Junny Ho（Datadance）談 AI 時代的個人資料資產網絡

0:00資料才是新時代的算力

0:00自從加密產業和 Bitcoin 問世以來，我們一直很著迷於挖礦算力、GPU、ASIC 礦機。近年來，隨著 AI 越來越受關注，我們又開始瘋狂追逐電力合約、晶片和新型運算能力。但今天我想談的不是大型模型或算力，而是資料。我們認為，資料將成為未來幾十年新的挖礦方式和資源。Since the moment we have this crypto industry and we have Bitcoin, we've been quite fascinated by mining power and GPUs and ASIC miners. And in recent years, as AI is gaining more and more attention, we started to, you know, be obsessed with power contracts, with chips, and with new kinds of computing power. So I think in today's talk, I want to talk to you about data, not the large models, not the computing power, but we think data is a new mining and a new resource for the next decades of years.

0:43我今天的主張是：資料就是新的算力。但現在有數十億筆資料被鎖在每個人的裝置裡，未來幾年 Datadance 要來解鎖這些資料。那我們就從大家都有的東西開始說起吧。The thesis statement I'm giving today is data is the new compute. Yet billions of data are locked in individual devices and to be unlocked by Datadance in the next few years. So let us start with something everyone already has, all right?

1:02個人資料的價值與現況

1:02想像這是我的手機，裡面有 Uber、Booking、Airbnb 等各種 App。這支手機可能有我九年的購物紀錄、上千筆訂單，還有六年來 Uber 叫車的紀錄，超過 800 次行程。還有外送、旅遊訂票，甚至我參加各種加密會議的紀錄。問題來了，這些資料我該怎麼辦？這些都是已經存在的歷史紀錄，不需要再重新產生。So imagine this is my phone. And obviously I have Uber, bookings, Airbnb, a lot of apps in there. And in this form, I may have nine years of shopping records with thousands of orders. I may have Uber or driver orders, the ride-hailing stuff for the past six years, over 800 trips. I have food deliveries. I have travel bookings. I made a lot of trips to different crypto conferences. I have all those kind of things. Now, the question is, what should I do with this data? These are histories that already exist, and I don't need to be generated as a new figure to the society.

1:42請大家理解，這些資料對企業有很大價值。比如我的購物紀錄可以讓商家知道我的品牌忠誠度和價格敏感度，叫車紀錄能看出我住哪、工作地點，外送紀錄反映我的生活習慣，旅遊訂票則透露我愛搭哪家航空、單趟會花多少錢。現在的行銷和定價團隊花很多時間猜測，或買市場上的分眾資料，甚至模仿類似族群來預測我的行為。但事實是，這些真實的歷史資料早就存在，為什麼不能直接買賣？So please understand that these data are containing a lot of value to business. For example, this behavior data tells a very illustrative figure of myself. The shopping history can tell the business owners about my brand loyalty and my price sensitivity. My riding history will show where I live and where I work at. The food delivery showed my routines. The travel bookings showed my favorite flight airlines and also how much I will spend on a single trip. So the current issue is we see these marketing and pricing teams, they spend a lot of time trying to make guesses, or they would buy segments from the market, or sometimes they even have to mimic the lookalikes to guess my behavior. The bitter truth is, while there are already years of observed real data and behavior, which can definitely beat any audience segment, why cannot people buy and sell this data directly?

2:49AI 時代資料規模法則

2:49為什麼這個時刻特別重要？我們認為資料的規模效應就像算力一樣。在 AI 時代，模型表現會隨著算力和資料量增加而提升，這就是所謂的 scaling law。每增加一點資料或算力，模型表現都會有提升，雖然邊際效益會遞減，但整體曲線還是持續上升。DeepMind Chinchilla 的研究證明，想讓模型持續進步，資料量必須和算力同步成長。So, firstly, why does this matter at this precise moment? We think that data also scales like compute. What does it actually mean? It means that in times of AI, we learn that the performance of models improve as you add on to more compute and data. So that is called the scaling law. And every addition always improves a little margin to it, or a little less every time, but the curves always keeps rising. The DeepMind Chinchilla work showed already that in order to keep improving the performance of the model, the data have to grow in step with the compute power as well.

3:33所以要訓練好模型，永遠都需要更多資料。這不只適用於聊天機器人或語言模型，Meta 的研究也發現推薦系統同樣有這種規模效應。推薦系統決定你每天看到什麼、被推薦什麼、價格怎麼訂。實際測試顯示，資料量提升能讓線上樣本的表現提升 12%，這是很大的進步。這個規律過去推動 AI 實驗室，現在也推動每一家消費型企業。And therefore, to train this model sufficiently, you will always need more data. And this is not just about chatbots and language models. The meta research also showed similar patterns when it comes to recommender systems, which are the engines that decide what you see, what you are offered, and at what prices on a daily basis. So, in real-life tests, they already showed about 12% lift in online samples, and this is a huge improvement. And therefore, there is a bridge, right? The same thing, the same law that drives AI labs in the past is now driving every single consumer business.

4:18簡單說，所有需要推薦、分眾、定價的消費型企業，只要有更多資料，就能做得更好。但如果你比較算力和資料這兩種資源，會發現很大不同。算力已經有成熟市場，你可以按小時計價租 GPU，價格公開透明、供應商競爭。但資料，尤其是行為資料，卻沒有這種市場。Basically, every consumer business that recommends, that targets, and that prices can benefit from having more data to have better results. But if you compare between these two inputs, if you compare between the compute and the data, you will see things very differently. For compute, you already have a mature market. You can literally rent a GPU at an hourly rate, where you have prices publicly, you have standard units. Maybe you can potentially rent from a few competing suppliers, right? But for data, especially behavioral data, there's no such functioning market.

4:52資料市場的缺口與機會

4:52每一筆資料都被鎖在你的手機或不同平台上。諷刺的是，任何企業都沒辦法合法取得用戶的歷史資料，因為會違反資料隱私法。所以這件事只能從用戶自己這一端發生。我們看到的機會不只是要有更多資料，而是要能以乾淨、驗證過、授權過的方式收集這些資料，這正是 Datadance 的使命。Every data is locked in your own phones or on different platforms. And literally, the sarcastic fact is that there's no possible way for any business to legally upload and acquire user historical data without violating the data privacy laws. And therefore, this can only happen from the user side themselves. What the opportunity we see here is not just about having more data, but we see that you have to collect this data in a clean and verified and authorized way. And this is something Datadance is missioned to do.

5:38有人會問，真的有這麼大的需求嗎？答案是肯定的。以零售媒體為例，他們每年花兩千億美元在用第一手購買資料做的廣告，這數字顯示企業多重視真實消費行為。反觀資料供給卻在萎縮，Apple 調查顯示，只有 35% iPhone 用戶願意同意跨 App 追蹤。總之，零售、旅遊、飯店、品牌、外送、本地服務等都願意花錢買這些資料，但取得管道卻一直被拒絕。Then people may ask. Do we really have such a strong demand to this? Well, we say yes, because if you look at the retail media, that they plan to spend around $200 billion on advertisements that are built with first-hand purchase data. And this is a number that shows how they value knowing what people actually have to buy. In contrast, the supply is shrinking. The Apple's own survey of iPhone users show that only 35% of users would give consent to cross-app tracking of their data on their phones. For whatever reasons, you don't have those data at all. So, in short, retailers, travel and hotels, consumer brands, food deliveries, local services. They all are willing to pay to get this data. There is a very strong demand, but a constant denial of access channels.

6:34個人資料類型與 AI 應用

6:34這不只關於購物資料。我們把個人資料分成十大類，每一類都有買家。例如消費資料賣給品牌和零售商，金融資料給銀行和放款機構，定位資料給房地產或供應鏈，技能資料給獵才公司，健康資料給保險和研究單位，甚至 Web3 原生資料給加密應用。這十類資料都有現成需求和流通管道。This is also not just about shopping, right? We basically categorize personal data into ten categories, and each of these are with buyers. For example, you can have consumer data going to brands and retailers. You can have financial data going to banks and lenders. Have location data, I don't know, to property owners or sellers and supply chains. You can have skills data to recruiters. You can also have health data to insurance and research, or even Web3 native data to crypto apps. So each of these ten categories represent existing demands and channels for this data.

7:16除了商業用途，資料也能流向 AI。想像過去瀏覽器用 cookies 記錄你的點擊，網站就能記住你。但這些資料只被網站保留。在 AI 時代，Datadance 錢包讓你一鍵收集歷史資料，只要授權給 AI 助理和代理人，他們就能真正理解你，而不是瞎猜。過去是網站記住你，現在是 AI 真的懂你，而且資料始終掌握在你手上，不會被平台留存。The other part beside business is that the data can also flow to AI. And think about it as you think of cookies on web service. So for now, on a browser, you click on certain items, and your browser will collect it with cookies, and then the website remembers you. However, this data is only retained by the website. And in AI times, things work very differently. The Datadance wallet allows for a one-tap collection of your historical data, and you only need to authorize that to your AI assistants and agents. And then your AI assistant and agents can understand you instead of guessing. So cookies in the past, in Web2 times, that websites remember you, while authorized data with Datadance. Let AI understand you, and the nice part of this is your data is no longer retained by the website or platform, but is always under your control.

8:13歷史資料的價值與 Datadance 策略

8:13我知道現在很流行 agentic payment（代理人支付）和 AI 實時產生的新資料，像 Coinbase Ventures、Linus Foundation 都在推這一塊，這確實有價值。但我們要注意，歷史資料動輒八九年，AI 代理人產生的新資料才不到一年。這些歷史資料的價值遠超過一年內的 agent payment。Now, I understand that there is a trending of agentic payment activities and, you know, real-time data generated by AI activities, especially proposed and propagated by Coinbase Ventures and Linus Foundation, right? And we definitely understand this is real value, of course. But can we note at scale that there are already a lot of historical data that has a storage of over eight or nine years, while the AI activities are only emerging for less than a year ago. So I think this is very important because these years of historical data per accounts. Weighs heavily against roughly one year of agent payments.

8:54另外，AI 代理人產生的資料最終都歸 AI 公司（像 ChatGPT）所有，Datadance 並不打算跟他們競爭。我們更專注於用戶手上大量現成的歷史資料，這些才是真正有價值的資產。Another comment here is that agents generate data accrues to AI companies such as ChatGPT first, and Datadance doesn't intend to compete against them. We prefer to focus and prioritize the far larger in-stock data, which we believe are held by users that contains much, much more value.

9:14資料流通的難題與 Datadance 實踐

9:14既然有需求也有供給，為什麼沒人把兩者連起來？因為這不只是技術問題，還有大量繁瑣的地面工作。用戶要從不同 App 匯出資料、上傳，Datadance 錢包要驗證資料真實性和所有權、排除詐騙、清理資料、記錄每一步授權。這些工作又慢又貴，這才是最大挑戰。If there is demand and supply already, then why nobody connects them together, right? Because we claim this is not just about the rails. The difficulty relies with the ground game, because this is a costly and extensive labor work that needs to be done. A user has to upload the, sorry, the data after exporting it from different apps. And then the wallets, Datadance have to verify this data are authentic and owned by the users themselves. They have to screen out the frauds. They have to clean the data to make sure this is something the buyer can actually use. And they have to license it to record the consent for every single actions. So this work is slow and costly, and that is precisely the most.

9:55我想強調，我今天不是在講什麼顛覆性技術，也不是說我們在做最強的資料協議。其實任何厲害的開發者四五個月就能做出不錯的資料協議，甚至很容易被 fork。重點是，過去九個月我們已經 onboard 很多用戶，讓他們主動貢獻歷史資料到我們平台，目前已經有 40 萬用戶。這不是我們單打獨鬥，而是透過 IPDAX、Meet48、Offgrid、Tattoo Life 等強大導流管道，這些社群有粉絲、AI 偶像、娛樂、精品生活需求，大家都願意授權歷史資料給我們，這才是真正的規模化地面戰。Now, I want to be very clear here. I don't think we are trying to talk about a disruptive technology in this speech. I'm not saying that we are building the best data protocol. In fact, any skillful developer may build a pretty good, you know, data protocol in four or five months, and any good data protocol may be forked easily, right? I think the argument we're making here is that in the past nine months, we're able to onboard a lot of users. To have active people who contribute their historical data to our platform, and 400,000 users is a milestone we have achieved. We didn't do this alone and acquire users one by one. We have very strong acquisition channels of IPDAX, of Meet48, of Offgrid, of Tattoo Life. And all these people who have resource networks of fans, of AI idols, of entertainment needs, of premium lifestyles contribute their historical data with users' consent to us. And this is the grand game at scale we are running.

10:56鏈上與隱私新標準

10:56接下來我要說明我們哪些資料會上鏈、哪些不會。我們不會把所有東西都上鏈，因為沒必要。把你的訂單、住址寫到鏈上不是擁有權，而是永久洩漏個資。我們上鏈的只有交易紀錄：誰因為授權什麼資料給誰而收到多少錢。這才是鏈上帳本該記錄的，和鏈下活動搭配才對。The next thing I want to talk about is that I want to be very accurate about what we actually put on-chain and what we do not. So I don't think that we are putting everything on-chain, simply because that is not necessary. And by the way, writing your orders and where you live on-chain, this is not ownership. This is permanent leak of your personal data. What we bring on-chain is simply the activity. Basically, who have been paid for what amounts by giving out what kind of data to whom. And this is a record that needs to be made on-chain. And we believe this is the right use of on-chain ledgers in together with off-chain activities.

11:33如果你是外部觀察者，只會看到一組簽章和新的 key hash，不會看到錢包地址，也就不會洩漏所有鏈上活動。Datadance 團隊很努力制定了 ERC-7962 新標準，用 key hash 取代公開地址，這樣就不會洩漏資訊。現在 ERC-20、ERC-721 都還是用公開地址。So if you are someone who sees a steal, what you can see on-chain is a signature and a new key hash. But you're not going to see the wallet address and therefore not leaking all your activities on-chain. It's very important to note that the Datadance team has worked very hard to make a new standard of ERC-7962. So basically, we replace a public address with a key hash, and with this key hash, you won't have to leak all your information. Currently, with ERC-20 and ERC-721 standard, you have to use your public address.

12:07我想補充一點，昨天才跟 Vitalik 和其他開發者聊到這個。現在有很多隱私協議，像 StarkNet 上的 STRC-20、Zama 推出的 FHE 標準（ERC-7972 或 7984）、Aztec 的新隱私協議等等，這些都很強大。但我們選擇的權衡是，不希望用戶 iPhone 上要跑零知識證明或重運算。我們認為只要把公開地址和擁有權分離就夠了，這才適合讓數萬用戶參與資料交易的市場。And I want to talk about something not on this slide, right? Because I kind of just talked about this to Vitalik and other developers yesterday. So I understand currently there's a lot of privacy protocols, right? For instance, you know, old and new ones. We have STRC-20 on StarkNet proposed these days. Zama recently released the new FHE standard tokens, which is ERC-7972 or 7984, I think, and this is very popular. Aztec making a new privacy-first protocol. So there are a lot of them. I want to make it very clear that they're all very powerful systems, but we are making the right trade-off simply because. We think we don't want any zero-knowledge proof to be operated or heavy computation to be done on the users' iPhones. And we think this is enough to delink or detach the public addresses from the ownership. And this is the standard we believe the most suitable for such a market to be built for tens of thousands of users who want to trade their data.

13:10Datadance 資料流通流程與願景

13:10總結一下，你帶著自己的歷史資料，用 Datadance 錢包一鍵授權，Datadance 會驗證、記錄你的同意，然後開始結算付款。這裡有兩條路：一是資料賣給消費型企業，企業每次購買我們都抽成，你也能持續獲得分潤；二是資料流向 AI，讓 AI 助理和代理人真正讀懂你的歷史，不再只是猜測。過去市場都在為算力建設，未來市場將為資料而生。So let's put it together, right? So basically that you come with your history and the stock data, and you authorize that with one tap on the Datadance wallet. The Datadance wallet verifies it, and they record your consent, and then they start your settlement payment. It basically do two things. Firstly, it makes sure it goes to the consumer businesses through the Datadance market. Where the consumer business buys access to your data, every time they buy it, we take a fee and you get paid. A second sale to a second business, you get paid again. Secondly, it goes to AI through AI Connector, where we make sure that now you have AI agents and AI assistants who read through and draw on this history that actually makes them know you instead of just guessing what you need. Basically, for the past years, the entire market has been building something for the compute, and we claim that the next market is for the data.

14:10最重要的是，最大量的資料供給不是明天才會產生，而是早已存在於數十億個帳號裡，只等著被授權。Datadance 要解鎖這些歷史，把它變成用戶自己掌控的資產，賣給企業、分享給 AI，並把收益回饋給資料創造者。接下來會有一段授權流程的 demo 影片，也歡迎大家到我們 PB517 攤位參觀，謝謝！Most importantly, we think the largest supply is not to be generated tomorrow. It is already sitting in billions of accounts waiting for permission. Datadance unlocks that history and turns it into some assets under the control of the owner, sold it to business, and shared with AI, and paid back to those who created it. And now you can see a video of the demo to see how our authorization happens. We also look forward to seeing you at our booth at PB517. Thank you very much.