Vitalik 用 AI 學密碼學，怎麼知道自己真懂？

0:00公鑰密碼學還撐得住嗎

0:00V 先生，很高興見到你。好，我們可能要來做 Justin Drake 那則推文和你回應的第二集。那則推文有三、四百萬次瀏覽，你也發文回應了。你覺得公鑰密碼學還會繼續存在嗎？我認為它會繼續存在。不過我確實覺得，它的具體安全性大概會受到一些打擊。我覺得比較實際的思考方式，是去看密碼學現在依賴的那些困難問題，過去發生了什麼事。以因數分解為例。很長一段時間，大家認為最快的分解方法基本上就是暴力破解，只是可以把難度降成平方根等級。Mr. V. Good to see you. Yeah, good. All right, so we've had — we might do part two of the Justin Drake tweet and your response. But what, I mean, that got three, four million views. What, I mean, you posted about it. Do you think public key cryptography continues? I think it continues. I do think it will probably take some hits to concrete security. And I think one of the ways that I think it's, like, realistic to think about this is to think about what's happened to existing hard problems that cryptography depends on, right? So things like factoring, for example, right? Like, for a long time, people thought that, you know, like — the fastest way to factor is basically, like, effectively brute force, but you sort of make it quadratically easier.

0:55做法是從下往上、同時從上往下夾擠，你得到一個平方根，這樣就相當容易了。後來有人發明了二次篩法，又有人發明了 GNFS（一般數域篩法），這基本上就是為什麼安全的 RSA 金鑰長度，從 256 位元一路漲到 3072 位元。所以它沒有被破解，但安全性下降了不少。我覺得有一個很合理的論點是：萬一橢圓曲線和格密碼裡面，也藏著差不多等量的「衣櫃裡的骷髏」呢？也就是說，它們也有各自對應的二次篩法、GNFS 之類的東西，只是我們還不夠聰明，還沒發現而已。You kind of do it from the bottom to the top and the top at the same time. You get a square root, and it's, like, pretty easy. And then people invented the quadratic sieve, people invented GNFS, and, like — that's basically why the size of a safe RSA key has gone up from, you know, 256 bits to 3072 bits, right? So not broken, but degraded by quite a bit. And I think, like, there is a very reasonable kind of point to make that basically, what if elliptic curves and lattices have about the same level of, like, extra skeletons in the closet in them? Like, they have their equivalent of, you know, quadratic sieve and GNFS and all of these. But it's just, well, we haven't been smart enough to discover them yet, right?

1:44用 AI 學最難的東西

1:44因為這些東西需要同時在腦中記住的事情，比我們習慣的還要多。但機器人可以。觀眾裡有很多人背景沒那麼技術，像我自己入門線性代數時，很喜歡 Gilbert Strang 的《Linear Algebra》。既然大家都得回頭從密碼學、數論、離散對數這些基礎重新學起，你推薦什麼？你是從哪些參考資料學的，或覺得哪些有用？我最近一次學習那種對我來說難到瘋狂的東西，大概是今年夏天發表的那一系列關於混淆的部落格文章。那有什麼經典參考書嗎？這就是我發現的事情。Because they just require holding more things in your head at the same time than we're used to, right? But the bots can. So, you know, one thing for a lot of people in the audience who aren't as technical, you know, when starting out with linear algebra, I like Gilbert Strang's Linear Algebra, for example. Where should, you know, given that everybody's going to have to go back to basics on cryptography, number theory, discrete log, all this kind of stuff, what do you like? What references did you learn from or do you find useful? So the latest experience I've had learning things that are, like, crazy difficult for me was probably the series of blog posts on obfuscation that I published over the summer. But what canonical references? So this is — okay, this is the thing that I discovered, right?

2:32我發現直接跟機器人對話，就是非常有效的學習方式。而且跟機器人對話有一套特定的方法。因為你很容易問一個問題、拿到答案，就以為自己懂了。寫永遠比讀學得更好。所以對我有用的做法是：一直問問題，最後用自己的話把所有東西重新解釋一遍，然後問它「我說得對還是錯？」通常 AI 都很會拍馬屁，所以它會說「太好了，你大致上是對的，但有幾個地方我會改進」。然後就一直來回，直到它不再說「但有幾個地方我會改進」為止。這很有意思。這有點像學數學時的主動閱讀，做得對的話，你會試著預測下一頁、再下一頁會寫什麼。I discovered that just talking to the bot is just, like, a very effective way to learn. And I think there's, like, a specific way you want to talk to the bot, right? Because it's very easy to ask a question, get the answer, think that you know. Like, you always learn better by writing than by reading, right? And so the thing that works for me is, you know, you keep asking questions, and then eventually you re-explain everything in your own words. And then you ask it, "Am I right or wrong?" And then usually, you know, the AIs are all sycophantic, and so they tell you, "Oh, yay, you're mostly right," "but here are some things that I'll improve," right? And just keep going back and forth until it stops saying, "but here are some things that I'll improve." Yeah, so it's interesting. I mean, it's kind of like, you know, when you're learning math, it's — like active reading, if you're doing it right, where you're trying to anticipate what's on the next page and the page after that.

3:28同樣地，我用這些模型的方式，是請它出選擇題或答案是整數之類的題目給我。你自己算出來、輸入答案，就知道自己沒有在騙自己、誤以為答對了。這是很棒的技巧。那你都用在哪些地方？用過哪些模型？Claude、Anthropic，全都用？全都用。我現在對這些模型的體驗是，我發現即使是開放模型，例如 Qwen3-8B，它就跑在我的筆電上，已經足以應付大多數日常任務。真的在筆電上跑的唯一缺點，就是速度慢。And similarly here, you know, the way that I use these models is I will ask it to ask me a multiple-choice or integer-valued answer or something like that. And you can compute it and type it in, and then you know that you're not, you know, fooling yourself into thinking that you got it right. Yeah, that's a great technique. Yeah. So what have you used it for? Like, you know, which models have you used for this kind of thing? Just everything, you know, Claude, Anthropic? Just everything. I mean, so my experience with the models right now is I've realized that, like, even the open ones, like even, like, Qwen3-8B, which runs on my freaking laptop, yeah, it's like — good enough for, like, a majority of day-to-day tasks. The only downside of literally running it on my laptop is that it's slow.

4:15開放模型夠用，區塊鏈該付錢請 AI 稽核

4:15前沿模型，像是 Fable、Opus 5.5、Astra，真正突出、真的非用不可的地方，基本上是最前沿的工作。也就是你在理解、甚至貢獻某個複雜領域的最前緣，真正推進它的時候。低於這個層級的事情，基本上什麼模型都行，連開放模型、本機模型都已經變得非常好。這很有意思。Haseeb 和我這幾天一直在聊一件事：就拿這些開放模型來說，它們現在已經相當好了。有個情境我想聽聽你的看法，就是把開放權重模型和區塊鏈節點的其他部分打包在一起。The places where the, like, frontier ones, like basically, you know, like Fable and Opus 5.5 and Astra, like, really stand out and you need them, basically, is the frontier stuff, right? Like the stuff where you're literally understanding or even contributing to, like actually advancing the very frontier of some, like, complicated field. For anything below that level, like... Basically, yeah, anything that works, like even the open ones, even the local ones are getting very good. It's interesting. You know, one thing that Haseeb and I have been talking about for a couple of days is, you know, let's take these open models. Now they're pretty good. There's a scenario I'd like to hear your thoughts on, where open-weight models are bundled with the rest of a blockchain node.

5:03為什麼？因為區塊鏈需要撥出一部分代幣，付錢持續稽核它的程式碼、它的數學假設等等。這就像國防部，你從挖出來的 ETH 或其他來源撥出一定比例，實際花在 OpenRouter 之類服務提供的 token 上，持續去探測這條區塊鏈底層的假設。新模型一出來，你就搶在別人之前用它攻擊自己。你怎麼看？我覺得我喜歡的版本，也覺得超有意思的，是不要把單一模型寫死在鏈上。當然。取而代之的是，很久以前，在區塊鏈領域非常早期、超過 10 年前，有個想法叫做「卓越證明」（proof of excellence）。Why? Because a blockchain will need to devote some fraction of its tokens to paying for constant audits of its codebase and its mathematical assumptions and so on. And it's like a defense department, where you have a certain percentage of, whether it's ETH that's mined or what have you, that is actually spent on, you know, tokens served up by OpenRouter or something like that, that are constantly going and probing the assumptions underneath that blockchain. As soon as a new model comes out, you're trying to attack yourself with it before somebody else. What are your thoughts on that? Yeah, I think the version of that that I like, I think is super interesting, is: don't enshrine, like, one single model in the chain. Yeah, sure. Instead, like — There was an idea from way back, like more than 10 years ago, in the very early blockchain space called proof of excellence.

5:55這個想法基本上是一種為成果付費的發行機制。你記得 Primecoin 嗎？很久以前了。對，Sunny King 做的。就是那個。它想做一種有用的工作量證明，去找下一個更大的質數。沒錯。所以我覺得真正可能合理的，是做一種有用的工作量證明，任務是改善區塊鏈自己的演算法。因為很多情況下，你可以把目標形式化地定義出來。這很有意思，因為很多數學成果感覺是非常不連續的。你要怎麼像算雜湊那樣有漸進式的進展？你的想法是什麼？And the idea is basically it's like an issuance mechanism that pays for results, right? So if you remember, do you remember Primecoin? Oh, a long time ago. Yeah, like Sunny King. But that was the one. It was trying to make a useful proof of work to find the next largest prime, right? Exactly. Yeah. And so what I think actually might make sense is to make a useful proof of work where the task is to optimize the blockchain's own algorithms. Because the thing that you can do is you can, like — In many cases, you can formally define the goals, right? So this is interesting because, you know, it feels like a lot of these mathematical results are very discontinuous. How would you have incremental progress in the same way you'd have a hash? Like, what are you thinking there?

6:40你不會真的拿它來做共識，而是拿它來做發行。既然是發行，那麼某一年裡由一個人貢獻 75% 的價值也沒關係，因為有時候事情就是這樣發生的。我懂了，有意思。其實你最近還發了另一個東西，就是以太坊路線圖。幾週前你寫了〈Make Ethereum Cyberpunk Again〉，接著又發了路線圖。「數學奇點」會影響你對路線圖的想法嗎？絕對會。我認為整個 Lean 路線圖沒有 AI 基本上不可能實現。沒有 AI 的話，以我們現在構想的形式、以我們現在設定的企圖心，這份路線圖很可能真的要拖到 2040 或 2050 年。Yeah, so you don't literally use it for consensus, right? Like, you use it as issuance. Like, and if it's issuance, then, like, it's fine if, you know, during one year one person provides 75% of the value, right? Because that's sometimes what happens. I see. Interesting. So, are you — are we going to do — actually, you did put out something else recently, which is the Ethereum roadmap, like about a... Like a few weeks ago, you'd done Make Ethereum Cyberpunk Again, and then you put out the roadmap. Does the math singularity, does that influence how you think about the roadmap? It definitely does. Like, I think the whole Lean roadmap, it would basically not be possible without AI. Like, I think without AI, the roadmap... As we now conceive it, with the level of ambition that we conceive it, like, plausibly would have taken until like literally 2040 or 2050.

7:29AI 加速以太坊研發與形式化驗證

7:29基本上，我們因為 AI 而大幅加速，因為 AI 做了很多研究，也開始寫很多程式。尤其是 Lean Ethereum 的很多東西。Lean Ethereum 這個名字是刻意的雙關，一方面是讓以太坊更精簡（leaner）、更快，另一方面所有東西都用 Lean 4 寫。喔，我不知道這件事，有意思。對，我們手上已經有非常進階版本的 EVM、遞迴證明、共識，這些元件都附有安全性的形式化證明。因為表達「這是安全的」這個概念的敘述，往往比實際演算法小上一百倍。Basically, yeah, we've been getting massive speedups from the fact that AI is doing a lot of research, AI is starting to do a lot of coding, and we've especially — a lot of the Lean Ethereum stuff, like, you know, Lean Ethereum is sort of intentionally a pun, right? Because it's, you know, making Ethereum leaner and faster, and then also everything is written in Lean 4. Oh, I didn't know that. Interesting. Okay. Yeah, and basically, yeah, like, what we already have is — like, very advanced versions of EVM, recursive proving, consensus, all of these different components with formal proofs of their security, right? Because the statement that, like, basically expresses the idea this is secure is, like, often a hundred times smaller than the actual algorithm.

8:21所以不再是人類得稽核 10,000 行程式碼、確保沒有 bug，而是只有 100 行左右。就算是這 100 行，機器人也會幫你。有意思，這有兩個層次。一個是理論上成立，有 Lean 證明。另一個是具體的機器實作，在某些機器上它可能會做四捨五入、浮點運算或其他奇怪的事。你也許還是想稽核程式碼，同時也看 Lean 證明。你同意嗎？還是你覺得光有 Lean 證明就夠了？我想這取決於形式化驗證是在抽象層次，還是跟機器相關。我個人相信，形式化驗證的未來是端到端的。And so instead of there being, you know, like — 10,000 lines that humans have to audit and be sure that there's no bugs. There's like 100 lines. And then even for 100 lines, those 100 lines, the bots help you. Well, it's interesting. There's two versions of that. One is that it's true in theory, and there's a Lean proof. And then there's the specific machine implementation, which, right, you know, sometimes on some machine it'll do rounding or floating point or weird things. You might still want to audit the code, but you also look at the Lean proof. Do you agree with that? I mean, or do you think the Lean proof would just be enough? I guess it depends if that formal verification is in an abstract context or if it's machine-dependent. Yeah, I mean, like, I personally, like, I believe that the future of formal verification is end-to-end.

9:07對我來說，這一切真正的奇點是：我相信我們會有以太坊客戶端安全性的形式化證明，被證明的範圍一路涵蓋整個作業系統，甚至硬體。我們真的會有一台用 Lean 做形式化驗證的電腦，唯一的假設只剩物理定律。有趣的是，AI 讓這件事變得可以想像。對，而且也變得有必要。沒錯。我還想問你一件事，然後我們來聊新創社會。Haseeb 在這點上跟我意見不同，但我確實認為純數學家現在很沮喪，可是未來幾年他們的薪水會漲 10 倍。因為只有他們能下 prompt 並驗證結果，也許不是只有他們，但可以說他們是做這件事最強的人。Like, to me, the real singularity of all of that is — I believe that we will have formal proofs of the security of our Ethereum clients that stretch all the way to including, in the thing that's being proven, the entire operating system, and even the hardware. Like, we'll literally have a Lean, a formally verified, a formally verified computer where the only assumptions are the laws of physics. You know, what's funny is that AI has made that possible to contemplate. Yes. And actually necessary to do. Exactly. Yeah, it's funny because — so, I wanted your thoughts on one more thing, and then let's talk startup societies. You know, Haseeb disagrees with me on this, but I do think that — pure mathematicians are going to see their salaries — they're despondent right now, but I think their salaries are going to go up 10x in the years to come because they're the only ones who can prompt and verify, and maybe not the only ones, but let's say the best ones who can do this.

10:09這種能力的價值現在就像核武密碼一樣，至少在加密貨幣領域是如此，可能在很多其他領域也是。特別是短期內，我真的同意。我還有一個有趣的觀察，同樣來自寫那三篇混淆文章的經驗：兩年前我根本寫不出來。兩年前，如果沒有花好幾週跟發明那些協議的人討論，我不可能理解那種題材。而這些人是非常稀缺的資源，他們的時間很少能拿來回答笨問題。所以我覺得寫那些文章，一方面價值少了四倍，因為就算我沒寫，很多情況下別人也能搞懂同樣的東西。And the value of that is now like, you know, the nuclear codes equivalent, at least for crypto and maybe for a lot of other things. Yeah, I mean, especially in the short term, I really agree. Like, I think one interesting observation I also had, again from writing those three posts on obfuscation that I did, is like two years ago, I would not have been able to do it, right? Basically, two years ago, I would not have been able to understand that kind of subject matter without, like, weeks of talking to the people who invented those protocols. And they're a very scarce resource, and their time is very scarce for answering dumb questions. So it felt to me like writing those posts was at the same time four times less useful because, well, if I never wrote them, like someone else could have, in many cases, figured out the same thing.

10:57但另一方面，速度也快了四倍。基本上我覺得很多人都在往上移一層。你做的工作仍然善用你的經驗、你累積的直覺、你的領域專業，但很多時候，你能以更純粹的形式運用這些東西，要做的無聊雜務也變少了。所以你的看法是，你覺得 AI 會不會……嗯，這要看領域。我認為在可驗證的領域，AI 不會遇到天花板，它會一直進步，因為它可以不斷產生東西。但在其他領域，它就……你怎麼看？例如寫作。But at the same time, it was also four times faster, right? And so... Yeah, but like, basically, I feel like a lot of people are kind of moving one level up the stack, right? And so you're doing work that's really still taking advantage of your experience, your acquired intuitions, your domain expertise, but often you're actually able to take advantage of those things in a more pure form. Those things in a more pure form, like, you have to do sort of less boring housekeeping. You know, and so is your view — do you think AI is going to — well, it depends on the domain. I think in a verifiable domain, AI does not asymptote. It just kind of keeps going because it can keep generating things. But in other domains, it... What is your view on that? Like, for example, writing.

11:46AI 寫作為什麼沒有變好

11:46我其實覺得 AI 寫作已經有一段時間沒進步了。你同意嗎？同意。我想部分原因是，寫程式碼時，那些公式化的 if-then-else 結構重複一點沒關係，但在散文裡就很笨拙。所以程式碼聽起來像機器寫的沒問題，散文就不行。還是說，其實我也不知道答案，你覺得是因為 Anthropic 和 OpenAI 沒有在寫作上下功夫，所以它還是一看就知道是機器人寫的？我猜，尤其是 Anthropic，確實在乎寫作品質。這感覺是他們品牌的核心，有點像是「我們是有人味的那一家」。I don't actually feel AI writing has gotten better in a while. You agree with that? Yeah, I agree. Partly, I think it's because with code, the formulaic if-then, you know, else constructions, it's fine to be repetitive. But in prose, that's clunky. You know, and so it's okay when it sounds machine-like when it's code, but not when it's prose, you know. Or do you think it's just because — actually, I don't know the answer to this — do you think it's just because Anthropic and OpenAI have not optimized writing, that it's still so detectably robotic? My guess would be that especially Anthropic does care about writing quality. Like, it feels, you know, very core to their brand that sort of we are the human ones, right?

12:35OpenAI 的話，我注意到一件事。你知道 ChatGPT 嗎，你請它解釋事情，它到處插表情符號、到處隨便加粗。從傳統那種古板的《紐約客》文風來看，這簡直是「天啊，他們在幹嘛？」但我其實覺得，他們在這點上發明了未來。原因是，如果你去看那些獨立於長篇散文之外演化出來的書寫形式，排前兩名的基本上就是數學和程式碼。拿數學來說，想想二次方程式的公式，你以前看過：負 b 加減根號 b 平方減 4ac，全部除以 2a。OpenAI, actually, one thing I've noticed about OpenAI is, you know how with ChatGPT, if you ask it to give you explanations of things, like, it inserts emojis everywhere; it inserts random bold everywhere. And, like, from a traditional, you know, like, stodgy New Yorker-register perspective, this is like, oh my God, what are they doing? Like, I actually think, like, they've, like, invented the future in that, right? Basically, the reason is that if you look at forms of writing that evolved independently from, you know, like, long-form prose, like, the top two that we have are basically math and code, right? And if you think about math, like, if you think about the quadratic equation, right? Like, you've seen it before, you know, like, minus b plus or minus square root of b squared minus 4ac, all over 2a.

13:25如果你把我剛才講的話逐字聽寫下來，照那樣擺在那裡，你根本看不懂那是什麼。但你看過二次方程式的公式，它看起來非常漂亮，這裡一個根號，那裡一個分數，這讓你理解起來容易三、四倍。還有語法上色，基本上就是說，好，這裡要縮排，不同的東西用不同顏色。所以你認為 AI 寫作的結構化、表情符號，其實會變成一種結構化的書寫形式？對，基本上我認為這些光靠人類也可能發生，但感覺我們其實是在快速通關，走向能把寫作的力氣和閱讀的力氣分開的那一天。Like, if you imagine, like — transcribing, like, what I just said just now and, like, sticking it in like that, like, you're not going to be able to understand what it is, right? But, you know, you've seen the quadratic formula before. Like, it looks really beautiful. You have, like, a square root over here, a fraction over here, and, like, that makes it three or four times easier for you to understand it, right? And, like, syntax highlighting and, like, basically saying, like, okay, we're going to actually have indenting over here. We're going to have different things be different colors. So you think AI writing and the structuring and the emojis will actually become sort of, like, a structured form of writing? Yeah, like, basically, like, I think all of that could have happened with just humans, but it does feel like, you know, we're actually sort of speedrunning to being able to essentially separate writing effort from reading effort.

14:14這很有趣，我的想法是這樣：在 ChatGPT 出現之前，每次你看到一篇有表情符號之類的貼文，代表作者花了更多心力。但現在每次看到這種東西，我都假設他們花的心力更少，是某種機器人垃圾訊息。所以奇怪的是，這到底還算不算是在跟另一端的人溝通？它幫寫的人省時間，卻沒幫讀的人省時間，這跟寫作本來該有的樣子正好相反。那它會被採用嗎？講究格調的那群人會讓步嗎？我反而認為，我們會看到「人類證明」的社群網路，完全禁止 AI 寫作。某種意義上，這些網路會位在模型的上游，因為它們產生新資訊，而模型不會。You know, it's funny because — so here's what I think. I think, you know, it used to be, pre-ChatGPT and so on, whenever you saw a post that had emojis and so on, people had put more work into it. But whenever I see that today, I assume they put less work into it, and it's some bot spam. And so it's weird that I think there's an aspect of, is it actually communication to someone on the other side? It's saving time for the writer, but not the reader, which is the opposite of what writing is supposed to be. So will it get adopted by — will the prestige kind of people break? I actually think instead we're going to see proof-of-human social networks where no AI writing is allowed at all, and those are going to actually be, in a sense, upstream of the models, because they're generating new information and the models aren't.

15:06不知道你怎麼看。我同意會有很多這種東西。但我覺得同時，我們也會從這些快速出現的新風格學到很多。我認為最後得到的永遠會是一種綜合。就算是現在，例如你做投影片，你做的投影片大概既不是你五年前做的那種，連風格都不同，也不是 Claude 預設會幫人做的那種。我現在用 AI 的方式，是用來摘要，不是用來創作。我現在很討厭的一件事是，我只要說「是 A，不是 B」，聽起來就像 AI。但對比式否定在英文裡是很有用的句型，所以你得重新設計句子，讓它永遠不出現這種結構。I don't know your thoughts on that. I agree we'll have a lot of that. I think at the same time, we are going to learn a lot from these new styles that are rapidly getting created. And, like, I think the thing that we're going to get will just always be a synthesis, right? Like, if you think about, like, right now even, like, when you make slides, for example, right? Like, the slides you make are probably both not the slides you made five years ago in, like, in terms of even style, and they're also not, like, the type of slides that Claude would make for someone by default, right? So, you know, here's how I use AI, at least right now, is summarization but not creation. And, you know, one of the things now I hate is if I even say A, not B, it sounds AI. But contrastive negation is a useful thing in English, so you have to re-architect your sentences to never include that.

16:00總之，我想我同意。所以說摘要，你拿一段 60 分鐘的語音備忘錄，對吧？你說。對，我想我同意。寫作確實是一門藝術，要讓東西對讀者盡可能友善，同時又不能只是讓機器人去做、自己草草帶過。因為如果你這樣做，我覺得問題不只是大家反感那種已經看過十萬次的風格。還有一點是，如果你沒有親自把內容走一遍，它非常擅長做出「假裝是內容」的東西，但其實經不起任何檢驗。But anyway, so — I think I agree. But yeah, so summarization: you know, you take a 60-minute voice memo, you know, right? Go ahead. Yeah, no, I think I agree with that. Like, I think, you know, there really is an art in writing in a way that, like, really does make things maximally friendly for people to read, but where at the same time you're not just kind of, like, letting the bot do things and glossing over it. Because if you do that, like, I don't even think it's just about people's aversion to the fact that they've seen that exact style a hundred thousand times before. I think it's also, you know, if you don't, like, actually walk over the content yourself, it's, like, very good at making things that pretend to be content, but, like, actually don't pass any kind of scrutiny, right?

16:52我有個想法。你知道那些真正的隨機性來源，也就是熵源嗎？有人會用一些奇怪的東西，像熔岩燈之類的。我確實認為會出現離線辦公室、封閉式圖書館，裡面的電腦不只沒有 AI，連稽核都確保電腦上沒有 AI，你必須像這樣一字一字打出所有東西。這些地方產生的資料，某種程度上是純淨、未受污染的，就像 2022 年以前的資料，像一道純人類資料的山泉，位在上游。不然在不可驗證的領域，整件事會變得非常循環。Well, you know, here's the thought. You know those, those, like, true randomness sources, like entropy sources? You know, like, people have weird things like lava clocks and stuff like that, right? I do think that there's going to be, like, offline offices, lockdown libraries where there's computers that have not only no AI, but the auditing of them is something where there's no AI on the computer, and you have to type everything like this. And the data generated by those things is sort of like pure and uncontaminated in a way. It's like pre-2022, like a mountain spring of purely human data, you know, that's upstream. Because otherwise the thing gets so circular in a non-verifiable domain.

17:38你不這麼認為？不，我覺得這絕對是對的。好，再一件事。在我們換話題之前，公鑰密碼學，你覺得會繼續存在？對，我認為會繼續存在，但每個人都得再做一、兩輪更新，讓參數再變大一、兩輪。基本上我們會多出幾次「Y2K」。網路安全也是一樣。我對最後走出來的結果是樂觀的，但我們正在穿越一條有障礙物的隧道，得像打電動一樣，跳來跳去閃過它們。很好。好，換個話題：Zuzalu。You don't — you don't think so? No. Yeah, yeah, no, I think that's definitely true. So, okay, all right. One more thing. So public key cryptography, before we move on, public key cryptography, you think continues? I think, yeah, I think it continues, but I think everyone will have to do like one or two more rounds of, like, refreshing things, one or two rounds of letting their parameters get bigger. Like, it's, you know, like, we basically get extra Y2Ks. And I mean, I think, you know, like cybersecurity is the same. Like, I am optimistic about where we come out the other end, but, you know, I think, you know, we are kind of going through, you know, like a tunnel that has obstacles in it that we have to, like, you know, like play the video game where you sort of, like, bounce and dodge them. Great. Okay, cool. All right, switching topics: Zuzalu.

18:24網路國家從理論走進實踐

18:24我們現在在 Network State 大會。我剛剛在會前傳訊息跟你說，我認為網路國家現在已經從理論走進實踐。這裡有幾件我們討論過的事，也許你有些想法。第一，你辦了 Zuzalu，也許可以跟大家說說 Zuzalu 是什麼，你從中學到了什麼。Zuzalu 基本上是我做的一個快閃城市實驗。最初那一次是在 2023 年，基本上就是找 200 個人，邀請他們全部到同一個地方，在蒙特內哥羅。我刻意混合不同社群，有一些以太坊的人、一些理性主義社群、一些長壽研究社群，還有其他人，基本上就是一起住兩個月，當一座快閃城市。So, you know, we're at Network State. I actually, one of the things I was just saying to you in text beforehand is I think Network State has now moved out of theory into practice. Here's a few things that we've discussed that maybe you have some thoughts on. One, like, you did Zuzalu; maybe you can describe to them what Zuzalu is and what you learned from doing that. Yeah, so Zuzalu was this — basically a pop-up city experiment that I did. I mean, the original one back in 2023, basically, like, get 200 people, invite them all into one place, this one place in Montenegro. And I was intentionally mixing different communities, like some Ethereum, some, like, rationalist communities, some longevity, some other things, and basically just stay for two months and be a pop-up city.

19:10我覺得它非常成功地證明了這整套模式是可行的。然後我認為它最後打開的大問題是：好，把 200 個人聚在一起是可能的，辦快閃城市是可能的，那接下來要往哪裡走？下一步是什麼？我覺得試著回答這個問題，很大一部分就是你辦這場大會、經營整個後設社群的目的。沒錯。而且我覺得我現在有了一個衡量這個領域的指標：幣有 CoinMarketCap，社群對應的就是「人月」。And I feel like that was still kind of very successful at just showing that that whole pattern is possible. And then I think, you know, the big question it ended up opening is that, okay, you know, it's possible to bring 200 people together, it's possible to make a pop-up, and... Like, where do you go from there? What is the next thing after that? And I feel like trying to answer that question is a big part of, like, what I imagine you're doing, like, this conference and your kind of whole meta-community for. That's right. And I think actually I've got a metric on the space now, which is: the equivalent of CoinMarketCap for coins is man-months for communities.

19:55傳統的人口普查，計算的是定居在一地的德國人、俄國人、日本人之類的人數。但如果是數位遊牧的世界，大家不在同一個地方。所以用人月數來算，例如 1,000 個人在 Zuzalu 待了兩個月，或不管實際數字多少，那就是一個 2,000 人月的社群。這非常容易量化，因為你有人類證明，也有位置證明。你可以說這些人至少證明自己曾經透過某個位置信標出現在這個地點，然後得到一個累積訊號，也就是那個社群或許多社群累積的人月數。接著你就可以開始幫這些社群排名，也有了一個可以持續往上衝的成長指標。So because, like, a traditional census counts the number of sessile human beings who are German or, you know, Russian, Japanese, whatever. But if you've got a digital nomadic world, people aren't in the same place. So number of man-months, let's say, for example, a thousand people spent two months at Zuzalu, or whatever the number is, then that's a 2,000 man-month community. And now you can, and that's very quantifiable because you have proof of human and you have proof of location. So you can say these humans at least attest that they were in this location with whatever location beacon, and you can have a cumulative signal, which is the cumulative number of man-months for that community or tons of communities. And now you can start ranking these things, and now you've got a growth metric you can start going up on.

20:43你也知道，比特幣出現好幾年後，才開始有比特幣價格。所以我認為推出這個指標會是很重要的一件事，我們很快就會推出。對，我覺得這絕對是個好方法，可以衡量我們從辦一場大型家庭派對，走到真正進入新創社會的下一個階段，進展到哪裡了。我想問你一個比較難量化的指標。你經營 Network School 已經兩年多了吧。幾乎剛好兩年。剛好兩年。這是真實存在的東西，你投入了全部心血。你經歷過高峰也經歷過低谷，全網路都知道。Because it took a couple of years, as you know, into Bitcoin before the Bitcoin price even existed. And so I think the introduction of that metric is going to be an important thing. We're going to be rolling that out soon. Yeah, no, I think it's definitely a good way of just measuring, you know, like how far along we are from basically having a big house party to having, like, something like actually the next stage of startup societies. Yeah. So I wanted to ask you just a less quantifiable metric. So you've been running Network School for more than two years now, I think. Just almost exactly two years. Exactly two years. It's an actual thing. You've put your heart and soul in it. You've had your highs and your lows that the entire internet knows about it.

21:29與哈薩克的合作

21:29你感覺如何？很棒。我真的感覺很棒，你會在最後看到一些原因，我想我就自己爆料吧。我們將能夠……對。我們已經和哈薩克這個國家結合在一起，有點像《聖戰士》（Voltron）那樣合體。還有大約十個其他國家為 Network School 提出了競標。所以對我來說，論點是這樣的：世界上大多數事情，跟政府之間的比分都是 0 比 0。意思是，比如你今天穿紫色襯衫，大部分政府都不在乎，所以他們不會對這件事表態。然後有些事情是政府反對的，你跟某個政府就是 0 比 1。How do you feel? Awesome. I actually do feel awesome, and you'll see some of that towards the end where — you know, I guess I'll scoop myself. So we're going to be able to — Yeah. So we have fused with the state of Kazakhstan, like, you know, Voltron or something like that. And there's like ten other countries that put in bids for Network School. And so to me, the thesis, I mean, see, the thing is, most things in the world have a score of 0-0 with governments. What I mean by that is whether, I don't know, you're wearing a purple shirt today, no government cares about that for the most part. So it's not something they're going to take a position on and so on. Then there are some things that governments oppose, and you can be 0 and 1 with a government.

22:20但也有些事情是其他政府支持的。任何事只要有些政府反對、有些政府支持，比分變成 1 比 1、3 比 1 或 5 比 3，它就是影響全世界的一個關鍵二元開關。就像老派的 C 語言，檔案最上面有一堆 #define，這些開關影響後面所有的行為。換成消費者的例子，就是深色模式或淺色模式，整個 app 都受它影響。同樣地，各國現在正在面對的問題是：他們要不要讓全球遊牧人口落腳在自己的土地上？我認為這件事不是那麼顯而易見，原因是各國已經習慣了觀光客，而觀光客隱含的意思是只待幾天。But then there's things that other governments support, and anything where some governments oppose it and other governments support it, where it's like 1 and 1 or 3 and 1 or 5 and 3, is a critical binary switch on the whole world. It's like, you know, in old-school C, you know, if you have #defines at the top of a file, and those switches influence all subsequent behavior, you know. Consumer example: dark mode versus light mode, you know, the entire app is influenced by that. In the same way, what countries are now coming to terms with is, do they want, like, global nomadic populations on their soil or not? And, you know, the reason I think this is non-obvious: countries have been used to having tourists, and a tourist is there for implicitly a few days.

23:12觀光客是個人，而且只是來休閒。但數位遊牧社群可能有好幾千人，可能待上好幾年，而且他們還在工作。所以這看起來像外國飛地，但在某些人眼中的外國飛地，在另一些人眼中是全球城市的雛形。所以那些真正想在自己國土上出現下一個紐約、下一個杜拜、下一個新加坡的國家，會發展得非常好。關鍵在於，至少這整件事讓我看清楚的是，想要的國家很多，多到這件事會贏。至少我是這麼認為。對，我同意。到目前為止，你覺得哈薩克的新節點怎麼樣？你應該來看看。你會說俄語。沒錯。他們會鋪紅地毯歡迎你。It's like an individual, and they're only recreating. But a digital nomad community, it could be thousands of people. They could be there for years, and they're also working. So it looks like a foreign enclave, but one man's foreign enclave is another man's global city in embryo, right? So the countries that actually want the next New York, the next Dubai, the next Singapore on their soil, those are going to do very well. And the crucial thing, at least the thing that this whole thing surfaced for me, is that there's a lot that do, enough that do, that this is going to win. That's what I think, at least. Yeah, no, I agree. How have you been finding the new node in Kazakhstan so far? You should come and visit. I mean, you speak Russian. That's true. That's right. And they'll roll out the red carpet.

24:01Olaf 在那裡，Shervin 在那裡，Mike Maples 也會過去，很多 Zcash 的人也會過去。有趣的是，我喜歡的國家很多，但我真的很愛哈薩克。因為至少就現在來看，前景也不錯，它是一個真的能蓋新創城市的地方。它的人口只有新加坡的三倍，面積卻是新加坡的 3,000 倍。所以我打算這樣做：我們先蓋自己的據點，從世界各地吸引人才，把創投帶進來，或由創投把人帶進來。我們把規模做到 10,000 人、100,000 人，達到之後，就可以有 Ethereum City、Coinbase City、Stripe City。We have Olaf there. We have Shervin there. We have, I mean, Mike Maples is going to come up there. A lot of Zcash people coming up there. You know, it's funny because, like, there are countries that I like, but I really love Kazakhstan because it, at least right now, and it looks good, is a place you can actually build startup cities because it's only three times the population of Singapore, but it's 3,000 times the space of Singapore. So, you know, we can have — so here's what I plan to do. We're going to build, you know, our hub. We could get talent from around the world. We could bring in the VCs, or the VCs bring in people. We get that to 10,000, 100,000 people, and once we've got that, then you have Ethereum City, you have Coinbase City, you have Stripe City.

24:48也許還有 Tempo City，不管是什麼，你也可以在哈薩克有 Shopify City。有了幾座這樣的城市之後，順帶一提，那裡也能成為這些公司安置全球人才的據點。然後你就可以把這個模式複製到各地，因為這就是新創城市育成中心。我認為過去缺少的關鍵，是一個同時擁有數位創新和建設空間的國家。新加坡我很愛，但它沒有空間。舊金山也沒有。數位加上實體，這是新的。我覺得這很有意思。關於這些東西的成熟度，我想問的另一個問題是：你覺得哈薩克節點日常的運作，以及大家繼續待在那裡的意願，現在有多大程度依賴你？Maybe Tempo City, whatever, you know, and you could have Shopify City in Kazakhstan. And then once you've got a few of those, that's also, by the way, a place that they can base talent from around the world. And then you can scale that model everywhere, right? Because the startup city incubator. And I think the key thing that was missing was a country that had digital innovation and the space to build. Yeah, because Singapore, I love it. It doesn't have that, right? San Francisco doesn't have that, right? Digital plus physical is new. I think that's interesting. So, in terms of, like, maturity of these things, the other kind of question I had in mind was: how much do you think does the Kazakhstan node, just, like, day-to-day, it running, and also, like, people's interest in continuing to be there — how much do you think that depends on you right now?

25:40「最少必要的 Balaji」與 Network School 的未來

25:40好問題。這是我們稱為「最少必要的 Balaji」的東西。好。「最少必要的 Balaji」的意思是：現在我必須親自做這件事。我很愛你們每一個人，別誤會，但基本上目標是把我自己去中心化，培養一百個新創社會的領導者，這樣到某個時間點，我就可以放鬆，回去做數學、生物和遺傳學，喝喝咖啡。但現在我確實覺得，我需要親自把這整件事啟動起來，為我們創造選項，因為舊世界正在崩解。具體來說，Network School 最早的第一版，有相當多人是因為所謂的名人效應而來的。So, great question. So this is something which we call minimum necessary Balaji, okay? Okay. And minimum necessary Balaji is: right now I have to do this, okay? And I love all you guys, don't get me wrong, but — basically, the goal is to decentralize myself and have a hundred startup society leaders, so at some point I can just chill and go back to just doing math and, you know, like biology and genetics, right, and just sip coffee. But right now, I do feel that I need to kind of bootstrap this whole thing to give us options because the old world is melting down. So actually, just on this concretely, you know, the very, very first version of Network School, there were a fair number of people there who were there for, let's call it celebrity.

26:29這也沒關係。但到了第二版，重點就完全是社群了。我大概一個月講一次課，甚至更少。大家真的不是為了我而來。就像你來新加坡，是為了李光耀嗎？不是，至少現在不是了。他建立了一套制度。這在幾個方面也是開明的自利。第一，你擴展了自己，不必事事親為。第二，這是別人可以複製去做的東西。這就是我的想法。我想問，你認為 Balaji 在哈薩克的 Network School 這個機構，壽命能不能超過你高度積極參與的時期？還是你的願景是，N 年後整個東西會分裂？Okay, which is fine, whatever. But then by V2, it was very much about community. And maybe I would give a talk like once a month, if that, you know. But it really wasn't — people weren't there for, in the same way, when you come to Singapore, are you there for Lee Kuan Yew? Not really, right? Not anymore. You know, he set up a system. And it's also enlightened self-interest in a few ways. A, you scale yourself, so you don't have to, right? B, it's something which others can copy and do, right? So that's how I think about that. I guess, do you see, like... the institution that is, you know, like Balaji's Network School in Kazakhstan as having something that, with a longevity that exceeds, like, your level of very active involvement, or is your vision that, like, after N years the whole thing, like, splits?

27:19喔，不是。我是這樣想的。哈佛一開始只是一間單一教室的學堂。它的角色是什麼？是建立麻薩諸塞灣殖民地，殖民地後來建立了麻州，麻州又在建立美國、贏得南北戰爭、贏得一戰、二戰和冷戰的過程中發揮影響力。所以基本上，哈佛輸出的是它的理念，也就是「濕體」（wetware）。我認為這是最被低估的一層。我們談硬體、談軟體，但濕體，也就是我們腦中的軟體，才是被低估的那一層。哈佛的濕體在大約 400 年間一直最具主導地位，直到最近 30 年，史丹佛開始變得更成功。為什麼？Oh, no. Yeah. Well, so here's how I think about this. Harvard started as a one-room schoolhouse, right? And what was its role? Its role was to build the Massachusetts Bay Colony, which then built Massachusetts, which then basically was influential in building the United States of America, in winning the Civil War, and then winning, you know, World War I, World War II, the Cold War. And so, you know, basically Harvard's export of its ideas, the wetware, right? That's, I think, the most underappreciated layer. We talk about hardware, we talk about software, but the wetware, the software in our heads, that's, you know, the underappreciated layer. So Harvard's wetware, basically over 400 years, was the most dominant until the last 30 years, where Stanford started becoming more successful. Why?

28:11因為那是網路從 20 世紀國家手中接手的時期，而史丹佛是建立在網路之上的。現在我認為我們需要 3.0 版。這是在回答你的問題，只是答得比較長。先是計程車，然後是 Uber，也就是數位叫車，現在則是沒有人類駕駛的完全自動駕駛。或者先是實體現金，然後是銀行帳戶裡的法幣，現在則是純數位現金、純網路。我們有哈佛，純離線、純法幣；有史丹佛，像是法幣和網路的混合體。現在我們需要一個屬於後網路秩序的 Network School。我確實認為網路會存在很長一段時間。如果是這樣，我們這些概念也會存在很久，它們會帶領大家建立這些社會，而這些社會又會去建立其他東西。Because that was the network taking over from the 20th-century state, and Stanford was an internet, you know, based thing. And now I think we need a version 3.0. So, just like you had taxis — this is just a long answer to your point. You had taxis, then you had Ubers, which are digital hail, and now you have full self-driving where there's no human. Or you had physical cash, and then you had fiat in a bank account, and then you have pure digital cash, pure internet. We had Harvard, which is pure offline, pure fiat. Stanford, which is like a fiat-internet hybrid. And now we need a Network School that is for the post-internet order. Right. And I do think the internet is going to be around for a long time. And so if that's true, then I think the kind of concepts that we have will be around for a long time, and they'll lead to people building these societies that will themselves build other things.

29:05Network School 的治理

29:05至少想法是這樣。我好奇的是，等 Network School 能自給自足到你可以放鬆的程度，你想像它的治理結構會長什麼樣子？好問題。我希望全世界有一百個類似 Network School 的地方，每個都有自己的領導者。這是簡短版本。我現在開始把這個叫做「去中心化民主」，這個詞不錯。你說。這有道理，但我想問的是，你在哈薩克創辦的那個實體據點，它之後會怎樣？還沒決定。我其實不知道，因為我得先把它做成某個東西。但希望我自己還有 30、40 年可以拚，然後……At least that's the idea. Yeah, I guess I'm curious, like once... Network School kind of becomes self-sustaining to the point where you can afford to chill, like, what do you envision the governance structure of the thing looking like? That's a great question. I want to have a hundred Network School equivalents around the world, which each have their own leaders. Right. That's the short version. So this is what I'm actually starting to call — this is a good term — decentralized democracy. Right. Right. Go ahead. That makes sense, but I guess — the one that is, like, the one that is in Kazakhstan, the physical one that you started, yeah. Like, what will happen to that? TBD. I actually don't know because, you know, I have to build it into something that's, you know, right. But, you know, hopefully I've got 30, 40 years in the tank myself, right, and then...

29:52到那個時候，下一步就會很明顯了。也許那時智慧合約和 AI 已經進步到可以讓它成為真正的 DAO，我不知道。這是一個重要的問題，因為一個東西在你離世之後要怎麼辦？傳承問題要怎麼處理？新加坡有人說過一句很好笑的話：「這是隨機的，多生幾個孩子吧。」因為你不知道誰會接班。這也是另一種去中心化。所以我還不知道答案，但我確實認為答案的一部分，是降低對「我走了之後 Network School 還能成功」的依賴，讓世界上有另外一百個替代選項。You know, at that point, the next step will be obvious. Maybe you have — maybe smart contracts and AI will advance to the point that you can have it be a true DAO. I don't know. You know, this is an important point because, like, what do you do with something after you've passed on? How do you handle the succession problem? You know, someone in Singapore said something funny, which is, "It's random, have more kids," right? Because you don't know, you know, who's going to. Another possibility, that's another way of decentralization. So I don't know the answer to that yet, but I do think part of the answer is basically reduce the dependence on Network School succeeding after I go, such that there's another hundred alternatives.

30:38對，我同意這是答案很重要的一部分。我覺得完全走向那種「快閃後解散」、分叉至上的治理方式，對大學這類東西其實可能非常有效。因為特別是有了現代工具和 AI，單一機構不需要累積長達一世紀的聲譽。但如果要做在場很多人在乎的其他事，例如成立一家飛往火星的火箭公司，那需要 10 年以上的實體和硬體投入。對，我們絕對會做這件事。我們在蓋一個大校園之類的，之後會再多說。你說。我注意到一件事，不管是在蒙特內哥羅辦最初的 Zuzalu，還是之後做的各種事情，成為眾人目光的焦點，過一陣子真的會造成創傷。Yeah, I agree that's an important part of the answer. I think, like, going full-on kind of, you know, like, pop-up-and-disappear sort of, like, fork-maximalist governance, like, I think for something like universities, it could actually work very well. Because especially with, you know, like, modern tools and AI, like, you don't need single things to have century-long reputations. If you want to do other things that I think a lot of people here care about, like, say, make a new rocket company that goes to Mars, like, that requires, you know, like, a 10-year physical and hardware commitment, like more than 10 years. Yeah, we're definitely doing that. We're doing a big campus and so on, but more on that soon. Go ahead. Yeah, like, no, I... One of the things that I've just noticed from, I mean, like, both running the original Zuzalu in Montenegro and doing, you know, like, lots of all kinds of things after that is, like, being at the center of attention, like it really gets traumatizing after a while.

31:39領導者、以太坊與社群

31:39對，你不喜歡那樣，我知道。我也不喜歡，但我的耐受度可能稍微高一點。基本上每個人都想要……我可以提出一個觀察嗎？當然。以以太坊來說，你可以直接推出一套純粹由規則構成的系統，然後大家就可以跟程式碼爭論等等。但在實體世界，你人在 Zuzalu，大家就把你當成 Zuzalu 的總統，即使你並不想當。所以就算你不是正式的領導者，你也是隱性的領導者，大家會緊抓著你說的每一句話。所以在實體世界，我不認為我們已經能擺脫領導，還沒辦法把它自動化掉。你同意嗎？同意。Yeah, you don't like that. I know. Yeah, I don't like it either, but I have a slightly higher tolerance, maybe. Like, basically everybody wanted — you can actually — can I make an observation? Sure. Yeah. So with Ethereum, you could just put out a system of pure law, and then people could just argue with the code and so on and so forth, right? In the physical world, you being at Zuzalu, people took you to be the president of Zuzalu, even if you didn't want to be the president of Zuzalu. And so even if you were not officially the leader, you were implicitly the leader. And then everybody would hang on one statement you said. And so in the physical world, I don't think we can do away with leadership yet. We can't automate it away yet, right? Do you agree with that? Yeah.

32:25正因如此，唯一比有領導者更糟的，是沒有任何領導者，而這正是美國正在走向的狀態：美式無政府。所以我是非常不情願地綁好鞋帶上場做這些事。主要是因為，如果沒有人提出對後網路未來的正面願景，那些緬懷法西斯主義、共產主義這些 20 世紀思想的人，是不會把它建立起來的。這就是我做這些事的原因。對，你得擅長培養和支持很多新的領導者。去中心化不是從一個領導者變成零個，而是從一個變成很多個。沒錯，完全正確。我覺得以太坊正開始走到那一步。And so because of that, you know, the only thing worse than having a leader is having no leaders, and, you know, like, which is what America's going to: American anarchy. And so it's like very reluctantly that I lace up the cleats and I do this stuff, mainly because it's kind of like, you know, if one doesn't have some positive vision of the post-internet future, it's not going to get built by people hearkening to fascism or communism or all these 20th-century things. So that's why I do what I'm doing. Yeah, you have to get good at, like, building up and supporting, like, many new leaders, right? Like decentralization, it's not about going from one leader to zero leaders. It's about going from one leader to many. Yeah, exactly. That's right. Yeah, like I feel like Ethereum is starting to get there.

33:15像 Justin 已經學會了一項重要技能，就是能跟我一樣有效地激怒 Crypto Twitter。這很重要。好。所以問題是，這些社群什麼時候能達到那種程度。這很難，因為 Ben Horowitz 也談過這件事：創辦人兼執行長這種人，很少會有一個跟自己相似的關鍵副手。偶爾會有 Bill Gates，然後是 Steve Ballmer，再來是 Satya Nadella，他其實有點像那樣，但這很罕見。通常的情況是，像 Steve Jobs 在 Apple 就沒有另一個 Steve Jobs。也許你可以說，那個原本會成為……算了，我不想點名。You know, like Justin has acquired the important skill of making Crypto Twitter angry as effectively as I can. So that's important. Yeah. Okay. So the question is, like, when any of these can get to, like, that kind of level. It's hard because, you know, one of the things is, you know, Ben Horowitz has talked about this also. The kind of person who is a founder CEO, it's relatively rare that they have a key lieutenant who's as similar. You know, once in a while you have Bill Gates and Steve Ballmer and then Satya Nadella, who's actually kind of like that, but it's rare. Right. Usually. It's something like, you know, like Steve Jobs did not have another Steve Jobs at Apple. Maybe you could argue, gosh, the guy who was going to become — well, I don't want to name names.

34:04有個人原本可能取代 Tim Cook 當上執行長，但正因為這樣，在網路時代的結果往往是，那個人會離開去做別的事，而不是留在既有的機構裡。所以傳承計畫也許就是去中心化。但以太坊的情況是，你其實沒辦法離開以太坊。不管你做什麼、說什麼，只要你還活著，大家基本上都會這樣看你，除非你跟以太坊公開決裂，而你不該這樣做，誰都不該這樣做。所以在這個意義上，你沒辦法擺脫你對它的影響力。對。我覺得對我來說，我一直在找的正面訊號基本上是：基金會和社群裡的其他人，能不能在沒有我參與的情況下，既推動它前進，又捍衛正確的價值？There was a guy who might have become CEO instead of Tim Cook, but because of that, it kind of turns out in the internet that that person goes and does something else rather than necessarily working within the existing institution. So that is maybe the succession plan is decentralization. But Ethereum is doing — the thing is, you can't actually leave Ethereum. Everybody, no matter what you would do or what you would say, as long as you're alive, people will basically say, unless you had some — which you shouldn't do. Do not do this for anybody — some public breakup with Ethereum, right? So you're not going to be able to go away from your influence over it in that sense. Yeah. I think, like, for me, the positive signs that I always look for is basically, are — other people, you know, in the foundation and in the community, able to kind of both push it forward and kind of defend the right values, even without my involvement?

34:58我覺得最近這種情況越來越多，而且發生得非常有效。很多人都在堅持說：「嘿，我們真的必須重視安全。」「嘿，我們應該重視隱私。」然後……你把話說完。對，基本上還是一樣，把去中心化理解成從一個變成很多個，而不是從一變成零。我覺得看到那一層的人能自己出現、自己把事情做好，不用等我開口，這真的是最關鍵的事。太棒了。好，我們去吃午餐吧。好，謝謝。謝謝你，Balaji。And I feel like recently that's been happening more and more, and it's been happening very effectively. Like, a lot of people are, you know, insisting that, you know, hey, we actually have to care about security. Like, hey, you know, we are supposed to value privacy here. And, like, and — Go finish what you're saying. Yeah, like, basically, you know, like the — again, it's like sort of focusing on decentralization not as going from one to zero, but as going from one to many. And I think, like, seeing the presence of that kind of layer able to just appear and do its thing, even without kind of waiting for me to say something, I think has, like, really been the key thing. Awesome. All right, great. Let's get some lunch. Good. Thanks. Thank you, Balaji.