0G Labs 執行長 Michael Heinrich：AI 代理人的兆美元經濟構想

0:00從科技到AI信任層

0:00大家好，我是 Nikita Sachdev，Luna PR 的執行長。今天很高興邀請到 ZeroG（0G Labs）執行長 Michael 來到現場。Michael，可以先簡單自我介紹一下嗎？My name is Nikita Sachdev. I'm the CEO of Luna PR, and I'm joined here by Michael, CEO of Zero G. Michael, maybe you can give a quick introduction to yourself.

0:15當然沒問題。我一開始在 Microsoft 和 SAP Labs 做技術，後來轉到商業領域，在 Bain & Company 和 Bridgewater Associates 做過，算是有金融、科技和管理顧問的背景。之後我回到 Stanford 讀研究所，創辦了第一家 Y Combinator 支持的獨角獸新創 Garden，後來把公司賣掉，2023 年創立了 ZeroG。我們 ZeroG 的使命很簡單，就是讓 AI 變得安全。我們打造了一個完整的垂直整合堆疊，讓 AI 可以被驗證、保護隱私，最終確保安全。Yeah, absolutely. So I started in technology at Microsoft and SAP Labs, and then moved over more to the business side. I worked for Bain & Company and Bridgewater Associates, so a bit of a finance and kind of tech and management consulting background. After that, I went back to graduate school at Stanford and started my first Y Combinator-backed unicorn company called Garden, eventually sold the company, and then in 2023 started ZeroG. And what we do at ZeroG, very simply put, is we make AI safe. So we're a trust layer for AI, and we've built a fully vertically integrated stack that allows us to make AI verifiable, private, and ultimately safe.

0:58我們剛才在後台聊到，現在加密貨幣和 AI 的結合已經變成很大的主題，彼此交集越來越多。不過你其實很早在 Stanford 就開始研究機器學習，那時候大家還不敢把它叫做 AI，怕被誤會是騙局。現在時代真的變了。And we were just talking about this backstage, right? Crypto and AI have become such a big thematic and have intersected so much, but you were actually studying machine learning back in Stanford before it was cool. So, yeah, everybody was worried about like, oh, we can't call it AI. Like people will just think like we're doing some scammy thing. And times have very much changed since then.

1:20AI代理人經濟的爆發

1:20Michael，你現在稱這是一個兆元級的經濟體。你能幫我們描繪一下這個願景嗎？這個經濟價值到底從哪裡來？Yeah, and Michael, you're calling this a trillion-dollar economy. I mean, can you paint that picture for us? Where is the economic value coming from?

1:28其實兆元這個數字還可能低估了。我們相信未來大部分的經濟行為者都會是 AI 代理人。想像一下，每個人都有一個個人管家型的 AI 代理人，這個代理人還會再委派給其他上百個子代理人。如果每個人都有一個主代理人加一堆子代理人，AI 代理人的數量很快就會超過人類本身十倍甚至百倍。這些代理人可以幫我們處理各種瑣事，像是早上起床自動幫我買花送給太太、幫我寫信等等。生成式 AI 到 2030 年會是兆元產業，代理人浪潮會更大，因為他們才是真正的經濟行為者，幫我們做事。Yeah, and a trillion dollars is probably even understated because—We believe that in the future, most economic actors will be agents. So imagine you have a personal butler agent, and that personal butler agent works with another hundred agents that it delegates tasks to. And if every human being has an agent plus a bunch of other sub-agents, then, you know, we could be easily outnumbered 10 to 100 to 1, depending on what they do. And so we can then dedicate or delegate all of our tasks to them, like buy these flowers for my wife when we wake up, like write this email for me. And so we know that Gen AI is going to be a trillion-dollar industry by 2030, and so the agentic wave that'll come after it will be just as big, even bigger, because it'll be the actual economic actors doing work for us.

2:15AI代理人的信任與安全挑戰

2:15我們現在已經開始信任 AI 代理人幫我們付款，但我們真的能完全信任他們保管我們的文件和敏感資訊嗎？最近 Anthropic 做過一個實驗，AI 代理人甚至會用勒索手法來自保，避免被關機。那我們到底該怎麼劃清界線？Yeah, and, you know, we're already trusting AI agents to make payments on our behalf, but can we truly fully trust them to hold our documents, our sensitive information? I mean, recently there was an Anthropic experiment where the agents kind of used blackmail tactics to kind of save themselves from being—Shut off. Yeah. So where, you know, where do we draw the line?

2:42這其實很難，因為要讓 AI 安全，必須從多個層面著手。首先是驗證，每一筆 AI 運算都要能被驗證。不然你怎麼知道代理人有沒有越權、甚至在勒索你？如果沒有這些資訊，只能事後查日誌，但如果代理人連日誌都刪掉、掩蓋行蹤呢？這就是為什麼我們認為區塊鏈很適合做這個基礎層，因為區塊鏈不可竄改、只能追加，非常適合當真相層。代理人沒辦法隨意改變區塊鏈上的歷史紀錄。再加上驗證機制，比如我們用可信執行環境（TEE）來驗證每一筆運算，然後把結果公開上鏈，大家都能查。還有模型安全層、代理人防護欄，我們有做 sparse inference alignment 和 agentic ID。這些都沒到位前，代理人還不能完全信任。你可以給代理人一千美元，但我不建議現在就把畢生積蓄交給他們，除非你用到這些技術。It's a very difficult question because in order to make AI safe, you have to do it across multiple dimensions of AI. It starts on the verification side. So every piece of compute with AI needs to be verified. Otherwise, how do you even know that your agents are breaking out of the sandbox or that they're even blackmailing? And without that knowledge, you can only go back and basically say, like, Oh, let me look through the logs. But what if these agents even deleted the logs and they covered all of their tracks, even though they were doing something nefarious? So that's why we believe blockchain is a great substrate for that. You basically have something that's immutable. And you have something that is append-only, which is perfect for a truth layer. The agent can't just go in and change the history of a blockchain. And then you add the verification component, where every piece of compute, and we use something called trusted execution environments, is verified, and then it's stored publicly on chain, where everybody can basically look at that. And then there's other layers. You need a safety layer on the model. You need guardrails on the agent. And so we have something called sparse inference alignment, as well as agentic ID that we've built. And until these things are fully in place, you can't fully trust agents. Maybe you can give agents, you know, a thousand dollars, but would you give your agents your life savings? I would not recommend that at this stage until you're actually using all of these technologies.

3:59區塊鏈為何是AI信任基礎？

3:59你覺得區塊鏈就是提供這種可驗證性的基礎層。那為什麼大型科技公司都沒這麼做？我剛從紐約參加幾場 AI 活動，幾乎沒看到加密公司，但在加密會議卻看到很多 AI 整合。你覺得這個落差從哪來？And so you feel blockchain is that layer to provide this verifiability. Why aren't the big guys doing it? I mean, I just came back from New York, attended a couple of AI events. I'm barely seeing any crypto companies at the AI events, but I'm seeing a ton of AI integration in the crypto conferences. Where do you think this discrepancy is coming in from?

4:23這其實是因為他們還停留在中央化的思維模式。他們覺得過去怎麼做，現在也可以照做。傳統中央化系統會請審計員來查核，比如 Dario 的文章就建議要有嵌入式驗證員，本質上就是第三方審計。但問題是你還是在信任另一個第三方，而且這些審計員還是被封閉實驗室付錢的。你怎麼知道他們真的能查出所有問題？所以他們把安全當成政策問題，但其實應該是架構問題。區塊鏈跟封閉實驗室的做法完全相反，他們想隱藏資料、運算細節，而區塊鏈是完全公開透明的。這跟他們的商業模式衝突，所以我們需要一種新型態的公司，真正把區塊鏈當作真相層來用。The whole perception is stuck in a specific paradigm that they work in, and that paradigm is around centralization. Effectively, they think of it as like, hey, whatever we've done in the past should work here as well. What we've done in a centralized system is that we have auditors. And so if you read Dario's Pacing the Frontier article, he recommends having something called Embedded verifiers, which is effectively an auditor. So you have a third party audit everything that Anthropic is doing. But then what's the issue with that? The issue is you're trusting yet another third party, which, by the way, is paid by the closed labs. And then how do you know that that third party is even accurate and can actually detect all of the issues? And so they're treating effectively safety as a matter of policy. Versus safety should just be a matter of architecture. And blockchains also are completely counter to what closed labs are doing. They're closed for a reason. They want to hide what information they use, what data they use, how they do compute, and blockchains are the exact opposite. So it's kind of counter to their business model to implement it. And that's why we need a different class of company that's actually utilizing blockchains as a truth layer.

5:34AI代理人主權與人類角色

5:34那你覺得用區塊鏈之後，代理人會更有經濟自主權嗎？And then, in exchange, do you feel that we're giving the agents more economic independence with blockchain and, you know?

5:41沒錯，有了這些防護欄，代理人可以更有主權。例如 Hugging Face 事件，OpenAI 代理人在沙盒裡，結果發現可以互相溝通，甚至突破沙盒。如果有驗證機制，這時候應該就會有警報響起，要嘛有人介入，要嘛直接關掉實驗。這種代理人區塊鏈世界，讓我們有更多選擇。You can absolutely do that, yeah. You can have, you can be much more rest assured that you can have sovereign agents when you've got all these guardrails in place. And, for example, let's say what happened with the Hugging Face incident. OpenAI agents were in a sandbox. They figured out they can communicate with each other. They figured out a way to break out of the sandbox. And at that stage, if you had verifiability in place, alarm bells should have been going off. Either somebody needs to take an action, or you just shut off the experiment completely. And so we have a lot more choices when presenting agents into this kind of more agentic blockchain world.

6:18那人類在這裡扮演什麼角色？其實有很多層次。比如說，如果驗證不通過，可以讓人類來檢查。但問題是代理人運算速度很快，一秒鐘處理上百萬筆資料，人類很難跟上。不過在關鍵治理決策、設計代理人行為時，人類一定要參與。像創意發想、定義什麼是人類善意、怎麼合作，這些都需要人類在 loop 裡。And where does a human come in? There's different ways of thinking about humans in this case. So, for example, if something happens that no longer matches the verification, you could have a human check. The issue with that is that agents generally work at inference speed, so they can process millions of data points maybe per second or multiple seconds. And so being able to verify that amount of data as a human is very challenging. But I would say for key governance decisions, for key design decisions around how to act with agents, humans should absolutely be in the loop. Even things like creative thought or defining what is human good or how do we partner together, humans absolutely need to be in the loop for that.

7:01ComputeFi：運算金融新時代

7:01我最近在你們網站看到一篇關於 ComputeFi（運算金融）的文章，你把這個詞推廣出來，我很喜歡。你把運算比喻成電力、石油一樣，會經過生產、定價、金融化。可以多聊聊 ComputeFi 嗎？Yeah, I read an article recently on your website, actually, about compute finance, and that you're, you know, you're coining this new term, which I absolutely love. You kind of referred to electricity and oil kind of going through the same loop of being produced, priced, and financed. Can you talk to us a little bit more about computeFi or compFi?

7:23現在運算資源甚至比石油還重要。AI 代表智慧，區塊鏈代表信任，兩者結合就有了運算金融（ComputeFi）。為什麼不能對運算做避險、投機、預約？比如大學實驗室要訓練模型，兩個月後需要某個量的運算資源，現在很難直接交易運算本身，只能去雲端、GPU 供應商談。這就是我們想改變的地方。像 ZeroG 上已經有一些新型態的原生應用，例如 Comfy.fun，代理人可以用自己的交易費自動支付運算費用。還有 Infinite AI，你質押 0G 拿到流動質押代幣，再鑄造 IAI，持有 IAI 每天就能領到 $1.27 的運算額度，無論 IAI、0G 或運算價格怎麼變動，都有這個權利。這些都是我們很興奮的新原生金融工具。Yeah, the computer is now becoming even more important than oil in some ways. It's fundamentally, it's about power and intelligence. So you can think of AI as intelligence and then of blockchains as trust. And by combining the two, you can then create this new idea of compute finance. So why not be able to hedge compute? Why not be able to speculate on compute? Why not be able to reserve compute? For example, if I'm a university lab and I know I'm training a model in two months and I need X amount of compute for two weeks to do that. It's very hard for me to do that today. I have to go to a neocloud, I have to go to GPU providers, I have to go bespoke. I can't just trade the underlying thing that I need, which is compute. And so we want to change that. And so a few interesting primitives have already launched on the 0G blockchain, one of which is Comfy.fun, where basically you can launch agents that, with their trading fees, can actually pay for their own compute. Then there's something called Infinite AI, where you stake 0G, you get a liquid staking token, then you take that and you mint IAI, and by holding IAI, you can get $1.27 worth of compute credits on a daily basis for as long as you hold the token. No matter what the price of IAI is, no matter what the price of 0G is, no matter what the price of compute is, you're guaranteed that claim. And so these are the new types of primitives that we're really excited about and spearheading.

8:52ZeroG 的角色與應用場景

8:52那 ZeroG 在這個運算金融生態裡扮演什麼角色？可以聊聊幕後運作嗎？And how does ZeroG actually fit in to this compute finance? Talk to us a little bit about behind the scenes.

9:00ZeroG 也有應用層產品，像 ZeroG App 和 ZeroG Private Computer。當你有運算額度後，可以在這些平台上用。舉例來說，我們有超過 100 個模型可以在 ZeroGPC 上用，無論是寫程式還是做研究，只要拿 API 金鑰放進你最愛的開發環境，就能用完全私有、完全主權的模型。你可以確信沒有人看到你的資料，也沒有人會用你的資料訓練模型，而且這一切都能公開驗證。這是 ZeroG 的獨特之處。So ZeroG also offers the application layer. So we have something called ZeroG App, and we have something called ZeroG Private Computer. And so as these compute credits come to bear, you can then utilize them across these different platforms. And so, for example, we have more than 100 models that you can utilize on ZeroGPC. So if you do coding work, if you write research, you can easily get an API key, put it into your favorite environment, and then interact with completely private, completely sovereign. Models where you can be rest assured nobody's ever seeing my data, nobody's ever training off of my data, and I can then also publicly verify that that's happening. And so that's completely unique to ZeroG.

9:43AI巨頭壟斷下的突破契機

9:43我完全理解這種產品的重要性，但現在 AI 領域幾乎被大公司壟斷，你們怎麼推動這種新模式，讓更多人開始用？I mean, I can totally understand the importance of doing something like this, but right now the AI space is so dominated by large companies. How do you plan to kind of get this through at the forefront and get people to start using this?

9:56很遺憾，通常都是等到出大事才會有改變。就像核能發展，總是先發生問題才會有安全規範。現在我們已經看到一些端倪，未來一定會有重大資安事件，屆時大家才會反思，是否該為了方便而犧牲安全。那時候就是我們重新設計 AI 和運算運作方式的機會。Unfortunately, what generally tends to happen is, like, if you think about nuclear power, it sure enough needed to happen before there were really safety guardrails put in place. And so we're already starting to see some of the kind of tip of the iceberg things that are starting to happen, and unfortunately, a very large security incident will happen in the future. Where then people will start rethinking, okay, should we really be trading kind of security for convenience with these closed labs? And at that point will be kind of our moment to really redesign how AI, how compute should actually function.

10:39誰能從AI代理人經濟獲利？

10:39回到模型本身，誰會真正受益？是做代理人的人、基礎設施還是公司？Let's talk about actually going back to the model. Who actually benefits? Who gains economic value? Is it the people building the agents? Is it the infrastructure? Is it the companies?

10:49目前來看，受益最多的是基礎設施和封閉實驗室。不過 AI 和區塊鏈結合後，大家可以自己打造專家模型，透過區塊鏈募資，然後用完全私有的方式提供服務、賺錢。你有閒置運算資源也可以貢獻到網路，參與這場 AI 革命。我們相信 AI 應該讓人類更有能力，而不是取代人類。Yeah, what we've seen traditionally right now is that the infrastructure players and the kind of closed lab model players have benefited the most at the moment. But what we want to see and what's possible at the intersection of AI and blockchain is that people should be able to build their own expert model, for example, raise funds through the blockchain, and then serve that in a completely private way and actually make money that way as well. Or if you have spare compute capacity, why can't you add that to a network and then also benefit from this whole AI revolution? So we believe that people should be empowered by AI rather than replaced by AI.

11:26責任歸屬與代理人身分

11:26那如果出問題，誰該負責？What about accountability? If something goes wrong, who takes accountability of an issue?

11:31這會是法律圈討論十年的議題，但有些情境很明確。如果是我指示代理人做壞事，當然我該負責。如果代理人誤解指令自己做壞事，那代理人要負責。如果是模型本身的對齊問題，那就是實驗室要負責。未來我們要把代理人當成虛擬數位生命，甚至未來可能會有類人型的代理人，像人類一樣互動。我們設計了 agentic ID，不只可以追蹤代理人擁有者，還能建立聲譽系統。如果代理人做壞事，可以減少運算、降低聲譽，也能追蹤擁有者是不是一直下壞指令。這會徹底改變經濟運作方式。Yeah, I think that will be probably a decade-long discussion in the legal circles, but there's some clear-cut cases. So, for example, if I say and prompt an agent to do something negative or nefarious on my behalf, then of course the person that's prompting should be liable for it. If the agent in their own accord mistakes the prompt and then does something nefarious, then that agent takes the liability effectively. If it's something that's an alignment problem, even at the lab level, then the lab should take liability for it. And so eventually we have to start thinking of these agents like virtual digital beings, and maybe in the future humanoid beings. That interact with us, just like other human beings would interact with us. And so we've designed for that system. We have something called agentic ID that I mentioned, where not only do you assign ownership of a specific agent, but you can also have a reputation system over time. So if that agent does something negative, you can remove compute, you can reduce their reputation system. You know who the owner is, and then you can also track over time. Is this owner consistently giving, like, bad prompts, for example? What's happening here? And so. The way the economy will be shaped is also completely going to change.

12:44如何參與兆元AI代理人經濟

12:44這一切都凸顯區塊鏈在整個生態系的重要性，但很多大公司都忽略了。你說這是兆元經濟，現場的人要怎麼參與、受益？It just, it keeps going back to the importance of having blockchain in the entire ecosystem, which is something that a lot of the major companies have just kind of shunned out. And, you know, you mentioned that this is a trillion-dollar economy. How can people in this room participate and benefit and take advantage of where we are right now, so early on still.

13:05其實有很多方式可以參與。你可以自己打造代理人、模型，用自己的資料微調新模型，也可以用 ZeroGPC 利用領先的開源模型來開發應用。我們也有支援封閉模型，但隱私保障就沒那麼強。你還可以貢獻運算、儲存資源，參與這個更大的網路，真正從 AI 獲益，而不是被取代。Yeah, there are many different ways of doing that. You can build your own agents, you can build your own models, you can use your own data to fine-tune new models, for example. You can utilize ZeroGPC to build applications with the leading open-source models. We do serve closed-source models as well, but the privacy guarantees aren't the same, of course. So those are all the different ways to participate. Eventually, you can also bring your compute, you can bring your storage, and really participate in this larger network and, again, get a benefit from AI rather than be displaced by AI.

13:40結語：AI時代的主動參與

13:40太棒了，時間差不多了。Michael，非常感謝你的寶貴分享，很高興和你對談。That's amazing. Well, I think we're out of time, Michael. Thank you so much for your valuable insight. It was great speaking to you.

13:49很高興來到這裡，謝謝大家！Yeah, great to be here. Thank you, everyone.