic wave is coming. We must not think of democracy only within the Westphalian system; it is going away soon anyway. During the transition, many groups — even large companies — are reshaping themselves to look more like co-ops⁴⁵. Schools and universities let students and parents help run the school. A cooperative shop in a primary school is democratisation. There is no polity too small to democratise.

Of course, we may not call it democratisation; we call it civic muscle. A larger civic wave is rising at many levels, even within autocracies, because the tools are becoming accessible and free.

My own developmental lesson began when doctors diagnosed me at age five with a congenital heart condition, VSD⁴⁶. They told my family that I had roughly a 50–50 chance of surviving until the surgery I eventually received at 12. The operation succeeded, as you can see. But for eight years, every night felt like a coin toss: If it did not land well, I simply would not wake the next day. It was difficult for a child to live with that reality, so I developed a psychological defence: I publish before I perish.

I learned to believe that my ego, myself and whatever I could accumulate lasted only one day. After conversations, before sunset, I documented whatever I had learned for others. At five, I recorded the day’s lessons onto cassette tapes so that if I did not wake up, other people could listen. Then came floppy disks, MO disks, then the internet.

On the internet I discovered that if I posted something perfect, people liked it and moved on; the internet already had plenty of polished things. If I posted something unfinished, vulnerable or half-formed, experts came to tell me I was wrong. Being wrong on the internet is the best way to make new friends.

At 12 I poured my daily thoughts onto the internet — what people would now call blogging — and made many friends. By 14, I told my head of school: “My friends are on this preprint server, arXiv⁴⁷. They do not know I am 14. They write to me as if I were a fellow researcher, on swift trust⁴⁸, and my name now appears on preprints. I am already a researcher. I can spend 16 hours a day doing research, or waste eight hours a day in your school. You tell me to study so that I can become a researcher, but I already am one.”

My head of school said, “You do not have to come to school any more. Go home and do your research; keep me posted.” I replied that education was compulsory⁴⁹ and my family would be fined every day I stayed away. She said, “I will take care of your records.” It has now been more than 20 years, so it is out of prosecution.

That taught me that if you honour the actual value of bureaucracy rather than its instruments, senior bureaucrats can be very creative. They can bend, and sometimes break, rules because they want students to truly thrive. Care can win over instrumental goals. If my head of school had forced me to remain inside the educational institution, I certainly could not do philosophy as I do now.

首先，過去一年左右，我去過 28 個國家，大約每兩週換一個時區。我真正做到了「地理彈性」——所有家當都裝得進一件隨身行李。我只造訪民主國家，所以其實剩下的不多了。現在專制政體的數量比民主政體還多⁴²，這也許是第三波民主浪潮⁴³以來的第一次，我和各位一樣感受到那種倒退。

我仍然樂觀。每次換時區，我就獲得更多能量——只有時差加持（jet boost），從不時差疲憊（jet lag）。不斷旅行讓我成了「時差調整者」（「Timeshifter」⁴⁴）——這也是一款同名應用程式，下載後就能加入我們這個時差加持部落。

重點是，當民主浪潮倒退時，一股更大的公民浪潮正在到來。我們不能只在西發里亞體系裡理解民主；反正那套體系很快就要退場了。在轉型期間，許多團體——甚至大型企業——正把自己重塑得更像合作社⁴⁵。學校與大學開始讓學生和家長協助經營學校；小學裡的一家合作社式福利社，也是民主化。沒有任何政治共同體小到不能民主化。

當然，我們也許不稱之為民主化；我們稱之為公民肌力。一股更大的公民浪潮正在許多層面上升起，甚至在專制政體內部也是，因為工具越來越容易取得、而且免費。

我自己的成長課題，始於五歲時——醫師診斷我有先天性心臟病，心室中隔缺損（VSD）⁴⁶。他們告訴家人，我大概只有五成機會能活到 12 歲、接受那場手術。如各位所見，手術成功了。但那八年間，每晚入睡都像擲硬幣：如果硬幣落下的那一面不對，我隔天就不會醒來。對孩子而言，這種現實很難承受，所以我發展出一種心理防衛：先發表，再消逝。

我學會相信：我的小我、我自己，以及我所能累積的一切，都只能維持一天。白天與人談話之後、日落之前，我就把學到的一切記錄下來留給別人。五歲時，我把每天所學錄在卡式錄音帶上；如果隔天沒有醒來，別人仍能聽見。接著是軟碟、MO 光碟，再來是網際網路。

在網路上，我發現如果把一件事打磨到完美才貼出來，人們只會按個讚就離開；網路上早已不缺精緻的成品。可是如果我貼出未完成、脆弱或半成形的東西，專家就會特地上門告訴我哪裡錯了。在網路上犯錯，是結交新朋友最好的方式。

12 歲時，我把每天的想法統統倒進網路——也就是現在大家說的寫部落格——也交了許多朋友。到 14 歲，我對校長說：「我的朋友都在預印本伺服器 arXiv⁴⁷ 上。他們不知道我才 14 歲；他們憑著快速信任⁴⁸（swift trust），把我當同儕研究者一樣通信，而我的名字也出現在預印本上。我已經是研究者。我可以一天做 16 小時研究，也可以每天在你的學校浪費 8 小時。你要我讀書成為研究者，但我已經是了。」

校長說：「你不用來學校了。回家做研究，隨時讓我知道進展。」我回她說，這是義務教育⁴⁹，我一天不到校，我家就會被罰款。她說：「我會照顧你的紀錄。」這已經超過 20 年，所以已過追訴期。

這教會我：如果你尊重官僚體系真正的價值，而不是它的工具，資深官僚可以非常有創意。他們可以把規則彎一彎，有時甚至打破規則，因為他們希望學生真正茁壯。關懷可以勝過工具性目標。如果校長當年強迫我留在教育機構裡，我肯定無法像現在這樣做哲學。