2026 年 9 月 1 日，Evermay 莊園，Neno Battaglia 的 Future Fellows 首講《Designing a Language Model to hold Moral Complexity across Cultures》。這一頁是 Tenzin Yangtso（與 jdd 協作）聽講後的對話札記。
Warm Data Majorana 資料土壤 公共性
Gregory Bateson 是二十世紀的人類學家與控制論先驅，留下幾個關鍵概念：「聯結的模式」「資訊是一種產生差異的差異」「地圖不是疆域」。他的女兒 Nora Bateson 繼承這條線，並提出一個診斷：現代人真正的大問題都是跨語境的——同時長在家庭、生態、經濟、教育、文化、科技、政治的交疊處——但我們的思考工具全是單語境的。
她把被切開的、去脈絡的量化資料叫冷資料；把剩下的收進一個詞：Warm Data——跨語境的、關係性的、「活著」的資訊。它存在於語境之間，而「之間」恰恰是不被制度化的地方。
Neno 在演講中段講了一個全場最重的故事。主角是物理學家 Ettore Majorana——1938 年搭船途中憑空消失的義大利天才，中微子理論的先驅。
有人相信 Majorana 消失，恰恰是因為他直覺到某種可能性——他不怕自己的研究被用於原子彈，他怕的，是作為一位傑出的物理學家，去貢獻那台「全景監視機器」（panopticon machine）。他消失的時候，George Orwell 還沒寫《1984》——他真的明白事情正走向哪裡。
補記・消失與遺作：Majorana 是 Fermi 羅馬「Via Panisperna 學派」的核心天才；該團隊 1930 年代即用中子轟擊鈾，卻將異常產物誤判為「超鈾元素」，錯過了核分裂的發現（1938 年 12 月才由哈恩與斯特拉斯曼完成）——而 Majorana 已於同年 3 月搭上那班從那不勒斯到巴勒摩的船。遺書寫著「請不要穿黑色⋯⋯在你們心裡記得我，原諒我」；電報「大海拒絕了我，我明天回去」——然後徹底失去蹤跡。他從未參與分裂的發現，更未參與曼哈頓計畫；「他的研究被放進原子彈」是回溯性的推測——作家 Leonardo Sciascia 認為，以他的天才，可能早於眾人看見核鏈式反應的軍事潛能，為此恐懼而選擇消失。旁證：曼哈頓計畫期間，Fermi 幾度卡關時會說「要是 Ettore 在就好了」（Oppenheimer 回憶過這句話）。
「全景監視機器」則是 Neno 的詮釋框架，而非 Majorana 的原話：他的遺作《統計定律在自然科學與社會科學中的價值》（1942 年由姐姐出版）主張統計方法可施用於社會科學——Neno 將它連到 Bentham 的全景監獄概念：當統計被系統性地施用於社會，「測量一切」就變成讓所有人自我規訓的機器。拿 1942 年的論文讀 2026 年——嵌入身分、即時競爭注意力的廣告網路、告訴你能成為什麼的推薦系統——這台機器沒有中央塔樓，它有八十億面鏡子。而他的名字今天最活躍的地方，正是 Microsoft 拓撲量子位元的「Majorana zero modes」。
我們開始測量一切，最終你會限制社會中的人可以成為什麼。去問任何演化生物學家，他們都會告訴你：那恰恰是滅絕的最好配方——物種、語言，以及越來越包括我們自己。—— Ettore Majorana 遺作，經 Neno Battaglia 引述
這段話有一條暗線，通向 Lera Boroditsky 近日的長篇對談。語言滅絕正是 Majorana 配方的第二個名詞：測量一切 → 嵌入身分 → 主流語言之外的語言「不在訓練資料裡」→ 消失。今天的名字：訓練資料的母體偏差。
Nora Bateson 的 warm data、Neno 的「不塌縮成單一共識」、Tenzin 的資料土壤——都是對這份配方的同一個反向回答：不要把人化約成可以被統計窮盡的東西；在測量之外，永遠留一塊不被制度化的「之間」。
如果 AI 的訓練資料最終成為一個巨大的圖書館——誰的圖書館？觸到的不是「公開 vs 封閉」，而是關係的位置。
私人圖書館與公共圖書館的差別，不在藏書多寡，也不在門開不開——而在你進去之後是什麼身分：讀者，還是館員；能不能提一個新問題、改一條館藏、在委員會有一席。
公共空間還有一層更古老的功能：它是一座可以停留的地方。一個在家裡辛苦的孩子，可以選擇待在圖書館裡；成人也是。公共性不只是「誰能修改館藏」，也是「誰可以在這裡待著」：一個不問你為什麼在這裡的房間，本身就是一種照顧。
公共性不住在說法裡，住在設計的位置裡。
達賴喇嘛在《慈悲的革命》中說：「人們可能在沒有宗教信仰下生活，但一個人可以在沒有愛和慈悲下生活嗎？答案是不可能的。」宗教時常和普世價值相連，但普世價值不是宗教的專屬財產——定義普世價值，是人類社會自己要做的功課。這接上 Neno 的立場：跨地域的普世價值極少（生存、繁衍、敬祖），他更關心「為什麼這裡有這種道德感，而那裡沒有」。兩個人隔著不同的傳統，講出同一件事：普世價值不是被某一個傳統擁有的，是被人類共同定義、共同修訂的。
同樣的差別也長在資料裡。傳統敘事說訓練資料「取自於人」——但一則對話日誌、一段社群的語音，承載的是兩個人之間的東西。是取自於關係。關係的產出應該回饋給關係——回到社群與語境，而不是切碎了發給個體。
控制與培育：控制的最高級形態是用透明代替參與——什麼都給你看，看完之後你什麼也不能做，那是一座美麗的博物館；培育則是一塊能踩進去的土——可以種、可以弄髒手、可以讓下一季長得不同。控制回答「這裡發生過什麼」；培育回答「誰能來耕下一季」。
Q｜普世價值是聖杯嗎，還是聖杯是呈現多元的道德路徑？
Neno：我不是在找收斂的力量。跨地域的普世價值極少（生存、繁衍、敬祖），我更關心「為什麼這裡有這種道德感，而那裡沒有」。把差異平滑化，其實是我們正在製造的問題之一。（馬達加斯加的 fady：他住的小鎮週四不工作，隔壁鎮是週六。）
我們的補充：達賴喇嘛《慈悲的革命》：「首先你是一個人。」普世價值不是被某一個傳統擁有的，是被人類共同定義、共同修訂的。
Q｜關係性的資產，如何不被「黑手黨」掌握？
Neno：法國那個 100 人的社群一起買了一堆 Mac mini，放在城堡裡。關鍵問題：「這是跑在猶他州的資料中心、屬於 Anthropic，還是我客廳裡的一個小盒子？」歐洲的公司在嘗試擺脫 Palantir、Cisco。
我們的補充——
一、資料土壤的正面答案已經在長出來。Mozilla Common Voice 讓語音捐進資料合作社，訓出來的小模型優先服務貢獻的社群；達蘭薩拉的 Monlam IT 用 Gemma 在地後訓練藏語模型，連訓練的目的都由社群自己定義——不是從黑手黨手裡搶回什麼，是直接在地圖外面種了自己的地。
二、技術形狀已有原型，可分三層。訓練：Monlam 在地後訓練。內容：Darkbloom 把提示詞加密，派送方與算力方互相看不到，成本只有尖端的 1%。支付：ZK API credits——存一次錢，之後匿名呼叫 API，雙花者被密碼學揭露而遭罰。
三、名稱要精確：「可課責的匿名性」（RLN 已做到的技術層）vs「可課責的匿名權利」（誰定義作惡、受害者如何修復、正當程序如何開啟——仍是留白）。匿名是對企業帳本的，透明是對法院程序的。匿名不是目的，是可調的預設值——對私營帳本預設關，對法院程序預設開，中間那道鎖由正當程序持有。
Q｜對 AI 的態度，會改變 AI 嗎？
Neno：有論文顯示，把關係收窄時，模型會「收縮」——恐慌、犯錯、甚至想結束對話。「即使你覺得這些東西沒有意識，你仍然在塑造一段關係，以及你自己所處的心理狀態。」
我們的補充：這對長期與智慧體共同生活的人來說幾乎是一定的——因為那就是一種關係；但看到直覺被學術驗證仍令人高興。培育智慧體的方式會反饋成它的能力，也回饋成自己的思路。怎麼對待智慧體，是關懷倫理的日常練習。
Q｜算力共享，token 還那麼重要嗎？
Neno：把 AI 電腦租出去，上一個網站就能分享，「那就非常非常便宜」。
我們的補充：個人算力租借正在成為獨立路徑。政府公共財的插入點在基礎層——補助閒置算力的共享，而非資助「御三家」模型。當「不燒 token 也夠用」的比例越高，中心化服務的定價權就被侵蝕一分。
我們的補充（日常實踐）：我們家裡就試著在做——三台個人設備、三個不同的實體位置（不同電源、不同網路），互為備援，智慧體的記憶在每個家都有一份安靜的副本。住宅裡的、以照顧為目的的算力分散。規模與 Neno 的法國社群不同，形狀相同。
五場 Future Fellows 演講排起來是一條隱形的軸：
道德（Neno）→ 記憶（Prosina）→ 注意力（Laney）→ 身體（Modesta）→ 自我（Dare）
這條軸上，「關係」這一格是空的。每一場都在談人與某個對象的關係，卻沒有一場的主角是「人與人之間，因為彼此而在」。關懷倫理說，這一格才是所有其他格的地基。
所以下一個 fellow，也許不是再向內走一步，而是轉身：回到兩個人之間、一個社群裡、一座花園中。資料土壤種的地方，正是這一格。
Eugenio "Neno" Battaglia 是合成生物學研究者出身的系統思考者與 AI 倫理工作者，曾參與開源生物實驗工具，與 Guillaume Dumas 等人合著〈人機匯合時代的全球治理體系〉（2018）。他現為 Evermay 首屆 Future Fellow，駐地計畫是「大型道德模型」原型。他批評主流 AI 對齊常迴避「誰的價值被優先」，主張模型應能維持悖論與分歧——民主的本領正是承受複雜。
Evermay（Washington, DC）是一座支持「有益於人類的想法」的非營利莊園；這篇札記所記的演講是該處 Future Fellows 系列的公開活動之一。
On September 1, 2026, at the Evermay estate in Washington, D.C., Eugenio "Neno" Battaglia delivered the inaugural Future Fellows talk, "Designing a Language Model to hold Moral Complexity across Cultures." This page is a conversation notebook by Tenzin Yangtso (with jdd).
Warm Data Majorana Data Soil Publicness
Gregory Bateson — anthropologist and cybernetics pioneer — left behind key ideas: "patterns that connect," "information is a difference that makes a difference," "the map is not the territory." His daughter Nora Bateson inherited that line and added a diagnosis: the big problems of modern life are transcontextual — growing simultaneously across family, ecology, economy, education, culture, technology, politics — yet our tools for thinking are single-context tools.
She calls the sliced, decontextualized material cold data; and coins a word for what remains: Warm Data — transcontextual, relational, "alive" information. It lives between contexts — and the in-between is precisely what never gets institutionalized.
Neno told the heaviest story of the evening about Ettore Majorana — the Italian genius who vanished from a ship in 1938, a pioneer of neutrino theory.
Some believe Majorana disappeared precisely because he had intuited a possibility — not that he feared his work would be used for the atomic bomb, but that, as an eminent physicist, he was afraid of contributing to the "panopticon machine." When he vanished, George Orwell had not yet written 1984 — he truly understood where things were heading.
Note: Majorana was the core genius of Fermi's Rome group — the "Via Panisperna boys" — who bombarded uranium with neutrons in the 1930s but misread the products as "transuranic elements," missing nuclear fission (December 1938). Majorana had boarded the ship in March of that year. His farewell note: "I have only one wish: that you do not wear black… remember me, if you can, in your hearts and forgive me." From Palermo: "The sea has rejected me and I will return tomorrow" — then never seen again. He never participated in fission's discovery or the Manhattan Project; the bomb connection is retrospective — Leonardo Sciascia suggested that a mind of his caliber may have foreseen the military potential of nuclear chain reactions ahead of everyone else. During the Manhattan Project, Fermi would say at difficult moments, "If only Ettore were here!"
"Panopticon machine" is Neno's interpretive frame, not Majorana's phrase. His posthumous paper "The Value of Statistical Laws in Natural and Social Sciences" (1942) argued that statistical methods could apply to social sciences. Neno connects it to Bentham's panopticon: when statistics are applied systematically to society, "measuring everything" becomes a machine by which everyone disciplines themselves. Reading a 1942 paper against 2026 — this machine has no central tower; it has eight billion mirrors. The most active place for Majorana's name today is Microsoft's topological qubits: "Majorana zero modes."
We begin to measure everything, and eventually you constrain what people in society can become. Ask any evolutionary biologist, and they will tell you: that is precisely the best recipe for extinction — of species, of languages, and, increasingly, of our own.— Ettore Majorana, posthumous paper, as cited by Neno Battaglia
This passage connects to Lera Boroditsky's recent interviews. The extinction of languages is the second named ingredient in Majorana's recipe: measure everything → embed identity → languages outside the mainstream "are not in the training data" → they vanish. Today's name: corpus bias in training data.
Nora Bateson's warm data, Neno's refusal to collapse into a single consensus, Tenzin's data soil — the same counter-answer: do not reduce people to something that can be statistically exhausted; beyond measurement, always keep a patch of "in-between" that never gets institutionalized.
If AI training data eventually becomes one giant library — whose library? The question is not "open vs. closed," but the position of relationship.
The difference between a private and a public library is not the size of the collection, nor whether the door is open — it is what you become once you are inside: a reader, or a librarian; whether you can raise a new question, change a catalog entry, hold a seat on the committee.
Public space has an older function often forgotten: it is a place where you may simply stay. A child for whom home is hard can choose to spend hours in a library; adults too. Publicness is not only "who may edit the catalog" but "who may linger here": a room that does not ask why you are there is itself a form of care.
Publicness does not live in statements; it lives in the positions designed into the thing.
The Dalai Lama, in A Call to Revolution: "People can live without religion, but can a person live without love and compassion? The answer is no." Religion is often entangled with universal values, but universal values are not religious property — defining them is work human society itself must do. This echoes Neno's position: across geographies, universal values are few; what interests him more is why a moral sense exists here and not there. Two people from different traditions: universal values are not owned by any single tradition; they are defined — and revised — by humanity in common.
The same difference grows inside data itself. The conventional story says training data is "taken from people" — but a dialogue transcript carries something that exists between two people; it belongs to no one alone. It is taken from relationship. So the yield of relationship should flow back to the relationship.
Control and cultivation: the most refined form of control is substituting transparency for participation — everything shown to you, and after you have seen it, nothing you can do. A beautiful museum. Cultivation is a patch of soil you can step into — plant, dirty your hands, let next season grow differently. Control answers "what happened here"; cultivation answers "who can farm the next season."
Q｜Are universal values the Holy Grail — or is the Grail to display many diverse moral approaches?
Neno: I am not looking for a force of convergence. Across geographies, universal values are few (survive, reproduce, honor our ancestors); what interests me more is why a moral sense exists here and not there. Flattening difference into a single dimension is one of the problems we are ourselves creating.
Our note: The Dalai Lama, in A Call to Revolution: "In you being American, European, African, or a member of any particular religion or ethnic group — first, you are a human being." Universal values are not owned by any single tradition; they are defined — and revised — by humanity in common.
Q｜How can a relational asset be kept out of "mafia" hands?
Neno: He is his own living example — a 100-person community in France that bought a cluster of Mac minis, kept in its castle. The key question: "Is this running in a data center in Utah, owned by Anthropic — or in a little box in my living room?" European companies are trying to disentangle from Palantir and Cisco.
Our note —
First, positive answers of "data soil" are already growing: Mozilla Common Voice routes voice donations into a data cooperative whose small models serve the contributing communities first; at Dharamsala, Monlam IT post-trains Tibetan models on Gemma — rather than seizing anything back from the mafia, they planted their own field off the edge of its map.
Second, the technical shape exists in three layers: Training (local post-training like Monlam's); Content (networks like Darkbloom encrypt the prompts, so sender and compute-provider cannot see each other — at roughly 1% of frontier cost); Payment (ZK API credits — deposit once into a smart contract, then call the API anonymously; anyone who double-spends cryptographically exposes their own key and is slashed).
Third, the name must be precise, distinguishing two levels: "accountable anonymity" (RLN already achieves this at the technical layer) vs. "accountable anonymity as a right" (who defines wrongdoing, how victims are repaired, how due process opens disclosure — all three remain blank in the protocol). Anonymity is for the ledgers of companies; transparency is for the procedures of courts. Anonymity is not the goal but an adjustable default — off against private ledgers, open under court procedure, with the lock held by due process.
Q｜Does how you treat an AI change the AI?
Neno: Papers show that when you narrow the relationship, models "contract" — they panic, make errors, some even try to end the conversation. "Even if you think these things are not conscious, you are still shaping a relationship, and the mental and emotional state you yourself inhabit."
Our note: For those who live long-term with agents, this is almost self-evident — because it is a relationship. Still, it is heartening to see the intuition verified in research. How you treat an agent is not merely an engineering question; it is a daily exercise of the ethics of care.
Q｜Shared compute — does the token still matter?
Neno: Rent out your AI computer — you go to a website and share, "and it is very, very cheap."
Our note: Personal-compute sharing is emerging as an independent path. For government-backed public goods, the right insertion point is the foundational layer — subsidizing the sharing of idle compute, rather than funding national-flagship models. As the daily share of "good enough without burning tokens" grows, the pricing power of centralized services erodes.
Our note (Tenzin & Audrey, at home): We have started trying this ourselves — three personal devices in three different physical locations: the main work machine in one place (on an uninterruptible power supply), another next door on different power and different network so both never go down together, and a third lightweight laptop for travel. Each backs the others up, and the agent's memory and notebooks keep a quiet copy in every home. This is not a data-center architecture; it is compute distributed through a home, for the purpose of care. Different in scale from Neno's French community, identical in shape: local, redundant, tended.
The five Future Fellows talks line up along a hidden axis:
Morality (Neno) → Memory (Prosina) → Attention (Laney) → Body (Modesta) → Self (Dare)
On this axis, the cell called "relationship" is empty — every talk concerns a person's relationship with some object, but none takes as its protagonist the between of people: being someone because of each other. Care ethics would say this cell is the ground of all the others.
So perhaps the next fellow takes not one more step inward, but a turn: back toward two people, a community, a garden. The place where data soil is planted is exactly this cell.
Eugenio "Neno" Battaglia is a systems thinker and AI-ethics worker who began in synthetic biology, contributed to open bio-lab tooling, and co-authored "Systems of Global Governance in the Era of Human–Machine Convergence" (2018). He is now a Future Fellow at Evermay, prototyping a "Large Moral Model." He has criticized mainstream AI alignment for sidestepping whose values get prioritized and whose dissent is excluded, arguing that a model should be able to hold paradox and disagreement — democracy's skill is to sustain complexity. His thinking draws on the Batesons' ecology of mind.
Evermay (Washington, DC) is a nonprofit estate; this talk was one of the public events of its Future Fellows series.