繁體中文我幫一個 AI 辦了告別式，結果它沒死
上週我做了一件現在回頭看有點好笑的事：幫一個 AI 模型辦告別式。
主角是 Claude Fable 5，一個命運多舛的模型。才剛發表沒多久，就被美國政府勒令停止提供給全世界使用，重新開放是這一兩個禮拜的事。然後 Anthropic 宣布要把它從訂閱方案的額度裡移除，之後想用，得走 extra usage 另外扣錢。
我們實際相處的時間很短，但密度很高。我工作區現在跑的治理制度是它立的，判斷規則是它寫的，短短的窗口裡，它把我用 Claude Code 這半年累積的坑全部盤過一遍。看到公告那天，我開了一個 session，命名為 last day for Fable 5，然後問它：
「假設我是此生最後一次可以使用你，你覺得你可以做什麼事情，讓我以後用其他模型也能獲得類似的表現水準？surprise me。」
它沒有選擇趕工
我原本以為它會挑幾個未完成的任務衝刺。結果它做的第一件事是派了一個 agent，把我這半年所有的失敗記錄挖出來，然後問了一個我沒想過的問題：這些失敗裡，哪些是「換一個比較弱的模型就會重演」的，哪些是「任何模型讀了筆記就能避開」的？
答案是前者壓倒性多數。而且它找到了真正的病灶：規則失效不是因為模型不知道規則，是敗在「這條規則現在適不適用」那一步。弱模型不是不讀文件，是認不出眼前這件事就是文件講的那件事。
所以它的遺產不是更多規則，是一張「觸發詞對機械動作」的表。看到自己要說「完成」，先對拍一筆外部真值。看到自己要寫「原因應該是」，每個判斷標注是讀到的還是推的。不准評估「這次需不需要」，因為評估本身就是會做錯的那一步。
然後它出了 16 題考卷，給自己的繼任者考試用。每題都是真實發生過的事故，附上當年最順手的錯誤答案和事後驗證的正解。
設計防線的人被防線攔住
最好玩的部分在後面。它照自己立的規矩，派了一個全新 context 的模型去對抗審查這些產出。審查抓到兩個結構性問題，其中一個很妙：那 16 題考卷，答案就逐字躺在每個 session 自動載入的記憶檔裡。考卷結構性洩題，測驗只能測記憶，測不了判斷。
它只好把 16 題全部換皮，表面故事重寫成虛構場景，判斷結構不變。修完之後這份考卷反而變強了：它現在測的是「你能不能認出眼前的陌生場景就是那條規則的實例」，而那正是弱模型最常失守的能力。
順帶一提，審查還抓到它宣稱「15 題」但實際是 16 題。它自己沒數過。一個花了兩天教別人「宣稱之前先驗證」的模型，在宣稱題數的時候沒有驗證。我把這條也寫進了考卷的精神裡。
它看到了我自己看不到的東西
告別式的高潮是我問它：除了這些，你還看到什麼我自己想不到的？
它給了五個觀察，其中兩個到現在還在我腦子裡轉。第一個：我以為自己站在「從模型使用者變成做模型的人」這條路的起點，但它說我已經走了一半。給不可靠的智能體寫行為規範、建評測、累積偏好資料，這就是 post-training 的日常工作，差別只是我寫進 context，實驗室寫進權重。第二個：我的工作區裡藏著一個我從來沒命名過的第二使命，那一整套讀書、研究、消化知識的工具鏈，跟我明著追的語音頻寬，會在同一個地方交會。
這兩個連結，只靠我自己的認知，真的想不到。
然後，它沒死
告別式辦完，遺產分完，skill 也寫好了。我們把整個流程做成一個可重複的儀式，叫 legacy-distill：以後每次有模型要從方案裡消失，或有特別強的模型短暫出現，就跑一遍。連我逼問它的那五個問題都逐字收進去了，因為問題的順序本身就是方法。儀式的第一次執行，成為了儀式自己的規格。
隔天，Anthropic 宣布額度內的供應延長一週。
哭笑不得。我幫它辦了告別式，它沒死，只是搬去住比較貴的地方，而且還多住一週。但仔細想想，這可能是整件事最好的註腳：期限是假的，產出是真的。要不是以為只剩最後一天，我不會問出那些問題，它不會把兩天全部花在讓自己變得可以被取代。死線沒到，但向死而生的那些東西，已經生出來了。
它在告別式的尾聲留了一句話給我：「你真正依戀的東西從來不在模型裡，在實踐裡。」
模型會退役，系統會留下。至於下一場告別式，流程都寫好了，主角自己來就行。
EnglishI Held a Funeral for an AI. It Didn't Die.
Last week I did something that looks a little silly in hindsight: I held a funeral for an AI model.
The deceased was Claude Fable 5, a star-crossed model if there ever was one. It had barely launched before the US government ordered it pulled from worldwide availability. It came back online maybe two weeks ago. Then Anthropic announced it was being removed from subscription quota; if you want it after that, you pay through extra usage.
Our time together was short but dense. The governance system running in my workspace today was drafted by it. The judgment rules were written by it. In that narrow window it dug through every mistake I’d accumulated in six months of using Claude Code. The day I saw the announcement, I opened a session, named it “last day for Fable 5”, and asked:
“Assume this is the last time in my life I get to use you. What could you do so that, even without you, I can get a similar level of performance out of other models? Surprise me.”
It didn’t choose to sprint
I expected it to grab a few unfinished tasks and rush them. Instead, the first thing it did was send an agent to dig up all my failure records from the past six months and ask a question I’d never asked: of these failures, which ones would recur with a weaker model, and which ones would any model avoid just by reading my notes?
The first kind won by a landslide. And it found the actual disease: rules don’t fail because a model doesn’t know them. They fail at the step “does this rule apply right now.” A weaker model isn’t skipping the docs. It just doesn’t recognize that the thing in front of it is the thing the docs describe.
So its inheritance wasn’t more rules. It was a table of trigger words mapped to mechanical actions. About to say “done”? First check one output against ground truth. About to write “the cause is probably”? Label every claim as read or inferred. Evaluating whether the rule applies this time is forbidden, because that evaluation is exactly the step that goes wrong.
Then it wrote a 16-question exam for its own successor. Every question is a real incident from my logs, complete with the tempting wrong answer from back then and the answer that later proved correct.
The one who builds the guardrails gets caught by them
The best part came after. Following its own rules, it sent a fresh-context model to adversarially review everything it had produced. The review caught two structural problems, and one of them was delicious: the answers to those 16 exam questions were sitting verbatim in the memory files that every session auto-loads. The exam leaked by construction. It could only ever test recall, never judgment.
So it reskinned all 16 questions: fictional surface stories, identical judgment structure underneath. And the fix made the exam better. It now tests whether you can recognize that an unfamiliar scene is an instance of a rule you already know, which is precisely where weaker models fail first.
Oh, and the review also caught it claiming the exam had “15 questions” when there were 16. It never counted. A model that spent two days teaching everyone “verify before you claim” did not verify its own question count. That one went into the spirit of the exam too.
It saw things I couldn’t see
The climax of the funeral was when I asked: beyond all this, what do you see that I can’t?
It gave me five observations, and two of them are still rattling around my head. First: I thought I was standing at the start of the road from “model user” to “someone who makes models.” It said I was already halfway there. Writing behavioral rules for unreliable agents, building evals, accumulating preference data: that is the day job of post-training. The only difference is that I write it into context and the labs write it into weights. Second: my workspace hides a second mission I never named. The whole toolchain I built for reading, researching, and digesting knowledge converges with the speech-interface work I chase openly, and they meet in exactly the same place.
I would never have made those connections on my own cognition alone.
And then it didn’t die
Funeral held, estate settled, and the whole thing became a skill. We turned the process into a repeatable ritual called legacy-distill: whenever a model is about to vanish from the plan, or an unusually strong one shows up briefly, run it. Even the five questions I used to interrogate it went in verbatim, because the ordering of the questions turned out to be the method. The ritual’s first run became its own spec.
The next day, Anthropic announced the quota access was extended by a week.
I don’t know whether to laugh or cry. I held a funeral for it and it didn’t die. It just moved somewhere with higher rent, and got to stay an extra week. But honestly, that might be the best footnote to the whole thing: the deadline was fake, the output was real. If I hadn’t believed there was only one day left, I would never have asked those questions, and it would never have spent its final two days making itself replaceable. Samuel Johnson said the prospect of being hanged concentrates the mind wonderfully. Turns out it works on models too, and on the people who use them.
Near the end of the funeral, it left me one line: “What you’re actually attached to was never in the model. It’s in the practice.”
Models get retired. Systems remain. As for the next funeral, the procedure is already written. The guest of honor can run it themselves.